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 
52 #define KMP_CANCEL_THREADS
53 #define KMP_THREAD_ATTR
54 
55 // Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being
56 // built on Android
57 #if defined(__ANDROID__)
58 #undef KMP_CANCEL_THREADS
59 #endif
60 
61 #include <signal.h>
62 #include <stdarg.h>
63 #include <stddef.h>
64 #include <stdio.h>
65 #include <stdlib.h>
66 #include <string.h>
67 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
68  Microsoft library. Some macros provided below to replace these functions */
69 #ifndef __ABSOFT_WIN
70 #include <sys/types.h>
71 #endif
72 #include <limits.h>
73 #include <time.h>
74 
75 #include <errno.h>
76 
77 #include "kmp_os.h"
78 
79 #include "kmp_safe_c_api.h"
80 
81 #if KMP_STATS_ENABLED
82 class kmp_stats_list;
83 #endif
84 
85 #if KMP_USE_HIER_SCHED
86 // Only include hierarchical scheduling if affinity is supported
87 #undef KMP_USE_HIER_SCHED
88 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
89 #endif
90 
91 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
92 #include "hwloc.h"
93 #ifndef HWLOC_OBJ_NUMANODE
94 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
95 #endif
96 #ifndef HWLOC_OBJ_PACKAGE
97 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
98 #endif
99 #endif
100 
101 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
102 #include <xmmintrin.h>
103 #endif
104 
105 #include "kmp_debug.h"
106 #include "kmp_lock.h"
107 #include "kmp_version.h"
108 #if USE_DEBUGGER
109 #include "kmp_debugger.h"
110 #endif
111 #include "kmp_i18n.h"
112 
113 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
114 
115 #include "kmp_wrapper_malloc.h"
116 #if KMP_OS_UNIX
117 #include <unistd.h>
118 #if !defined NSIG && defined _NSIG
119 #define NSIG _NSIG
120 #endif
121 #endif
122 
123 #if KMP_OS_LINUX
124 #pragma weak clock_gettime
125 #endif
126 
127 #if OMPT_SUPPORT
128 #include "ompt-internal.h"
129 #endif
130 
131 #if OMP_50_ENABLED
132 // Affinity format function
133 #include "kmp_str.h"
134 #endif
135 
136 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
137 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
138 // free lists of limited size.
139 #ifndef USE_FAST_MEMORY
140 #define USE_FAST_MEMORY 3
141 #endif
142 
143 #ifndef KMP_NESTED_HOT_TEAMS
144 #define KMP_NESTED_HOT_TEAMS 0
145 #define USE_NESTED_HOT_ARG(x)
146 #else
147 #if KMP_NESTED_HOT_TEAMS
148 #if OMP_40_ENABLED
149 #define USE_NESTED_HOT_ARG(x) , x
150 #else
151 // Nested hot teams feature depends on omp 4.0, disable it for earlier versions
152 #undef KMP_NESTED_HOT_TEAMS
153 #define KMP_NESTED_HOT_TEAMS 0
154 #define USE_NESTED_HOT_ARG(x)
155 #endif
156 #else
157 #define USE_NESTED_HOT_ARG(x)
158 #endif
159 #endif
160 
161 // Assume using BGET compare_exchange instruction instead of lock by default.
162 #ifndef USE_CMP_XCHG_FOR_BGET
163 #define USE_CMP_XCHG_FOR_BGET 1
164 #endif
165 
166 // Test to see if queuing lock is better than bootstrap lock for bget
167 // #ifndef USE_QUEUING_LOCK_FOR_BGET
168 // #define USE_QUEUING_LOCK_FOR_BGET
169 // #endif
170 
171 #define KMP_NSEC_PER_SEC 1000000000L
172 #define KMP_USEC_PER_SEC 1000000L
173 
182 enum {
187  /* 0x04 is no longer used */
196  KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
197  KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
198  KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
199 
200  KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
201  KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
202 
214  KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
215  KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
216  KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
217  KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
218 };
219 
223 typedef struct ident {
224  kmp_int32 reserved_1;
225  kmp_int32 flags;
227  kmp_int32 reserved_2;
228 #if USE_ITT_BUILD
229 /* but currently used for storing region-specific ITT */
230 /* contextual information. */
231 #endif /* USE_ITT_BUILD */
232  kmp_int32 reserved_3;
233  char const *psource;
237 } ident_t;
242 // Some forward declarations.
243 typedef union kmp_team kmp_team_t;
244 typedef struct kmp_taskdata kmp_taskdata_t;
245 typedef union kmp_task_team kmp_task_team_t;
246 typedef union kmp_team kmp_team_p;
247 typedef union kmp_info kmp_info_p;
248 typedef union kmp_root kmp_root_p;
249 
250 #ifdef __cplusplus
251 extern "C" {
252 #endif
253 
254 /* ------------------------------------------------------------------------ */
255 
256 /* Pack two 32-bit signed integers into a 64-bit signed integer */
257 /* ToDo: Fix word ordering for big-endian machines. */
258 #define KMP_PACK_64(HIGH_32, LOW_32) \
259  ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
260 
261 // Generic string manipulation macros. Assume that _x is of type char *
262 #define SKIP_WS(_x) \
263  { \
264  while (*(_x) == ' ' || *(_x) == '\t') \
265  (_x)++; \
266  }
267 #define SKIP_DIGITS(_x) \
268  { \
269  while (*(_x) >= '0' && *(_x) <= '9') \
270  (_x)++; \
271  }
272 #define SKIP_TOKEN(_x) \
273  { \
274  while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
275  (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
276  (_x)++; \
277  }
278 #define SKIP_TO(_x, _c) \
279  { \
280  while (*(_x) != '\0' && *(_x) != (_c)) \
281  (_x)++; \
282  }
283 
284 /* ------------------------------------------------------------------------ */
285 
286 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
287 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
288 
289 /* ------------------------------------------------------------------------ */
290 /* Enumeration types */
291 
292 enum kmp_state_timer {
293  ts_stop,
294  ts_start,
295  ts_pause,
296 
297  ts_last_state
298 };
299 
300 enum dynamic_mode {
301  dynamic_default,
302 #ifdef USE_LOAD_BALANCE
303  dynamic_load_balance,
304 #endif /* USE_LOAD_BALANCE */
305  dynamic_random,
306  dynamic_thread_limit,
307  dynamic_max
308 };
309 
310 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
311  * not include it here */
312 #ifndef KMP_SCHED_TYPE_DEFINED
313 #define KMP_SCHED_TYPE_DEFINED
314 typedef enum kmp_sched {
315  kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
316  // Note: need to adjust __kmp_sch_map global array in case enum is changed
317  kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
318  kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
319  kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
320  kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
321  kmp_sched_upper_std = 5, // upper bound for standard schedules
322  kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
323  kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
324 #if KMP_STATIC_STEAL_ENABLED
325  kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
326 #endif
327  kmp_sched_upper,
328  kmp_sched_default = kmp_sched_static // default scheduling
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 #if OMP_45_ENABLED
356  /* static with chunk adjustment (e.g., simd) */
357  kmp_sch_static_balanced_chunked = 45,
358  kmp_sch_guided_simd = 46,
359  kmp_sch_runtime_simd = 47,
360 #endif
361 
362  /* accessible only through KMP_SCHEDULE environment variable */
366  kmp_ord_static_chunked = 65,
368  kmp_ord_dynamic_chunked = 67,
369  kmp_ord_guided_chunked = 68,
370  kmp_ord_runtime = 69,
372  kmp_ord_trapezoidal = 71,
375 #if OMP_40_ENABLED
376  /* Schedules for Distribute construct */
379 #endif
380 
381  /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
382  single iteration/chunk, even if the loop is serialized. For the schedule
383  types listed above, the entire iteration vector is returned if the loop is
384  serialized. This doesn't work for gcc/gcomp sections. */
385  kmp_nm_lower = 160,
387  kmp_nm_static_chunked =
388  (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
390  kmp_nm_dynamic_chunked = 163,
392  kmp_nm_runtime = 165,
393  kmp_nm_auto = 166,
394  kmp_nm_trapezoidal = 167,
395 
396  /* accessible only through KMP_SCHEDULE environment variable */
397  kmp_nm_static_greedy = 168,
398  kmp_nm_static_balanced = 169,
399  /* accessible only through KMP_SCHEDULE environment variable */
400  kmp_nm_guided_iterative_chunked = 170,
401  kmp_nm_guided_analytical_chunked = 171,
402  kmp_nm_static_steal =
403  172, /* accessible only through OMP_SCHEDULE environment variable */
404 
405  kmp_nm_ord_static_chunked = 193,
407  kmp_nm_ord_dynamic_chunked = 195,
408  kmp_nm_ord_guided_chunked = 196,
409  kmp_nm_ord_runtime = 197,
411  kmp_nm_ord_trapezoidal = 199,
414 #if OMP_45_ENABLED
415  /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
416  we need to distinguish the three possible cases (no modifier, monotonic
417  modifier, nonmonotonic modifier), we need separate bits for each modifier.
418  The absence of monotonic does not imply nonmonotonic, especially since 4.5
419  says that the behaviour of the "no modifier" case is implementation defined
420  in 4.5, but will become "nonmonotonic" in 5.0.
421 
422  Since we're passing a full 32 bit value, we can use a couple of high bits
423  for these flags; out of paranoia we avoid the sign bit.
424 
425  These modifiers can be or-ed into non-static schedules by the compiler to
426  pass the additional information. They will be stripped early in the
427  processing in __kmp_dispatch_init when setting up schedules, so most of the
428  code won't ever see schedules with these bits set. */
429  kmp_sch_modifier_monotonic =
430  (1 << 29),
431  kmp_sch_modifier_nonmonotonic =
432  (1 << 30),
434 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
435  (enum sched_type)( \
436  (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))
437 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
438 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
439 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
440  (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
441 #else
442 /* By doing this we hope to avoid multiple tests on OMP_45_ENABLED. Compilers
443  can now eliminate tests on compile time constants and dead code that results
444  from them, so we can leave code guarded by such an if in place. */
445 #define SCHEDULE_WITHOUT_MODIFIERS(s) (s)
446 #define SCHEDULE_HAS_MONOTONIC(s) false
447 #define SCHEDULE_HAS_NONMONOTONIC(s) false
448 #define SCHEDULE_HAS_NO_MODIFIERS(s) true
449 #endif
450 
452 };
453 
454 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
455 typedef union kmp_r_sched {
456  struct {
457  enum sched_type r_sched_type;
458  int chunk;
459  };
460  kmp_int64 sched;
461 } kmp_r_sched_t;
462 
463 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
464 // internal schedule types
465 
466 enum library_type {
467  library_none,
468  library_serial,
469  library_turnaround,
470  library_throughput
471 };
472 
473 #if KMP_OS_LINUX
474 enum clock_function_type {
475  clock_function_gettimeofday,
476  clock_function_clock_gettime
477 };
478 #endif /* KMP_OS_LINUX */
479 
480 #if KMP_MIC_SUPPORTED
481 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
482 #endif
483 
484 /* -- fast reduction stuff ------------------------------------------------ */
485 
486 #undef KMP_FAST_REDUCTION_BARRIER
487 #define KMP_FAST_REDUCTION_BARRIER 1
488 
489 #undef KMP_FAST_REDUCTION_CORE_DUO
490 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
491 #define KMP_FAST_REDUCTION_CORE_DUO 1
492 #endif
493 
494 enum _reduction_method {
495  reduction_method_not_defined = 0,
496  critical_reduce_block = (1 << 8),
497  atomic_reduce_block = (2 << 8),
498  tree_reduce_block = (3 << 8),
499  empty_reduce_block = (4 << 8)
500 };
501 
502 // Description of the packed_reduction_method variable:
503 // The packed_reduction_method variable consists of two enum types variables
504 // that are packed together into 0-th byte and 1-st byte:
505 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
506 // barrier that will be used in fast reduction: bs_plain_barrier or
507 // bs_reduction_barrier
508 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
509 // be used in fast reduction;
510 // Reduction method is of 'enum _reduction_method' type and it's defined the way
511 // so that the bits of 0-th byte are empty, so no need to execute a shift
512 // instruction while packing/unpacking
513 
514 #if KMP_FAST_REDUCTION_BARRIER
515 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
516  ((reduction_method) | (barrier_type))
517 
518 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
519  ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
520 
521 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
522  ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
523 #else
524 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
525  (reduction_method)
526 
527 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
528  (packed_reduction_method)
529 
530 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
531 #endif
532 
533 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
534  ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
535  (which_reduction_block))
536 
537 #if KMP_FAST_REDUCTION_BARRIER
538 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
539  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
540 
541 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
542  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
543 #endif
544 
545 typedef int PACKED_REDUCTION_METHOD_T;
546 
547 /* -- end of fast reduction stuff ----------------------------------------- */
548 
549 #if KMP_OS_WINDOWS
550 #define USE_CBLKDATA
551 #if KMP_MSVC_COMPAT
552 #pragma warning(push)
553 #pragma warning(disable : 271 310)
554 #endif
555 #include <windows.h>
556 #if KMP_MSVC_COMPAT
557 #pragma warning(pop)
558 #endif
559 #endif
560 
561 #if KMP_OS_UNIX
562 #include <dlfcn.h>
563 #include <pthread.h>
564 #endif
565 
566 /* Only Linux* OS and Windows* OS support thread affinity. */
567 #if KMP_AFFINITY_SUPPORTED
568 
569 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
570 #if KMP_OS_WINDOWS
571 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
572 typedef struct GROUP_AFFINITY {
573  KAFFINITY Mask;
574  WORD Group;
575  WORD Reserved[3];
576 } GROUP_AFFINITY;
577 #endif /* _MSC_VER < 1600 */
578 #if KMP_GROUP_AFFINITY
579 extern int __kmp_num_proc_groups;
580 #else
581 static const int __kmp_num_proc_groups = 1;
582 #endif /* KMP_GROUP_AFFINITY */
583 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
584 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
585 
586 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
587 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
588 
589 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
590 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
591 
592 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
593  GROUP_AFFINITY *);
594 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
595 #endif /* KMP_OS_WINDOWS */
596 
597 #if KMP_USE_HWLOC
598 extern hwloc_topology_t __kmp_hwloc_topology;
599 extern int __kmp_hwloc_error;
600 extern int __kmp_numa_detected;
601 extern int __kmp_tile_depth;
602 #endif
603 
604 extern size_t __kmp_affin_mask_size;
605 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
606 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
607 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
608 #define KMP_CPU_SET_ITERATE(i, mask) \
609  for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
610 #define KMP_CPU_SET(i, mask) (mask)->set(i)
611 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
612 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
613 #define KMP_CPU_ZERO(mask) (mask)->zero()
614 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
615 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
616 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
617 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
618 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
619 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
620 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
621 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
622 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
623 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
624 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
625 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
626  (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
627 #define KMP_CPU_FREE_ARRAY(arr, n) \
628  __kmp_affinity_dispatch->deallocate_mask_array(arr)
629 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
630 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
631 #define __kmp_get_system_affinity(mask, abort_bool) \
632  (mask)->get_system_affinity(abort_bool)
633 #define __kmp_set_system_affinity(mask, abort_bool) \
634  (mask)->set_system_affinity(abort_bool)
635 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
636 
637 class KMPAffinity {
638 public:
639  class Mask {
640  public:
641  void *operator new(size_t n);
642  void operator delete(void *p);
643  void *operator new[](size_t n);
644  void operator delete[](void *p);
645  virtual ~Mask() {}
646  // Set bit i to 1
647  virtual void set(int i) {}
648  // Return bit i
649  virtual bool is_set(int i) const { return false; }
650  // Set bit i to 0
651  virtual void clear(int i) {}
652  // Zero out entire mask
653  virtual void zero() {}
654  // Copy src into this mask
655  virtual void copy(const Mask *src) {}
656  // this &= rhs
657  virtual void bitwise_and(const Mask *rhs) {}
658  // this |= rhs
659  virtual void bitwise_or(const Mask *rhs) {}
660  // this = ~this
661  virtual void bitwise_not() {}
662  // API for iterating over an affinity mask
663  // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
664  virtual int begin() const { return 0; }
665  virtual int end() const { return 0; }
666  virtual int next(int previous) const { return 0; }
667  // Set the system's affinity to this affinity mask's value
668  virtual int set_system_affinity(bool abort_on_error) const { return -1; }
669  // Set this affinity mask to the current system affinity
670  virtual int get_system_affinity(bool abort_on_error) { return -1; }
671  // Only 1 DWORD in the mask should have any procs set.
672  // Return the appropriate index, or -1 for an invalid mask.
673  virtual int get_proc_group() const { return -1; }
674  };
675  void *operator new(size_t n);
676  void operator delete(void *p);
677  // Need virtual destructor
678  virtual ~KMPAffinity() = default;
679  // Determine if affinity is capable
680  virtual void determine_capable(const char *env_var) {}
681  // Bind the current thread to os proc
682  virtual void bind_thread(int proc) {}
683  // Factory functions to allocate/deallocate a mask
684  virtual Mask *allocate_mask() { return nullptr; }
685  virtual void deallocate_mask(Mask *m) {}
686  virtual Mask *allocate_mask_array(int num) { return nullptr; }
687  virtual void deallocate_mask_array(Mask *m) {}
688  virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
689  static void pick_api();
690  static void destroy_api();
691  enum api_type {
692  NATIVE_OS
693 #if KMP_USE_HWLOC
694  ,
695  HWLOC
696 #endif
697  };
698  virtual api_type get_api_type() const {
699  KMP_ASSERT(0);
700  return NATIVE_OS;
701  }
702 
703 private:
704  static bool picked_api;
705 };
706 
707 typedef KMPAffinity::Mask kmp_affin_mask_t;
708 extern KMPAffinity *__kmp_affinity_dispatch;
709 
710 // Declare local char buffers with this size for printing debug and info
711 // messages, using __kmp_affinity_print_mask().
712 #define KMP_AFFIN_MASK_PRINT_LEN 1024
713 
714 enum affinity_type {
715  affinity_none = 0,
716  affinity_physical,
717  affinity_logical,
718  affinity_compact,
719  affinity_scatter,
720  affinity_explicit,
721  affinity_balanced,
722  affinity_disabled, // not used outsize the env var parser
723  affinity_default
724 };
725 
726 enum affinity_gran {
727  affinity_gran_fine = 0,
728  affinity_gran_thread,
729  affinity_gran_core,
730  affinity_gran_tile,
731  affinity_gran_numa,
732  affinity_gran_package,
733  affinity_gran_node,
734 #if KMP_GROUP_AFFINITY
735  // The "group" granularity isn't necesssarily coarser than all of the
736  // other levels, but we put it last in the enum.
737  affinity_gran_group,
738 #endif /* KMP_GROUP_AFFINITY */
739  affinity_gran_default
740 };
741 
742 enum affinity_top_method {
743  affinity_top_method_all = 0, // try all (supported) methods, in order
744 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
745  affinity_top_method_apicid,
746  affinity_top_method_x2apicid,
747 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
748  affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
749 #if KMP_GROUP_AFFINITY
750  affinity_top_method_group,
751 #endif /* KMP_GROUP_AFFINITY */
752  affinity_top_method_flat,
753 #if KMP_USE_HWLOC
754  affinity_top_method_hwloc,
755 #endif
756  affinity_top_method_default
757 };
758 
759 #define affinity_respect_mask_default (-1)
760 
761 extern enum affinity_type __kmp_affinity_type; /* Affinity type */
762 extern enum affinity_gran __kmp_affinity_gran; /* Affinity granularity */
763 extern int __kmp_affinity_gran_levels; /* corresponding int value */
764 extern int __kmp_affinity_dups; /* Affinity duplicate masks */
765 extern enum affinity_top_method __kmp_affinity_top_method;
766 extern int __kmp_affinity_compact; /* Affinity 'compact' value */
767 extern int __kmp_affinity_offset; /* Affinity offset value */
768 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */
769 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */
770 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask?
771 extern char *__kmp_affinity_proclist; /* proc ID list */
772 extern kmp_affin_mask_t *__kmp_affinity_masks;
773 extern unsigned __kmp_affinity_num_masks;
774 extern void __kmp_affinity_bind_thread(int which);
775 
776 extern kmp_affin_mask_t *__kmp_affin_fullMask;
777 extern char *__kmp_cpuinfo_file;
778 
779 #endif /* KMP_AFFINITY_SUPPORTED */
780 
781 #if OMP_40_ENABLED
782 
783 // This needs to be kept in sync with the values in omp.h !!!
784 typedef enum kmp_proc_bind_t {
785  proc_bind_false = 0,
786  proc_bind_true,
787  proc_bind_master,
788  proc_bind_close,
789  proc_bind_spread,
790  proc_bind_intel, // use KMP_AFFINITY interface
791  proc_bind_default
792 } kmp_proc_bind_t;
793 
794 typedef struct kmp_nested_proc_bind_t {
795  kmp_proc_bind_t *bind_types;
796  int size;
797  int used;
798 } kmp_nested_proc_bind_t;
799 
800 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
801 
802 #endif /* OMP_40_ENABLED */
803 
804 #if OMP_50_ENABLED
805 extern int __kmp_display_affinity;
806 extern char *__kmp_affinity_format;
807 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
808 #endif // OMP_50_ENABLED
809 
810 #if KMP_AFFINITY_SUPPORTED
811 #define KMP_PLACE_ALL (-1)
812 #define KMP_PLACE_UNDEFINED (-2)
813 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
814 #define KMP_AFFINITY_NON_PROC_BIND \
815  ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
816  __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
817  (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
818 #endif /* KMP_AFFINITY_SUPPORTED */
819 
820 extern int __kmp_affinity_num_places;
821 
822 #if OMP_40_ENABLED
823 typedef enum kmp_cancel_kind_t {
824  cancel_noreq = 0,
825  cancel_parallel = 1,
826  cancel_loop = 2,
827  cancel_sections = 3,
828  cancel_taskgroup = 4
829 } kmp_cancel_kind_t;
830 #endif // OMP_40_ENABLED
831 
832 // KMP_HW_SUBSET support:
833 typedef struct kmp_hws_item {
834  int num;
835  int offset;
836 } kmp_hws_item_t;
837 
838 extern kmp_hws_item_t __kmp_hws_socket;
839 extern kmp_hws_item_t __kmp_hws_node;
840 extern kmp_hws_item_t __kmp_hws_tile;
841 extern kmp_hws_item_t __kmp_hws_core;
842 extern kmp_hws_item_t __kmp_hws_proc;
843 extern int __kmp_hws_requested;
844 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
845 
846 /* ------------------------------------------------------------------------ */
847 
848 #define KMP_PAD(type, sz) \
849  (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
850 
851 // We need to avoid using -1 as a GTID as +1 is added to the gtid
852 // when storing it in a lock, and the value 0 is reserved.
853 #define KMP_GTID_DNE (-2) /* Does not exist */
854 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
855 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
856 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
857 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
858 
859 #if OMP_50_ENABLED
860 /* OpenMP 5.0 Memory Management support */
861 
862 #ifndef __OMP_H
863 // Duplicate type definitios from omp.h
864 typedef uintptr_t omp_uintptr_t;
865 
866 typedef enum {
867  OMP_ATK_THREADMODEL = 1,
868  OMP_ATK_ALIGNMENT = 2,
869  OMP_ATK_ACCESS = 3,
870  OMP_ATK_POOL_SIZE = 4,
871  OMP_ATK_FALLBACK = 5,
872  OMP_ATK_FB_DATA = 6,
873  OMP_ATK_PINNED = 7,
874  OMP_ATK_PARTITION = 8
875 } omp_alloctrait_key_t;
876 
877 typedef enum {
878  OMP_ATV_FALSE = 0,
879  OMP_ATV_TRUE = 1,
880  OMP_ATV_DEFAULT = 2,
881  OMP_ATV_CONTENDED = 3,
882  OMP_ATV_UNCONTENDED = 4,
883  OMP_ATV_SEQUENTIAL = 5,
884  OMP_ATV_PRIVATE = 6,
885  OMP_ATV_ALL = 7,
886  OMP_ATV_THREAD = 8,
887  OMP_ATV_PTEAM = 9,
888  OMP_ATV_CGROUP = 10,
889  OMP_ATV_DEFAULT_MEM_FB = 11,
890  OMP_ATV_NULL_FB = 12,
891  OMP_ATV_ABORT_FB = 13,
892  OMP_ATV_ALLOCATOR_FB = 14,
893  OMP_ATV_ENVIRONMENT = 15,
894  OMP_ATV_NEAREST = 16,
895  OMP_ATV_BLOCKED = 17,
896  OMP_ATV_INTERLEAVED = 18
897 } omp_alloctrait_value_t;
898 
899 typedef void *omp_memspace_handle_t;
900 extern omp_memspace_handle_t const omp_default_mem_space;
901 extern omp_memspace_handle_t const omp_large_cap_mem_space;
902 extern omp_memspace_handle_t const omp_const_mem_space;
903 extern omp_memspace_handle_t const omp_high_bw_mem_space;
904 extern omp_memspace_handle_t const omp_low_lat_mem_space;
905 
906 typedef struct {
907  omp_alloctrait_key_t key;
908  omp_uintptr_t value;
909 } omp_alloctrait_t;
910 
911 typedef void *omp_allocator_handle_t;
912 extern omp_allocator_handle_t const omp_null_allocator;
913 extern omp_allocator_handle_t const omp_default_mem_alloc;
914 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
915 extern omp_allocator_handle_t const omp_const_mem_alloc;
916 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
917 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
918 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
919 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
920 extern omp_allocator_handle_t const omp_thread_mem_alloc;
921 extern omp_allocator_handle_t const kmp_max_mem_alloc;
922 extern omp_allocator_handle_t __kmp_def_allocator;
923 
924 // end of duplicate type definitios from omp.h
925 #endif
926 
927 extern int __kmp_memkind_available;
928 
929 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
930 
931 typedef struct kmp_allocator_t {
932  omp_memspace_handle_t memspace;
933  void **memkind; // pointer to memkind
934  int alignment;
935  omp_alloctrait_value_t fb;
936  kmp_allocator_t *fb_data;
937  kmp_uint64 pool_size;
938  kmp_uint64 pool_used;
939 } kmp_allocator_t;
940 
941 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
942  omp_memspace_handle_t,
943  int ntraits,
944  omp_alloctrait_t traits[]);
945 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
946 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
947 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
948 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
949 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
950 
951 extern void __kmp_init_memkind();
952 extern void __kmp_fini_memkind();
953 #endif // OMP_50_ENABLED
954 
955 /* ------------------------------------------------------------------------ */
956 
957 #define KMP_UINT64_MAX \
958  (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
959 
960 #define KMP_MIN_NTH 1
961 
962 #ifndef KMP_MAX_NTH
963 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
964 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
965 #else
966 #define KMP_MAX_NTH INT_MAX
967 #endif
968 #endif /* KMP_MAX_NTH */
969 
970 #ifdef PTHREAD_STACK_MIN
971 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
972 #else
973 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
974 #endif
975 
976 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
977 
978 #if KMP_ARCH_X86
979 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
980 #elif KMP_ARCH_X86_64
981 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
982 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
983 #else
984 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
985 #endif
986 
987 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
988 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
989 #define KMP_MAX_MALLOC_POOL_INCR \
990  (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
991 
992 #define KMP_MIN_STKOFFSET (0)
993 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
994 #if KMP_OS_DARWIN
995 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
996 #else
997 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
998 #endif
999 
1000 #define KMP_MIN_STKPADDING (0)
1001 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1002 
1003 #define KMP_BLOCKTIME_MULTIPLIER \
1004  (1000) /* number of blocktime units per second */
1005 #define KMP_MIN_BLOCKTIME (0)
1006 #define KMP_MAX_BLOCKTIME \
1007  (INT_MAX) /* Must be this for "infinite" setting the work */
1008 #define KMP_DEFAULT_BLOCKTIME (200) /* __kmp_blocktime is in milliseconds */
1009 
1010 #if KMP_USE_MONITOR
1011 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1012 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1013 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1014 
1015 /* Calculate new number of monitor wakeups for a specific block time based on
1016  previous monitor_wakeups. Only allow increasing number of wakeups */
1017 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1018  (((blocktime) == KMP_MAX_BLOCKTIME) \
1019  ? (monitor_wakeups) \
1020  : ((blocktime) == KMP_MIN_BLOCKTIME) \
1021  ? KMP_MAX_MONITOR_WAKEUPS \
1022  : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1023  ? (monitor_wakeups) \
1024  : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1025 
1026 /* Calculate number of intervals for a specific block time based on
1027  monitor_wakeups */
1028 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1029  (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1030  (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1031 #else
1032 #define KMP_BLOCKTIME(team, tid) \
1033  (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1034 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1035 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1036 extern kmp_uint64 __kmp_ticks_per_msec;
1037 #if KMP_COMPILER_ICC
1038 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1039 #else
1040 #define KMP_NOW() __kmp_hardware_timestamp()
1041 #endif
1042 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1043 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1044  (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1045 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1046 #else
1047 // System time is retrieved sporadically while blocking.
1048 extern kmp_uint64 __kmp_now_nsec();
1049 #define KMP_NOW() __kmp_now_nsec()
1050 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1051 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1052  (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1053 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1054 #endif
1055 #endif // KMP_USE_MONITOR
1056 
1057 #define KMP_MIN_STATSCOLS 40
1058 #define KMP_MAX_STATSCOLS 4096
1059 #define KMP_DEFAULT_STATSCOLS 80
1060 
1061 #define KMP_MIN_INTERVAL 0
1062 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1063 #define KMP_DEFAULT_INTERVAL 0
1064 
1065 #define KMP_MIN_CHUNK 1
1066 #define KMP_MAX_CHUNK (INT_MAX - 1)
1067 #define KMP_DEFAULT_CHUNK 1
1068 
1069 #define KMP_DFLT_DISP_NUM_BUFF 7
1070 #define KMP_MAX_ORDERED 8
1071 
1072 #define KMP_MAX_FIELDS 32
1073 
1074 #define KMP_MAX_BRANCH_BITS 31
1075 
1076 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1077 
1078 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1079 
1080 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1081 
1082 /* Minimum number of threads before switch to TLS gtid (experimentally
1083  determined) */
1084 /* josh TODO: what about OS X* tuning? */
1085 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1086 #define KMP_TLS_GTID_MIN 5
1087 #else
1088 #define KMP_TLS_GTID_MIN INT_MAX
1089 #endif
1090 
1091 #define KMP_MASTER_TID(tid) ((tid) == 0)
1092 #define KMP_WORKER_TID(tid) ((tid) != 0)
1093 
1094 #define KMP_MASTER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) == 0)
1095 #define KMP_WORKER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) != 0)
1096 #define KMP_INITIAL_GTID(gtid) ((gtid) == 0)
1097 
1098 #ifndef TRUE
1099 #define FALSE 0
1100 #define TRUE (!FALSE)
1101 #endif
1102 
1103 /* NOTE: all of the following constants must be even */
1104 
1105 #if KMP_OS_WINDOWS
1106 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1107 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1108 #elif KMP_OS_CNK
1109 #define KMP_INIT_WAIT 16U /* initial number of spin-tests */
1110 #define KMP_NEXT_WAIT 8U /* susequent number of spin-tests */
1111 #elif KMP_OS_LINUX
1112 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1113 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1114 #elif KMP_OS_DARWIN
1115 /* TODO: tune for KMP_OS_DARWIN */
1116 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1117 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1118 #elif KMP_OS_DRAGONFLY
1119 /* TODO: tune for KMP_OS_DRAGONFLY */
1120 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1121 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1122 #elif KMP_OS_FREEBSD
1123 /* TODO: tune for KMP_OS_FREEBSD */
1124 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1125 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1126 #elif KMP_OS_NETBSD
1127 /* TODO: tune for KMP_OS_NETBSD */
1128 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1129 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1130 #elif KMP_OS_HURD
1131 /* TODO: tune for KMP_OS_HURD */
1132 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1133 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1134 #elif KMP_OS_OPENBSD
1135 /* TODO: tune for KMP_OS_OPENBSD */
1136 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1137 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1138 #elif KMP_OS_KFREEBSD
1139 /* TODO: tune for KMP_OS_KFREEBSD */
1140 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1141 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1142 #endif
1143 
1144 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1145 typedef struct kmp_cpuid {
1146  kmp_uint32 eax;
1147  kmp_uint32 ebx;
1148  kmp_uint32 ecx;
1149  kmp_uint32 edx;
1150 } kmp_cpuid_t;
1151 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1152 #if KMP_ARCH_X86
1153 extern void __kmp_x86_pause(void);
1154 #elif KMP_MIC
1155 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1156 // regression after removal of extra PAUSE from spin loops. Changing
1157 // the delay from 100 to 300 showed even better performance than double PAUSE
1158 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1159 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1160 #else
1161 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1162 #endif
1163 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1164 #elif KMP_ARCH_PPC64
1165 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1166 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1167 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1168 #define KMP_CPU_PAUSE() \
1169  do { \
1170  KMP_PPC64_PRI_LOW(); \
1171  KMP_PPC64_PRI_MED(); \
1172  KMP_PPC64_PRI_LOC_MB(); \
1173  } while (0)
1174 #else
1175 #define KMP_CPU_PAUSE() /* nothing to do */
1176 #endif
1177 
1178 #define KMP_INIT_YIELD(count) \
1179  { (count) = __kmp_yield_init; }
1180 
1181 #define KMP_OVERSUBSCRIBED \
1182  (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1183 
1184 #define KMP_TRY_YIELD \
1185  ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1186 
1187 #define KMP_TRY_YIELD_OVERSUB \
1188  ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1189 
1190 #define KMP_YIELD(cond) \
1191  { \
1192  KMP_CPU_PAUSE(); \
1193  if ((cond) && (KMP_TRY_YIELD)) \
1194  __kmp_yield(); \
1195  }
1196 
1197 #define KMP_YIELD_OVERSUB() \
1198  { \
1199  KMP_CPU_PAUSE(); \
1200  if ((KMP_TRY_YIELD_OVERSUB)) \
1201  __kmp_yield(); \
1202  }
1203 
1204 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1205 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1206 #define KMP_YIELD_SPIN(count) \
1207  { \
1208  KMP_CPU_PAUSE(); \
1209  if (KMP_TRY_YIELD) { \
1210  (count) -= 2; \
1211  if (!(count)) { \
1212  __kmp_yield(); \
1213  (count) = __kmp_yield_next; \
1214  } \
1215  } \
1216  }
1217 
1218 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count) \
1219  { \
1220  KMP_CPU_PAUSE(); \
1221  if ((KMP_TRY_YIELD_OVERSUB)) \
1222  __kmp_yield(); \
1223  else if (__kmp_use_yield == 1) { \
1224  (count) -= 2; \
1225  if (!(count)) { \
1226  __kmp_yield(); \
1227  (count) = __kmp_yield_next; \
1228  } \
1229  } \
1230  }
1231 
1232 /* ------------------------------------------------------------------------ */
1233 /* Support datatypes for the orphaned construct nesting checks. */
1234 /* ------------------------------------------------------------------------ */
1235 
1236 enum cons_type {
1237  ct_none,
1238  ct_parallel,
1239  ct_pdo,
1240  ct_pdo_ordered,
1241  ct_psections,
1242  ct_psingle,
1243  ct_critical,
1244  ct_ordered_in_parallel,
1245  ct_ordered_in_pdo,
1246  ct_master,
1247  ct_reduce,
1248  ct_barrier
1249 };
1250 
1251 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1252 
1253 struct cons_data {
1254  ident_t const *ident;
1255  enum cons_type type;
1256  int prev;
1257  kmp_user_lock_p
1258  name; /* address exclusively for critical section name comparison */
1259 };
1260 
1261 struct cons_header {
1262  int p_top, w_top, s_top;
1263  int stack_size, stack_top;
1264  struct cons_data *stack_data;
1265 };
1266 
1267 struct kmp_region_info {
1268  char *text;
1269  int offset[KMP_MAX_FIELDS];
1270  int length[KMP_MAX_FIELDS];
1271 };
1272 
1273 /* ---------------------------------------------------------------------- */
1274 /* ---------------------------------------------------------------------- */
1275 
1276 #if KMP_OS_WINDOWS
1277 typedef HANDLE kmp_thread_t;
1278 typedef DWORD kmp_key_t;
1279 #endif /* KMP_OS_WINDOWS */
1280 
1281 #if KMP_OS_UNIX
1282 typedef pthread_t kmp_thread_t;
1283 typedef pthread_key_t kmp_key_t;
1284 #endif
1285 
1286 extern kmp_key_t __kmp_gtid_threadprivate_key;
1287 
1288 typedef struct kmp_sys_info {
1289  long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1290  long minflt; /* the number of page faults serviced without any I/O */
1291  long majflt; /* the number of page faults serviced that required I/O */
1292  long nswap; /* the number of times a process was "swapped" out of memory */
1293  long inblock; /* the number of times the file system had to perform input */
1294  long oublock; /* the number of times the file system had to perform output */
1295  long nvcsw; /* the number of times a context switch was voluntarily */
1296  long nivcsw; /* the number of times a context switch was forced */
1297 } kmp_sys_info_t;
1298 
1299 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1300 typedef struct kmp_cpuinfo {
1301  int initialized; // If 0, other fields are not initialized.
1302  int signature; // CPUID(1).EAX
1303  int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1304  int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1305  // Model << 4 ) + Model)
1306  int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1307  int sse2; // 0 if SSE2 instructions are not supported, 1 otherwise.
1308  int rtm; // 0 if RTM instructions are not supported, 1 otherwise.
1309  int cpu_stackoffset;
1310  int apic_id;
1311  int physical_id;
1312  int logical_id;
1313  kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1314  char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1315 } kmp_cpuinfo_t;
1316 #endif
1317 
1318 #if USE_ITT_BUILD
1319 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1320 // required type here. Later we will check the type meets requirements.
1321 typedef int kmp_itt_mark_t;
1322 #define KMP_ITT_DEBUG 0
1323 #endif /* USE_ITT_BUILD */
1324 
1325 typedef kmp_int32 kmp_critical_name[8];
1326 
1336 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1337 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1338  ...);
1339 
1344 /* ---------------------------------------------------------------------------
1345  */
1346 /* Threadprivate initialization/finalization function declarations */
1347 
1348 /* for non-array objects: __kmpc_threadprivate_register() */
1349 
1354 typedef void *(*kmpc_ctor)(void *);
1355 
1360 typedef void (*kmpc_dtor)(
1361  void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1362  compiler */
1367 typedef void *(*kmpc_cctor)(void *, void *);
1368 
1369 /* for array objects: __kmpc_threadprivate_register_vec() */
1370 /* First arg: "this" pointer */
1371 /* Last arg: number of array elements */
1377 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1383 typedef void (*kmpc_dtor_vec)(void *, size_t);
1389 typedef void *(*kmpc_cctor_vec)(void *, void *,
1390  size_t); /* function unused by compiler */
1391 
1396 /* keeps tracked of threadprivate cache allocations for cleanup later */
1397 typedef struct kmp_cached_addr {
1398  void **addr; /* address of allocated cache */
1399  void ***compiler_cache; /* pointer to compiler's cache */
1400  void *data; /* pointer to global data */
1401  struct kmp_cached_addr *next; /* pointer to next cached address */
1402 } kmp_cached_addr_t;
1403 
1404 struct private_data {
1405  struct private_data *next; /* The next descriptor in the list */
1406  void *data; /* The data buffer for this descriptor */
1407  int more; /* The repeat count for this descriptor */
1408  size_t size; /* The data size for this descriptor */
1409 };
1410 
1411 struct private_common {
1412  struct private_common *next;
1413  struct private_common *link;
1414  void *gbl_addr;
1415  void *par_addr; /* par_addr == gbl_addr for MASTER thread */
1416  size_t cmn_size;
1417 };
1418 
1419 struct shared_common {
1420  struct shared_common *next;
1421  struct private_data *pod_init;
1422  void *obj_init;
1423  void *gbl_addr;
1424  union {
1425  kmpc_ctor ctor;
1426  kmpc_ctor_vec ctorv;
1427  } ct;
1428  union {
1429  kmpc_cctor cctor;
1430  kmpc_cctor_vec cctorv;
1431  } cct;
1432  union {
1433  kmpc_dtor dtor;
1434  kmpc_dtor_vec dtorv;
1435  } dt;
1436  size_t vec_len;
1437  int is_vec;
1438  size_t cmn_size;
1439 };
1440 
1441 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1442 #define KMP_HASH_TABLE_SIZE \
1443  (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1444 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1445 #define KMP_HASH(x) \
1446  ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1447 
1448 struct common_table {
1449  struct private_common *data[KMP_HASH_TABLE_SIZE];
1450 };
1451 
1452 struct shared_table {
1453  struct shared_common *data[KMP_HASH_TABLE_SIZE];
1454 };
1455 
1456 /* ------------------------------------------------------------------------ */
1457 
1458 #if KMP_USE_HIER_SCHED
1459 // Shared barrier data that exists inside a single unit of the scheduling
1460 // hierarchy
1461 typedef struct kmp_hier_private_bdata_t {
1462  kmp_int32 num_active;
1463  kmp_uint64 index;
1464  kmp_uint64 wait_val[2];
1465 } kmp_hier_private_bdata_t;
1466 #endif
1467 
1468 typedef struct kmp_sched_flags {
1469  unsigned ordered : 1;
1470  unsigned nomerge : 1;
1471  unsigned contains_last : 1;
1472 #if KMP_USE_HIER_SCHED
1473  unsigned use_hier : 1;
1474  unsigned unused : 28;
1475 #else
1476  unsigned unused : 29;
1477 #endif
1478 } kmp_sched_flags_t;
1479 
1480 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1481 
1482 #if KMP_STATIC_STEAL_ENABLED
1483 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1484  kmp_int32 count;
1485  kmp_int32 ub;
1486  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1487  kmp_int32 lb;
1488  kmp_int32 st;
1489  kmp_int32 tc;
1490  kmp_int32 static_steal_counter; /* for static_steal only; maybe better to put
1491  after ub */
1492 
1493  // KMP_ALIGN( 16 ) ensures ( if the KMP_ALIGN macro is turned on )
1494  // a) parm3 is properly aligned and
1495  // b) all parm1-4 are in the same cache line.
1496  // Because of parm1-4 are used together, performance seems to be better
1497  // if they are in the same line (not measured though).
1498 
1499  struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1500  kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1501  kmp_int32 parm2; // make no real change at least while padding is off.
1502  kmp_int32 parm3;
1503  kmp_int32 parm4;
1504  };
1505 
1506  kmp_uint32 ordered_lower;
1507  kmp_uint32 ordered_upper;
1508 #if KMP_OS_WINDOWS
1509  // This var can be placed in the hole between 'tc' and 'parm1', instead of
1510  // 'static_steal_counter'. It would be nice to measure execution times.
1511  // Conditional if/endif can be removed at all.
1512  kmp_int32 last_upper;
1513 #endif /* KMP_OS_WINDOWS */
1514 } dispatch_private_info32_t;
1515 
1516 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1517  kmp_int64 count; // current chunk number for static & static-steal scheduling
1518  kmp_int64 ub; /* upper-bound */
1519  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1520  kmp_int64 lb; /* lower-bound */
1521  kmp_int64 st; /* stride */
1522  kmp_int64 tc; /* trip count (number of iterations) */
1523  kmp_int64 static_steal_counter; /* for static_steal only; maybe better to put
1524  after ub */
1525 
1526  /* parm[1-4] are used in different ways by different scheduling algorithms */
1527 
1528  // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1529  // a) parm3 is properly aligned and
1530  // b) all parm1-4 are in the same cache line.
1531  // Because of parm1-4 are used together, performance seems to be better
1532  // if they are in the same line (not measured though).
1533 
1534  struct KMP_ALIGN(32) {
1535  kmp_int64 parm1;
1536  kmp_int64 parm2;
1537  kmp_int64 parm3;
1538  kmp_int64 parm4;
1539  };
1540 
1541  kmp_uint64 ordered_lower;
1542  kmp_uint64 ordered_upper;
1543 #if KMP_OS_WINDOWS
1544  // This var can be placed in the hole between 'tc' and 'parm1', instead of
1545  // 'static_steal_counter'. It would be nice to measure execution times.
1546  // Conditional if/endif can be removed at all.
1547  kmp_int64 last_upper;
1548 #endif /* KMP_OS_WINDOWS */
1549 } dispatch_private_info64_t;
1550 #else /* KMP_STATIC_STEAL_ENABLED */
1551 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1552  kmp_int32 lb;
1553  kmp_int32 ub;
1554  kmp_int32 st;
1555  kmp_int32 tc;
1556 
1557  kmp_int32 parm1;
1558  kmp_int32 parm2;
1559  kmp_int32 parm3;
1560  kmp_int32 parm4;
1561 
1562  kmp_int32 count;
1563 
1564  kmp_uint32 ordered_lower;
1565  kmp_uint32 ordered_upper;
1566 #if KMP_OS_WINDOWS
1567  kmp_int32 last_upper;
1568 #endif /* KMP_OS_WINDOWS */
1569 } dispatch_private_info32_t;
1570 
1571 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1572  kmp_int64 lb; /* lower-bound */
1573  kmp_int64 ub; /* upper-bound */
1574  kmp_int64 st; /* stride */
1575  kmp_int64 tc; /* trip count (number of iterations) */
1576 
1577  /* parm[1-4] are used in different ways by different scheduling algorithms */
1578  kmp_int64 parm1;
1579  kmp_int64 parm2;
1580  kmp_int64 parm3;
1581  kmp_int64 parm4;
1582 
1583  kmp_int64 count; /* current chunk number for static scheduling */
1584 
1585  kmp_uint64 ordered_lower;
1586  kmp_uint64 ordered_upper;
1587 #if KMP_OS_WINDOWS
1588  kmp_int64 last_upper;
1589 #endif /* KMP_OS_WINDOWS */
1590 } dispatch_private_info64_t;
1591 #endif /* KMP_STATIC_STEAL_ENABLED */
1592 
1593 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1594  union private_info {
1595  dispatch_private_info32_t p32;
1596  dispatch_private_info64_t p64;
1597  } u;
1598  enum sched_type schedule; /* scheduling algorithm */
1599  kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1600  kmp_int32 ordered_bumped;
1601  // To retain the structure size after making ordered_iteration scalar
1602  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 3];
1603  // Stack of buffers for nest of serial regions
1604  struct dispatch_private_info *next;
1605  kmp_int32 type_size; /* the size of types in private_info */
1606 #if KMP_USE_HIER_SCHED
1607  kmp_int32 hier_id;
1608  void *parent; /* hierarchical scheduling parent pointer */
1609 #endif
1610  enum cons_type pushed_ws;
1611 } dispatch_private_info_t;
1612 
1613 typedef struct dispatch_shared_info32 {
1614  /* chunk index under dynamic, number of idle threads under static-steal;
1615  iteration index otherwise */
1616  volatile kmp_uint32 iteration;
1617  volatile kmp_uint32 num_done;
1618  volatile kmp_uint32 ordered_iteration;
1619  // Dummy to retain the structure size after making ordered_iteration scalar
1620  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1621 } dispatch_shared_info32_t;
1622 
1623 typedef struct dispatch_shared_info64 {
1624  /* chunk index under dynamic, number of idle threads under static-steal;
1625  iteration index otherwise */
1626  volatile kmp_uint64 iteration;
1627  volatile kmp_uint64 num_done;
1628  volatile kmp_uint64 ordered_iteration;
1629  // Dummy to retain the structure size after making ordered_iteration scalar
1630  kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1631 } dispatch_shared_info64_t;
1632 
1633 typedef struct dispatch_shared_info {
1634  union shared_info {
1635  dispatch_shared_info32_t s32;
1636  dispatch_shared_info64_t s64;
1637  } u;
1638  volatile kmp_uint32 buffer_index;
1639 #if OMP_45_ENABLED
1640  volatile kmp_int32 doacross_buf_idx; // teamwise index
1641  volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1642  kmp_int32 doacross_num_done; // count finished threads
1643 #endif
1644 #if KMP_USE_HIER_SCHED
1645  void *hier;
1646 #endif
1647 #if KMP_USE_HWLOC
1648  // When linking with libhwloc, the ORDERED EPCC test slows down on big
1649  // machines (> 48 cores). Performance analysis showed that a cache thrash
1650  // was occurring and this padding helps alleviate the problem.
1651  char padding[64];
1652 #endif
1653 } dispatch_shared_info_t;
1654 
1655 typedef struct kmp_disp {
1656  /* Vector for ORDERED SECTION */
1657  void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1658  /* Vector for END ORDERED SECTION */
1659  void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1660 
1661  dispatch_shared_info_t *th_dispatch_sh_current;
1662  dispatch_private_info_t *th_dispatch_pr_current;
1663 
1664  dispatch_private_info_t *th_disp_buffer;
1665  kmp_int32 th_disp_index;
1666 #if OMP_45_ENABLED
1667  kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1668  volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1669  union { // we can use union here because doacross cannot be used in
1670  // nonmonotonic loops
1671  kmp_int64 *th_doacross_info; // info on loop bounds
1672  kmp_lock_t *th_steal_lock; // lock used for chunk stealing (8-byte variable)
1673  };
1674 #else
1675 #if KMP_STATIC_STEAL_ENABLED
1676  kmp_lock_t *th_steal_lock; // lock used for chunk stealing (8-byte variable)
1677  void *dummy_padding[1]; // make it 64 bytes on Intel(R) 64
1678 #else
1679  void *dummy_padding[2]; // make it 64 bytes on Intel(R) 64
1680 #endif
1681 #endif
1682 #if KMP_USE_INTERNODE_ALIGNMENT
1683  char more_padding[INTERNODE_CACHE_LINE];
1684 #endif
1685 } kmp_disp_t;
1686 
1687 /* ------------------------------------------------------------------------ */
1688 /* Barrier stuff */
1689 
1690 /* constants for barrier state update */
1691 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1692 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1693 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1694 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1695 
1696 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1697 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1698 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1699 
1700 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1701 #error "Barrier sleep bit must be smaller than barrier bump bit"
1702 #endif
1703 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1704 #error "Barrier unused bit must be smaller than barrier bump bit"
1705 #endif
1706 
1707 // Constants for release barrier wait state: currently, hierarchical only
1708 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1709 #define KMP_BARRIER_OWN_FLAG \
1710  1 // Normal state; worker waiting on own b_go flag in release
1711 #define KMP_BARRIER_PARENT_FLAG \
1712  2 // Special state; worker waiting on parent's b_go flag in release
1713 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1714  3 // Special state; tells worker to shift from parent to own b_go
1715 #define KMP_BARRIER_SWITCHING \
1716  4 // Special state; worker resets appropriate flag on wake-up
1717 
1718 #define KMP_NOT_SAFE_TO_REAP \
1719  0 // Thread th_reap_state: not safe to reap (tasking)
1720 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1721 
1722 enum barrier_type {
1723  bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1724  barriers if enabled) */
1725  bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1726 #if KMP_FAST_REDUCTION_BARRIER
1727  bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1728 #endif // KMP_FAST_REDUCTION_BARRIER
1729  bs_last_barrier /* Just a placeholder to mark the end */
1730 };
1731 
1732 // to work with reduction barriers just like with plain barriers
1733 #if !KMP_FAST_REDUCTION_BARRIER
1734 #define bs_reduction_barrier bs_plain_barrier
1735 #endif // KMP_FAST_REDUCTION_BARRIER
1736 
1737 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1738  bp_linear_bar =
1739  0, /* Single level (degenerate) tree */
1740  bp_tree_bar =
1741  1, /* Balanced tree with branching factor 2^n */
1742  bp_hyper_bar =
1743  2, /* Hypercube-embedded tree with min branching
1744  factor 2^n */
1745  bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1746  bp_last_bar /* Placeholder to mark the end */
1747 } kmp_bar_pat_e;
1748 
1749 #define KMP_BARRIER_ICV_PUSH 1
1750 
1751 /* Record for holding the values of the internal controls stack records */
1752 typedef struct kmp_internal_control {
1753  int serial_nesting_level; /* corresponds to the value of the
1754  th_team_serialized field */
1755  kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1756  thread) */
1757  kmp_int8
1758  bt_set; /* internal control for whether blocktime is explicitly set */
1759  int blocktime; /* internal control for blocktime */
1760 #if KMP_USE_MONITOR
1761  int bt_intervals; /* internal control for blocktime intervals */
1762 #endif
1763  int nproc; /* internal control for #threads for next parallel region (per
1764  thread) */
1765  int thread_limit; /* internal control for thread-limit-var */
1766  int max_active_levels; /* internal control for max_active_levels */
1767  kmp_r_sched_t
1768  sched; /* internal control for runtime schedule {sched,chunk} pair */
1769 #if OMP_40_ENABLED
1770  kmp_proc_bind_t proc_bind; /* internal control for affinity */
1771  kmp_int32 default_device; /* internal control for default device */
1772 #endif // OMP_40_ENABLED
1773  struct kmp_internal_control *next;
1774 } kmp_internal_control_t;
1775 
1776 static inline void copy_icvs(kmp_internal_control_t *dst,
1777  kmp_internal_control_t *src) {
1778  *dst = *src;
1779 }
1780 
1781 /* Thread barrier needs volatile barrier fields */
1782 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1783  // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1784  // uses of it). It is not explicitly aligned below, because we *don't* want
1785  // it to be padded -- instead, we fit b_go into the same cache line with
1786  // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1787  kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1788  // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1789  // same NGO store
1790  volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1791  KMP_ALIGN_CACHE volatile kmp_uint64
1792  b_arrived; // STATE => task reached synch point.
1793  kmp_uint32 *skip_per_level;
1794  kmp_uint32 my_level;
1795  kmp_int32 parent_tid;
1796  kmp_int32 old_tid;
1797  kmp_uint32 depth;
1798  struct kmp_bstate *parent_bar;
1799  kmp_team_t *team;
1800  kmp_uint64 leaf_state;
1801  kmp_uint32 nproc;
1802  kmp_uint8 base_leaf_kids;
1803  kmp_uint8 leaf_kids;
1804  kmp_uint8 offset;
1805  kmp_uint8 wait_flag;
1806  kmp_uint8 use_oncore_barrier;
1807 #if USE_DEBUGGER
1808  // The following field is intended for the debugger solely. Only the worker
1809  // thread itself accesses this field: the worker increases it by 1 when it
1810  // arrives to a barrier.
1811  KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1812 #endif /* USE_DEBUGGER */
1813 } kmp_bstate_t;
1814 
1815 union KMP_ALIGN_CACHE kmp_barrier_union {
1816  double b_align; /* use worst case alignment */
1817  char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1818  kmp_bstate_t bb;
1819 };
1820 
1821 typedef union kmp_barrier_union kmp_balign_t;
1822 
1823 /* Team barrier needs only non-volatile arrived counter */
1824 union KMP_ALIGN_CACHE kmp_barrier_team_union {
1825  double b_align; /* use worst case alignment */
1826  char b_pad[CACHE_LINE];
1827  struct {
1828  kmp_uint64 b_arrived; /* STATE => task reached synch point. */
1829 #if USE_DEBUGGER
1830  // The following two fields are indended for the debugger solely. Only
1831  // master of the team accesses these fields: the first one is increased by
1832  // 1 when master arrives to a barrier, the second one is increased by one
1833  // when all the threads arrived.
1834  kmp_uint b_master_arrived;
1835  kmp_uint b_team_arrived;
1836 #endif
1837  };
1838 };
1839 
1840 typedef union kmp_barrier_team_union kmp_balign_team_t;
1841 
1842 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
1843  threads when a condition changes. This is to workaround an NPTL bug where
1844  padding was added to pthread_cond_t which caused the initialization routine
1845  to write outside of the structure if compiled on pre-NPTL threads. */
1846 #if KMP_OS_WINDOWS
1847 typedef struct kmp_win32_mutex {
1848  /* The Lock */
1849  CRITICAL_SECTION cs;
1850 } kmp_win32_mutex_t;
1851 
1852 typedef struct kmp_win32_cond {
1853  /* Count of the number of waiters. */
1854  int waiters_count_;
1855 
1856  /* Serialize access to <waiters_count_> */
1857  kmp_win32_mutex_t waiters_count_lock_;
1858 
1859  /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
1860  int release_count_;
1861 
1862  /* Keeps track of the current "generation" so that we don't allow */
1863  /* one thread to steal all the "releases" from the broadcast. */
1864  int wait_generation_count_;
1865 
1866  /* A manual-reset event that's used to block and release waiting threads. */
1867  HANDLE event_;
1868 } kmp_win32_cond_t;
1869 #endif
1870 
1871 #if KMP_OS_UNIX
1872 
1873 union KMP_ALIGN_CACHE kmp_cond_union {
1874  double c_align;
1875  char c_pad[CACHE_LINE];
1876  pthread_cond_t c_cond;
1877 };
1878 
1879 typedef union kmp_cond_union kmp_cond_align_t;
1880 
1881 union KMP_ALIGN_CACHE kmp_mutex_union {
1882  double m_align;
1883  char m_pad[CACHE_LINE];
1884  pthread_mutex_t m_mutex;
1885 };
1886 
1887 typedef union kmp_mutex_union kmp_mutex_align_t;
1888 
1889 #endif /* KMP_OS_UNIX */
1890 
1891 typedef struct kmp_desc_base {
1892  void *ds_stackbase;
1893  size_t ds_stacksize;
1894  int ds_stackgrow;
1895  kmp_thread_t ds_thread;
1896  volatile int ds_tid;
1897  int ds_gtid;
1898 #if KMP_OS_WINDOWS
1899  volatile int ds_alive;
1900  DWORD ds_thread_id;
1901 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
1902  However, debugger support (libomp_db) cannot work with handles, because they
1903  uncomparable. For example, debugger requests info about thread with handle h.
1904  h is valid within debugger process, and meaningless within debugee process.
1905  Even if h is duped by call to DuplicateHandle(), so the result h' is valid
1906  within debugee process, but it is a *new* handle which does *not* equal to
1907  any other handle in debugee... The only way to compare handles is convert
1908  them to system-wide ids. GetThreadId() function is available only in
1909  Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
1910  on all Windows* OS flavours (including Windows* 95). Thus, we have to get
1911  thread id by call to GetCurrentThreadId() from within the thread and save it
1912  to let libomp_db identify threads. */
1913 #endif /* KMP_OS_WINDOWS */
1914 } kmp_desc_base_t;
1915 
1916 typedef union KMP_ALIGN_CACHE kmp_desc {
1917  double ds_align; /* use worst case alignment */
1918  char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
1919  kmp_desc_base_t ds;
1920 } kmp_desc_t;
1921 
1922 typedef struct kmp_local {
1923  volatile int this_construct; /* count of single's encountered by thread */
1924  void *reduce_data;
1925 #if KMP_USE_BGET
1926  void *bget_data;
1927  void *bget_list;
1928 #if !USE_CMP_XCHG_FOR_BGET
1929 #ifdef USE_QUEUING_LOCK_FOR_BGET
1930  kmp_lock_t bget_lock; /* Lock for accessing bget free list */
1931 #else
1932  kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
1933 // bootstrap lock so we can use it at library
1934 // shutdown.
1935 #endif /* USE_LOCK_FOR_BGET */
1936 #endif /* ! USE_CMP_XCHG_FOR_BGET */
1937 #endif /* KMP_USE_BGET */
1938 
1939  PACKED_REDUCTION_METHOD_T
1940  packed_reduction_method; /* stored by __kmpc_reduce*(), used by
1941  __kmpc_end_reduce*() */
1942 
1943 } kmp_local_t;
1944 
1945 #define KMP_CHECK_UPDATE(a, b) \
1946  if ((a) != (b)) \
1947  (a) = (b)
1948 #define KMP_CHECK_UPDATE_SYNC(a, b) \
1949  if ((a) != (b)) \
1950  TCW_SYNC_PTR((a), (b))
1951 
1952 #define get__blocktime(xteam, xtid) \
1953  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
1954 #define get__bt_set(xteam, xtid) \
1955  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
1956 #if KMP_USE_MONITOR
1957 #define get__bt_intervals(xteam, xtid) \
1958  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
1959 #endif
1960 
1961 #define get__dynamic_2(xteam, xtid) \
1962  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
1963 #define get__nproc_2(xteam, xtid) \
1964  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
1965 #define get__sched_2(xteam, xtid) \
1966  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
1967 
1968 #define set__blocktime_team(xteam, xtid, xval) \
1969  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
1970  (xval))
1971 
1972 #if KMP_USE_MONITOR
1973 #define set__bt_intervals_team(xteam, xtid, xval) \
1974  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
1975  (xval))
1976 #endif
1977 
1978 #define set__bt_set_team(xteam, xtid, xval) \
1979  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
1980 
1981 #define set__dynamic(xthread, xval) \
1982  (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
1983 #define get__dynamic(xthread) \
1984  (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
1985 
1986 #define set__nproc(xthread, xval) \
1987  (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
1988 
1989 #define set__thread_limit(xthread, xval) \
1990  (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
1991 
1992 #define set__max_active_levels(xthread, xval) \
1993  (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
1994 
1995 #define get__max_active_levels(xthread) \
1996  ((xthread)->th.th_current_task->td_icvs.max_active_levels)
1997 
1998 #define set__sched(xthread, xval) \
1999  (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2000 
2001 #if OMP_40_ENABLED
2002 
2003 #define set__proc_bind(xthread, xval) \
2004  (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2005 #define get__proc_bind(xthread) \
2006  ((xthread)->th.th_current_task->td_icvs.proc_bind)
2007 
2008 #endif /* OMP_40_ENABLED */
2009 
2010 // OpenMP tasking data structures
2011 
2012 typedef enum kmp_tasking_mode {
2013  tskm_immediate_exec = 0,
2014  tskm_extra_barrier = 1,
2015  tskm_task_teams = 2,
2016  tskm_max = 2
2017 } kmp_tasking_mode_t;
2018 
2019 extern kmp_tasking_mode_t
2020  __kmp_tasking_mode; /* determines how/when to execute tasks */
2021 extern int __kmp_task_stealing_constraint;
2022 #if OMP_40_ENABLED
2023 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2024 // specified, defaults to 0 otherwise
2025 #endif
2026 #if OMP_45_ENABLED
2027 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2028 extern kmp_int32 __kmp_max_task_priority;
2029 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2030 extern kmp_uint64 __kmp_taskloop_min_tasks;
2031 #endif
2032 
2033 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2034  taskdata first */
2035 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2036 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2037 
2038 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2039 // were spawned and queued since the previous barrier release.
2040 #define KMP_TASKING_ENABLED(task_team) \
2041  (TCR_SYNC_4((task_team)->tt.tt_found_tasks) == TRUE)
2042 
2049 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2050 
2051 #if OMP_40_ENABLED || OMP_45_ENABLED
2052 typedef union kmp_cmplrdata {
2053 #if OMP_45_ENABLED
2054  kmp_int32 priority;
2055 #endif // OMP_45_ENABLED
2056 #if OMP_40_ENABLED
2057  kmp_routine_entry_t
2058  destructors; /* pointer to function to invoke deconstructors of
2059  firstprivate C++ objects */
2060 #endif // OMP_40_ENABLED
2061  /* future data */
2062 } kmp_cmplrdata_t;
2063 #endif
2064 
2065 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2068 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2069  void *shareds;
2070  kmp_routine_entry_t
2071  routine;
2072  kmp_int32 part_id;
2073 #if OMP_40_ENABLED || OMP_45_ENABLED
2074  kmp_cmplrdata_t
2075  data1; /* Two known optional additions: destructors and priority */
2076  kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2077 /* future data */
2078 #endif
2079  /* private vars */
2080 } kmp_task_t;
2081 
2086 #if OMP_40_ENABLED
2087 typedef struct kmp_taskgroup {
2088  std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2089  std::atomic<kmp_int32>
2090  cancel_request; // request for cancellation of this taskgroup
2091  struct kmp_taskgroup *parent; // parent taskgroup
2092 #if OMP_50_ENABLED
2093  // Block of data to perform task reduction
2094  void *reduce_data; // reduction related info
2095  kmp_int32 reduce_num_data; // number of data items to reduce
2096 #endif
2097 } kmp_taskgroup_t;
2098 
2099 // forward declarations
2100 typedef union kmp_depnode kmp_depnode_t;
2101 typedef struct kmp_depnode_list kmp_depnode_list_t;
2102 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2103 
2104 // Compiler sends us this info:
2105 typedef struct kmp_depend_info {
2106  kmp_intptr_t base_addr;
2107  size_t len;
2108  struct {
2109  bool in : 1;
2110  bool out : 1;
2111  bool mtx : 1;
2112  } flags;
2113 } kmp_depend_info_t;
2114 
2115 // Internal structures to work with task dependencies:
2116 struct kmp_depnode_list {
2117  kmp_depnode_t *node;
2118  kmp_depnode_list_t *next;
2119 };
2120 
2121 // Max number of mutexinoutset dependencies per node
2122 #define MAX_MTX_DEPS 4
2123 
2124 typedef struct kmp_base_depnode {
2125  kmp_depnode_list_t *successors; /* used under lock */
2126  kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2127  kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2128  kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2129  kmp_lock_t lock; /* guards shared fields: task, successors */
2130 #if KMP_SUPPORT_GRAPH_OUTPUT
2131  kmp_uint32 id;
2132 #endif
2133  std::atomic<kmp_int32> npredecessors;
2134  std::atomic<kmp_int32> nrefs;
2135 } kmp_base_depnode_t;
2136 
2137 union KMP_ALIGN_CACHE kmp_depnode {
2138  double dn_align; /* use worst case alignment */
2139  char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2140  kmp_base_depnode_t dn;
2141 };
2142 
2143 struct kmp_dephash_entry {
2144  kmp_intptr_t addr;
2145  kmp_depnode_t *last_out;
2146  kmp_depnode_list_t *last_ins;
2147  kmp_depnode_list_t *last_mtxs;
2148  kmp_int32 last_flag;
2149  kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2150  kmp_dephash_entry_t *next_in_bucket;
2151 };
2152 
2153 typedef struct kmp_dephash {
2154  kmp_dephash_entry_t **buckets;
2155  size_t size;
2156 #ifdef KMP_DEBUG
2157  kmp_uint32 nelements;
2158  kmp_uint32 nconflicts;
2159 #endif
2160 } kmp_dephash_t;
2161 
2162 #if OMP_50_ENABLED
2163 typedef struct kmp_task_affinity_info {
2164  kmp_intptr_t base_addr;
2165  size_t len;
2166  struct {
2167  bool flag1 : 1;
2168  bool flag2 : 1;
2169  kmp_int32 reserved : 30;
2170  } flags;
2171 } kmp_task_affinity_info_t;
2172 #endif
2173 
2174 #endif
2175 
2176 #ifdef BUILD_TIED_TASK_STACK
2177 
2178 /* Tied Task stack definitions */
2179 typedef struct kmp_stack_block {
2180  kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2181  struct kmp_stack_block *sb_next;
2182  struct kmp_stack_block *sb_prev;
2183 } kmp_stack_block_t;
2184 
2185 typedef struct kmp_task_stack {
2186  kmp_stack_block_t ts_first_block; // first block of stack entries
2187  kmp_taskdata_t **ts_top; // pointer to the top of stack
2188  kmp_int32 ts_entries; // number of entries on the stack
2189 } kmp_task_stack_t;
2190 
2191 #endif // BUILD_TIED_TASK_STACK
2192 
2193 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2194  /* Compiler flags */ /* Total compiler flags must be 16 bits */
2195  unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2196  unsigned final : 1; /* task is final(1) so execute immediately */
2197  unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2198  code path */
2199 #if OMP_40_ENABLED
2200  unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2201  invoke destructors from the runtime */
2202 #if OMP_45_ENABLED
2203  unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2204  context of the RTL) */
2205  unsigned priority_specified : 1; /* set if the compiler provides priority
2206  setting for the task */
2207  unsigned reserved : 10; /* reserved for compiler use */
2208 #else
2209  unsigned reserved : 12; /* reserved for compiler use */
2210 #endif
2211 #else // OMP_40_ENABLED
2212  unsigned reserved : 13; /* reserved for compiler use */
2213 #endif // OMP_40_ENABLED
2214 
2215  /* Library flags */ /* Total library flags must be 16 bits */
2216  unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2217  unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2218  unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2219  // (1) or may be deferred (0)
2220  unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2221  // (0) [>= 2 threads]
2222  /* If either team_serial or tasking_ser is set, task team may be NULL */
2223  /* Task State Flags: */
2224  unsigned started : 1; /* 1==started, 0==not started */
2225  unsigned executing : 1; /* 1==executing, 0==not executing */
2226  unsigned complete : 1; /* 1==complete, 0==not complete */
2227  unsigned freed : 1; /* 1==freed, 0==allocateed */
2228  unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2229  unsigned reserved31 : 7; /* reserved for library use */
2230 
2231 } kmp_tasking_flags_t;
2232 
2233 struct kmp_taskdata { /* aligned during dynamic allocation */
2234  kmp_int32 td_task_id; /* id, assigned by debugger */
2235  kmp_tasking_flags_t td_flags; /* task flags */
2236  kmp_team_t *td_team; /* team for this task */
2237  kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2238  /* Currently not used except for perhaps IDB */
2239  kmp_taskdata_t *td_parent; /* parent task */
2240  kmp_int32 td_level; /* task nesting level */
2241  std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2242  ident_t *td_ident; /* task identifier */
2243  // Taskwait data.
2244  ident_t *td_taskwait_ident;
2245  kmp_uint32 td_taskwait_counter;
2246  kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2247  KMP_ALIGN_CACHE kmp_internal_control_t
2248  td_icvs; /* Internal control variables for the task */
2249  KMP_ALIGN_CACHE std::atomic<kmp_int32>
2250  td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2251  deallocated */
2252  std::atomic<kmp_int32>
2253  td_incomplete_child_tasks; /* Child tasks not yet complete */
2254 #if OMP_40_ENABLED
2255  kmp_taskgroup_t
2256  *td_taskgroup; // Each task keeps pointer to its current taskgroup
2257  kmp_dephash_t
2258  *td_dephash; // Dependencies for children tasks are tracked from here
2259  kmp_depnode_t
2260  *td_depnode; // Pointer to graph node if this task has dependencies
2261 #endif // OMP_40_ENABLED
2262 #if OMP_45_ENABLED
2263  kmp_task_team_t *td_task_team;
2264  kmp_int32 td_size_alloc; // The size of task structure, including shareds etc.
2265 #if defined(KMP_GOMP_COMPAT)
2266  // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2267  kmp_int32 td_size_loop_bounds;
2268 #endif
2269 #endif // OMP_45_ENABLED
2270  kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2271 #if defined(KMP_GOMP_COMPAT) && OMP_45_ENABLED
2272  // GOMP sends in a copy function for copy constructors
2273  void (*td_copy_func)(void *, void *);
2274 #endif
2275 #if OMPT_SUPPORT
2276  ompt_task_info_t ompt_task_info;
2277 #endif
2278 }; // struct kmp_taskdata
2279 
2280 // Make sure padding above worked
2281 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2282 
2283 // Data for task team but per thread
2284 typedef struct kmp_base_thread_data {
2285  kmp_info_p *td_thr; // Pointer back to thread info
2286  // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2287  // queued?
2288  kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2289  kmp_taskdata_t *
2290  *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2291  kmp_int32 td_deque_size; // Size of deck
2292  kmp_uint32 td_deque_head; // Head of deque (will wrap)
2293  kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2294  kmp_int32 td_deque_ntasks; // Number of tasks in deque
2295  // GEH: shouldn't this be volatile since used in while-spin?
2296  kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2297 #ifdef BUILD_TIED_TASK_STACK
2298  kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2299 // scheduling constraint
2300 #endif // BUILD_TIED_TASK_STACK
2301 } kmp_base_thread_data_t;
2302 
2303 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2304 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2305 
2306 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2307 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2308 
2309 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2310  kmp_base_thread_data_t td;
2311  double td_align; /* use worst case alignment */
2312  char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2313 } kmp_thread_data_t;
2314 
2315 // Data for task teams which are used when tasking is enabled for the team
2316 typedef struct kmp_base_task_team {
2317  kmp_bootstrap_lock_t
2318  tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2319  /* must be bootstrap lock since used at library shutdown*/
2320  kmp_task_team_t *tt_next; /* For linking the task team free list */
2321  kmp_thread_data_t
2322  *tt_threads_data; /* Array of per-thread structures for task team */
2323  /* Data survives task team deallocation */
2324  kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2325  executing this team? */
2326  /* TRUE means tt_threads_data is set up and initialized */
2327  kmp_int32 tt_nproc; /* #threads in team */
2328  kmp_int32
2329  tt_max_threads; /* number of entries allocated for threads_data array */
2330 #if OMP_45_ENABLED
2331  kmp_int32
2332  tt_found_proxy_tasks; /* Have we found proxy tasks since last barrier */
2333 #endif
2334  kmp_int32 tt_untied_task_encountered;
2335 
2336  KMP_ALIGN_CACHE
2337  std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2338 
2339  KMP_ALIGN_CACHE
2340  volatile kmp_uint32
2341  tt_active; /* is the team still actively executing tasks */
2342 } kmp_base_task_team_t;
2343 
2344 union KMP_ALIGN_CACHE kmp_task_team {
2345  kmp_base_task_team_t tt;
2346  double tt_align; /* use worst case alignment */
2347  char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2348 };
2349 
2350 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2351 // Free lists keep same-size free memory slots for fast memory allocation
2352 // routines
2353 typedef struct kmp_free_list {
2354  void *th_free_list_self; // Self-allocated tasks free list
2355  void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2356  // threads
2357  void *th_free_list_other; // Non-self free list (to be returned to owner's
2358  // sync list)
2359 } kmp_free_list_t;
2360 #endif
2361 #if KMP_NESTED_HOT_TEAMS
2362 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2363 // are not put in teams pool, and they don't put threads in threads pool.
2364 typedef struct kmp_hot_team_ptr {
2365  kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2366  kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2367 } kmp_hot_team_ptr_t;
2368 #endif
2369 #if OMP_40_ENABLED
2370 typedef struct kmp_teams_size {
2371  kmp_int32 nteams; // number of teams in a league
2372  kmp_int32 nth; // number of threads in each team of the league
2373 } kmp_teams_size_t;
2374 #endif
2375 
2376 // This struct stores a thread that acts as a "root" for a contention
2377 // group. Contention groups are rooted at kmp_root threads, but also at
2378 // each master thread of each team created in the teams construct.
2379 // This struct therefore also stores a thread_limit associated with
2380 // that contention group, and a counter to track the number of threads
2381 // active in that contention group. Each thread has a list of these: CG
2382 // root threads have an entry in their list in which cg_root refers to
2383 // the thread itself, whereas other workers in the CG will have a
2384 // single entry where cg_root is same as the entry containing their CG
2385 // root. When a thread encounters a teams construct, it will add a new
2386 // entry to the front of its list, because it now roots a new CG.
2387 typedef struct kmp_cg_root {
2388  kmp_info_p *cg_root; // "root" thread for a contention group
2389  // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2390  // thread_limit clause for teams masters
2391  kmp_int32 cg_thread_limit;
2392  kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2393  struct kmp_cg_root *up; // pointer to higher level CG root in list
2394 } kmp_cg_root_t;
2395 
2396 // OpenMP thread data structures
2397 
2398 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2399  /* Start with the readonly data which is cache aligned and padded. This is
2400  written before the thread starts working by the master. Uber masters may
2401  update themselves later. Usage does not consider serialized regions. */
2402  kmp_desc_t th_info;
2403  kmp_team_p *th_team; /* team we belong to */
2404  kmp_root_p *th_root; /* pointer to root of task hierarchy */
2405  kmp_info_p *th_next_pool; /* next available thread in the pool */
2406  kmp_disp_t *th_dispatch; /* thread's dispatch data */
2407  int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2408 
2409  /* The following are cached from the team info structure */
2410  /* TODO use these in more places as determined to be needed via profiling */
2411  int th_team_nproc; /* number of threads in a team */
2412  kmp_info_p *th_team_master; /* the team's master thread */
2413  int th_team_serialized; /* team is serialized */
2414 #if OMP_40_ENABLED
2415  microtask_t th_teams_microtask; /* save entry address for teams construct */
2416  int th_teams_level; /* save initial level of teams construct */
2417 /* it is 0 on device but may be any on host */
2418 #endif
2419 
2420 /* The blocktime info is copied from the team struct to the thread sruct */
2421 /* at the start of a barrier, and the values stored in the team are used */
2422 /* at points in the code where the team struct is no longer guaranteed */
2423 /* to exist (from the POV of worker threads). */
2424 #if KMP_USE_MONITOR
2425  int th_team_bt_intervals;
2426  int th_team_bt_set;
2427 #else
2428  kmp_uint64 th_team_bt_intervals;
2429 #endif
2430 
2431 #if KMP_AFFINITY_SUPPORTED
2432  kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2433 #endif
2434 #if OMP_50_ENABLED
2435  omp_allocator_handle_t th_def_allocator; /* default allocator */
2436 #endif
2437  /* The data set by the master at reinit, then R/W by the worker */
2438  KMP_ALIGN_CACHE int
2439  th_set_nproc; /* if > 0, then only use this request for the next fork */
2440 #if KMP_NESTED_HOT_TEAMS
2441  kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2442 #endif
2443 #if OMP_40_ENABLED
2444  kmp_proc_bind_t
2445  th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2446  kmp_teams_size_t
2447  th_teams_size; /* number of teams/threads in teams construct */
2448 #if KMP_AFFINITY_SUPPORTED
2449  int th_current_place; /* place currently bound to */
2450  int th_new_place; /* place to bind to in par reg */
2451  int th_first_place; /* first place in partition */
2452  int th_last_place; /* last place in partition */
2453 #endif
2454 #endif
2455 #if OMP_50_ENABLED
2456  int th_prev_level; /* previous level for affinity format */
2457  int th_prev_num_threads; /* previous num_threads for affinity format */
2458 #endif
2459 #if USE_ITT_BUILD
2460  kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2461  kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2462  kmp_uint64 th_frame_time; /* frame timestamp */
2463 #endif /* USE_ITT_BUILD */
2464  kmp_local_t th_local;
2465  struct private_common *th_pri_head;
2466 
2467  /* Now the data only used by the worker (after initial allocation) */
2468  /* TODO the first serial team should actually be stored in the info_t
2469  structure. this will help reduce initial allocation overhead */
2470  KMP_ALIGN_CACHE kmp_team_p
2471  *th_serial_team; /*serialized team held in reserve*/
2472 
2473 #if OMPT_SUPPORT
2474  ompt_thread_info_t ompt_thread_info;
2475 #endif
2476 
2477  /* The following are also read by the master during reinit */
2478  struct common_table *th_pri_common;
2479 
2480  volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2481  /* while awaiting queuing lock acquire */
2482 
2483  volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2484 
2485  ident_t *th_ident;
2486  unsigned th_x; // Random number generator data
2487  unsigned th_a; // Random number generator data
2488 
2489  /* Tasking-related data for the thread */
2490  kmp_task_team_t *th_task_team; // Task team struct
2491  kmp_taskdata_t *th_current_task; // Innermost Task being executed
2492  kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2493  kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2494  // at nested levels
2495  kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2496  kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2497  kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2498  // tasking, thus safe to reap
2499 
2500  /* More stuff for keeping track of active/sleeping threads (this part is
2501  written by the worker thread) */
2502  kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2503  int th_active; // ! sleeping; 32 bits for TCR/TCW
2504  struct cons_header *th_cons; // used for consistency check
2505 #if KMP_USE_HIER_SCHED
2506  // used for hierarchical scheduling
2507  kmp_hier_private_bdata_t *th_hier_bar_data;
2508 #endif
2509 
2510  /* Add the syncronizing data which is cache aligned and padded. */
2511  KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2512 
2513  KMP_ALIGN_CACHE volatile kmp_int32
2514  th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2515 
2516 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2517 #define NUM_LISTS 4
2518  kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2519 // allocation routines
2520 #endif
2521 
2522 #if KMP_OS_WINDOWS
2523  kmp_win32_cond_t th_suspend_cv;
2524  kmp_win32_mutex_t th_suspend_mx;
2525  int th_suspend_init;
2526 #endif
2527 #if KMP_OS_UNIX
2528  kmp_cond_align_t th_suspend_cv;
2529  kmp_mutex_align_t th_suspend_mx;
2530  int th_suspend_init_count;
2531 #endif
2532 
2533 #if USE_ITT_BUILD
2534  kmp_itt_mark_t th_itt_mark_single;
2535 // alignment ???
2536 #endif /* USE_ITT_BUILD */
2537 #if KMP_STATS_ENABLED
2538  kmp_stats_list *th_stats;
2539 #endif
2540 #if KMP_OS_UNIX
2541  std::atomic<bool> th_blocking;
2542 #endif
2543  kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2544 } kmp_base_info_t;
2545 
2546 typedef union KMP_ALIGN_CACHE kmp_info {
2547  double th_align; /* use worst case alignment */
2548  char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2549  kmp_base_info_t th;
2550 } kmp_info_t;
2551 
2552 // OpenMP thread team data structures
2553 
2554 typedef struct kmp_base_data { volatile kmp_uint32 t_value; } kmp_base_data_t;
2555 
2556 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2557  double dt_align; /* use worst case alignment */
2558  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2559  kmp_base_data_t dt;
2560 } kmp_sleep_team_t;
2561 
2562 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2563  double dt_align; /* use worst case alignment */
2564  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2565  kmp_base_data_t dt;
2566 } kmp_ordered_team_t;
2567 
2568 typedef int (*launch_t)(int gtid);
2569 
2570 /* Minimum number of ARGV entries to malloc if necessary */
2571 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2572 
2573 // Set up how many argv pointers will fit in cache lines containing
2574 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2575 // larger value for more space between the master write/worker read section and
2576 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2577 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2578 #define KMP_INLINE_ARGV_BYTES \
2579  (4 * CACHE_LINE - \
2580  ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2581  sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2582  CACHE_LINE))
2583 #else
2584 #define KMP_INLINE_ARGV_BYTES \
2585  (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2586 #endif
2587 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2588 
2589 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2590  // Synchronization Data
2591  // ---------------------------------------------------------------------------
2592  KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2593  kmp_balign_team_t t_bar[bs_last_barrier];
2594  std::atomic<int> t_construct; // count of single directive encountered by team
2595  char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2596 
2597  // Master only
2598  // ---------------------------------------------------------------------------
2599  KMP_ALIGN_CACHE int t_master_tid; // tid of master in parent team
2600  int t_master_this_cons; // "this_construct" single counter of master in parent
2601  // team
2602  ident_t *t_ident; // if volatile, have to change too much other crud to
2603  // volatile too
2604  kmp_team_p *t_parent; // parent team
2605  kmp_team_p *t_next_pool; // next free team in the team pool
2606  kmp_disp_t *t_dispatch; // thread's dispatch data
2607  kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2608 #if OMP_40_ENABLED
2609  kmp_proc_bind_t t_proc_bind; // bind type for par region
2610 #endif // OMP_40_ENABLED
2611 #if USE_ITT_BUILD
2612  kmp_uint64 t_region_time; // region begin timestamp
2613 #endif /* USE_ITT_BUILD */
2614 
2615  // Master write, workers read
2616  // --------------------------------------------------------------------------
2617  KMP_ALIGN_CACHE void **t_argv;
2618  int t_argc;
2619  int t_nproc; // number of threads in team
2620  microtask_t t_pkfn;
2621  launch_t t_invoke; // procedure to launch the microtask
2622 
2623 #if OMPT_SUPPORT
2624  ompt_team_info_t ompt_team_info;
2625  ompt_lw_taskteam_t *ompt_serialized_team_info;
2626 #endif
2627 
2628 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2629  kmp_int8 t_fp_control_saved;
2630  kmp_int8 t_pad2b;
2631  kmp_int16 t_x87_fpu_control_word; // FP control regs
2632  kmp_uint32 t_mxcsr;
2633 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2634 
2635  void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2636 
2637  KMP_ALIGN_CACHE kmp_info_t **t_threads;
2638  kmp_taskdata_t
2639  *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2640  int t_level; // nested parallel level
2641 
2642  KMP_ALIGN_CACHE int t_max_argc;
2643  int t_max_nproc; // max threads this team can handle (dynamicly expandable)
2644  int t_serialized; // levels deep of serialized teams
2645  dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2646  int t_id; // team's id, assigned by debugger.
2647  int t_active_level; // nested active parallel level
2648  kmp_r_sched_t t_sched; // run-time schedule for the team
2649 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2650  int t_first_place; // first & last place in parent thread's partition.
2651  int t_last_place; // Restore these values to master after par region.
2652 #endif // OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2653 #if OMP_50_ENABLED
2654  int t_display_affinity;
2655 #endif
2656  int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2657 // omp_set_num_threads() call
2658 #if OMP_50_ENABLED
2659  omp_allocator_handle_t t_def_allocator; /* default allocator */
2660 #endif
2661 
2662 // Read/write by workers as well
2663 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2664  // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2665  // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2666  // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2667  // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2668  char dummy_padding[1024];
2669 #endif
2670  // Internal control stack for additional nested teams.
2671  KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2672 // for SERIALIZED teams nested 2 or more levels deep
2673 #if OMP_40_ENABLED
2674  // typed flag to store request state of cancellation
2675  std::atomic<kmp_int32> t_cancel_request;
2676 #endif
2677  int t_master_active; // save on fork, restore on join
2678  void *t_copypriv_data; // team specific pointer to copyprivate data array
2679 #if KMP_OS_WINDOWS
2680  std::atomic<kmp_uint32> t_copyin_counter;
2681 #endif
2682 #if USE_ITT_BUILD
2683  void *t_stack_id; // team specific stack stitching id (for ittnotify)
2684 #endif /* USE_ITT_BUILD */
2685 } kmp_base_team_t;
2686 
2687 union KMP_ALIGN_CACHE kmp_team {
2688  kmp_base_team_t t;
2689  double t_align; /* use worst case alignment */
2690  char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2691 };
2692 
2693 typedef union KMP_ALIGN_CACHE kmp_time_global {
2694  double dt_align; /* use worst case alignment */
2695  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2696  kmp_base_data_t dt;
2697 } kmp_time_global_t;
2698 
2699 typedef struct kmp_base_global {
2700  /* cache-aligned */
2701  kmp_time_global_t g_time;
2702 
2703  /* non cache-aligned */
2704  volatile int g_abort;
2705  volatile int g_done;
2706 
2707  int g_dynamic;
2708  enum dynamic_mode g_dynamic_mode;
2709 } kmp_base_global_t;
2710 
2711 typedef union KMP_ALIGN_CACHE kmp_global {
2712  kmp_base_global_t g;
2713  double g_align; /* use worst case alignment */
2714  char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2715 } kmp_global_t;
2716 
2717 typedef struct kmp_base_root {
2718  // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2719  // (r_in_parallel>= 0)
2720  // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2721  // the synch overhead or keeping r_active
2722  volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2723  // keeps a count of active parallel regions per root
2724  std::atomic<int> r_in_parallel;
2725  // GEH: This is misnamed, should be r_active_levels
2726  kmp_team_t *r_root_team;
2727  kmp_team_t *r_hot_team;
2728  kmp_info_t *r_uber_thread;
2729  kmp_lock_t r_begin_lock;
2730  volatile int r_begin;
2731  int r_blocktime; /* blocktime for this root and descendants */
2732 } kmp_base_root_t;
2733 
2734 typedef union KMP_ALIGN_CACHE kmp_root {
2735  kmp_base_root_t r;
2736  double r_align; /* use worst case alignment */
2737  char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2738 } kmp_root_t;
2739 
2740 struct fortran_inx_info {
2741  kmp_int32 data;
2742 };
2743 
2744 /* ------------------------------------------------------------------------ */
2745 
2746 extern int __kmp_settings;
2747 extern int __kmp_duplicate_library_ok;
2748 #if USE_ITT_BUILD
2749 extern int __kmp_forkjoin_frames;
2750 extern int __kmp_forkjoin_frames_mode;
2751 #endif
2752 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2753 extern int __kmp_determ_red;
2754 
2755 #ifdef KMP_DEBUG
2756 extern int kmp_a_debug;
2757 extern int kmp_b_debug;
2758 extern int kmp_c_debug;
2759 extern int kmp_d_debug;
2760 extern int kmp_e_debug;
2761 extern int kmp_f_debug;
2762 #endif /* KMP_DEBUG */
2763 
2764 /* For debug information logging using rotating buffer */
2765 #define KMP_DEBUG_BUF_LINES_INIT 512
2766 #define KMP_DEBUG_BUF_LINES_MIN 1
2767 
2768 #define KMP_DEBUG_BUF_CHARS_INIT 128
2769 #define KMP_DEBUG_BUF_CHARS_MIN 2
2770 
2771 extern int
2772  __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2773 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2774 extern int
2775  __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2776 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2777  entry pointer */
2778 
2779 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2780 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2781  printed in buffer so far */
2782 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2783  recommended in warnings */
2784 /* end rotating debug buffer */
2785 
2786 #ifdef KMP_DEBUG
2787 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2788 
2789 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2790 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2791 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2792 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2793 extern int __kmp_par_range_lb;
2794 extern int __kmp_par_range_ub;
2795 #endif
2796 
2797 /* For printing out dynamic storage map for threads and teams */
2798 extern int
2799  __kmp_storage_map; /* True means print storage map for threads and teams */
2800 extern int __kmp_storage_map_verbose; /* True means storage map includes
2801  placement info */
2802 extern int __kmp_storage_map_verbose_specified;
2803 
2804 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2805 extern kmp_cpuinfo_t __kmp_cpuinfo;
2806 #endif
2807 
2808 extern volatile int __kmp_init_serial;
2809 extern volatile int __kmp_init_gtid;
2810 extern volatile int __kmp_init_common;
2811 extern volatile int __kmp_init_middle;
2812 extern volatile int __kmp_init_parallel;
2813 #if KMP_USE_MONITOR
2814 extern volatile int __kmp_init_monitor;
2815 #endif
2816 extern volatile int __kmp_init_user_locks;
2817 extern int __kmp_init_counter;
2818 extern int __kmp_root_counter;
2819 extern int __kmp_version;
2820 
2821 /* list of address of allocated caches for commons */
2822 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2823 
2824 /* Barrier algorithm types and options */
2825 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
2826 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
2827 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
2828 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
2829 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
2830 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
2831 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
2832 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
2833 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
2834 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
2835 extern char const *__kmp_barrier_type_name[bs_last_barrier];
2836 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
2837 
2838 /* Global Locks */
2839 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
2840 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
2841 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
2842 extern kmp_bootstrap_lock_t
2843  __kmp_exit_lock; /* exit() is not always thread-safe */
2844 #if KMP_USE_MONITOR
2845 extern kmp_bootstrap_lock_t
2846  __kmp_monitor_lock; /* control monitor thread creation */
2847 #endif
2848 extern kmp_bootstrap_lock_t
2849  __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
2850  __kmp_threads expansion to co-exist */
2851 
2852 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
2853 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
2854 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
2855 
2856 extern enum library_type __kmp_library;
2857 
2858 extern enum sched_type __kmp_sched; /* default runtime scheduling */
2859 extern enum sched_type __kmp_static; /* default static scheduling method */
2860 extern enum sched_type __kmp_guided; /* default guided scheduling method */
2861 extern enum sched_type __kmp_auto; /* default auto scheduling method */
2862 extern int __kmp_chunk; /* default runtime chunk size */
2863 
2864 extern size_t __kmp_stksize; /* stack size per thread */
2865 #if KMP_USE_MONITOR
2866 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
2867 #endif
2868 extern size_t __kmp_stkoffset; /* stack offset per thread */
2869 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
2870 
2871 extern size_t
2872  __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
2873 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
2874 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
2875 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
2876 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
2877 extern int __kmp_generate_warnings; /* should we issue warnings? */
2878 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
2879 
2880 #ifdef DEBUG_SUSPEND
2881 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
2882 #endif
2883 
2884 extern kmp_int32 __kmp_use_yield;
2885 extern kmp_int32 __kmp_use_yield_exp_set;
2886 extern kmp_uint32 __kmp_yield_init;
2887 extern kmp_uint32 __kmp_yield_next;
2888 
2889 /* ------------------------------------------------------------------------- */
2890 extern int __kmp_allThreadsSpecified;
2891 
2892 extern size_t __kmp_align_alloc;
2893 /* following data protected by initialization routines */
2894 extern int __kmp_xproc; /* number of processors in the system */
2895 extern int __kmp_avail_proc; /* number of processors available to the process */
2896 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
2897 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
2898 // maximum total number of concurrently-existing threads on device
2899 extern int __kmp_max_nth;
2900 // maximum total number of concurrently-existing threads in a contention group
2901 extern int __kmp_cg_max_nth;
2902 extern int __kmp_teams_max_nth; // max threads used in a teams construct
2903 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
2904  __kmp_root */
2905 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
2906  region a la OMP_NUM_THREADS */
2907 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
2908  initialization */
2909 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
2910  used (fixed) */
2911 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
2912  (__kmpc_threadprivate_cached()) */
2913 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
2914  blocking (env setting) */
2915 #if KMP_USE_MONITOR
2916 extern int
2917  __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
2918 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
2919  blocking */
2920 #endif
2921 #ifdef KMP_ADJUST_BLOCKTIME
2922 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
2923 #endif /* KMP_ADJUST_BLOCKTIME */
2924 #ifdef KMP_DFLT_NTH_CORES
2925 extern int __kmp_ncores; /* Total number of cores for threads placement */
2926 #endif
2927 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
2928 extern int __kmp_abort_delay;
2929 
2930 extern int __kmp_need_register_atfork_specified;
2931 extern int
2932  __kmp_need_register_atfork; /* At initialization, call pthread_atfork to
2933  install fork handler */
2934 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
2935  0 - not set, will be set at runtime
2936  1 - using stack search
2937  2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
2938  X*) or TlsGetValue(Windows* OS))
2939  3 - static TLS (__declspec(thread) __kmp_gtid),
2940  Linux* OS .so only. */
2941 extern int
2942  __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
2943 #ifdef KMP_TDATA_GTID
2944 extern KMP_THREAD_LOCAL int __kmp_gtid;
2945 #endif
2946 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
2947 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
2948 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2949 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
2950 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
2951 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
2952 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2953 
2954 // max_active_levels for nested parallelism enabled by default via
2955 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
2956 extern int __kmp_dflt_max_active_levels;
2957 // Indicates whether value of __kmp_dflt_max_active_levels was already
2958 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
2959 extern bool __kmp_dflt_max_active_levels_set;
2960 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
2961  concurrent execution per team */
2962 #if KMP_NESTED_HOT_TEAMS
2963 extern int __kmp_hot_teams_mode;
2964 extern int __kmp_hot_teams_max_level;
2965 #endif
2966 
2967 #if KMP_OS_LINUX
2968 extern enum clock_function_type __kmp_clock_function;
2969 extern int __kmp_clock_function_param;
2970 #endif /* KMP_OS_LINUX */
2971 
2972 #if KMP_MIC_SUPPORTED
2973 extern enum mic_type __kmp_mic_type;
2974 #endif
2975 
2976 #ifdef USE_LOAD_BALANCE
2977 extern double __kmp_load_balance_interval; // load balance algorithm interval
2978 #endif /* USE_LOAD_BALANCE */
2979 
2980 // OpenMP 3.1 - Nested num threads array
2981 typedef struct kmp_nested_nthreads_t {
2982  int *nth;
2983  int size;
2984  int used;
2985 } kmp_nested_nthreads_t;
2986 
2987 extern kmp_nested_nthreads_t __kmp_nested_nth;
2988 
2989 #if KMP_USE_ADAPTIVE_LOCKS
2990 
2991 // Parameters for the speculative lock backoff system.
2992 struct kmp_adaptive_backoff_params_t {
2993  // Number of soft retries before it counts as a hard retry.
2994  kmp_uint32 max_soft_retries;
2995  // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
2996  // the right
2997  kmp_uint32 max_badness;
2998 };
2999 
3000 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3001 
3002 #if KMP_DEBUG_ADAPTIVE_LOCKS
3003 extern const char *__kmp_speculative_statsfile;
3004 #endif
3005 
3006 #endif // KMP_USE_ADAPTIVE_LOCKS
3007 
3008 #if OMP_40_ENABLED
3009 extern int __kmp_display_env; /* TRUE or FALSE */
3010 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3011 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3012 #endif
3013 
3014 /* ------------------------------------------------------------------------- */
3015 
3016 /* the following are protected by the fork/join lock */
3017 /* write: lock read: anytime */
3018 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3019 /* read/write: lock */
3020 extern volatile kmp_team_t *__kmp_team_pool;
3021 extern volatile kmp_info_t *__kmp_thread_pool;
3022 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3023 
3024 // total num threads reachable from some root thread including all root threads
3025 extern volatile int __kmp_nth;
3026 /* total number of threads reachable from some root thread including all root
3027  threads, and those in the thread pool */
3028 extern volatile int __kmp_all_nth;
3029 extern std::atomic<int> __kmp_thread_pool_active_nth;
3030 
3031 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3032 /* end data protected by fork/join lock */
3033 /* ------------------------------------------------------------------------- */
3034 
3035 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3036 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3037 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3038 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3039 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3040 
3041 // AT: Which way is correct?
3042 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3043 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3044 #define __kmp_get_team_num_threads(gtid) \
3045  (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3046 
3047 static inline bool KMP_UBER_GTID(int gtid) {
3048  KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3049  KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3050  return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3051  __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3052 }
3053 
3054 static inline int __kmp_tid_from_gtid(int gtid) {
3055  KMP_DEBUG_ASSERT(gtid >= 0);
3056  return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3057 }
3058 
3059 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3060  KMP_DEBUG_ASSERT(tid >= 0 && team);
3061  return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3062 }
3063 
3064 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3065  KMP_DEBUG_ASSERT(thr);
3066  return thr->th.th_info.ds.ds_gtid;
3067 }
3068 
3069 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3070  KMP_DEBUG_ASSERT(gtid >= 0);
3071  return __kmp_threads[gtid];
3072 }
3073 
3074 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3075  KMP_DEBUG_ASSERT(gtid >= 0);
3076  return __kmp_threads[gtid]->th.th_team;
3077 }
3078 
3079 /* ------------------------------------------------------------------------- */
3080 
3081 extern kmp_global_t __kmp_global; /* global status */
3082 
3083 extern kmp_info_t __kmp_monitor;
3084 // For Debugging Support Library
3085 extern std::atomic<kmp_int32> __kmp_team_counter;
3086 // For Debugging Support Library
3087 extern std::atomic<kmp_int32> __kmp_task_counter;
3088 
3089 #if USE_DEBUGGER
3090 #define _KMP_GEN_ID(counter) \
3091  (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3092 #else
3093 #define _KMP_GEN_ID(counter) (~0)
3094 #endif /* USE_DEBUGGER */
3095 
3096 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3097 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3098 
3099 /* ------------------------------------------------------------------------ */
3100 
3101 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3102  size_t size, char const *format, ...);
3103 
3104 extern void __kmp_serial_initialize(void);
3105 extern void __kmp_middle_initialize(void);
3106 extern void __kmp_parallel_initialize(void);
3107 
3108 extern void __kmp_internal_begin(void);
3109 extern void __kmp_internal_end_library(int gtid);
3110 extern void __kmp_internal_end_thread(int gtid);
3111 extern void __kmp_internal_end_atexit(void);
3112 extern void __kmp_internal_end_fini(void);
3113 extern void __kmp_internal_end_dtor(void);
3114 extern void __kmp_internal_end_dest(void *);
3115 
3116 extern int __kmp_register_root(int initial_thread);
3117 extern void __kmp_unregister_root(int gtid);
3118 
3119 extern int __kmp_ignore_mppbeg(void);
3120 extern int __kmp_ignore_mppend(void);
3121 
3122 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3123 extern void __kmp_exit_single(int gtid);
3124 
3125 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3126 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3127 
3128 #ifdef USE_LOAD_BALANCE
3129 extern int __kmp_get_load_balance(int);
3130 #endif
3131 
3132 extern int __kmp_get_global_thread_id(void);
3133 extern int __kmp_get_global_thread_id_reg(void);
3134 extern void __kmp_exit_thread(int exit_status);
3135 extern void __kmp_abort(char const *format, ...);
3136 extern void __kmp_abort_thread(void);
3137 KMP_NORETURN extern void __kmp_abort_process(void);
3138 extern void __kmp_warn(char const *format, ...);
3139 
3140 extern void __kmp_set_num_threads(int new_nth, int gtid);
3141 
3142 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3143 // registered.
3144 static inline kmp_info_t *__kmp_entry_thread() {
3145  int gtid = __kmp_entry_gtid();
3146 
3147  return __kmp_threads[gtid];
3148 }
3149 
3150 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3151 extern int __kmp_get_max_active_levels(int gtid);
3152 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3153 extern int __kmp_get_team_size(int gtid, int level);
3154 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3155 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3156 
3157 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3158 extern void __kmp_init_random(kmp_info_t *thread);
3159 
3160 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3161 extern void __kmp_adjust_num_threads(int new_nproc);
3162 
3163 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3164 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3165 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3166 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3167 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3168 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3169 
3170 #if USE_FAST_MEMORY
3171 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3172  size_t size KMP_SRC_LOC_DECL);
3173 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3174 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3175 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3176 #define __kmp_fast_allocate(this_thr, size) \
3177  ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3178 #define __kmp_fast_free(this_thr, ptr) \
3179  ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3180 #endif
3181 
3182 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3183 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3184  size_t elsize KMP_SRC_LOC_DECL);
3185 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3186  size_t size KMP_SRC_LOC_DECL);
3187 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3188 #define __kmp_thread_malloc(th, size) \
3189  ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3190 #define __kmp_thread_calloc(th, nelem, elsize) \
3191  ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3192 #define __kmp_thread_realloc(th, ptr, size) \
3193  ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3194 #define __kmp_thread_free(th, ptr) \
3195  ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3196 
3197 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3198 #define KMP_INTERNAL_FREE(p) free(p)
3199 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3200 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3201 
3202 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3203 
3204 #if OMP_40_ENABLED
3205 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3206  kmp_proc_bind_t proc_bind);
3207 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3208  int num_threads);
3209 #endif
3210 
3211 extern void __kmp_yield();
3212 
3213 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3214  enum sched_type schedule, kmp_int32 lb,
3215  kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3216 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3217  enum sched_type schedule, kmp_uint32 lb,
3218  kmp_uint32 ub, kmp_int32 st,
3219  kmp_int32 chunk);
3220 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3221  enum sched_type schedule, kmp_int64 lb,
3222  kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3223 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3224  enum sched_type schedule, kmp_uint64 lb,
3225  kmp_uint64 ub, kmp_int64 st,
3226  kmp_int64 chunk);
3227 
3228 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3229  kmp_int32 *p_last, kmp_int32 *p_lb,
3230  kmp_int32 *p_ub, kmp_int32 *p_st);
3231 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3232  kmp_int32 *p_last, kmp_uint32 *p_lb,
3233  kmp_uint32 *p_ub, kmp_int32 *p_st);
3234 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3235  kmp_int32 *p_last, kmp_int64 *p_lb,
3236  kmp_int64 *p_ub, kmp_int64 *p_st);
3237 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3238  kmp_int32 *p_last, kmp_uint64 *p_lb,
3239  kmp_uint64 *p_ub, kmp_int64 *p_st);
3240 
3241 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3242 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3243 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3244 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3245 
3246 #ifdef KMP_GOMP_COMPAT
3247 
3248 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3249  enum sched_type schedule, kmp_int32 lb,
3250  kmp_int32 ub, kmp_int32 st,
3251  kmp_int32 chunk, int push_ws);
3252 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3253  enum sched_type schedule, kmp_uint32 lb,
3254  kmp_uint32 ub, kmp_int32 st,
3255  kmp_int32 chunk, int push_ws);
3256 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3257  enum sched_type schedule, kmp_int64 lb,
3258  kmp_int64 ub, kmp_int64 st,
3259  kmp_int64 chunk, int push_ws);
3260 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3261  enum sched_type schedule, kmp_uint64 lb,
3262  kmp_uint64 ub, kmp_int64 st,
3263  kmp_int64 chunk, int push_ws);
3264 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3265 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3266 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3267 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3268 
3269 #endif /* KMP_GOMP_COMPAT */
3270 
3271 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3272 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3273 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3274 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3275 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3276 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3277  kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3278  void *obj);
3279 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3280  kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3281 
3282 class kmp_flag_32;
3283 class kmp_flag_64;
3284 class kmp_flag_oncore;
3285 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64 *flag,
3286  int final_spin
3287 #if USE_ITT_BUILD
3288  ,
3289  void *itt_sync_obj
3290 #endif
3291  );
3292 extern void __kmp_release_64(kmp_flag_64 *flag);
3293 
3294 extern void __kmp_infinite_loop(void);
3295 
3296 extern void __kmp_cleanup(void);
3297 
3298 #if KMP_HANDLE_SIGNALS
3299 extern int __kmp_handle_signals;
3300 extern void __kmp_install_signals(int parallel_init);
3301 extern void __kmp_remove_signals(void);
3302 #endif
3303 
3304 extern void __kmp_clear_system_time(void);
3305 extern void __kmp_read_system_time(double *delta);
3306 
3307 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3308 
3309 extern void __kmp_expand_host_name(char *buffer, size_t size);
3310 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3311 
3312 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3313 extern void
3314 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3315 #endif
3316 
3317 extern void
3318 __kmp_runtime_initialize(void); /* machine specific initialization */
3319 extern void __kmp_runtime_destroy(void);
3320 
3321 #if KMP_AFFINITY_SUPPORTED
3322 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3323  kmp_affin_mask_t *mask);
3324 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3325  kmp_affin_mask_t *mask);
3326 extern void __kmp_affinity_initialize(void);
3327 extern void __kmp_affinity_uninitialize(void);
3328 extern void __kmp_affinity_set_init_mask(
3329  int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3330 #if OMP_40_ENABLED
3331 extern void __kmp_affinity_set_place(int gtid);
3332 #endif
3333 extern void __kmp_affinity_determine_capable(const char *env_var);
3334 extern int __kmp_aux_set_affinity(void **mask);
3335 extern int __kmp_aux_get_affinity(void **mask);
3336 extern int __kmp_aux_get_affinity_max_proc();
3337 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3338 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3339 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3340 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3341 #if KMP_OS_LINUX
3342 extern int kmp_set_thread_affinity_mask_initial(void);
3343 #endif
3344 #endif /* KMP_AFFINITY_SUPPORTED */
3345 #if OMP_50_ENABLED
3346 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3347 // format string is for affinity, so platforms that do not support
3348 // affinity can still use the other fields, e.g., %n for num_threads
3349 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3350  kmp_str_buf_t *buffer);
3351 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3352 #endif
3353 
3354 extern void __kmp_cleanup_hierarchy();
3355 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3356 
3357 #if KMP_USE_FUTEX
3358 
3359 extern int __kmp_futex_determine_capable(void);
3360 
3361 #endif // KMP_USE_FUTEX
3362 
3363 extern void __kmp_gtid_set_specific(int gtid);
3364 extern int __kmp_gtid_get_specific(void);
3365 
3366 extern double __kmp_read_cpu_time(void);
3367 
3368 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3369 
3370 #if KMP_USE_MONITOR
3371 extern void __kmp_create_monitor(kmp_info_t *th);
3372 #endif
3373 
3374 extern void *__kmp_launch_thread(kmp_info_t *thr);
3375 
3376 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3377 
3378 #if KMP_OS_WINDOWS
3379 extern int __kmp_still_running(kmp_info_t *th);
3380 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3381 extern void __kmp_free_handle(kmp_thread_t tHandle);
3382 #endif
3383 
3384 #if KMP_USE_MONITOR
3385 extern void __kmp_reap_monitor(kmp_info_t *th);
3386 #endif
3387 extern void __kmp_reap_worker(kmp_info_t *th);
3388 extern void __kmp_terminate_thread(int gtid);
3389 
3390 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3391 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3392 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3393 
3394 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag);
3395 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag);
3396 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
3397 extern void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag);
3398 extern void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag);
3399 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
3400 
3401 extern void __kmp_elapsed(double *);
3402 extern void __kmp_elapsed_tick(double *);
3403 
3404 extern void __kmp_enable(int old_state);
3405 extern void __kmp_disable(int *old_state);
3406 
3407 extern void __kmp_thread_sleep(int millis);
3408 
3409 extern void __kmp_common_initialize(void);
3410 extern void __kmp_common_destroy(void);
3411 extern void __kmp_common_destroy_gtid(int gtid);
3412 
3413 #if KMP_OS_UNIX
3414 extern void __kmp_register_atfork(void);
3415 #endif
3416 extern void __kmp_suspend_initialize(void);
3417 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3418 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3419 
3420 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3421  int tid);
3422 #if OMP_40_ENABLED
3423 extern kmp_team_t *
3424 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3425 #if OMPT_SUPPORT
3426  ompt_data_t ompt_parallel_data,
3427 #endif
3428  kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3429  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3430 #else
3431 extern kmp_team_t *
3432 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3433 #if OMPT_SUPPORT
3434  ompt_id_t ompt_parallel_id,
3435 #endif
3436  kmp_internal_control_t *new_icvs,
3437  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3438 #endif // OMP_40_ENABLED
3439 extern void __kmp_free_thread(kmp_info_t *);
3440 extern void __kmp_free_team(kmp_root_t *,
3441  kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3442 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3443 
3444 /* ------------------------------------------------------------------------ */
3445 
3446 extern void __kmp_initialize_bget(kmp_info_t *th);
3447 extern void __kmp_finalize_bget(kmp_info_t *th);
3448 
3449 KMP_EXPORT void *kmpc_malloc(size_t size);
3450 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3451 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3452 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3453 KMP_EXPORT void kmpc_free(void *ptr);
3454 
3455 /* declarations for internal use */
3456 
3457 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3458  size_t reduce_size, void *reduce_data,
3459  void (*reduce)(void *, void *));
3460 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3461 extern int __kmp_barrier_gomp_cancel(int gtid);
3462 
3467 enum fork_context_e {
3468  fork_context_gnu,
3470  fork_context_intel,
3471  fork_context_last
3472 };
3473 extern int __kmp_fork_call(ident_t *loc, int gtid,
3474  enum fork_context_e fork_context, kmp_int32 argc,
3475  microtask_t microtask, launch_t invoker,
3476 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
3477 #if (KMP_ARCH_ARM || KMP_ARCH_X86_64 || KMP_ARCH_AARCH64) && KMP_OS_LINUX
3478  va_list *ap
3479 #else
3480  va_list ap
3481 #endif
3482  );
3483 
3484 extern void __kmp_join_call(ident_t *loc, int gtid
3485 #if OMPT_SUPPORT
3486  ,
3487  enum fork_context_e fork_context
3488 #endif
3489 #if OMP_40_ENABLED
3490  ,
3491  int exit_teams = 0
3492 #endif
3493  );
3494 
3495 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3496 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3497 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3498 extern int __kmp_invoke_task_func(int gtid);
3499 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3500  kmp_info_t *this_thr,
3501  kmp_team_t *team);
3502 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3503  kmp_info_t *this_thr,
3504  kmp_team_t *team);
3505 
3506 // should never have been exported
3507 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3508 #if OMP_40_ENABLED
3509 extern int __kmp_invoke_teams_master(int gtid);
3510 extern void __kmp_teams_master(int gtid);
3511 extern int __kmp_aux_get_team_num();
3512 extern int __kmp_aux_get_num_teams();
3513 #endif
3514 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3515 extern void __kmp_user_set_library(enum library_type arg);
3516 extern void __kmp_aux_set_library(enum library_type arg);
3517 extern void __kmp_aux_set_stacksize(size_t arg);
3518 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3519 extern void __kmp_aux_set_defaults(char const *str, int len);
3520 
3521 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3522 void kmpc_set_blocktime(int arg);
3523 void ompc_set_nested(int flag);
3524 void ompc_set_dynamic(int flag);
3525 void ompc_set_num_threads(int arg);
3526 
3527 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3528  kmp_team_t *team, int tid);
3529 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3530 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3531  kmp_tasking_flags_t *flags,
3532  size_t sizeof_kmp_task_t,
3533  size_t sizeof_shareds,
3534  kmp_routine_entry_t task_entry);
3535 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3536  kmp_team_t *team, int tid,
3537  int set_curr_task);
3538 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3539 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3540 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
3541  kmp_flag_32 *flag, int final_spin,
3542  int *thread_finished,
3543 #if USE_ITT_BUILD
3544  void *itt_sync_obj,
3545 #endif /* USE_ITT_BUILD */
3546  kmp_int32 is_constrained);
3547 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
3548  kmp_flag_64 *flag, int final_spin,
3549  int *thread_finished,
3550 #if USE_ITT_BUILD
3551  void *itt_sync_obj,
3552 #endif /* USE_ITT_BUILD */
3553  kmp_int32 is_constrained);
3554 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
3555  kmp_flag_oncore *flag, int final_spin,
3556  int *thread_finished,
3557 #if USE_ITT_BUILD
3558  void *itt_sync_obj,
3559 #endif /* USE_ITT_BUILD */
3560  kmp_int32 is_constrained);
3561 
3562 extern void __kmp_free_task_team(kmp_info_t *thread,
3563  kmp_task_team_t *task_team);
3564 extern void __kmp_reap_task_teams(void);
3565 extern void __kmp_wait_to_unref_task_teams(void);
3566 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3567  int always);
3568 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3569 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3570 #if USE_ITT_BUILD
3571  ,
3572  void *itt_sync_obj
3573 #endif /* USE_ITT_BUILD */
3574  ,
3575  int wait = 1);
3576 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3577  int gtid);
3578 
3579 extern int __kmp_is_address_mapped(void *addr);
3580 extern kmp_uint64 __kmp_hardware_timestamp(void);
3581 
3582 #if KMP_OS_UNIX
3583 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3584 #endif
3585 
3586 /* ------------------------------------------------------------------------ */
3587 //
3588 // Assembly routines that have no compiler intrinsic replacement
3589 //
3590 
3591 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3592 
3593 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
3594 
3595 #define __kmp_load_mxcsr(p) _mm_setcsr(*(p))
3596 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
3597 
3598 extern void __kmp_load_x87_fpu_control_word(kmp_int16 *p);
3599 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
3600 extern void __kmp_clear_x87_fpu_status_word();
3601 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
3602 
3603 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3604 
3605 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3606  void *argv[]
3607 #if OMPT_SUPPORT
3608  ,
3609  void **exit_frame_ptr
3610 #endif
3611  );
3612 
3613 /* ------------------------------------------------------------------------ */
3614 
3615 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3616 KMP_EXPORT void __kmpc_end(ident_t *);
3617 
3618 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3619  kmpc_ctor_vec ctor,
3620  kmpc_cctor_vec cctor,
3621  kmpc_dtor_vec dtor,
3622  size_t vector_length);
3623 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3624  kmpc_ctor ctor, kmpc_cctor cctor,
3625  kmpc_dtor dtor);
3626 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3627  void *data, size_t size);
3628 
3629 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3630 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3631 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3632 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3633 
3634 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3635 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3636  kmpc_micro microtask, ...);
3637 
3638 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3639 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3640 
3641 KMP_EXPORT void __kmpc_flush(ident_t *);
3642 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3643 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3644 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3645 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3646 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3647 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3648  kmp_critical_name *);
3649 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3650  kmp_critical_name *);
3651 
3652 #if OMP_45_ENABLED
3653 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3654  kmp_critical_name *, uint32_t hint);
3655 #endif
3656 
3657 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3658 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3659 
3660 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3661  kmp_int32 global_tid);
3662 
3663 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3664 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3665 
3666 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3667  kmp_int32 schedtype, kmp_int32 *plastiter,
3668  kmp_int *plower, kmp_int *pupper,
3669  kmp_int *pstride, kmp_int incr,
3670  kmp_int chunk);
3671 
3672 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3673 
3674 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3675  size_t cpy_size, void *cpy_data,
3676  void (*cpy_func)(void *, void *),
3677  kmp_int32 didit);
3678 
3679 extern void KMPC_SET_NUM_THREADS(int arg);
3680 extern void KMPC_SET_DYNAMIC(int flag);
3681 extern void KMPC_SET_NESTED(int flag);
3682 
3683 /* OMP 3.0 tasking interface routines */
3684 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3685  kmp_task_t *new_task);
3686 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3687  kmp_int32 flags,
3688  size_t sizeof_kmp_task_t,
3689  size_t sizeof_shareds,
3690  kmp_routine_entry_t task_entry);
3691 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3692  kmp_task_t *task);
3693 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3694  kmp_task_t *task);
3695 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3696  kmp_task_t *new_task);
3697 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3698 
3699 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3700  int end_part);
3701 
3702 #if TASK_UNUSED
3703 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3704 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3705  kmp_task_t *task);
3706 #endif // TASK_UNUSED
3707 
3708 /* ------------------------------------------------------------------------ */
3709 
3710 #if OMP_40_ENABLED
3711 
3712 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3713 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3714 
3715 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3716  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3717  kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3718  kmp_depend_info_t *noalias_dep_list);
3719 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3720  kmp_int32 ndeps,
3721  kmp_depend_info_t *dep_list,
3722  kmp_int32 ndeps_noalias,
3723  kmp_depend_info_t *noalias_dep_list);
3724 
3725 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3726  bool serialize_immediate);
3727 
3728 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3729  kmp_int32 cncl_kind);
3730 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3731  kmp_int32 cncl_kind);
3732 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3733 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3734 
3735 #if OMP_45_ENABLED
3736 
3737 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3738 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3739 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3740  kmp_int32 if_val, kmp_uint64 *lb,
3741  kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3742  kmp_int32 sched, kmp_uint64 grainsize,
3743  void *task_dup);
3744 #endif
3745 #if OMP_50_ENABLED
3746 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3747 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3748 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
3749  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3750  kmp_task_affinity_info_t *affin_list);
3751 #endif
3752 
3753 #endif
3754 
3755 /* Lock interface routines (fast versions with gtid passed in) */
3756 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3757  void **user_lock);
3758 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3759  void **user_lock);
3760 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3761  void **user_lock);
3762 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3763  void **user_lock);
3764 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3765 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3766  void **user_lock);
3767 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3768  void **user_lock);
3769 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3770  void **user_lock);
3771 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3772 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3773  void **user_lock);
3774 
3775 #if OMP_45_ENABLED
3776 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3777  void **user_lock, uintptr_t hint);
3778 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3779  void **user_lock,
3780  uintptr_t hint);
3781 #endif
3782 
3783 /* Interface to fast scalable reduce methods routines */
3784 
3785 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3786  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3787  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3788  kmp_critical_name *lck);
3789 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3790  kmp_critical_name *lck);
3791 KMP_EXPORT kmp_int32 __kmpc_reduce(
3792  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3793  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3794  kmp_critical_name *lck);
3795 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3796  kmp_critical_name *lck);
3797 
3798 /* Internal fast reduction routines */
3799 
3800 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3801  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3802  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3803  kmp_critical_name *lck);
3804 
3805 // this function is for testing set/get/determine reduce method
3806 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3807 
3808 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3809 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3810 
3811 // C++ port
3812 // missing 'extern "C"' declarations
3813 
3814 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3815 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3816 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3817  kmp_int32 num_threads);
3818 
3819 #if OMP_40_ENABLED
3820 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3821  int proc_bind);
3822 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3823  kmp_int32 num_teams,
3824  kmp_int32 num_threads);
3825 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3826  kmpc_micro microtask, ...);
3827 #endif
3828 #if OMP_45_ENABLED
3829 struct kmp_dim { // loop bounds info casted to kmp_int64
3830  kmp_int64 lo; // lower
3831  kmp_int64 up; // upper
3832  kmp_int64 st; // stride
3833 };
3834 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
3835  kmp_int32 num_dims,
3836  const struct kmp_dim *dims);
3837 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
3838  const kmp_int64 *vec);
3839 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
3840  const kmp_int64 *vec);
3841 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
3842 #endif
3843 
3844 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
3845  void *data, size_t size,
3846  void ***cache);
3847 
3848 // Symbols for MS mutual detection.
3849 extern int _You_must_link_with_exactly_one_OpenMP_library;
3850 extern int _You_must_link_with_Intel_OpenMP_library;
3851 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3852 extern int _You_must_link_with_Microsoft_OpenMP_library;
3853 #endif
3854 
3855 // The routines below are not exported.
3856 // Consider making them 'static' in corresponding source files.
3857 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
3858  void *data_addr, size_t pc_size);
3859 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
3860  void *data_addr,
3861  size_t pc_size);
3862 void __kmp_threadprivate_resize_cache(int newCapacity);
3863 void __kmp_cleanup_threadprivate_caches();
3864 
3865 // ompc_, kmpc_ entries moved from omp.h.
3866 #if KMP_OS_WINDOWS
3867 #define KMPC_CONVENTION __cdecl
3868 #else
3869 #define KMPC_CONVENTION
3870 #endif
3871 
3872 #ifndef __OMP_H
3873 typedef enum omp_sched_t {
3874  omp_sched_static = 1,
3875  omp_sched_dynamic = 2,
3876  omp_sched_guided = 3,
3877  omp_sched_auto = 4
3878 } omp_sched_t;
3879 typedef void *kmp_affinity_mask_t;
3880 #endif
3881 
3882 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
3883 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
3884 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
3885 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
3886 KMP_EXPORT int KMPC_CONVENTION
3887 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
3888 KMP_EXPORT int KMPC_CONVENTION
3889 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
3890 KMP_EXPORT int KMPC_CONVENTION
3891 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
3892 
3893 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
3894 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
3895 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
3896 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
3897 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
3898 
3899 #if OMP_50_ENABLED
3900 enum kmp_target_offload_kind {
3901  tgt_disabled = 0,
3902  tgt_default = 1,
3903  tgt_mandatory = 2
3904 };
3905 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
3906 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
3907 extern kmp_target_offload_kind_t __kmp_target_offload;
3908 extern int __kmpc_get_target_offload();
3909 #endif
3910 
3911 #if OMP_40_ENABLED
3912 // Constants used in libomptarget
3913 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
3914 #define KMP_HOST_DEVICE -10 // This is what it is in libomptarget, go figure.
3915 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
3916 #endif // OMP_40_ENABLED
3917 
3918 #if OMP_50_ENABLED
3919 // OMP Pause Resource
3920 
3921 // The following enum is used both to set the status in __kmp_pause_status, and
3922 // as the internal equivalent of the externally-visible omp_pause_resource_t.
3923 typedef enum kmp_pause_status_t {
3924  kmp_not_paused = 0, // status is not paused, or, requesting resume
3925  kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
3926  kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
3927 } kmp_pause_status_t;
3928 
3929 // This stores the pause state of the runtime
3930 extern kmp_pause_status_t __kmp_pause_status;
3931 extern int __kmpc_pause_resource(kmp_pause_status_t level);
3932 extern int __kmp_pause_resource(kmp_pause_status_t level);
3933 // Soft resume sets __kmp_pause_status, and wakes up all threads.
3934 extern void __kmp_resume_if_soft_paused();
3935 // Hard resume simply resets the status to not paused. Library will appear to
3936 // be uninitialized after hard pause. Let OMP constructs trigger required
3937 // initializations.
3938 static inline void __kmp_resume_if_hard_paused() {
3939  if (__kmp_pause_status == kmp_hard_paused) {
3940  __kmp_pause_status = kmp_not_paused;
3941  }
3942 }
3943 #endif // OMP_50_ENABLED
3944 
3945 #ifdef __cplusplus
3946 }
3947 #endif
3948 
3949 #endif /* KMP_H */
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:227
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:1360
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:1383
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:224
void *(* kmpc_ctor_vec)(void *, size_t)
Definition: kmp.h:1377
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:232
void *(* kmpc_cctor_vec)(void *, void *, size_t)
Definition: kmp.h:1389
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)
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:1367
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 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_ordered(ident_t *, kmp_int32 global_tid)
sched_type
Definition: kmp.h:336
KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
Definition: kmp.h:223
KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid)
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 *)
void *(* kmpc_ctor)(void *)
Definition: kmp.h:1354
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)
void(* kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
Definition: kmp.h:1336
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:233
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:225
struct ident ident_t