LLVM OpenMP* Runtime Library
Loading...
Searching...
No Matches
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#define KMP_WEIGHTED_ITERATIONS_SUPPORTED \
31 (KMP_AFFINITY_SUPPORTED && KMP_STATIC_STEAL_ENABLED && \
32 (KMP_ARCH_X86 || KMP_ARCH_X86_64))
33
34#define TASK_CURRENT_NOT_QUEUED 0
35#define TASK_CURRENT_QUEUED 1
36
37#define TASK_NOT_PUSHED 1
38#define TASK_SUCCESSFULLY_PUSHED 0
39#define TASK_TIED 1
40#define TASK_UNTIED 0
41#define TASK_EXPLICIT 1
42#define TASK_IMPLICIT 0
43#define TASK_PROXY 1
44#define TASK_FULL 0
45#define TASK_DETACHABLE 1
46#define TASK_UNDETACHABLE 0
47
48#define KMP_THREAD_ATTR
49
50#if !KMP_OS_WASI
51#include <signal.h>
52#endif
53#include <stdarg.h>
54#include <stddef.h>
55#include <stdio.h>
56#include <stdlib.h>
57#include <string.h>
58#include <limits>
59#include <type_traits>
60/* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
61 Microsoft library. Some macros provided below to replace these functions */
62#ifndef __ABSOFT_WIN
63#include <sys/types.h>
64#endif
65#include <limits.h>
66#include <time.h>
67
68#include <errno.h>
69
70#include "kmp_os.h"
71
72#include "kmp_safe_c_api.h"
73
74#if KMP_STATS_ENABLED
75class kmp_stats_list;
76#endif
77
78#if KMP_USE_HIER_SCHED
79// Only include hierarchical scheduling if affinity is supported
80#undef KMP_USE_HIER_SCHED
81#define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
82#endif
83
84// OMPD_SKIP_HWLOC used in libompd/omp-icv.cpp to avoid OMPD depending on hwloc
85#if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED && !defined(OMPD_SKIP_HWLOC)
86#include "hwloc.h"
87#define KMP_HWLOC_ENABLED 1
88#ifndef HWLOC_OBJ_NUMANODE
89#define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
90#endif
91#ifndef HWLOC_OBJ_PACKAGE
92#define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
93#endif
94#else
95#define KMP_HWLOC_ENABLED 0
96#endif
97
98#if KMP_ARCH_X86 || KMP_ARCH_X86_64
99#include <xmmintrin.h>
100#endif
101
102// The below has to be defined before including "kmp_barrier.h".
103#define KMP_INTERNAL_MALLOC(sz) malloc(sz)
104#define KMP_INTERNAL_FREE(p) free(p)
105#define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
106#define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
107
108#include "kmp_debug.h"
109#include "kmp_lock.h"
110#include "kmp_version.h"
111#include "kmp_barrier.h"
112#if USE_DEBUGGER
113#include "kmp_debugger.h"
114#endif
115#include "kmp_i18n.h"
116
117#define KMP_HANDLE_SIGNALS ((KMP_OS_UNIX && !KMP_OS_WASI) || KMP_OS_WINDOWS)
118
119#include "kmp_wrapper_malloc.h"
120#if KMP_OS_UNIX
121#include <unistd.h>
122#if !defined NSIG && defined _NSIG
123#define NSIG _NSIG
124#endif
125#endif
126
127#if KMP_OS_LINUX
128#pragma weak clock_gettime
129#endif
130
131#if OMPT_SUPPORT
132#include "ompt-internal.h"
133#endif
134
135#if OMPD_SUPPORT
136#include "ompd-specific.h"
137#endif
138
139#ifndef UNLIKELY
140#define UNLIKELY(x) (x)
141#endif
142
143// Affinity format function
144#include "kmp_str.h"
145
146// 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
147// 3 - fast allocation using sync, non-sync free lists of any size, non-self
148// free lists of limited size.
149#ifndef USE_FAST_MEMORY
150#define USE_FAST_MEMORY 3
151#endif
152
153// Assume using BGET compare_exchange instruction instead of lock by default.
154#ifndef USE_CMP_XCHG_FOR_BGET
155#define USE_CMP_XCHG_FOR_BGET 1
156#endif
157
158// Test to see if queuing lock is better than bootstrap lock for bget
159// #ifndef USE_QUEUING_LOCK_FOR_BGET
160// #define USE_QUEUING_LOCK_FOR_BGET
161// #endif
162
163#define KMP_NSEC_PER_SEC 1000000000L
164#define KMP_USEC_PER_SEC 1000000L
165#define KMP_NSEC_PER_USEC 1000L
166
171
175enum {
180 /* 0x04 is no longer used */
189 KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
190 KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
191 KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
192
193 KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
194 KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
195
207 KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
208 KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
209 KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
210 KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
211 KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
212};
213
217typedef struct ident {
218 kmp_int32 reserved_1;
219 kmp_int32 flags;
221 kmp_int32 reserved_2;
222#if USE_ITT_BUILD
223/* but currently used for storing region-specific ITT */
224/* contextual information. */
225#endif /* USE_ITT_BUILD */
226 kmp_int32 reserved_3;
227 char const *psource;
231 // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
232 kmp_int32 get_openmp_version() {
233 return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
234 }
239
240// Some forward declarations.
241typedef union kmp_team kmp_team_t;
242typedef struct kmp_taskdata kmp_taskdata_t;
243typedef union kmp_task_team kmp_task_team_t;
244typedef union kmp_team kmp_team_p;
245typedef union kmp_info kmp_info_p;
246typedef union kmp_root kmp_root_p;
247
248template <bool C = false, bool S = true> class kmp_flag_32;
249template <bool C = false, bool S = true> class kmp_flag_64;
250template <bool C = false, bool S = true> class kmp_atomic_flag_64;
251class kmp_flag_oncore;
252
253#ifdef __cplusplus
254extern "C" {
255#endif
256
257/* ------------------------------------------------------------------------ */
258
259/* Pack two 32-bit signed integers into a 64-bit signed integer */
260/* ToDo: Fix word ordering for big-endian machines. */
261#define KMP_PACK_64(HIGH_32, LOW_32) \
262 ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
263
264// Generic string manipulation macros. Assume that _x is of type char *
265#define SKIP_WS(_x) \
266 { \
267 while (*(_x) == ' ' || *(_x) == '\t') \
268 (_x)++; \
269 }
270#define SKIP_DIGITS(_x) \
271 { \
272 while (*(_x) >= '0' && *(_x) <= '9') \
273 (_x)++; \
274 }
275#define SKIP_TOKEN(_x) \
276 { \
277 while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
278 (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
279 (_x)++; \
280 }
281#define SKIP_TO(_x, _c) \
282 { \
283 while (*(_x) != '\0' && *(_x) != (_c)) \
284 (_x)++; \
285 }
286
287/* ------------------------------------------------------------------------ */
288
289#define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
290#define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
291
292/* ------------------------------------------------------------------------ */
293/* Enumeration types */
294
295enum kmp_state_timer {
296 ts_stop,
297 ts_start,
298 ts_pause,
299
300 ts_last_state
301};
302
303enum dynamic_mode {
304 dynamic_default,
305#ifdef USE_LOAD_BALANCE
306 dynamic_load_balance,
307#endif /* USE_LOAD_BALANCE */
308 dynamic_random,
309 dynamic_thread_limit,
310 dynamic_max
311};
312
313/* external schedule constants, duplicate enum omp_sched in omp.h in order to
314 * not include it here */
315#ifndef KMP_SCHED_TYPE_DEFINED
316#define KMP_SCHED_TYPE_DEFINED
317typedef enum kmp_sched {
318 kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
319 // Note: need to adjust __kmp_sch_map global array in case enum is changed
320 kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
321 kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
322 kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
323 kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
324 kmp_sched_upper_std = 5, // upper bound for standard schedules
325 kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
326 kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
327#if KMP_STATIC_STEAL_ENABLED
328 kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
329#endif
330 kmp_sched_upper,
331 kmp_sched_default = kmp_sched_static, // default scheduling
332 kmp_sched_monotonic = 0x80000000
333} kmp_sched_t;
334#endif
335
340enum sched_type : kmp_int32 {
342 kmp_sch_static_chunked = 33,
344 kmp_sch_dynamic_chunked = 35,
346 kmp_sch_runtime = 37,
348 kmp_sch_trapezoidal = 39,
349
350 /* accessible only through KMP_SCHEDULE environment variable */
351 kmp_sch_static_greedy = 40,
352 kmp_sch_static_balanced = 41,
353 /* accessible only through KMP_SCHEDULE environment variable */
354 kmp_sch_guided_iterative_chunked = 42,
355 kmp_sch_guided_analytical_chunked = 43,
356 /* accessible only through KMP_SCHEDULE environment variable */
357 kmp_sch_static_steal = 44,
358
359 /* static with chunk adjustment (e.g., simd) */
360 kmp_sch_static_balanced_chunked = 45,
363
364 /* accessible only through KMP_SCHEDULE environment variable */
366
368 kmp_ord_static_chunked = 65,
370 kmp_ord_dynamic_chunked = 67,
371 kmp_ord_guided_chunked = 68,
372 kmp_ord_runtime = 69,
374 kmp_ord_trapezoidal = 71,
376
377 /* Schedules for Distribute construct */
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. */
386
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,
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,
413
414 /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
415 we need to distinguish the three possible cases (no modifier, monotonic
416 modifier, nonmonotonic modifier), we need separate bits for each modifier.
417 The absence of monotonic does not imply nonmonotonic, especially since 4.5
418 says that the behaviour of the "no modifier" case is implementation defined
419 in 4.5, but will become "nonmonotonic" in 5.0.
420
421 Since we're passing a full 32 bit value, we can use a couple of high bits
422 for these flags; out of paranoia we avoid the sign bit.
423
424 These modifiers can be or-ed into non-static schedules by the compiler to
425 pass the additional information. They will be stripped early in the
426 processing in __kmp_dispatch_init when setting up schedules, so most of the
427 code won't ever see schedules with these bits set. */
429 (1 << 29),
431 (1 << 30),
432
433#define SCHEDULE_WITHOUT_MODIFIERS(s) \
434 (enum sched_type)( \
436#define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
437#define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
438#define SCHEDULE_HAS_NO_MODIFIERS(s) \
439 (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
440#define SCHEDULE_GET_MODIFIERS(s) \
441 ((enum sched_type)( \
442 (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
443#define SCHEDULE_SET_MODIFIERS(s, m) \
444 (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
445#define SCHEDULE_NONMONOTONIC 0
446#define SCHEDULE_MONOTONIC 1
447
449};
450
451// Apply modifiers on internal kind to standard kind
452static inline void
453__kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
454 enum sched_type internal_kind) {
455 if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
456 *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
457 }
458}
459
460// Apply modifiers on standard kind to internal kind
461static inline void
462__kmp_sched_apply_mods_intkind(kmp_sched_t kind,
463 enum sched_type *internal_kind) {
464 if ((int)kind & (int)kmp_sched_monotonic) {
465 *internal_kind = (enum sched_type)((int)*internal_kind |
467 }
468}
469
470// Get standard schedule without modifiers
471static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
472 return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
473}
474
475/* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
476typedef union kmp_r_sched {
477 struct {
478 enum sched_type r_sched_type;
479 int chunk;
480 };
481 kmp_int64 sched;
482} kmp_r_sched_t;
483
484extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
485// internal schedule types
486
487enum library_type {
488 library_none,
489 library_serial,
490 library_turnaround,
491 library_throughput
492};
493
494#if KMP_MIC_SUPPORTED
495enum mic_type { non_mic, mic1, mic2, mic3, dummy };
496#endif
497
498// OpenMP 3.1 - Nested num threads array
499typedef struct kmp_nested_nthreads_t {
500 int *nth;
501 int size;
502 int used;
503} kmp_nested_nthreads_t;
504
505extern kmp_nested_nthreads_t __kmp_nested_nth;
506
507/* -- fast reduction stuff ------------------------------------------------ */
508
509#undef KMP_FAST_REDUCTION_BARRIER
510#define KMP_FAST_REDUCTION_BARRIER 1
511
512#undef KMP_FAST_REDUCTION_CORE_DUO
513#if KMP_ARCH_X86 || KMP_ARCH_X86_64
514#define KMP_FAST_REDUCTION_CORE_DUO 1
515#endif
516
517enum _reduction_method {
518 reduction_method_not_defined = 0,
519 critical_reduce_block = (1 << 8),
520 atomic_reduce_block = (2 << 8),
521 tree_reduce_block = (3 << 8),
522 empty_reduce_block = (4 << 8)
523};
524
525// Description of the packed_reduction_method variable:
526// The packed_reduction_method variable consists of two enum types variables
527// that are packed together into 0-th byte and 1-st byte:
528// 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
529// barrier that will be used in fast reduction: bs_plain_barrier or
530// bs_reduction_barrier
531// 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
532// be used in fast reduction;
533// Reduction method is of 'enum _reduction_method' type and it's defined the way
534// so that the bits of 0-th byte are empty, so no need to execute a shift
535// instruction while packing/unpacking
536
537#if KMP_FAST_REDUCTION_BARRIER
538#define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
539 ((reduction_method) | (barrier_type))
540
541#define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
542 ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
543
544#define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
545 ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
546#else
547#define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
548 (reduction_method)
549
550#define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
551 (packed_reduction_method)
552
553#define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
554#endif
555
556#define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
557 ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
558 (which_reduction_block))
559
560#if KMP_FAST_REDUCTION_BARRIER
561#define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
562 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
563
564#define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
565 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
566#endif
567
568typedef int PACKED_REDUCTION_METHOD_T;
569
570/* -- end of fast reduction stuff ----------------------------------------- */
571
572#if KMP_OS_WINDOWS
573#define USE_CBLKDATA
574#if KMP_MSVC_COMPAT
575#pragma warning(push)
576#pragma warning(disable : 271 310)
577#endif
578#include <windows.h>
579#if KMP_MSVC_COMPAT
580#pragma warning(pop)
581#endif
582#endif
583
584#if KMP_OS_UNIX
585#if !KMP_OS_WASI
586#include <dlfcn.h>
587#endif
588#include <pthread.h>
589#endif
590
591enum kmp_hw_t : int {
592 KMP_HW_UNKNOWN = -1,
593 KMP_HW_SOCKET = 0,
594 KMP_HW_PROC_GROUP,
595 KMP_HW_NUMA,
596 KMP_HW_DIE,
597 KMP_HW_LLC,
598 KMP_HW_L3,
599 KMP_HW_TILE,
600 KMP_HW_MODULE,
601 KMP_HW_L2,
602 KMP_HW_L1,
603 KMP_HW_CORE,
604 KMP_HW_THREAD,
605 KMP_HW_LAST
606};
607
608typedef enum kmp_hw_core_type_t {
609 KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
610#if KMP_ARCH_X86 || KMP_ARCH_X86_64
611 KMP_HW_CORE_TYPE_ATOM = 0x20,
612 KMP_HW_CORE_TYPE_CORE = 0x40,
613 KMP_HW_MAX_NUM_CORE_TYPES = 3,
614#else
615 KMP_HW_MAX_NUM_CORE_TYPES = 1,
616#endif
617} kmp_hw_core_type_t;
618
619#define KMP_HW_MAX_NUM_CORE_EFFS 8
620
621#define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \
622 KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
623#define KMP_ASSERT_VALID_HW_TYPE(type) \
624 KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
625
626#define KMP_FOREACH_HW_TYPE(type) \
627 for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \
628 type = (kmp_hw_t)((int)type + 1))
629
630const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
631const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
632const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
633
634/* Only Linux* OS and Windows* OS support thread affinity. */
635#if KMP_AFFINITY_SUPPORTED
636
637// GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
638#if KMP_OS_WINDOWS
639#if _MSC_VER < 1600 && KMP_MSVC_COMPAT
640typedef struct GROUP_AFFINITY {
641 KAFFINITY Mask;
642 WORD Group;
643 WORD Reserved[3];
644} GROUP_AFFINITY;
645#endif /* _MSC_VER < 1600 */
646#if KMP_GROUP_AFFINITY
647extern int __kmp_num_proc_groups;
648#else
649static const int __kmp_num_proc_groups = 1;
650#endif /* KMP_GROUP_AFFINITY */
651typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
652extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
653
654typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
655extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
656
657typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
658extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
659
660typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
661 GROUP_AFFINITY *);
662extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
663#endif /* KMP_OS_WINDOWS */
664
665#if KMP_HWLOC_ENABLED
666extern hwloc_topology_t __kmp_hwloc_topology;
667extern int __kmp_hwloc_error;
668#endif // KMP_HWLOC_ENABLED
669
670extern size_t __kmp_affin_mask_size;
671#define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
672#define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
673#define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
674#define KMP_CPU_SET_ITERATE(i, mask) \
675 for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
676#define KMP_CPU_SET(i, mask) (mask)->set(i)
677#define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
678#define KMP_CPU_CLR(i, mask) (mask)->clear(i)
679#define KMP_CPU_ZERO(mask) (mask)->zero()
680#define KMP_CPU_ISEMPTY(mask) (mask)->empty()
681#define KMP_CPU_COPY(dest, src) (dest)->copy(src)
682#define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
683#define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
684#define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
685#define KMP_CPU_EQUAL(dest, src) (dest)->is_equal(src)
686#define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
687#define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
688#define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
689#define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
690#define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
691#define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
692#define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
693#define KMP_CPU_ALLOC_ARRAY(arr, n) \
694 (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
695#define KMP_CPU_FREE_ARRAY(arr, n) \
696 __kmp_affinity_dispatch->deallocate_mask_array(arr)
697#define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
698#define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
699#define __kmp_get_system_affinity(mask, abort_bool) \
700 (mask)->get_system_affinity(abort_bool)
701#define __kmp_set_system_affinity(mask, abort_bool) \
702 (mask)->set_system_affinity(abort_bool)
703#define __kmp_get_proc_group(mask) (mask)->get_proc_group()
704
705class KMPAffinity {
706public:
707 class Mask {
708 public:
709 void *operator new(size_t n);
710 void operator delete(void *p);
711 void *operator new[](size_t n);
712 void operator delete[](void *p);
713 virtual ~Mask() {}
714 // Set bit i to 1
715 virtual void set(int i) {}
716 // Return bit i
717 virtual bool is_set(int i) const { return false; }
718 // Set bit i to 0
719 virtual void clear(int i) {}
720 // Zero out entire mask
721 virtual void zero() {}
722 // Check whether mask is empty
723 virtual bool empty() const { return true; }
724 // Copy src into this mask
725 virtual void copy(const Mask *src) {}
726 // this &= rhs
727 virtual void bitwise_and(const Mask *rhs) {}
728 // this |= rhs
729 virtual void bitwise_or(const Mask *rhs) {}
730 // this = ~this
731 virtual void bitwise_not() {}
732 // this == rhs
733 virtual bool is_equal(const Mask *rhs) const { return false; }
734 // API for iterating over an affinity mask
735 // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
736 virtual int begin() const { return 0; }
737 virtual int end() const { return 0; }
738 virtual int next(int previous) const { return 0; }
739#if KMP_OS_WINDOWS
740 virtual int set_process_affinity(bool abort_on_error) const { return -1; }
741#endif
742 // Set the system's affinity to this affinity mask's value
743 virtual int set_system_affinity(bool abort_on_error) const { return -1; }
744 // Set this affinity mask to the current system affinity
745 virtual int get_system_affinity(bool abort_on_error) { return -1; }
746 // Only 1 DWORD in the mask should have any procs set.
747 // Return the appropriate index, or -1 for an invalid mask.
748 virtual int get_proc_group() const { return -1; }
749 int get_max_cpu() const {
750 int cpu;
751 int max_cpu = -1;
752 KMP_CPU_SET_ITERATE(cpu, this) {
753 if (cpu > max_cpu)
754 max_cpu = cpu;
755 }
756 return max_cpu;
757 }
758 };
759 void *operator new(size_t n);
760 void operator delete(void *p);
761 // Need virtual destructor
762 virtual ~KMPAffinity() = default;
763 // Determine if affinity is capable
764 virtual void determine_capable(const char *env_var) {}
765 // Bind the current thread to os proc
766 virtual void bind_thread(int proc) {}
767 // Factory functions to allocate/deallocate a mask
768 virtual Mask *allocate_mask() { return nullptr; }
769 virtual void deallocate_mask(Mask *m) {}
770 virtual Mask *allocate_mask_array(int num) { return nullptr; }
771 virtual void deallocate_mask_array(Mask *m) {}
772 virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
773 static void pick_api();
774 static void destroy_api();
775 enum api_type {
776 NATIVE_OS
777#if KMP_HWLOC_ENABLED
778 ,
779 HWLOC
780#endif // KMP_HWLOC_ENABLED
781 };
782 virtual api_type get_api_type() const {
783 KMP_ASSERT(0);
784 return NATIVE_OS;
785 }
786
787private:
788 static bool picked_api;
789};
790
791typedef KMPAffinity::Mask kmp_affin_mask_t;
792extern KMPAffinity *__kmp_affinity_dispatch;
793
794#if !KMP_OS_AIX
795class kmp_affinity_raii_t {
796 kmp_affin_mask_t *mask;
797 bool restored;
798
799public:
800 kmp_affinity_raii_t(const kmp_affin_mask_t *new_mask = nullptr)
801 : mask(nullptr), restored(false) {
802 if (KMP_AFFINITY_CAPABLE()) {
803 KMP_CPU_ALLOC(mask);
804 KMP_ASSERT(mask != NULL);
805 __kmp_get_system_affinity(mask, /*abort_on_error=*/true);
806 if (new_mask)
807 __kmp_set_system_affinity(new_mask, /*abort_on_error=*/true);
808 }
809 }
810 void restore() {
811 if (mask && KMP_AFFINITY_CAPABLE() && !restored) {
812 __kmp_set_system_affinity(mask, /*abort_on_error=*/true);
813 KMP_CPU_FREE(mask);
814 }
815 restored = true;
816 }
817 ~kmp_affinity_raii_t() { restore(); }
818};
819#endif // !KMP_OS_AIX
820
821// Declare local char buffers with this size for printing debug and info
822// messages, using __kmp_affinity_print_mask().
823#define KMP_AFFIN_MASK_PRINT_LEN 1024
824
825enum affinity_type {
826 affinity_none = 0,
827 affinity_physical,
828 affinity_logical,
829 affinity_compact,
830 affinity_scatter,
831 affinity_explicit,
832 affinity_balanced,
833 affinity_disabled, // not used outsize the env var parser
834 affinity_default
835};
836
837enum affinity_top_method {
838 affinity_top_method_all = 0, // try all (supported) methods, in order
839#if KMP_ARCH_X86 || KMP_ARCH_X86_64
840 affinity_top_method_apicid,
841 affinity_top_method_x2apicid,
842 affinity_top_method_x2apicid_1f,
843#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
844 affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
845#if KMP_GROUP_AFFINITY
846 affinity_top_method_group,
847#endif /* KMP_GROUP_AFFINITY */
848 affinity_top_method_flat,
849#if KMP_HWLOC_ENABLED
850 affinity_top_method_hwloc,
851#endif // KMP_HWLOC_ENABLED
852 affinity_top_method_default
853};
854
855#define affinity_respect_mask_default (2)
856
857typedef struct kmp_affinity_flags_t {
858 unsigned dups : 1;
859 unsigned verbose : 1;
860 unsigned warnings : 1;
861 unsigned respect : 2;
862 unsigned reset : 1;
863 unsigned initialized : 1;
864 unsigned core_types_gran : 1;
865 unsigned core_effs_gran : 1;
866 unsigned omp_places : 1;
867 unsigned reserved : 22;
868} kmp_affinity_flags_t;
869KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4);
870
871typedef struct kmp_affinity_ids_t {
872 int os_id;
873 int ids[KMP_HW_LAST];
874} kmp_affinity_ids_t;
875
876typedef struct kmp_affinity_attrs_t {
877 int core_type : 8;
878 int core_eff : 8;
879 unsigned valid : 1;
880 unsigned reserved : 15;
881} kmp_affinity_attrs_t;
882#define KMP_AFFINITY_ATTRS_UNKNOWN \
883 { KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 }
884
885typedef struct kmp_affinity_t {
886 char *proclist;
887 enum affinity_type type;
888 kmp_hw_t gran;
889 int gran_levels;
890 kmp_affinity_attrs_t core_attr_gran;
891 int compact;
892 int offset;
893 kmp_affinity_flags_t flags;
894 unsigned num_masks;
895 kmp_affin_mask_t *masks;
896 kmp_affinity_ids_t *ids;
897 kmp_affinity_attrs_t *attrs;
898 unsigned num_os_id_masks;
899 kmp_affin_mask_t *os_id_masks;
900 const char *env_var;
901} kmp_affinity_t;
902
903#define KMP_AFFINITY_INIT(env) \
904 { \
905 nullptr, affinity_default, KMP_HW_UNKNOWN, -1, KMP_AFFINITY_ATTRS_UNKNOWN, \
906 0, 0, \
907 {TRUE, FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE, \
908 FALSE, FALSE, FALSE}, \
909 0, nullptr, nullptr, nullptr, 0, nullptr, env \
910 }
911
912extern enum affinity_top_method __kmp_affinity_top_method;
913extern kmp_affinity_t __kmp_affinity;
914extern kmp_affinity_t __kmp_hh_affinity;
915extern kmp_affinity_t *__kmp_affinities[2];
916
917extern void __kmp_affinity_bind_thread(int which);
918
919extern kmp_affin_mask_t *__kmp_affin_fullMask;
920extern kmp_affin_mask_t *__kmp_affin_origMask;
921extern char *__kmp_cpuinfo_file;
922
923#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
924extern int __kmp_first_osid_with_ecore;
925#endif
926
927#endif /* KMP_AFFINITY_SUPPORTED */
928
929// This needs to be kept in sync with the values in omp.h !!!
930typedef enum kmp_proc_bind_t {
931 proc_bind_false = 0,
932 proc_bind_true,
933 proc_bind_primary,
934 proc_bind_close,
935 proc_bind_spread,
936 proc_bind_intel, // use KMP_AFFINITY interface
937 proc_bind_default
938} kmp_proc_bind_t;
939
940typedef struct kmp_nested_proc_bind_t {
941 kmp_proc_bind_t *bind_types;
942 int size;
943 int used;
944} kmp_nested_proc_bind_t;
945
946extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
947extern kmp_proc_bind_t __kmp_teams_proc_bind;
948
949extern int __kmp_display_affinity;
950extern char *__kmp_affinity_format;
951static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
952#if OMPT_SUPPORT
953extern int __kmp_tool;
954extern char *__kmp_tool_libraries;
955#endif // OMPT_SUPPORT
956
957#if KMP_AFFINITY_SUPPORTED
958#define KMP_PLACE_ALL (-1)
959#define KMP_PLACE_UNDEFINED (-2)
960// Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
961#define KMP_AFFINITY_NON_PROC_BIND \
962 ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
963 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
964 (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced))
965#endif /* KMP_AFFINITY_SUPPORTED */
966
967extern int __kmp_affinity_num_places;
968
969typedef enum kmp_cancel_kind_t {
970 cancel_noreq = 0,
971 cancel_parallel = 1,
972 cancel_loop = 2,
973 cancel_sections = 3,
974 cancel_taskgroup = 4
975} kmp_cancel_kind_t;
976
977// KMP_HW_SUBSET support:
978typedef struct kmp_hws_item {
979 int num;
980 int offset;
981} kmp_hws_item_t;
982
983extern kmp_hws_item_t __kmp_hws_socket;
984extern kmp_hws_item_t __kmp_hws_die;
985extern kmp_hws_item_t __kmp_hws_node;
986extern kmp_hws_item_t __kmp_hws_tile;
987extern kmp_hws_item_t __kmp_hws_core;
988extern kmp_hws_item_t __kmp_hws_proc;
989extern int __kmp_hws_requested;
990extern int __kmp_hws_abs_flag; // absolute or per-item number requested
991
992/* ------------------------------------------------------------------------ */
993
994#define KMP_PAD(type, sz) \
995 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
996
997// We need to avoid using -1 as a GTID as +1 is added to the gtid
998// when storing it in a lock, and the value 0 is reserved.
999#define KMP_GTID_DNE (-2) /* Does not exist */
1000#define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
1001#define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
1002#define KMP_GTID_UNKNOWN (-5) /* Is not known */
1003#define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
1004
1005/* OpenMP 5.0 Memory Management support */
1006
1007#ifndef __OMP_H
1008// Duplicate type definitions from omp.h
1009typedef uintptr_t omp_uintptr_t;
1010
1011typedef enum {
1012 omp_atk_sync_hint = 1,
1013 omp_atk_alignment = 2,
1014 omp_atk_access = 3,
1015 omp_atk_pool_size = 4,
1016 omp_atk_fallback = 5,
1017 omp_atk_fb_data = 6,
1018 omp_atk_pinned = 7,
1019 omp_atk_partition = 8,
1020 omp_atk_pin_device = 9,
1021 omp_atk_preferred_device = 10,
1022 omp_atk_device_access = 11,
1023 omp_atk_target_access = 12,
1024 omp_atk_atomic_scope = 13,
1025 omp_atk_part_size = 14
1026} omp_alloctrait_key_t;
1027
1028typedef enum {
1029 omp_atv_false = 0,
1030 omp_atv_true = 1,
1031 omp_atv_contended = 3,
1032 omp_atv_uncontended = 4,
1033 omp_atv_serialized = 5,
1034 omp_atv_sequential = omp_atv_serialized, // (deprecated)
1035 omp_atv_private = 6,
1036 omp_atv_device = 7,
1037 omp_atv_thread = 8,
1038 omp_atv_pteam = 9,
1039 omp_atv_cgroup = 10,
1040 omp_atv_default_mem_fb = 11,
1041 omp_atv_null_fb = 12,
1042 omp_atv_abort_fb = 13,
1043 omp_atv_allocator_fb = 14,
1044 omp_atv_environment = 15,
1045 omp_atv_nearest = 16,
1046 omp_atv_blocked = 17,
1047 omp_atv_interleaved = 18,
1048 omp_atv_all = 19,
1049 omp_atv_single = 20,
1050 omp_atv_multiple = 21,
1051 omp_atv_memspace = 22
1052} omp_alloctrait_value_t;
1053#define omp_atv_default ((omp_uintptr_t)-1)
1054
1055typedef void *omp_memspace_handle_t;
1056extern omp_memspace_handle_t const omp_null_mem_space;
1057extern omp_memspace_handle_t const omp_default_mem_space;
1058extern omp_memspace_handle_t const omp_large_cap_mem_space;
1059extern omp_memspace_handle_t const omp_const_mem_space;
1060extern omp_memspace_handle_t const omp_high_bw_mem_space;
1061extern omp_memspace_handle_t const omp_low_lat_mem_space;
1062extern omp_memspace_handle_t const omp_cgroup_mem_space;
1063extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
1064extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
1065extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
1066extern omp_memspace_handle_t const kmp_max_mem_space;
1067
1068typedef struct {
1069 omp_alloctrait_key_t key;
1070 omp_uintptr_t value;
1071} omp_alloctrait_t;
1072
1073typedef void *omp_allocator_handle_t;
1074extern omp_allocator_handle_t const omp_null_allocator;
1075extern omp_allocator_handle_t const omp_default_mem_alloc;
1076extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
1077extern omp_allocator_handle_t const omp_const_mem_alloc;
1078extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
1079extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
1080extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
1081extern omp_allocator_handle_t const omp_pteam_mem_alloc;
1082extern omp_allocator_handle_t const omp_thread_mem_alloc;
1083extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
1084extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
1085extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
1086extern omp_allocator_handle_t const kmp_max_mem_alloc;
1087extern omp_allocator_handle_t __kmp_def_allocator;
1088
1089typedef enum omp_access_t {
1090 omp_access_cgroup = 0,
1091 omp_access_pteam = 1
1092} omp_access_t;
1093
1094// declare the host API entry points for the Fortran interface to pick 'em up
1095extern void *omp_get_dyn_gprivate_ptr(size_t offset, omp_access_t access_group);
1096extern void *omp_get_dyn_gprivate_nofb_ptr(size_t offset,
1097 omp_access_t access_group);
1098extern size_t omp_get_dyn_gprivate_size(omp_access_t access_group);
1099extern omp_memspace_handle_t
1100omp_get_dyn_gprivate_memspace(omp_access_t access_group);
1101
1102// end of duplicate type definitions from omp.h
1103#endif
1104
1105extern int __kmp_memkind_available;
1106extern bool __kmp_hwloc_available;
1107
1109typedef struct kmp_memspace_t {
1110 omp_memspace_handle_t memspace; // predefined input memory space
1111 int num_resources = 0; // number of available resources
1112 int *resources = nullptr; // available resources
1113 kmp_memspace_t *next = nullptr; // next memory space handle
1115
1117typedef struct kmp_allocator_t {
1118 omp_memspace_handle_t memspace;
1119 void **memkind; // pointer to memkind
1120 size_t alignment;
1121 omp_alloctrait_value_t fb;
1122 kmp_allocator_t *fb_data;
1123 kmp_uint64 pool_size;
1124 kmp_uint64 pool_used;
1125 bool pinned;
1126 omp_alloctrait_value_t partition;
1127 int pin_device;
1128 int preferred_device;
1129 omp_alloctrait_value_t target_access;
1130 omp_alloctrait_value_t atomic_scope;
1131 size_t part_size;
1132#if KMP_HWLOC_ENABLED
1133 omp_alloctrait_value_t membind;
1134#endif // KMP_HWLOC_ENABLED
1136
1137extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1138 omp_memspace_handle_t,
1139 int ntraits,
1140 omp_alloctrait_t traits[]);
1141extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1142extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1143extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1144// external interfaces, may be used by compiler
1145extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1146extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1147 omp_allocator_handle_t al);
1148extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1149 omp_allocator_handle_t al);
1150extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1151 omp_allocator_handle_t al,
1152 omp_allocator_handle_t free_al);
1153extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1154// internal interfaces, contain real implementation
1155extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1156 omp_allocator_handle_t al);
1157extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1158 omp_allocator_handle_t al);
1159extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1160 omp_allocator_handle_t al,
1161 omp_allocator_handle_t free_al);
1162extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1163
1164extern void __kmp_init_memkind();
1165extern void __kmp_fini_memkind();
1166extern void __kmp_init_target_mem();
1167extern void __kmp_fini_target_mem();
1168
1169// OpenMP 6.0 (TR11) Memory Management support
1170extern omp_memspace_handle_t __kmp_get_devices_memspace(int ndevs,
1171 const int *devs,
1172 omp_memspace_handle_t,
1173 int host);
1174extern omp_allocator_handle_t __kmp_get_devices_allocator(int ndevs,
1175 const int *devs,
1176 omp_memspace_handle_t,
1177 int host);
1178extern int __kmp_get_memspace_num_resources(omp_memspace_handle_t memspace);
1179extern omp_memspace_handle_t
1180__kmp_get_submemspace(omp_memspace_handle_t memspace, int num_resources,
1181 int *resources);
1182
1183/* ------------------------------------------------------------------------ */
1184
1185#if ENABLE_LIBOMPTARGET
1186extern void __kmp_init_target_task();
1187#endif
1188
1189/* ------------------------------------------------------------------------ */
1190
1191#define KMP_UINT64_MAX \
1192 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1193
1194#define KMP_MIN_NTH 1
1195
1196#ifndef KMP_MAX_NTH
1197#if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1198#define KMP_MAX_NTH PTHREAD_THREADS_MAX
1199#else
1200#ifdef __ve__
1201// VE's pthread supports only up to 64 threads per a VE process.
1202// Please check p. 14 of following documentation for more details.
1203// https://sxauroratsubasa.sakura.ne.jp/documents/veos/en/VEOS_high_level_design.pdf
1204#define KMP_MAX_NTH 64
1205#else
1206#define KMP_MAX_NTH INT_MAX
1207#endif
1208#endif
1209#endif /* KMP_MAX_NTH */
1210
1211#ifdef PTHREAD_STACK_MIN
1212#define KMP_MIN_STKSIZE ((size_t)PTHREAD_STACK_MIN)
1213#else
1214#define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1215#endif
1216
1217#if KMP_OS_AIX && KMP_ARCH_PPC
1218#define KMP_MAX_STKSIZE 0x10000000 /* 256Mb max size on 32-bit AIX */
1219#else
1220#define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1221#endif
1222
1223#if KMP_ARCH_X86
1224#define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1225#elif KMP_ARCH_X86_64
1226#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1227#define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1228#elif KMP_ARCH_VE
1229// Minimum stack size for pthread for VE is 4MB.
1230// https://www.hpc.nec/documents/veos/en/glibc/Difference_Points_glibc.htm
1231#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1232#elif KMP_OS_AIX
1233// The default stack size for worker threads on AIX is 4MB.
1234#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1235#else
1236#define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1237#endif
1238
1239#define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1240#define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1241#define KMP_MAX_MALLOC_POOL_INCR \
1242 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1243
1244#define KMP_MIN_STKOFFSET (0)
1245#define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1246#if KMP_OS_DARWIN
1247#define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1248#else
1249#define KMP_DEFAULT_STKOFFSET CACHE_LINE
1250#endif
1251
1252#define KMP_MIN_STKPADDING (0)
1253#define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1254
1255#define KMP_BLOCKTIME_MULTIPLIER \
1256 (1000000) /* number of blocktime units per second */
1257#define KMP_MIN_BLOCKTIME (0)
1258#define KMP_MAX_BLOCKTIME \
1259 (INT_MAX) /* Must be this for "infinite" setting the work */
1260
1261/* __kmp_blocktime is in microseconds */
1262#define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200000))
1263
1264#if KMP_USE_MONITOR
1265#define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1266#define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1267#define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1268
1269/* Calculate new number of monitor wakeups for a specific block time based on
1270 previous monitor_wakeups. Only allow increasing number of wakeups */
1271#define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1272 (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \
1273 : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \
1274 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1275 ? (monitor_wakeups) \
1276 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1277
1278/* Calculate number of intervals for a specific block time based on
1279 monitor_wakeups */
1280#define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1281 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1282 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1283#else
1284#define KMP_BLOCKTIME(team, tid) \
1285 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1286#if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1287// HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1288extern kmp_uint64 __kmp_ticks_per_msec;
1289extern kmp_uint64 __kmp_ticks_per_usec;
1290#if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1291#define KMP_NOW() ((kmp_uint64)_rdtsc())
1292#else
1293#define KMP_NOW() __kmp_hardware_timestamp()
1294#endif
1295#define KMP_BLOCKTIME_INTERVAL(team, tid) \
1296 ((kmp_uint64)KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_usec)
1297#define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1298#else
1299// System time is retrieved sporadically while blocking.
1300extern kmp_uint64 __kmp_now_nsec();
1301#define KMP_NOW() __kmp_now_nsec()
1302#define KMP_BLOCKTIME_INTERVAL(team, tid) \
1303 ((kmp_uint64)KMP_BLOCKTIME(team, tid) * (kmp_uint64)KMP_NSEC_PER_USEC)
1304#define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1305#endif
1306#endif // KMP_USE_MONITOR
1307
1308#define KMP_MIN_STATSCOLS 40
1309#define KMP_MAX_STATSCOLS 4096
1310#define KMP_DEFAULT_STATSCOLS 80
1311
1312#define KMP_MIN_INTERVAL 0
1313#define KMP_MAX_INTERVAL (INT_MAX - 1)
1314#define KMP_DEFAULT_INTERVAL 0
1315
1316#define KMP_MIN_CHUNK 1
1317#define KMP_MAX_CHUNK (INT_MAX - 1)
1318#define KMP_DEFAULT_CHUNK 1
1319
1320#define KMP_MIN_DISP_NUM_BUFF 1
1321#define KMP_DFLT_DISP_NUM_BUFF 7
1322#define KMP_MAX_DISP_NUM_BUFF 4096
1323
1324#define KMP_MAX_ORDERED 8
1325
1326#define KMP_MAX_FIELDS 32
1327
1328#define KMP_MAX_BRANCH_BITS 31
1329
1330#define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1331
1332#define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1333
1334#define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1335
1336/* Minimum number of threads before switch to TLS gtid (experimentally
1337 determined) */
1338/* josh TODO: what about OS X* tuning? */
1339#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1340#define KMP_TLS_GTID_MIN 5
1341#else
1342#define KMP_TLS_GTID_MIN INT_MAX
1343#endif
1344
1345#define KMP_MASTER_TID(tid) (0 == (tid))
1346#define KMP_WORKER_TID(tid) (0 != (tid))
1347
1348#define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1349#define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1350#define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1351
1352#ifndef TRUE
1353#define FALSE 0
1354#define TRUE (!FALSE)
1355#endif
1356
1357/* NOTE: all of the following constants must be even */
1358
1359#if KMP_OS_WINDOWS
1360#define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1361#define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1362#elif KMP_OS_LINUX
1363#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1364#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1365#elif KMP_OS_DARWIN
1366/* TODO: tune for KMP_OS_DARWIN */
1367#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1368#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1369#elif KMP_OS_DRAGONFLY
1370/* TODO: tune for KMP_OS_DRAGONFLY */
1371#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1372#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1373#elif KMP_OS_FREEBSD
1374/* TODO: tune for KMP_OS_FREEBSD */
1375#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1376#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1377#elif KMP_OS_NETBSD
1378/* TODO: tune for KMP_OS_NETBSD */
1379#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1380#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1381#elif KMP_OS_OPENBSD
1382/* TODO: tune for KMP_OS_OPENBSD */
1383#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1384#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1385#elif KMP_OS_HAIKU
1386/* TODO: tune for KMP_OS_HAIKU */
1387#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1388#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1389#elif KMP_OS_HURD
1390/* TODO: tune for KMP_OS_HURD */
1391#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1392#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1393#elif KMP_OS_SOLARIS
1394/* TODO: tune for KMP_OS_SOLARIS */
1395#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1396#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1397#elif KMP_OS_WASI
1398/* TODO: tune for KMP_OS_WASI */
1399#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1400#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1401#elif KMP_OS_AIX
1402/* TODO: tune for KMP_OS_AIX */
1403#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1404#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1405#endif
1406
1407#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1408typedef struct kmp_cpuid {
1409 kmp_uint32 eax;
1410 kmp_uint32 ebx;
1411 kmp_uint32 ecx;
1412 kmp_uint32 edx;
1413} kmp_cpuid_t;
1414
1415typedef struct kmp_cpuinfo_flags_t {
1416 unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1417 unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1418 unsigned hybrid : 1;
1419 unsigned reserved : 29; // Ensure size of 32 bits
1420} kmp_cpuinfo_flags_t;
1421
1422typedef struct kmp_cpuinfo {
1423 int initialized; // If 0, other fields are not initialized.
1424 int signature; // CPUID(1).EAX
1425 int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1426 int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1427 // Model << 4 ) + Model)
1428 int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1429 kmp_cpuinfo_flags_t flags;
1430 int apic_id;
1431 kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1432 char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1433} kmp_cpuinfo_t;
1434
1435extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1436
1437#if KMP_OS_UNIX
1438// subleaf is only needed for cache and topology discovery and can be set to
1439// zero in most cases
1440static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1441 __asm__ __volatile__("cpuid"
1442 : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1443 : "a"(leaf), "c"(subleaf));
1444}
1445// Load p into FPU control word
1446static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1447 __asm__ __volatile__("fldcw %0" : : "m"(*p));
1448}
1449// Store FPU control word into p
1450static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1451 __asm__ __volatile__("fstcw %0" : "=m"(*p));
1452}
1453static inline void __kmp_clear_x87_fpu_status_word() {
1454#if KMP_MIC
1455 // 32-bit protected mode x87 FPU state
1456 struct x87_fpu_state {
1457 unsigned cw;
1458 unsigned sw;
1459 unsigned tw;
1460 unsigned fip;
1461 unsigned fips;
1462 unsigned fdp;
1463 unsigned fds;
1464 };
1465 struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1466 __asm__ __volatile__("fstenv %0\n\t" // store FP env
1467 "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1468 "fldenv %0\n\t" // load FP env back
1469 : "+m"(fpu_state), "+m"(fpu_state.sw));
1470#else
1471 __asm__ __volatile__("fnclex");
1472#endif // KMP_MIC
1473}
1474#if __SSE__
1475static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1476static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1477#else
1478static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1479static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1480#endif
1481#else
1482// Windows still has these as external functions in assembly file
1483extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1484extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1485extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1486extern void __kmp_clear_x87_fpu_status_word();
1487static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1488static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1489#endif // KMP_OS_UNIX
1490
1491#define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1492
1493// User-level Monitor/Mwait
1494#if KMP_HAVE_UMWAIT
1495// We always try for UMWAIT first
1496#if KMP_HAVE_WAITPKG_INTRINSICS
1497#if KMP_HAVE_IMMINTRIN_H
1498#include <immintrin.h>
1499#elif KMP_HAVE_INTRIN_H
1500#include <intrin.h>
1501#endif
1502#endif // KMP_HAVE_WAITPKG_INTRINSICS
1503
1504KMP_ATTRIBUTE_TARGET_WAITPKG
1505static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1506#if !KMP_HAVE_WAITPKG_INTRINSICS
1507 uint32_t timeHi = uint32_t(counter >> 32);
1508 uint32_t timeLo = uint32_t(counter & 0xffffffff);
1509 char flag;
1510 __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1511 "setb %0"
1512 // The "=q" restraint means any register accessible as rl
1513 // in 32-bit mode: a, b, c, and d;
1514 // in 64-bit mode: any integer register
1515 : "=q"(flag)
1516 : "a"(timeLo), "d"(timeHi), "c"(hint)
1517 :);
1518 return flag;
1519#else
1520 return _tpause(hint, counter);
1521#endif
1522}
1523KMP_ATTRIBUTE_TARGET_WAITPKG
1524static inline void __kmp_umonitor(void *cacheline) {
1525#if !KMP_HAVE_WAITPKG_INTRINSICS
1526 __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1527 :
1528 : "a"(cacheline)
1529 :);
1530#else
1531 _umonitor(cacheline);
1532#endif
1533}
1534KMP_ATTRIBUTE_TARGET_WAITPKG
1535static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1536#if !KMP_HAVE_WAITPKG_INTRINSICS
1537 uint32_t timeHi = uint32_t(counter >> 32);
1538 uint32_t timeLo = uint32_t(counter & 0xffffffff);
1539 char flag;
1540 __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1541 "setb %0"
1542 // The "=q" restraint means any register accessible as rl
1543 // in 32-bit mode: a, b, c, and d;
1544 // in 64-bit mode: any integer register
1545 : "=q"(flag)
1546 : "a"(timeLo), "d"(timeHi), "c"(hint)
1547 :);
1548 return flag;
1549#else
1550 return _umwait(hint, counter);
1551#endif
1552}
1553#elif KMP_HAVE_MWAIT
1554#if KMP_OS_UNIX
1555#include <pmmintrin.h>
1556#else
1557#include <intrin.h>
1558#endif
1559#if KMP_OS_UNIX
1560__attribute__((target("sse3")))
1561#endif
1562static inline void
1563__kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1564 _mm_monitor(cacheline, extensions, hints);
1565}
1566#if KMP_OS_UNIX
1567__attribute__((target("sse3")))
1568#endif
1569static inline void
1570__kmp_mm_mwait(unsigned extensions, unsigned hints) {
1571 _mm_mwait(extensions, hints);
1572}
1573#endif // KMP_HAVE_UMWAIT
1574
1575#if KMP_ARCH_X86
1576extern void __kmp_x86_pause(void);
1577#elif KMP_MIC
1578// Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1579// regression after removal of extra PAUSE from spin loops. Changing
1580// the delay from 100 to 300 showed even better performance than double PAUSE
1581// on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1582static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1583#else
1584static inline void __kmp_x86_pause(void) { _mm_pause(); }
1585#endif
1586#define KMP_CPU_PAUSE() __kmp_x86_pause()
1587#elif KMP_ARCH_PPC64
1588#define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1589#define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1590#define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1591#define KMP_CPU_PAUSE() \
1592 do { \
1593 KMP_PPC64_PRI_LOW(); \
1594 KMP_PPC64_PRI_MED(); \
1595 KMP_PPC64_PRI_LOC_MB(); \
1596 } while (0)
1597#else
1598#define KMP_CPU_PAUSE() /* nothing to do */
1599#endif
1600
1601#define KMP_INIT_YIELD(count) \
1602 { (count) = __kmp_yield_init; }
1603
1604#define KMP_INIT_BACKOFF(time) \
1605 { (time) = __kmp_pause_init; }
1606
1607#define KMP_OVERSUBSCRIBED \
1608 (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1609
1610#define KMP_TRY_YIELD \
1611 ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1612
1613#define KMP_TRY_YIELD_OVERSUB \
1614 ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1615
1616#define KMP_YIELD(cond) \
1617 { \
1618 KMP_CPU_PAUSE(); \
1619 if ((cond) && (KMP_TRY_YIELD)) \
1620 __kmp_yield(); \
1621 }
1622
1623#define KMP_YIELD_OVERSUB() \
1624 { \
1625 KMP_CPU_PAUSE(); \
1626 if ((KMP_TRY_YIELD_OVERSUB)) \
1627 __kmp_yield(); \
1628 }
1629
1630// Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1631// there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1632#define KMP_YIELD_SPIN(count) \
1633 { \
1634 KMP_CPU_PAUSE(); \
1635 if (KMP_TRY_YIELD) { \
1636 (count) -= 2; \
1637 if (!(count)) { \
1638 __kmp_yield(); \
1639 (count) = __kmp_yield_next; \
1640 } \
1641 } \
1642 }
1643
1644// If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1645// (C0.2) state, which improves performance of other SMT threads on the same
1646// core, otherwise, use the fast (C0.1) default state, or whatever the user has
1647// requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1648// available, fall back to the regular CPU pause and yield combination.
1649#if KMP_HAVE_UMWAIT
1650#define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF)
1651#define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1652 { \
1653 if (__kmp_tpause_enabled) { \
1654 if (KMP_OVERSUBSCRIBED) { \
1655 __kmp_tpause(0, (time)); \
1656 } else { \
1657 __kmp_tpause(__kmp_tpause_hint, (time)); \
1658 } \
1659 (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK; \
1660 } else { \
1661 KMP_CPU_PAUSE(); \
1662 if ((KMP_TRY_YIELD_OVERSUB)) { \
1663 __kmp_yield(); \
1664 } else if (__kmp_use_yield == 1) { \
1665 (count) -= 2; \
1666 if (!(count)) { \
1667 __kmp_yield(); \
1668 (count) = __kmp_yield_next; \
1669 } \
1670 } \
1671 } \
1672 }
1673#else
1674#define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1675 { \
1676 KMP_CPU_PAUSE(); \
1677 if ((KMP_TRY_YIELD_OVERSUB)) \
1678 __kmp_yield(); \
1679 else if (__kmp_use_yield == 1) { \
1680 (count) -= 2; \
1681 if (!(count)) { \
1682 __kmp_yield(); \
1683 (count) = __kmp_yield_next; \
1684 } \
1685 } \
1686 }
1687#endif // KMP_HAVE_UMWAIT
1688
1689/* ------------------------------------------------------------------------ */
1690/* Support datatypes for the orphaned construct nesting checks. */
1691/* ------------------------------------------------------------------------ */
1692
1693/* When adding to this enum, add its corresponding string in cons_text_c[]
1694 * array in kmp_error.cpp */
1695enum cons_type {
1696 ct_none,
1697 ct_parallel,
1698 ct_pdo,
1699 ct_pdo_ordered,
1700 ct_psections,
1701 ct_psingle,
1702 ct_critical,
1703 ct_ordered_in_parallel,
1704 ct_ordered_in_pdo,
1705 ct_master,
1706 ct_reduce,
1707 ct_barrier,
1708 ct_masked
1709};
1710
1711#define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1712
1713struct cons_data {
1714 ident_t const *ident;
1715 enum cons_type type;
1716 int prev;
1717 kmp_user_lock_p
1718 name; /* address exclusively for critical section name comparison */
1719};
1720
1721struct cons_header {
1722 int p_top, w_top, s_top;
1723 int stack_size, stack_top;
1724 struct cons_data *stack_data;
1725};
1726
1727struct kmp_region_info {
1728 char *text;
1729 int offset[KMP_MAX_FIELDS];
1730 int length[KMP_MAX_FIELDS];
1731};
1732
1733/* ---------------------------------------------------------------------- */
1734/* ---------------------------------------------------------------------- */
1735
1736#if KMP_OS_WINDOWS
1737typedef HANDLE kmp_thread_t;
1738typedef DWORD kmp_key_t;
1739#endif /* KMP_OS_WINDOWS */
1740
1741#if KMP_OS_UNIX
1742typedef pthread_t kmp_thread_t;
1743typedef pthread_key_t kmp_key_t;
1744#endif
1745
1746extern kmp_key_t __kmp_gtid_threadprivate_key;
1747
1748typedef struct kmp_sys_info {
1749 long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1750 long minflt; /* the number of page faults serviced without any I/O */
1751 long majflt; /* the number of page faults serviced that required I/O */
1752 long nswap; /* the number of times a process was "swapped" out of memory */
1753 long inblock; /* the number of times the file system had to perform input */
1754 long oublock; /* the number of times the file system had to perform output */
1755 long nvcsw; /* the number of times a context switch was voluntarily */
1756 long nivcsw; /* the number of times a context switch was forced */
1757} kmp_sys_info_t;
1758
1759#if USE_ITT_BUILD
1760// We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1761// required type here. Later we will check the type meets requirements.
1762typedef int kmp_itt_mark_t;
1763#define KMP_ITT_DEBUG 0
1764#endif /* USE_ITT_BUILD */
1765
1766typedef kmp_int32 kmp_critical_name[8];
1767
1777typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1778typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1779 ...);
1780
1785/* ---------------------------------------------------------------------------
1786 */
1787/* Threadprivate initialization/finalization function declarations */
1788
1789/* for non-array objects: __kmpc_threadprivate_register() */
1790
1795typedef void *(*kmpc_ctor)(void *);
1796
1801typedef void (*kmpc_dtor)(
1802 void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1803 compiler */
1808typedef void *(*kmpc_cctor)(void *, void *);
1809
1810/* for array objects: __kmpc_threadprivate_register_vec() */
1811/* First arg: "this" pointer */
1812/* Last arg: number of array elements */
1818typedef void *(*kmpc_ctor_vec)(void *, size_t);
1824typedef void (*kmpc_dtor_vec)(void *, size_t);
1830typedef void *(*kmpc_cctor_vec)(void *, void *,
1831 size_t); /* function unused by compiler */
1832
1836
1837/* keeps tracked of threadprivate cache allocations for cleanup later */
1838typedef struct kmp_cached_addr {
1839 void **addr; /* address of allocated cache */
1840 void ***compiler_cache; /* pointer to compiler's cache */
1841 void *data; /* pointer to global data */
1842 struct kmp_cached_addr *next; /* pointer to next cached address */
1843} kmp_cached_addr_t;
1844
1845struct private_data {
1846 struct private_data *next; /* The next descriptor in the list */
1847 void *data; /* The data buffer for this descriptor */
1848 int more; /* The repeat count for this descriptor */
1849 size_t size; /* The data size for this descriptor */
1850};
1851
1852struct private_common {
1853 struct private_common *next;
1854 struct private_common *link;
1855 void *gbl_addr;
1856 void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1857 size_t cmn_size;
1858};
1859
1860struct shared_common {
1861 struct shared_common *next;
1862 struct private_data *pod_init;
1863 void *obj_init;
1864 void *gbl_addr;
1865 union {
1866 kmpc_ctor ctor;
1867 kmpc_ctor_vec ctorv;
1868 } ct;
1869 union {
1870 kmpc_cctor cctor;
1871 kmpc_cctor_vec cctorv;
1872 } cct;
1873 union {
1874 kmpc_dtor dtor;
1875 kmpc_dtor_vec dtorv;
1876 } dt;
1877 size_t vec_len;
1878 int is_vec;
1879 size_t cmn_size;
1880};
1881
1882#define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1883#define KMP_HASH_TABLE_SIZE \
1884 (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1885#define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1886#define KMP_HASH(x) \
1887 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1888
1889struct common_table {
1890 struct private_common *data[KMP_HASH_TABLE_SIZE];
1891};
1892
1893struct shared_table {
1894 struct shared_common *data[KMP_HASH_TABLE_SIZE];
1895};
1896
1897/* ------------------------------------------------------------------------ */
1898
1899#if KMP_USE_HIER_SCHED
1900// Shared barrier data that exists inside a single unit of the scheduling
1901// hierarchy
1902typedef struct kmp_hier_private_bdata_t {
1903 kmp_int32 num_active;
1904 kmp_uint64 index;
1905 kmp_uint64 wait_val[2];
1906} kmp_hier_private_bdata_t;
1907#endif
1908
1909typedef struct kmp_sched_flags {
1910 unsigned ordered : 1;
1911 unsigned nomerge : 1;
1912 unsigned contains_last : 1;
1913 unsigned use_hier : 1; // Used in KMP_USE_HIER_SCHED code
1914 unsigned use_hybrid : 1; // Used in KMP_WEIGHTED_ITERATIONS_SUPPORTED code
1915 unsigned unused : 27;
1916} kmp_sched_flags_t;
1917
1918KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1919
1920#if KMP_STATIC_STEAL_ENABLED
1921typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1922 kmp_int32 count;
1923 kmp_int32 ub;
1924 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1925 kmp_int32 lb;
1926 kmp_int32 st;
1927 kmp_int32 tc;
1928 kmp_lock_t *steal_lock; // lock used for chunk stealing
1929
1930 kmp_uint32 ordered_lower;
1931 kmp_uint32 ordered_upper;
1932
1933 // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1934 // a) parm3 is properly aligned and
1935 // b) all parm1-4 are on the same cache line.
1936 // Because of parm1-4 are used together, performance seems to be better
1937 // if they are on the same cache line (not measured though).
1938
1939 struct KMP_ALIGN(32) {
1940 kmp_int32 parm1;
1941 kmp_int32 parm2;
1942 kmp_int32 parm3;
1943 kmp_int32 parm4;
1944 };
1945
1946#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
1947 kmp_uint32 pchunks;
1948 kmp_uint32 num_procs_with_pcore;
1949 kmp_int32 first_thread_with_ecore;
1950#endif
1951#if KMP_OS_WINDOWS
1952 kmp_int32 last_upper;
1953#endif /* KMP_OS_WINDOWS */
1954} dispatch_private_info32_t;
1955
1956#if CACHE_LINE <= 128
1957KMP_BUILD_ASSERT(sizeof(dispatch_private_info32_t) <= 128);
1958#endif
1959
1960typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1961 kmp_int64 count; // current chunk number for static & static-steal scheduling
1962 kmp_int64 ub; /* upper-bound */
1963 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1964 kmp_int64 lb; /* lower-bound */
1965 kmp_int64 st; /* stride */
1966 kmp_int64 tc; /* trip count (number of iterations) */
1967 kmp_lock_t *steal_lock; // lock used for chunk stealing
1968
1969 kmp_uint64 ordered_lower;
1970 kmp_uint64 ordered_upper;
1971 /* parm[1-4] are used in different ways by different scheduling algorithms */
1972
1973 // KMP_ALIGN(32) ensures ( if the KMP_ALIGN macro is turned on )
1974 // a) parm3 is properly aligned and
1975 // b) all parm1-4 are in the same cache line.
1976 // Because of parm1-4 are used together, performance seems to be better
1977 // if they are in the same line (not measured though).
1978 struct KMP_ALIGN(32) {
1979 kmp_int64 parm1;
1980 kmp_int64 parm2;
1981 kmp_int64 parm3;
1982 kmp_int64 parm4;
1983 };
1984
1985#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
1986 kmp_uint64 pchunks;
1987 kmp_uint64 num_procs_with_pcore;
1988 kmp_int64 first_thread_with_ecore;
1989#endif
1990
1991#if KMP_OS_WINDOWS
1992 kmp_int64 last_upper;
1993#endif /* KMP_OS_WINDOWS */
1994} dispatch_private_info64_t;
1995
1996#if CACHE_LINE <= 128
1997KMP_BUILD_ASSERT(sizeof(dispatch_private_info64_t) <= 128);
1998#endif
1999
2000#else /* KMP_STATIC_STEAL_ENABLED */
2001typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
2002 kmp_int32 lb;
2003 kmp_int32 ub;
2004 kmp_int32 st;
2005 kmp_int32 tc;
2006
2007 kmp_int32 parm1;
2008 kmp_int32 parm2;
2009 kmp_int32 parm3;
2010 kmp_int32 parm4;
2011
2012 kmp_int32 count;
2013
2014 kmp_uint32 ordered_lower;
2015 kmp_uint32 ordered_upper;
2016#if KMP_OS_WINDOWS
2017 kmp_int32 last_upper;
2018#endif /* KMP_OS_WINDOWS */
2019} dispatch_private_info32_t;
2020
2021typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
2022 kmp_int64 lb; /* lower-bound */
2023 kmp_int64 ub; /* upper-bound */
2024 kmp_int64 st; /* stride */
2025 kmp_int64 tc; /* trip count (number of iterations) */
2026
2027 /* parm[1-4] are used in different ways by different scheduling algorithms */
2028 kmp_int64 parm1;
2029 kmp_int64 parm2;
2030 kmp_int64 parm3;
2031 kmp_int64 parm4;
2032
2033 kmp_int64 count; /* current chunk number for static scheduling */
2034
2035 kmp_uint64 ordered_lower;
2036 kmp_uint64 ordered_upper;
2037#if KMP_OS_WINDOWS
2038 kmp_int64 last_upper;
2039#endif /* KMP_OS_WINDOWS */
2040} dispatch_private_info64_t;
2041#endif /* KMP_STATIC_STEAL_ENABLED */
2042
2043typedef struct KMP_ALIGN_CACHE dispatch_private_info {
2044 union private_info {
2045 dispatch_private_info32_t p32;
2046 dispatch_private_info64_t p64;
2047 } u;
2048 enum sched_type schedule; /* scheduling algorithm */
2049 kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
2050 std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
2051 kmp_int32 ordered_bumped;
2052 // Stack of buffers for nest of serial regions
2053 struct dispatch_private_info *next;
2054 kmp_int32 type_size; /* the size of types in private_info */
2055#if KMP_USE_HIER_SCHED
2056 kmp_int32 hier_id;
2057 void *parent; /* hierarchical scheduling parent pointer */
2058#endif
2059 enum cons_type pushed_ws;
2060} dispatch_private_info_t;
2061
2062typedef struct dispatch_shared_info32 {
2063 /* chunk index under dynamic, number of idle threads under static-steal;
2064 iteration index otherwise */
2065 volatile kmp_uint32 iteration;
2066 volatile kmp_int32 num_done;
2067 volatile kmp_uint32 ordered_iteration;
2068 // Dummy to retain the structure size after making ordered_iteration scalar
2069 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
2070} dispatch_shared_info32_t;
2071
2072typedef struct dispatch_shared_info64 {
2073 /* chunk index under dynamic, number of idle threads under static-steal;
2074 iteration index otherwise */
2075 volatile kmp_uint64 iteration;
2076 volatile kmp_int64 num_done;
2077 volatile kmp_uint64 ordered_iteration;
2078 // Dummy to retain the structure size after making ordered_iteration scalar
2079 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
2080} dispatch_shared_info64_t;
2081
2082typedef struct dispatch_shared_info {
2083 union shared_info {
2084 dispatch_shared_info32_t s32;
2085 dispatch_shared_info64_t s64;
2086 } u;
2087 volatile kmp_uint32 buffer_index;
2088 volatile kmp_int32 doacross_buf_idx; // teamwise index
2089 volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
2090 kmp_int32 doacross_num_done; // count finished threads
2091#if KMP_USE_HIER_SCHED
2092 void *hier;
2093#endif
2094#if KMP_HWLOC_ENABLED
2095 // When linking with libhwloc, the ORDERED EPCC test slows down on big
2096 // machines (> 48 cores). Performance analysis showed that a cache thrash
2097 // was occurring and this padding helps alleviate the problem.
2098 char padding[64];
2099#endif // KMP_HWLOC_ENABLED
2100} dispatch_shared_info_t;
2101
2102typedef struct kmp_disp {
2103 /* Vector for ORDERED SECTION */
2104 void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
2105 /* Vector for END ORDERED SECTION */
2106 void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
2107
2108 dispatch_shared_info_t *th_dispatch_sh_current;
2109 dispatch_private_info_t *th_dispatch_pr_current;
2110
2111 dispatch_private_info_t *th_disp_buffer;
2112 kmp_uint32 th_disp_index;
2113 kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
2114 volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
2115 kmp_int64 *th_doacross_info; // info on loop bounds
2116#if KMP_USE_INTERNODE_ALIGNMENT
2117 char more_padding[INTERNODE_CACHE_LINE];
2118#endif
2119} kmp_disp_t;
2120
2121/* ------------------------------------------------------------------------ */
2122/* Barrier stuff */
2123
2124/* constants for barrier state update */
2125#define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
2126#define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
2127#define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
2128#define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
2129
2130#define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
2131#define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
2132#define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
2133
2134#if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
2135#error "Barrier sleep bit must be smaller than barrier bump bit"
2136#endif
2137#if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
2138#error "Barrier unused bit must be smaller than barrier bump bit"
2139#endif
2140
2141// Constants for release barrier wait state: currently, hierarchical only
2142#define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
2143#define KMP_BARRIER_OWN_FLAG \
2144 1 // Normal state; worker waiting on own b_go flag in release
2145#define KMP_BARRIER_PARENT_FLAG \
2146 2 // Special state; worker waiting on parent's b_go flag in release
2147#define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
2148 3 // Special state; tells worker to shift from parent to own b_go
2149#define KMP_BARRIER_SWITCHING \
2150 4 // Special state; worker resets appropriate flag on wake-up
2151
2152#define KMP_NOT_SAFE_TO_REAP \
2153 0 // Thread th_reap_state: not safe to reap (tasking)
2154#define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
2155
2156// The flag_type describes the storage used for the flag.
2157enum flag_type {
2158 flag32,
2159 flag64,
2160 atomic_flag64,
2161 flag_oncore,
2162 flag_unset
2163};
2164
2165enum barrier_type {
2166 bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
2167 barriers if enabled) */
2168 bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
2169#if KMP_FAST_REDUCTION_BARRIER
2170 bs_reduction_barrier, /* 2, All barriers that are used in reduction */
2171#endif // KMP_FAST_REDUCTION_BARRIER
2172 bs_last_barrier /* Just a placeholder to mark the end */
2173};
2174
2175// to work with reduction barriers just like with plain barriers
2176#if !KMP_FAST_REDUCTION_BARRIER
2177#define bs_reduction_barrier bs_plain_barrier
2178#endif // KMP_FAST_REDUCTION_BARRIER
2179
2180typedef enum kmp_bar_pat { /* Barrier communication patterns */
2181 bp_linear_bar =
2182 0, /* Single level (degenerate) tree */
2183 bp_tree_bar =
2184 1, /* Balanced tree with branching factor 2^n */
2185 bp_hyper_bar = 2, /* Hypercube-embedded tree with min
2186 branching factor 2^n */
2187 bp_hierarchical_bar = 3, /* Machine hierarchy tree */
2188 bp_dist_bar = 4, /* Distributed barrier */
2189 bp_last_bar /* Placeholder to mark the end */
2190} kmp_bar_pat_e;
2191
2192#define KMP_BARRIER_ICV_PUSH 1
2193
2194/* Record for holding the values of the internal controls stack records */
2195typedef struct kmp_internal_control {
2196 int serial_nesting_level; /* corresponds to the value of the
2197 th_team_serialized field */
2198 kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
2199 thread) */
2200 kmp_int8
2201 bt_set; /* internal control for whether blocktime is explicitly set */
2202 int blocktime; /* internal control for blocktime */
2203#if KMP_USE_MONITOR
2204 int bt_intervals; /* internal control for blocktime intervals */
2205#endif
2206 int nproc; /* internal control for #threads for next parallel region (per
2207 thread) */
2208 int thread_limit; /* internal control for thread-limit-var */
2209 int task_thread_limit; /* internal control for thread-limit-var of a task*/
2210 int max_active_levels; /* internal control for max_active_levels */
2211 kmp_r_sched_t
2212 sched; /* internal control for runtime schedule {sched,chunk} pair */
2213 kmp_proc_bind_t proc_bind; /* internal control for affinity */
2214 kmp_int32 default_device; /* internal control for default device */
2215 struct kmp_internal_control *next;
2216} kmp_internal_control_t;
2217
2218static inline void copy_icvs(kmp_internal_control_t *dst,
2219 kmp_internal_control_t *src) {
2220 *dst = *src;
2221}
2222
2223/* Thread barrier needs volatile barrier fields */
2224typedef struct KMP_ALIGN_CACHE kmp_bstate {
2225 // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2226 // uses of it). It is not explicitly aligned below, because we *don't* want
2227 // it to be padded -- instead, we fit b_go into the same cache line with
2228 // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2229 kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2230 // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2231 // same NGO store
2232 volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2233 KMP_ALIGN_CACHE volatile kmp_uint64
2234 b_arrived; // STATE => task reached synch point.
2235 kmp_uint32 *skip_per_level;
2236 kmp_uint32 my_level;
2237 kmp_int32 parent_tid;
2238 kmp_int32 old_tid;
2239 kmp_uint32 depth;
2240 struct kmp_bstate *parent_bar;
2241 kmp_team_t *team;
2242 kmp_uint64 leaf_state;
2243 kmp_uint32 nproc;
2244 kmp_uint8 base_leaf_kids;
2245 kmp_uint8 leaf_kids;
2246 kmp_uint8 offset;
2247 kmp_uint8 wait_flag;
2248 kmp_uint8 use_oncore_barrier;
2249#if USE_DEBUGGER
2250 // The following field is intended for the debugger solely. Only the worker
2251 // thread itself accesses this field: the worker increases it by 1 when it
2252 // arrives to a barrier.
2253 KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2254#endif /* USE_DEBUGGER */
2255} kmp_bstate_t;
2256
2257union KMP_ALIGN_CACHE kmp_barrier_union {
2258 double b_align; /* use worst case alignment */
2259 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2260 kmp_bstate_t bb;
2261};
2262
2263typedef union kmp_barrier_union kmp_balign_t;
2264
2265/* Team barrier needs only non-volatile arrived counter */
2266union KMP_ALIGN_CACHE kmp_barrier_team_union {
2267 double b_align; /* use worst case alignment */
2268 char b_pad[CACHE_LINE];
2269 struct {
2270 kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2271#if USE_DEBUGGER
2272 // The following two fields are indended for the debugger solely. Only
2273 // primary thread of the team accesses these fields: the first one is
2274 // increased by 1 when the primary thread arrives to a barrier, the second
2275 // one is increased by one when all the threads arrived.
2276 kmp_uint b_master_arrived;
2277 kmp_uint b_team_arrived;
2278#endif
2279 };
2280};
2281
2282typedef union kmp_barrier_team_union kmp_balign_team_t;
2283
2284/* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2285 threads when a condition changes. This is to workaround an NPTL bug where
2286 padding was added to pthread_cond_t which caused the initialization routine
2287 to write outside of the structure if compiled on pre-NPTL threads. */
2288#if KMP_OS_WINDOWS
2289typedef struct kmp_win32_mutex {
2290 /* The Lock */
2291 CRITICAL_SECTION cs;
2292} kmp_win32_mutex_t;
2293
2294typedef struct kmp_win32_cond {
2295 /* Count of the number of waiters. */
2296 int waiters_count_;
2297
2298 /* Serialize access to <waiters_count_> */
2299 kmp_win32_mutex_t waiters_count_lock_;
2300
2301 /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2302 int release_count_;
2303
2304 /* Keeps track of the current "generation" so that we don't allow */
2305 /* one thread to steal all the "releases" from the broadcast. */
2306 int wait_generation_count_;
2307
2308 /* A manual-reset event that's used to block and release waiting threads. */
2309 HANDLE event_;
2310} kmp_win32_cond_t;
2311#endif
2312
2313#if KMP_OS_UNIX
2314
2315union KMP_ALIGN_CACHE kmp_cond_union {
2316 double c_align;
2317 char c_pad[CACHE_LINE];
2318 pthread_cond_t c_cond;
2319};
2320
2321typedef union kmp_cond_union kmp_cond_align_t;
2322
2323union KMP_ALIGN_CACHE kmp_mutex_union {
2324 double m_align;
2325 char m_pad[CACHE_LINE];
2326 pthread_mutex_t m_mutex;
2327};
2328
2329typedef union kmp_mutex_union kmp_mutex_align_t;
2330
2331#endif /* KMP_OS_UNIX */
2332
2333typedef struct kmp_desc_base {
2334 void *ds_stackbase;
2335 size_t ds_stacksize;
2336 int ds_stackgrow;
2337 kmp_thread_t ds_thread;
2338 volatile int ds_tid;
2339 int ds_gtid;
2340#if KMP_OS_WINDOWS
2341 volatile int ds_alive;
2342 DWORD ds_thread_id;
2343/* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2344 However, debugger support (libomp_db) cannot work with handles, because they
2345 uncomparable. For example, debugger requests info about thread with handle h.
2346 h is valid within debugger process, and meaningless within debugee process.
2347 Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2348 within debugee process, but it is a *new* handle which does *not* equal to
2349 any other handle in debugee... The only way to compare handles is convert
2350 them to system-wide ids. GetThreadId() function is available only in
2351 Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2352 on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2353 thread id by call to GetCurrentThreadId() from within the thread and save it
2354 to let libomp_db identify threads. */
2355#endif /* KMP_OS_WINDOWS */
2356} kmp_desc_base_t;
2357
2358typedef union KMP_ALIGN_CACHE kmp_desc {
2359 double ds_align; /* use worst case alignment */
2360 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2361 kmp_desc_base_t ds;
2362} kmp_desc_t;
2363
2364typedef struct kmp_local {
2365 volatile int this_construct; /* count of single's encountered by thread */
2366 void *reduce_data;
2367#if KMP_USE_BGET
2368 void *bget_data;
2369 void *bget_list;
2370#if !USE_CMP_XCHG_FOR_BGET
2371#ifdef USE_QUEUING_LOCK_FOR_BGET
2372 kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2373#else
2374 kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2375// bootstrap lock so we can use it at library
2376// shutdown.
2377#endif /* USE_LOCK_FOR_BGET */
2378#endif /* ! USE_CMP_XCHG_FOR_BGET */
2379#endif /* KMP_USE_BGET */
2380
2381 PACKED_REDUCTION_METHOD_T
2382 packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2383 __kmpc_end_reduce*() */
2384
2385} kmp_local_t;
2386
2387#define KMP_CHECK_UPDATE(a, b) \
2388 if ((a) != (b)) \
2389 (a) = (b)
2390#define KMP_CHECK_UPDATE_SYNC(a, b) \
2391 if ((a) != (b)) \
2392 TCW_SYNC_PTR((a), (b))
2393
2394#define get__blocktime(xteam, xtid) \
2395 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2396#define get__bt_set(xteam, xtid) \
2397 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2398#if KMP_USE_MONITOR
2399#define get__bt_intervals(xteam, xtid) \
2400 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2401#endif
2402
2403#define get__dynamic_2(xteam, xtid) \
2404 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2405#define get__nproc_2(xteam, xtid) \
2406 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2407#define get__sched_2(xteam, xtid) \
2408 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2409
2410#define set__blocktime_team(xteam, xtid, xval) \
2411 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2412 (xval))
2413
2414#if KMP_USE_MONITOR
2415#define set__bt_intervals_team(xteam, xtid, xval) \
2416 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2417 (xval))
2418#endif
2419
2420#define set__bt_set_team(xteam, xtid, xval) \
2421 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2422
2423#define set__dynamic(xthread, xval) \
2424 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2425#define get__dynamic(xthread) \
2426 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2427
2428#define set__nproc(xthread, xval) \
2429 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2430
2431#define set__thread_limit(xthread, xval) \
2432 (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2433
2434#define set__max_active_levels(xthread, xval) \
2435 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2436
2437#define get__max_active_levels(xthread) \
2438 ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2439
2440#define set__sched(xthread, xval) \
2441 (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2442
2443#define set__proc_bind(xthread, xval) \
2444 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2445#define get__proc_bind(xthread) \
2446 ((xthread)->th.th_current_task->td_icvs.proc_bind)
2447
2448// OpenMP tasking data structures
2449
2450typedef enum kmp_tasking_mode {
2451 tskm_immediate_exec = 0,
2452 tskm_extra_barrier = 1,
2453 tskm_task_teams = 2,
2454 tskm_max = 2
2455} kmp_tasking_mode_t;
2456
2457extern kmp_tasking_mode_t
2458 __kmp_tasking_mode; /* determines how/when to execute tasks */
2459extern int __kmp_task_stealing_constraint;
2460extern int __kmp_enable_task_throttling;
2461extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2462// specified, defaults to 0 otherwise
2463// Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2464extern kmp_int32 __kmp_max_task_priority;
2465// Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2466extern kmp_uint64 __kmp_taskloop_min_tasks;
2467
2468/* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2469 taskdata first */
2470#define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2471#define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2472
2473// The tt_found_tasks flag is a signal to all threads in the team that tasks
2474// were spawned and queued since the previous barrier release.
2475#define KMP_TASKING_ENABLED(task_team) \
2476 (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2481
2484typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2485
2486typedef union kmp_cmplrdata {
2487 kmp_int32 priority;
2488 kmp_routine_entry_t
2489 destructors; /* pointer to function to invoke deconstructors of
2490 firstprivate C++ objects */
2491 /* future data */
2492} kmp_cmplrdata_t;
2493
2494/* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2497typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2498 void *shareds;
2499 kmp_routine_entry_t
2500 routine;
2501 kmp_int32 part_id;
2502 kmp_cmplrdata_t
2503 data1; /* Two known optional additions: destructors and priority */
2504 kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2505 /* future data */
2506 /* private vars */
2507} kmp_task_t;
2508
2512
2513typedef struct kmp_taskgroup {
2514 std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2515 std::atomic<kmp_int32>
2516 cancel_request; // request for cancellation of this taskgroup
2517 struct kmp_taskgroup *parent; // parent taskgroup
2518 // Block of data to perform task reduction
2519 void *reduce_data; // reduction related info
2520 kmp_int32 reduce_num_data; // number of data items to reduce
2521 uintptr_t *gomp_data; // gomp reduction data
2522} kmp_taskgroup_t;
2523
2524// forward declarations
2525typedef union kmp_depnode kmp_depnode_t;
2526typedef struct kmp_depnode_list kmp_depnode_list_t;
2527typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2528
2529// macros for checking dep flag as an integer
2530#define KMP_DEP_IN 0x1
2531#define KMP_DEP_OUT 0x2
2532#define KMP_DEP_INOUT 0x3
2533#define KMP_DEP_MTX 0x4
2534#define KMP_DEP_SET 0x8
2535#define KMP_DEP_ALL 0x80
2536// Compiler sends us this info. Note: some test cases contain an explicit copy
2537// of this struct and should be in sync with any changes here.
2538typedef struct kmp_depend_info {
2539 kmp_intptr_t base_addr;
2540 size_t len;
2541 union {
2542 kmp_uint8 flag; // flag as an unsigned char
2543 struct { // flag as a set of 8 bits
2544#if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2545 /* Same fields as in the #else branch, but in reverse order */
2546 unsigned all : 1;
2547 unsigned unused : 3;
2548 unsigned set : 1;
2549 unsigned mtx : 1;
2550 unsigned out : 1;
2551 unsigned in : 1;
2552#else
2553 unsigned in : 1;
2554 unsigned out : 1;
2555 unsigned mtx : 1;
2556 unsigned set : 1;
2557 unsigned unused : 3;
2558 unsigned all : 1;
2559#endif
2560 } flags;
2561 };
2562} kmp_depend_info_t;
2563
2564// Internal structures to work with task dependencies:
2565struct kmp_depnode_list {
2566 kmp_depnode_t *node;
2567 kmp_depnode_list_t *next;
2568};
2569
2570// Max number of mutexinoutset dependencies per node
2571#define MAX_MTX_DEPS 4
2572
2573typedef struct kmp_base_depnode {
2574 kmp_depnode_list_t *successors; /* used under lock */
2575 kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2576 kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2577 kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2578 kmp_lock_t lock; /* guards shared fields: task, successors */
2579#if KMP_SUPPORT_GRAPH_OUTPUT
2580 kmp_uint32 id;
2581#endif
2582 std::atomic<kmp_int32> npredecessors;
2583 std::atomic<kmp_int32> nrefs;
2584} kmp_base_depnode_t;
2585
2586union KMP_ALIGN_CACHE kmp_depnode {
2587 double dn_align; /* use worst case alignment */
2588 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2589 kmp_base_depnode_t dn;
2590};
2591
2592struct kmp_dephash_entry {
2593 kmp_intptr_t addr;
2594 kmp_depnode_t *last_out;
2595 kmp_depnode_list_t *last_set;
2596 kmp_depnode_list_t *prev_set;
2597 kmp_uint8 last_flag;
2598 kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2599 kmp_dephash_entry_t *next_in_bucket;
2600};
2601
2602typedef struct kmp_dephash {
2603 kmp_dephash_entry_t **buckets;
2604 size_t size;
2605 kmp_depnode_t *last_all;
2606 size_t generation;
2607 kmp_uint32 nelements;
2608 kmp_uint32 nconflicts;
2609} kmp_dephash_t;
2610
2611typedef struct kmp_task_affinity_info {
2612 kmp_intptr_t base_addr;
2613 size_t len;
2614 struct {
2615 bool flag1 : 1;
2616 bool flag2 : 1;
2617 kmp_int32 reserved : 30;
2618 } flags;
2619} kmp_task_affinity_info_t;
2620
2621typedef enum kmp_event_type_t {
2622 KMP_EVENT_UNINITIALIZED = 0,
2623 KMP_EVENT_ALLOW_COMPLETION = 1
2624} kmp_event_type_t;
2625
2626typedef struct {
2627 kmp_event_type_t type;
2628 kmp_tas_lock_t lock;
2629 union {
2630 kmp_task_t *task;
2631 } ed;
2632} kmp_event_t;
2633
2634#if OMP_TASKGRAPH_EXPERIMENTAL
2635
2637typedef enum kmp_tdg_status {
2638 KMP_TDG_NONE = 0,
2639 KMP_TDG_RECORDING = 1,
2640 KMP_TDG_READY = 2
2641} kmp_tdg_status_t;
2642
2643#endif
2644
2645typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2646#if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2647 /* Same fields as in the #else branch, but in reverse order */
2648 unsigned reserved31 : 5;
2649 unsigned hidden_helper : 1;
2650 unsigned target : 1;
2651 unsigned native : 1;
2652 unsigned freed : 1;
2653 unsigned complete : 1;
2654 unsigned executing : 1;
2655 unsigned started : 1;
2656 unsigned team_serial : 1;
2657 unsigned tasking_ser : 1;
2658 unsigned task_serial : 1;
2659 unsigned tasktype : 1;
2660 unsigned reserved : 7;
2661 unsigned transparent : 1;
2662 unsigned free_agent_eligible : 1;
2663 unsigned detachable : 1;
2664 unsigned priority_specified : 1;
2665 unsigned proxy : 1;
2666 unsigned destructors_thunk : 1;
2667 unsigned merged_if0 : 1;
2668 unsigned final : 1;
2669 unsigned tiedness : 1;
2670#else
2671 /* Compiler flags */ /* Total compiler flags must be 16 bits */
2672 unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2673 unsigned final : 1; /* task is final(1) so execute immediately */
2674 unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2675 code path */
2676 unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2677 invoke destructors from the runtime */
2678 unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2679 context of the RTL) */
2680 unsigned priority_specified : 1; /* set if the compiler provides priority
2681 setting for the task */
2682 unsigned detachable : 1; /* 1 == can detach */
2683 unsigned free_agent_eligible : 1; /* set if task can be executed by a
2684 free-agent thread */
2685 unsigned transparent : 1; /* transparent task support (compiler uses this) */
2686 unsigned reserved : 7; /* reserved for compiler use */
2687
2688 /* Library flags */ /* Total library flags must be 16 bits */
2689 unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2690 unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2691 unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2692 // (1) or may be deferred (0)
2693 unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2694 // (0) [>= 2 threads]
2695 /* If either team_serial or tasking_ser is set, task team may be NULL */
2696 /* Task State Flags: */
2697 unsigned started : 1; /* 1==started, 0==not started */
2698 unsigned executing : 1; /* 1==executing, 0==not executing */
2699 unsigned complete : 1; /* 1==complete, 0==not complete */
2700 unsigned freed : 1; /* 1==freed, 0==allocated */
2701 unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2702 unsigned target : 1;
2703 unsigned hidden_helper : 1; /* 1 == hidden helper task */
2704 unsigned reserved31 : 5; /* reserved for library use */
2705#endif
2706} kmp_tasking_flags_t;
2707
2708typedef struct kmp_target_data {
2709 void *async_handle; // libomptarget async handle for task completion query
2710} kmp_target_data_t;
2711
2712struct kmp_taskdata { /* aligned during dynamic allocation */
2713 kmp_int32 td_task_id; /* id, assigned by debugger */
2714 kmp_tasking_flags_t td_flags; /* task flags */
2715 kmp_team_t *td_team; /* team for this task */
2716 kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2717 /* Currently not used except for perhaps IDB */
2718 kmp_taskdata_t *td_parent; /* parent task */
2719 kmp_int32 td_level; /* task nesting level */
2720 std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2721 ident_t *td_ident; /* task identifier */
2722 // Taskwait data.
2723 ident_t *td_taskwait_ident;
2724 kmp_uint32 td_taskwait_counter;
2725 kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2726 KMP_ALIGN_CACHE kmp_internal_control_t
2727 td_icvs; /* Internal control variables for the task */
2728 KMP_ALIGN_CACHE std::atomic<kmp_int32>
2729 td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2730 deallocated */
2731 std::atomic<kmp_int32>
2732 td_incomplete_child_tasks; /* Child tasks not yet complete */
2733 kmp_taskgroup_t
2734 *td_taskgroup; // Each task keeps pointer to its current taskgroup
2735 kmp_dephash_t
2736 *td_dephash; // Dependencies for children tasks are tracked from here
2737 kmp_depnode_t
2738 *td_depnode; // Pointer to graph node if this task has dependencies
2739 kmp_task_team_t *td_task_team;
2740 size_t td_size_alloc; // Size of task structure, including shareds etc.
2741#if defined(KMP_GOMP_COMPAT)
2742 // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2743 kmp_int32 td_size_loop_bounds;
2744#endif
2745 kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2746#if defined(KMP_GOMP_COMPAT)
2747 // GOMP sends in a copy function for copy constructors
2748 void (*td_copy_func)(void *, void *);
2749#endif
2750 kmp_event_t td_allow_completion_event;
2751#if OMPT_SUPPORT
2752 ompt_task_info_t ompt_task_info;
2753#endif
2754#if OMP_TASKGRAPH_EXPERIMENTAL
2755#endif
2756 kmp_target_data_t td_target_data;
2757}; // struct kmp_taskdata
2758
2759// Make sure padding above worked
2760KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2761
2762// Data for task team but per thread
2763typedef struct kmp_base_thread_data {
2764 kmp_info_p *td_thr; // Pointer back to thread info
2765 // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2766 // queued?
2767 kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2768 kmp_taskdata_t *
2769 *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2770 kmp_int32 td_deque_size; // Size of deck
2771 kmp_uint32 td_deque_head; // Head of deque (will wrap)
2772 kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2773 kmp_int32 td_deque_ntasks; // Number of tasks in deque
2774 // GEH: shouldn't this be volatile since used in while-spin?
2775 kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2776} kmp_base_thread_data_t;
2777
2778#define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2779#define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2780
2781#define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2782#define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2783
2784typedef union KMP_ALIGN_CACHE kmp_thread_data {
2785 kmp_base_thread_data_t td;
2786 double td_align; /* use worst case alignment */
2787 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2788} kmp_thread_data_t;
2789
2790typedef struct kmp_task_pri {
2791 kmp_thread_data_t td;
2792 kmp_int32 priority;
2793 kmp_task_pri *next;
2794} kmp_task_pri_t;
2795
2796// Data for task teams which are used when tasking is enabled for the team
2797typedef struct kmp_base_task_team {
2798 kmp_bootstrap_lock_t
2799 tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2800 /* must be bootstrap lock since used at library shutdown*/
2801
2802 // TODO: check performance vs kmp_tas_lock_t
2803 kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */
2804 kmp_task_pri_t *tt_task_pri_list;
2805
2806 kmp_task_team_t *tt_next; /* For linking the task team free list */
2807 kmp_thread_data_t
2808 *tt_threads_data; /* Array of per-thread structures for task team */
2809 /* Data survives task team deallocation */
2810 kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2811 executing this team? */
2812 /* TRUE means tt_threads_data is set up and initialized */
2813 kmp_int32 tt_nproc; /* #threads in team */
2814 kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2815 kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2816 kmp_int32 tt_untied_task_encountered;
2817 std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued
2818 // There is hidden helper thread encountered in this task team so that we must
2819 // wait when waiting on task team
2820 kmp_int32 tt_hidden_helper_task_encountered;
2821
2822 KMP_ALIGN_CACHE
2823 std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2824
2825 KMP_ALIGN_CACHE
2826 volatile kmp_uint32
2827 tt_active; /* is the team still actively executing tasks */
2828} kmp_base_task_team_t;
2829
2830union KMP_ALIGN_CACHE kmp_task_team {
2831 kmp_base_task_team_t tt;
2832 double tt_align; /* use worst case alignment */
2833 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2834};
2835
2836typedef struct kmp_task_team_list_t {
2837 kmp_task_team_t *task_team;
2838 kmp_task_team_list_t *next;
2839} kmp_task_team_list_t;
2840
2841#if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2842// Free lists keep same-size free memory slots for fast memory allocation
2843// routines
2844typedef struct kmp_free_list {
2845 void *th_free_list_self; // Self-allocated tasks free list
2846 void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2847 // threads
2848 void *th_free_list_other; // Non-self free list (to be returned to owner's
2849 // sync list)
2850} kmp_free_list_t;
2851#endif
2852// Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2853// are not put in teams pool, and they don't put threads in threads pool.
2854typedef struct kmp_hot_team_ptr {
2855 kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2856 kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2857} kmp_hot_team_ptr_t;
2858typedef struct kmp_teams_size {
2859 kmp_int32 nteams; // number of teams in a league
2860 kmp_int32 nth; // number of threads in each team of the league
2861} kmp_teams_size_t;
2862
2863// This struct stores a thread that acts as a "root" for a contention
2864// group. Contention groups are rooted at kmp_root threads, but also at
2865// each primary thread of each team created in the teams construct.
2866// This struct therefore also stores a thread_limit associated with
2867// that contention group, and a counter to track the number of threads
2868// active in that contention group. Each thread has a list of these: CG
2869// root threads have an entry in their list in which cg_root refers to
2870// the thread itself, whereas other workers in the CG will have a
2871// single entry where cg_root is same as the entry containing their CG
2872// root. When a thread encounters a teams construct, it will add a new
2873// entry to the front of its list, because it now roots a new CG.
2874typedef struct kmp_cg_root {
2875 kmp_info_p *cg_root; // "root" thread for a contention group
2876 // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2877 // thread_limit clause for teams primary threads
2878 kmp_int32 cg_thread_limit;
2879 kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2880 struct kmp_cg_root *up; // pointer to higher level CG root in list
2881} kmp_cg_root_t;
2882
2883// OpenMP thread data structures
2884
2885typedef struct KMP_ALIGN_CACHE kmp_base_info {
2886 /* Start with the readonly data which is cache aligned and padded. This is
2887 written before the thread starts working by the primary thread. Uber
2888 masters may update themselves later. Usage does not consider serialized
2889 regions. */
2890 kmp_desc_t th_info;
2891 kmp_team_p *th_team; /* team we belong to */
2892 kmp_root_p *th_root; /* pointer to root of task hierarchy */
2893 kmp_info_p *th_next_pool; /* next available thread in the pool */
2894 kmp_disp_t *th_dispatch; /* thread's dispatch data */
2895 int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2896
2897 /* The following are cached from the team info structure */
2898 /* TODO use these in more places as determined to be needed via profiling */
2899 int th_team_nproc; /* number of threads in a team */
2900 kmp_info_p *th_team_master; /* the team's primary thread */
2901 int th_team_serialized; /* team is serialized */
2902 microtask_t th_teams_microtask; /* save entry address for teams construct */
2903 int th_teams_level; /* save initial level of teams construct */
2904/* it is 0 on device but may be any on host */
2905
2906/* The blocktime info is copied from the team struct to the thread struct */
2907/* at the start of a barrier, and the values stored in the team are used */
2908/* at points in the code where the team struct is no longer guaranteed */
2909/* to exist (from the POV of worker threads). */
2910#if KMP_USE_MONITOR
2911 int th_team_bt_intervals;
2912 int th_team_bt_set;
2913#else
2914 kmp_uint64 th_team_bt_intervals;
2915#endif
2916
2917#if KMP_AFFINITY_SUPPORTED
2918 kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2919 kmp_affinity_ids_t th_topology_ids; /* thread's current topology ids */
2920 kmp_affinity_attrs_t th_topology_attrs; /* thread's current topology attrs */
2921#endif
2922 omp_allocator_handle_t th_def_allocator; /* default allocator */
2923 /* The data set by the primary thread at reinit, then R/W by the worker */
2924 KMP_ALIGN_CACHE int
2925 th_set_nproc; /* if > 0, then only use this request for the next fork */
2926 int *th_set_nested_nth;
2927 bool th_nt_strict; // num_threads clause has strict modifier
2928 ident_t *th_nt_loc; // loc for strict modifier
2929 int th_nt_sev; // error severity for strict modifier
2930 const char *th_nt_msg; // error message for strict modifier
2931 int th_set_nested_nth_sz;
2932 kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2933 kmp_proc_bind_t
2934 th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2935 kmp_teams_size_t
2936 th_teams_size; /* number of teams/threads in teams construct */
2937#if KMP_AFFINITY_SUPPORTED
2938 int th_current_place; /* place currently bound to */
2939 int th_new_place; /* place to bind to in par reg */
2940 int th_first_place; /* first place in partition */
2941 int th_last_place; /* last place in partition */
2942#endif
2943 int th_prev_level; /* previous level for affinity format */
2944 int th_prev_num_threads; /* previous num_threads for affinity format */
2945#if USE_ITT_BUILD
2946 kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2947 kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2948 kmp_uint64 th_frame_time; /* frame timestamp */
2949#endif /* USE_ITT_BUILD */
2950 kmp_local_t th_local;
2951 struct private_common *th_pri_head;
2952
2953 /* Now the data only used by the worker (after initial allocation) */
2954 /* TODO the first serial team should actually be stored in the info_t
2955 structure. this will help reduce initial allocation overhead */
2956 KMP_ALIGN_CACHE kmp_team_p
2957 *th_serial_team; /*serialized team held in reserve*/
2958
2959#if OMPT_SUPPORT
2960 ompt_thread_info_t ompt_thread_info;
2961#endif
2962
2963 /* The following are also read by the primary thread during reinit */
2964 struct common_table *th_pri_common;
2965
2966 volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2967 /* while awaiting queuing lock acquire */
2968
2969 volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2970 flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2971
2972 ident_t *th_ident;
2973 unsigned th_x; // Random number generator data
2974 unsigned th_a; // Random number generator data
2975
2976 /* Tasking-related data for the thread */
2977 kmp_task_team_t *th_task_team; // Task team struct
2978 kmp_taskdata_t *th_current_task; // Innermost Task being executed
2979 kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2980 kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2981 // tasking, thus safe to reap
2982
2983 /* More stuff for keeping track of active/sleeping threads (this part is
2984 written by the worker thread) */
2985 kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2986 int th_active; // ! sleeping; 32 bits for TCR/TCW
2987 std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
2988 // 0 = not used in team; 1 = used in team;
2989 // 2 = transitioning to not used in team; 3 = transitioning to used in team
2990 struct cons_header *th_cons; // used for consistency check
2991#if KMP_USE_HIER_SCHED
2992 // used for hierarchical scheduling
2993 kmp_hier_private_bdata_t *th_hier_bar_data;
2994#endif
2995
2996 /* Add the syncronizing data which is cache aligned and padded. */
2997 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2998
2999 KMP_ALIGN_CACHE volatile kmp_int32
3000 th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
3001
3002#if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
3003#define NUM_LISTS 4
3004 kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
3005// allocation routines
3006#endif
3007
3008#if KMP_OS_WINDOWS
3009 kmp_win32_cond_t th_suspend_cv;
3010 kmp_win32_mutex_t th_suspend_mx;
3011 std::atomic<int> th_suspend_init;
3012#endif
3013#if KMP_OS_UNIX
3014 kmp_cond_align_t th_suspend_cv;
3015 kmp_mutex_align_t th_suspend_mx;
3016 std::atomic<int> th_suspend_init_count;
3017#endif
3018
3019#if USE_ITT_BUILD
3020 kmp_itt_mark_t th_itt_mark_single;
3021// alignment ???
3022#endif /* USE_ITT_BUILD */
3023#if KMP_STATS_ENABLED
3024 kmp_stats_list *th_stats;
3025#endif
3026#if KMP_OS_UNIX
3027 std::atomic<bool> th_blocking;
3028#endif
3029 kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
3030} kmp_base_info_t;
3031
3032typedef union KMP_ALIGN_CACHE kmp_info {
3033 double th_align; /* use worst case alignment */
3034 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
3035 kmp_base_info_t th;
3036} kmp_info_t;
3037
3038// OpenMP thread team data structures
3039
3040typedef struct kmp_base_data {
3041 volatile kmp_uint32 t_value;
3042} kmp_base_data_t;
3043
3044typedef union KMP_ALIGN_CACHE kmp_sleep_team {
3045 double dt_align; /* use worst case alignment */
3046 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3047 kmp_base_data_t dt;
3048} kmp_sleep_team_t;
3049
3050typedef union KMP_ALIGN_CACHE kmp_ordered_team {
3051 double dt_align; /* use worst case alignment */
3052 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3053 kmp_base_data_t dt;
3054} kmp_ordered_team_t;
3055
3056typedef int (*launch_t)(int gtid);
3057
3058/* Minimum number of ARGV entries to malloc if necessary */
3059#define KMP_MIN_MALLOC_ARGV_ENTRIES 100
3060
3061// Set up how many argv pointers will fit in cache lines containing
3062// t_inline_argv. Historically, we have supported at least 96 bytes. Using a
3063// larger value for more space between the primary write/worker read section and
3064// read/write by all section seems to buy more performance on EPCC PARALLEL.
3065#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3066#define KMP_INLINE_ARGV_BYTES \
3067 (4 * CACHE_LINE - \
3068 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
3069 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
3070 CACHE_LINE))
3071#else
3072#define KMP_INLINE_ARGV_BYTES \
3073 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
3074#endif
3075#define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
3076
3077typedef struct KMP_ALIGN_CACHE kmp_base_team {
3078 // Synchronization Data
3079 // ---------------------------------------------------------------------------
3080 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
3081 kmp_balign_team_t t_bar[bs_last_barrier];
3082 std::atomic<int> t_construct; // count of single directive encountered by team
3083 char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
3084
3085 // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
3086 std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
3087 std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
3088
3089 // Primary thread only
3090 // ---------------------------------------------------------------------------
3091 KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
3092 int t_master_this_cons; // "this_construct" single counter of primary thread
3093 // in parent team
3094 ident_t *t_ident; // if volatile, have to change too much other crud to
3095 // volatile too
3096 kmp_team_p *t_parent; // parent team
3097 kmp_team_p *t_next_pool; // next free team in the team pool
3098 kmp_disp_t *t_dispatch; // thread's dispatch data
3099 kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
3100 kmp_proc_bind_t t_proc_bind; // bind type for par region
3101 int t_primary_task_state; // primary thread's task state saved
3102#if USE_ITT_BUILD
3103 kmp_uint64 t_region_time; // region begin timestamp
3104#endif /* USE_ITT_BUILD */
3105
3106 // Primary thread write, workers read
3107 // --------------------------------------------------------------------------
3108 KMP_ALIGN_CACHE void **t_argv;
3109 int t_argc;
3110 int t_nproc; // number of threads in team
3111 microtask_t t_pkfn;
3112 launch_t t_invoke; // procedure to launch the microtask
3113
3114#if OMPT_SUPPORT
3115 ompt_team_info_t ompt_team_info;
3116 ompt_lw_taskteam_t *ompt_serialized_team_info;
3117#endif
3118
3119#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3120 kmp_int8 t_fp_control_saved;
3121 kmp_int8 t_pad2b;
3122 kmp_int16 t_x87_fpu_control_word; // FP control regs
3123 kmp_uint32 t_mxcsr;
3124#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3125
3126 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
3127
3128 KMP_ALIGN_CACHE kmp_info_t **t_threads;
3129 kmp_taskdata_t
3130 *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
3131 int t_level; // nested parallel level
3132
3133 KMP_ALIGN_CACHE int t_max_argc;
3134 int t_max_nproc; // max threads this team can handle (dynamically expandable)
3135 int t_serialized; // levels deep of serialized teams
3136 dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
3137 int t_id; // team's id, assigned by debugger.
3138 int t_active_level; // nested active parallel level
3139 kmp_r_sched_t t_sched; // run-time schedule for the team
3140#if KMP_AFFINITY_SUPPORTED
3141 int t_first_place; // first & last place in parent thread's partition.
3142 int t_last_place; // Restore these values to primary thread after par region.
3143#endif // KMP_AFFINITY_SUPPORTED
3144 int t_display_affinity;
3145 int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
3146 // omp_set_num_threads() call
3147 omp_allocator_handle_t t_def_allocator; /* default allocator */
3148
3149// Read/write by workers as well
3150#if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
3151 // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
3152 // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
3153 // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
3154 // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
3155 char dummy_padding[1024];
3156#endif
3157 // Internal control stack for additional nested teams.
3158 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
3159 // for SERIALIZED teams nested 2 or more levels deep
3160 // typed flag to store request state of cancellation
3161 std::atomic<kmp_int32> t_cancel_request;
3162 int t_master_active; // save on fork, restore on join
3163 void *t_copypriv_data; // team specific pointer to copyprivate data array
3164#if KMP_OS_WINDOWS
3165 std::atomic<kmp_uint32> t_copyin_counter;
3166#endif
3167#if USE_ITT_BUILD
3168 void *t_stack_id; // team specific stack stitching id (for ittnotify)
3169#endif /* USE_ITT_BUILD */
3170 distributedBarrier *b; // Distributed barrier data associated with team
3171 kmp_nested_nthreads_t *t_nested_nth;
3172} kmp_base_team_t;
3173
3174// Assert that the list structure fits and aligns within
3175// the double task team pointer
3176KMP_BUILD_ASSERT(sizeof(kmp_task_team_t *[2]) == sizeof(kmp_task_team_list_t));
3177KMP_BUILD_ASSERT(alignof(kmp_task_team_t *[2]) ==
3178 alignof(kmp_task_team_list_t));
3179
3180union KMP_ALIGN_CACHE kmp_team {
3181 kmp_base_team_t t;
3182 double t_align; /* use worst case alignment */
3183 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
3184};
3185
3186typedef union KMP_ALIGN_CACHE kmp_time_global {
3187 double dt_align; /* use worst case alignment */
3188 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3189 kmp_base_data_t dt;
3190} kmp_time_global_t;
3191
3192typedef struct kmp_base_global {
3193 /* cache-aligned */
3194 kmp_time_global_t g_time;
3195
3196 /* non cache-aligned */
3197 volatile int g_abort;
3198 volatile int g_done;
3199
3200 int g_dynamic;
3201 enum dynamic_mode g_dynamic_mode;
3202} kmp_base_global_t;
3203
3204typedef union KMP_ALIGN_CACHE kmp_global {
3205 kmp_base_global_t g;
3206 double g_align; /* use worst case alignment */
3207 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
3208} kmp_global_t;
3209
3210typedef struct kmp_base_root {
3211 // TODO: GEH - combine r_active with r_in_parallel then r_active ==
3212 // (r_in_parallel>= 0)
3213 // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
3214 // the synch overhead or keeping r_active
3215 volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
3216 // keeps a count of active parallel regions per root
3217 std::atomic<int> r_in_parallel;
3218 // GEH: This is misnamed, should be r_active_levels
3219 kmp_team_t *r_root_team;
3220 kmp_team_t *r_hot_team;
3221 kmp_info_t *r_uber_thread;
3222 kmp_lock_t r_begin_lock;
3223 volatile int r_begin;
3224 int r_blocktime; /* blocktime for this root and descendants */
3225#if KMP_AFFINITY_SUPPORTED
3226 int r_affinity_assigned;
3227#endif // KMP_AFFINITY_SUPPORTED
3228} kmp_base_root_t;
3229
3230typedef union KMP_ALIGN_CACHE kmp_root {
3231 kmp_base_root_t r;
3232 double r_align; /* use worst case alignment */
3233 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
3234} kmp_root_t;
3235
3236struct fortran_inx_info {
3237 kmp_int32 data;
3238};
3239
3240// This list type exists to hold old __kmp_threads arrays so that
3241// old references to them may complete while reallocation takes place when
3242// expanding the array. The items in this list are kept alive until library
3243// shutdown.
3244typedef struct kmp_old_threads_list_t {
3245 kmp_info_t **threads;
3246 struct kmp_old_threads_list_t *next;
3247} kmp_old_threads_list_t;
3248
3249/* ------------------------------------------------------------------------ */
3250
3251extern int __kmp_settings;
3252extern int __kmp_duplicate_library_ok;
3253#if USE_ITT_BUILD
3254extern int __kmp_forkjoin_frames;
3255extern int __kmp_forkjoin_frames_mode;
3256#endif
3257extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3258extern int __kmp_determ_red;
3259
3260#ifdef KMP_DEBUG
3261extern int kmp_a_debug;
3262extern int kmp_b_debug;
3263extern int kmp_c_debug;
3264extern int kmp_d_debug;
3265extern int kmp_e_debug;
3266extern int kmp_f_debug;
3267#endif /* KMP_DEBUG */
3268
3269/* For debug information logging using rotating buffer */
3270#define KMP_DEBUG_BUF_LINES_INIT 512
3271#define KMP_DEBUG_BUF_LINES_MIN 1
3272
3273#define KMP_DEBUG_BUF_CHARS_INIT 128
3274#define KMP_DEBUG_BUF_CHARS_MIN 2
3275
3276extern int
3277 __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3278extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3279extern int
3280 __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3281extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3282 entry pointer */
3283
3284extern char *__kmp_debug_buffer; /* Debug buffer itself */
3285extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3286 printed in buffer so far */
3287extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3288 recommended in warnings */
3289/* end rotating debug buffer */
3290
3291#ifdef KMP_DEBUG
3292extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3293
3294#define KMP_PAR_RANGE_ROUTINE_LEN 1024
3295extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3296#define KMP_PAR_RANGE_FILENAME_LEN 1024
3297extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3298extern int __kmp_par_range_lb;
3299extern int __kmp_par_range_ub;
3300#endif
3301
3302/* For printing out dynamic storage map for threads and teams */
3303extern int
3304 __kmp_storage_map; /* True means print storage map for threads and teams */
3305extern int __kmp_storage_map_verbose; /* True means storage map includes
3306 placement info */
3307extern int __kmp_storage_map_verbose_specified;
3308
3309#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3310extern kmp_cpuinfo_t __kmp_cpuinfo;
3311static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3312#elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3313static inline bool __kmp_is_hybrid_cpu() { return true; }
3314#else
3315static inline bool __kmp_is_hybrid_cpu() { return false; }
3316#endif
3317
3318extern volatile int __kmp_init_serial;
3319extern volatile int __kmp_init_gtid;
3320extern volatile int __kmp_init_common;
3321extern volatile int __kmp_need_register_serial;
3322extern volatile int __kmp_init_middle;
3323extern volatile int __kmp_init_parallel;
3324#if KMP_USE_MONITOR
3325extern volatile int __kmp_init_monitor;
3326#endif
3327extern volatile int __kmp_init_user_locks;
3328extern volatile int __kmp_init_hidden_helper_threads;
3329extern int __kmp_init_counter;
3330extern int __kmp_root_counter;
3331extern int __kmp_version;
3332
3333/* list of address of allocated caches for commons */
3334extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3335
3336/* Barrier algorithm types and options */
3337extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3338extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3339extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3340extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3341extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3342extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3343extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3344extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3345extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3346extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3347extern char const *__kmp_barrier_type_name[bs_last_barrier];
3348extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3349
3350/* Global Locks */
3351extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3352extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3353extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3354extern kmp_bootstrap_lock_t
3355 __kmp_exit_lock; /* exit() is not always thread-safe */
3356#if KMP_USE_MONITOR
3357extern kmp_bootstrap_lock_t
3358 __kmp_monitor_lock; /* control monitor thread creation */
3359#endif
3360extern kmp_bootstrap_lock_t
3361 __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3362 __kmp_threads expansion to co-exist */
3363
3364extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
3365
3366extern enum library_type __kmp_library;
3367
3368extern enum sched_type __kmp_sched; /* default runtime scheduling */
3369extern enum sched_type __kmp_static; /* default static scheduling method */
3370extern enum sched_type __kmp_guided; /* default guided scheduling method */
3371extern enum sched_type __kmp_auto; /* default auto scheduling method */
3372extern int __kmp_chunk; /* default runtime chunk size */
3373extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3374
3375extern size_t __kmp_stksize; /* stack size per thread */
3376#if KMP_USE_MONITOR
3377extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3378#endif
3379extern size_t __kmp_stkoffset; /* stack offset per thread */
3380extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3381
3382extern size_t
3383 __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3384extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3385extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3386extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
3387extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3388extern int __kmp_generate_warnings; /* should we issue warnings? */
3389extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3390
3391#ifdef DEBUG_SUSPEND
3392extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3393#endif
3394
3395extern kmp_int32 __kmp_use_yield;
3396extern kmp_int32 __kmp_use_yield_exp_set;
3397extern kmp_uint32 __kmp_yield_init;
3398extern kmp_uint32 __kmp_yield_next;
3399extern kmp_uint64 __kmp_pause_init;
3400
3401/* ------------------------------------------------------------------------- */
3402extern int __kmp_allThreadsSpecified;
3403
3404extern size_t __kmp_align_alloc;
3405/* following data protected by initialization routines */
3406extern int __kmp_xproc; /* number of processors in the system */
3407extern int __kmp_avail_proc; /* number of processors available to the process */
3408extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3409extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3410// maximum total number of concurrently-existing threads on device
3411extern int __kmp_max_nth;
3412// maximum total number of concurrently-existing threads in a contention group
3413extern int __kmp_cg_max_nth;
3414extern int __kmp_task_max_nth; // max threads used in a task
3415extern int __kmp_teams_max_nth; // max threads used in a teams construct
3416extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3417 __kmp_root */
3418extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3419 region a la OMP_NUM_THREADS */
3420extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3421 initialization */
3422extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3423 used (fixed) */
3424extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3425 (__kmpc_threadprivate_cached()) */
3426extern int __kmp_dflt_blocktime; /* number of microseconds to wait before
3427 blocking (env setting) */
3428extern char __kmp_blocktime_units; /* 'm' or 'u' to note units specified */
3429extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */
3430
3431// Convert raw blocktime from ms to us if needed.
3432static inline void __kmp_aux_convert_blocktime(int *bt) {
3433 if (__kmp_blocktime_units == 'm') {
3434 if (*bt > INT_MAX / 1000) {
3435 *bt = INT_MAX / 1000;
3436 KMP_INFORM(MaxValueUsing, "kmp_set_blocktime(ms)", bt);
3437 }
3438 *bt = *bt * 1000;
3439 }
3440}
3441
3442#if KMP_USE_MONITOR
3443extern int
3444 __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3445extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3446 blocking */
3447#endif
3448#ifdef KMP_ADJUST_BLOCKTIME
3449extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3450#endif /* KMP_ADJUST_BLOCKTIME */
3451#ifdef KMP_DFLT_NTH_CORES
3452extern int __kmp_ncores; /* Total number of cores for threads placement */
3453#endif
3454/* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3455extern int __kmp_abort_delay;
3456
3457extern int __kmp_need_register_atfork_specified;
3458extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3459 to install fork handler */
3460extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3461 0 - not set, will be set at runtime
3462 1 - using stack search
3463 2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3464 X*) or TlsGetValue(Windows* OS))
3465 3 - static TLS (__declspec(thread) __kmp_gtid),
3466 Linux* OS .so only. */
3467extern int
3468 __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3469#ifdef KMP_TDATA_GTID
3470extern KMP_THREAD_LOCAL int __kmp_gtid;
3471#endif
3472extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3473extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3474#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3475extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3476extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3477extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3478#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3479
3480// max_active_levels for nested parallelism enabled by default via
3481// OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3482extern int __kmp_dflt_max_active_levels;
3483// Indicates whether value of __kmp_dflt_max_active_levels was already
3484// explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3485extern bool __kmp_dflt_max_active_levels_set;
3486extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3487 concurrent execution per team */
3488extern int __kmp_hot_teams_mode;
3489extern int __kmp_hot_teams_max_level;
3490
3491#if KMP_MIC_SUPPORTED
3492extern enum mic_type __kmp_mic_type;
3493#endif
3494
3495#ifdef USE_LOAD_BALANCE
3496extern double __kmp_load_balance_interval; // load balance algorithm interval
3497#endif /* USE_LOAD_BALANCE */
3498
3499#if KMP_USE_ADAPTIVE_LOCKS
3500
3501// Parameters for the speculative lock backoff system.
3502struct kmp_adaptive_backoff_params_t {
3503 // Number of soft retries before it counts as a hard retry.
3504 kmp_uint32 max_soft_retries;
3505 // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3506 // the right
3507 kmp_uint32 max_badness;
3508};
3509
3510extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3511
3512#if KMP_DEBUG_ADAPTIVE_LOCKS
3513extern const char *__kmp_speculative_statsfile;
3514#endif
3515
3516#endif // KMP_USE_ADAPTIVE_LOCKS
3517
3518extern int __kmp_display_env; /* TRUE or FALSE */
3519extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3520extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3521extern int __kmp_nteams;
3522extern int __kmp_teams_thread_limit;
3523
3524/* ------------------------------------------------------------------------- */
3525
3526/* the following are protected by the fork/join lock */
3527/* write: lock read: anytime */
3528extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3529/* Holds old arrays of __kmp_threads until library shutdown */
3530extern kmp_old_threads_list_t *__kmp_old_threads_list;
3531/* read/write: lock */
3532extern volatile kmp_team_t *__kmp_team_pool;
3533extern volatile kmp_info_t *__kmp_thread_pool;
3534extern kmp_info_t *__kmp_thread_pool_insert_pt;
3535
3536// total num threads reachable from some root thread including all root threads
3537extern volatile int __kmp_nth;
3538/* total number of threads reachable from some root thread including all root
3539 threads, and those in the thread pool */
3540extern volatile int __kmp_all_nth;
3541extern std::atomic<int> __kmp_thread_pool_active_nth;
3542
3543extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3544/* end data protected by fork/join lock */
3545/* ------------------------------------------------------------------------- */
3546
3547#define __kmp_get_gtid() __kmp_get_global_thread_id()
3548#define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3549#define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3550#define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3551#define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3552
3553// AT: Which way is correct?
3554// AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3555// AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3556#define __kmp_get_team_num_threads(gtid) \
3557 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3558
3559static inline bool KMP_UBER_GTID(int gtid) {
3560 KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3561 KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3562 return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3563 __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3564}
3565
3566static inline int __kmp_tid_from_gtid(int gtid) {
3567 KMP_DEBUG_ASSERT(gtid >= 0);
3568 return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3569}
3570
3571static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3572 KMP_DEBUG_ASSERT(tid >= 0 && team);
3573 return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3574}
3575
3576static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3577 KMP_DEBUG_ASSERT(thr);
3578 return thr->th.th_info.ds.ds_gtid;
3579}
3580
3581static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3582 KMP_DEBUG_ASSERT(gtid >= 0);
3583 return __kmp_threads[gtid];
3584}
3585
3586static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3587 KMP_DEBUG_ASSERT(gtid >= 0);
3588 return __kmp_threads[gtid]->th.th_team;
3589}
3590
3591static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3592 if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3593 KMP_FATAL(ThreadIdentInvalid);
3594}
3595
3596#if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3597extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3598extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3599extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3600extern int __kmp_mwait_hints; // Hints to pass in to mwait
3601#endif
3602
3603#if KMP_HAVE_UMWAIT
3604extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3605extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3606extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3607extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3608#endif
3609
3610/* ------------------------------------------------------------------------- */
3611
3612extern kmp_global_t __kmp_global; /* global status */
3613
3614extern kmp_info_t __kmp_monitor;
3615// For Debugging Support Library
3616extern std::atomic<kmp_int32> __kmp_team_counter;
3617// For Debugging Support Library
3618extern std::atomic<kmp_int32> __kmp_task_counter;
3619
3620#if USE_DEBUGGER
3621#define _KMP_GEN_ID(counter) \
3622 (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3623#else
3624#define _KMP_GEN_ID(counter) (~0)
3625#endif /* USE_DEBUGGER */
3626
3627#define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3628#define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3629
3630/* ------------------------------------------------------------------------ */
3631
3632extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3633 size_t size, char const *format, ...);
3634
3635extern void __kmp_serial_initialize(void);
3636extern void __kmp_middle_initialize(void);
3637extern void __kmp_parallel_initialize(void);
3638
3639extern void __kmp_internal_begin(void);
3640extern void __kmp_internal_end_library(int gtid);
3641extern void __kmp_internal_end_thread(int gtid);
3642extern void __kmp_internal_end_atexit(void);
3643extern void __kmp_internal_end_dtor(void);
3644extern void __kmp_internal_end_dest(void *);
3645
3646extern int __kmp_register_root(int initial_thread);
3647extern void __kmp_unregister_root(int gtid);
3648extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3649
3650extern int __kmp_ignore_mppbeg(void);
3651extern int __kmp_ignore_mppend(void);
3652
3653extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3654extern void __kmp_exit_single(int gtid);
3655
3656extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3657extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3658
3659#ifdef USE_LOAD_BALANCE
3660extern int __kmp_get_load_balance(int);
3661#endif
3662
3663extern int __kmp_get_global_thread_id(void);
3664extern int __kmp_get_global_thread_id_reg(void);
3665extern void __kmp_exit_thread(int exit_status);
3666extern void __kmp_abort(char const *format, ...);
3667extern void __kmp_abort_thread(void);
3668KMP_NORETURN extern void __kmp_abort_process(void);
3669extern void __kmp_warn(char const *format, ...);
3670
3671extern void __kmp_set_num_threads(int new_nth, int gtid);
3672
3673extern bool __kmp_detect_shm();
3674extern bool __kmp_detect_tmp();
3675
3676// Returns current thread (pointer to kmp_info_t). Current thread *must* be
3677// registered.
3678static inline kmp_info_t *__kmp_entry_thread() {
3679 int gtid = __kmp_entry_gtid();
3680
3681 return __kmp_threads[gtid];
3682}
3683
3684extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3685extern int __kmp_get_max_active_levels(int gtid);
3686extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3687extern int __kmp_get_team_size(int gtid, int level);
3688extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3689extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3690
3691extern unsigned short __kmp_get_random(kmp_info_t *thread);
3692extern void __kmp_init_random(kmp_info_t *thread);
3693
3694extern kmp_r_sched_t __kmp_get_schedule_global(void);
3695extern void __kmp_adjust_num_threads(int new_nproc);
3696extern void __kmp_check_stksize(size_t *val);
3697
3698extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3699extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3700extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3701#define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3702#define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3703#define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3704
3705#if USE_FAST_MEMORY
3706extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3707 size_t size KMP_SRC_LOC_DECL);
3708extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3709extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3710extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3711#define __kmp_fast_allocate(this_thr, size) \
3712 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3713#define __kmp_fast_free(this_thr, ptr) \
3714 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3715#endif
3716
3717extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3718extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3719 size_t elsize KMP_SRC_LOC_DECL);
3720extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3721 size_t size KMP_SRC_LOC_DECL);
3722extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3723#define __kmp_thread_malloc(th, size) \
3724 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3725#define __kmp_thread_calloc(th, nelem, elsize) \
3726 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3727#define __kmp_thread_realloc(th, ptr, size) \
3728 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3729#define __kmp_thread_free(th, ptr) \
3730 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3731
3732extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3733extern void __kmp_push_num_threads_list(ident_t *loc, int gtid,
3734 kmp_uint32 list_length,
3735 int *num_threads_list);
3736extern void __kmp_set_strict_num_threads(ident_t *loc, int gtid, int sev,
3737 const char *msg);
3738
3739extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3740 kmp_proc_bind_t proc_bind);
3741extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3742 int num_threads);
3743extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3744 int num_teams_ub, int num_threads);
3745
3746extern void __kmp_yield();
3747
3748extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3749 enum sched_type schedule, kmp_int32 lb,
3750 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3751extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3752 enum sched_type schedule, kmp_uint32 lb,
3753 kmp_uint32 ub, kmp_int32 st,
3754 kmp_int32 chunk);
3755extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3756 enum sched_type schedule, kmp_int64 lb,
3757 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3758extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3759 enum sched_type schedule, kmp_uint64 lb,
3760 kmp_uint64 ub, kmp_int64 st,
3761 kmp_int64 chunk);
3762
3763extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3764 kmp_int32 *p_last, kmp_int32 *p_lb,
3765 kmp_int32 *p_ub, kmp_int32 *p_st);
3766extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3767 kmp_int32 *p_last, kmp_uint32 *p_lb,
3768 kmp_uint32 *p_ub, kmp_int32 *p_st);
3769extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3770 kmp_int32 *p_last, kmp_int64 *p_lb,
3771 kmp_int64 *p_ub, kmp_int64 *p_st);
3772extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3773 kmp_int32 *p_last, kmp_uint64 *p_lb,
3774 kmp_uint64 *p_ub, kmp_int64 *p_st);
3775
3776extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3777extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3778extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3779extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3780
3781extern void __kmpc_dispatch_deinit(ident_t *loc, kmp_int32 gtid);
3782
3783#ifdef KMP_GOMP_COMPAT
3784
3785extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3786 enum sched_type schedule, kmp_int32 lb,
3787 kmp_int32 ub, kmp_int32 st,
3788 kmp_int32 chunk, int push_ws);
3789extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3790 enum sched_type schedule, kmp_uint32 lb,
3791 kmp_uint32 ub, kmp_int32 st,
3792 kmp_int32 chunk, int push_ws);
3793extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3794 enum sched_type schedule, kmp_int64 lb,
3795 kmp_int64 ub, kmp_int64 st,
3796 kmp_int64 chunk, int push_ws);
3797extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3798 enum sched_type schedule, kmp_uint64 lb,
3799 kmp_uint64 ub, kmp_int64 st,
3800 kmp_int64 chunk, int push_ws);
3801extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3802extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3803extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3804extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3805
3806#endif /* KMP_GOMP_COMPAT */
3807
3808extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3809extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3810extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3811extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3812extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3813extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3814 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3815 void *obj);
3816extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3817 kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3818
3819extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3820 int final_spin
3821#if USE_ITT_BUILD
3822 ,
3823 void *itt_sync_obj
3824#endif
3825);
3826extern void __kmp_release_64(kmp_flag_64<> *flag);
3827
3828extern void __kmp_infinite_loop(void);
3829
3830extern void __kmp_cleanup(void);
3831
3832#if KMP_HANDLE_SIGNALS
3833extern int __kmp_handle_signals;
3834extern void __kmp_install_signals(int parallel_init);
3835extern void __kmp_remove_signals(void);
3836#endif
3837
3838extern void __kmp_clear_system_time(void);
3839extern void __kmp_read_system_time(double *delta);
3840
3841extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3842
3843extern void __kmp_expand_host_name(char *buffer, size_t size);
3844extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3845
3846#if KMP_ARCH_X86 || KMP_ARCH_X86_64 || \
3847 (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM || KMP_ARCH_ARM64EC))
3848extern void
3849__kmp_initialize_system_tick(void); /* Initialize timer tick value */
3850#endif
3851
3852extern void
3853__kmp_runtime_initialize(void); /* machine specific initialization */
3854extern void __kmp_runtime_destroy(void);
3855
3856#if KMP_AFFINITY_SUPPORTED
3857extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3858 kmp_affin_mask_t *mask);
3859extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3860 kmp_affin_mask_t *mask);
3861extern void __kmp_affinity_initialize(kmp_affinity_t &affinity);
3862extern void __kmp_affinity_uninitialize(void);
3863extern void __kmp_affinity_set_init_mask(
3864 int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3865void __kmp_affinity_bind_init_mask(int gtid);
3866extern void __kmp_affinity_bind_place(int gtid);
3867extern void __kmp_affinity_determine_capable(const char *env_var);
3868extern int __kmp_aux_set_affinity(void **mask);
3869extern int __kmp_aux_get_affinity(void **mask);
3870extern int __kmp_aux_get_affinity_max_proc();
3871extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3872extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3873extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3874extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3875#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
3876extern int __kmp_get_first_osid_with_ecore(void);
3877#endif
3878#if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
3879 KMP_OS_AIX
3880extern int kmp_set_thread_affinity_mask_initial(void);
3881#endif
3882static inline void __kmp_assign_root_init_mask() {
3883 int gtid = __kmp_entry_gtid();
3884 kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3885 if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3886 __kmp_affinity_set_init_mask(gtid, /*isa_root=*/TRUE);
3887 __kmp_affinity_bind_init_mask(gtid);
3888 r->r.r_affinity_assigned = TRUE;
3889 }
3890}
3891static inline void __kmp_reset_root_init_mask(int gtid) {
3892 if (!KMP_AFFINITY_CAPABLE())
3893 return;
3894 kmp_info_t *th = __kmp_threads[gtid];
3895 kmp_root_t *r = th->th.th_root;
3896 if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3897 __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3898 KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3899 r->r.r_affinity_assigned = FALSE;
3900 }
3901}
3902#else /* KMP_AFFINITY_SUPPORTED */
3903#define __kmp_assign_root_init_mask() /* Nothing */
3904static inline void __kmp_reset_root_init_mask(int gtid) {}
3905#endif /* KMP_AFFINITY_SUPPORTED */
3906// No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3907// format string is for affinity, so platforms that do not support
3908// affinity can still use the other fields, e.g., %n for num_threads
3909extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3910 kmp_str_buf_t *buffer);
3911extern void __kmp_aux_display_affinity(int gtid, const char *format);
3912
3913extern void __kmp_cleanup_hierarchy();
3914extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3915
3916#if KMP_USE_FUTEX
3917
3918extern int __kmp_futex_determine_capable(void);
3919
3920#endif // KMP_USE_FUTEX
3921
3922extern void __kmp_gtid_set_specific(int gtid);
3923extern int __kmp_gtid_get_specific(void);
3924
3925extern double __kmp_read_cpu_time(void);
3926
3927extern int __kmp_read_system_info(struct kmp_sys_info *info);
3928
3929#if KMP_USE_MONITOR
3930extern void __kmp_create_monitor(kmp_info_t *th);
3931#endif
3932
3933extern void *__kmp_launch_thread(kmp_info_t *thr);
3934
3935extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3936
3937#if KMP_OS_WINDOWS
3938extern int __kmp_still_running(kmp_info_t *th);
3939extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3940extern void __kmp_free_handle(kmp_thread_t tHandle);
3941#endif
3942
3943#if KMP_USE_MONITOR
3944extern void __kmp_reap_monitor(kmp_info_t *th);
3945#endif
3946extern void __kmp_reap_worker(kmp_info_t *th);
3947extern void __kmp_terminate_thread(int gtid);
3948
3949extern int __kmp_try_suspend_mx(kmp_info_t *th);
3950extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3951extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3952
3953extern void __kmp_elapsed(double *);
3954extern void __kmp_elapsed_tick(double *);
3955
3956extern void __kmp_enable(int old_state);
3957extern void __kmp_disable(int *old_state);
3958
3959extern void __kmp_thread_sleep(int millis);
3960
3961extern void __kmp_common_initialize(void);
3962extern void __kmp_common_destroy(void);
3963extern void __kmp_common_destroy_gtid(int gtid);
3964
3965#if KMP_OS_UNIX
3966extern void __kmp_register_atfork(void);
3967#endif
3968extern void __kmp_suspend_initialize(void);
3969extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3970extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3971
3972extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3973 int tid);
3974extern kmp_team_t *__kmp_allocate_team(kmp_root_t *root, int new_nproc,
3975 int max_nproc,
3976#if OMPT_SUPPORT
3977 ompt_data_t ompt_parallel_data,
3978#endif
3979 kmp_proc_bind_t proc_bind,
3980 kmp_internal_control_t *new_icvs,
3981 int argc, kmp_info_t *thr);
3982extern void __kmp_free_thread(kmp_info_t *);
3983extern void __kmp_free_team(kmp_root_t *, kmp_team_t *, kmp_info_t *);
3984extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3985
3986/* ------------------------------------------------------------------------ */
3987
3988extern void __kmp_initialize_bget(kmp_info_t *th);
3989extern void __kmp_finalize_bget(kmp_info_t *th);
3990
3991KMP_EXPORT void *kmpc_malloc(size_t size);
3992KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3993KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3994KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3995KMP_EXPORT void kmpc_free(void *ptr);
3996
3997/* declarations for internal use */
3998
3999extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
4000 size_t reduce_size, void *reduce_data,
4001 void (*reduce)(void *, void *));
4002extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
4003extern int __kmp_barrier_gomp_cancel(int gtid);
4004
4009enum fork_context_e {
4010 fork_context_gnu,
4012 fork_context_intel,
4013 fork_context_last
4014};
4015extern int __kmp_fork_call(ident_t *loc, int gtid,
4016 enum fork_context_e fork_context, kmp_int32 argc,
4017 microtask_t microtask, launch_t invoker,
4018 kmp_va_list ap);
4019
4020extern void __kmp_join_call(ident_t *loc, int gtid
4021#if OMPT_SUPPORT
4022 ,
4023 enum fork_context_e fork_context
4024#endif
4025 ,
4026 int exit_teams = 0);
4027
4028extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
4029extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
4030extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
4031extern int __kmp_invoke_task_func(int gtid);
4032extern void __kmp_run_before_invoked_task(int gtid, int tid,
4033 kmp_info_t *this_thr,
4034 kmp_team_t *team);
4035extern void __kmp_run_after_invoked_task(int gtid, int tid,
4036 kmp_info_t *this_thr,
4037 kmp_team_t *team);
4038
4039// should never have been exported
4040KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
4041extern int __kmp_invoke_teams_master(int gtid);
4042extern void __kmp_teams_master(int gtid);
4043extern int __kmp_aux_get_team_num();
4044extern int __kmp_aux_get_num_teams();
4045extern void __kmp_save_internal_controls(kmp_info_t *thread);
4046extern void __kmp_user_set_library(enum library_type arg);
4047extern void __kmp_aux_set_library(enum library_type arg);
4048extern void __kmp_aux_set_stacksize(size_t arg);
4049extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
4050extern void __kmp_aux_set_defaults(char const *str, size_t len);
4051
4052/* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
4053void kmpc_set_blocktime(int arg);
4054void ompc_set_nested(int flag);
4055void ompc_set_dynamic(int flag);
4056void ompc_set_num_threads(int arg);
4057void kmpc_set_num_threads_8(kmp_int64 arg);
4058
4059extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
4060 kmp_team_t *team, int tid);
4061extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
4062extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4063 kmp_tasking_flags_t *flags,
4064 size_t sizeof_kmp_task_t,
4065 size_t sizeof_shareds,
4066 kmp_routine_entry_t task_entry);
4067extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
4068 kmp_team_t *team, int tid,
4069 int set_curr_task);
4070extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
4071extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
4072
4073extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
4074 int gtid,
4075 kmp_task_t *task);
4076extern void __kmp_fulfill_event(kmp_event_t *event);
4077
4078extern void __kmp_free_task_team(kmp_info_t *thread,
4079 kmp_task_team_t *task_team);
4080extern void __kmp_reap_task_teams(void);
4081extern void __kmp_push_task_team_node(kmp_info_t *thread, kmp_team_t *team);
4082extern void __kmp_pop_task_team_node(kmp_info_t *thread, kmp_team_t *team);
4083extern void __kmp_wait_to_unref_task_teams(void);
4084extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team);
4085extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
4086extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
4087#if USE_ITT_BUILD
4088 ,
4089 void *itt_sync_obj
4090#endif /* USE_ITT_BUILD */
4091 ,
4092 int wait = 1);
4093extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
4094 int gtid);
4095#if KMP_DEBUG
4096#define KMP_DEBUG_ASSERT_TASKTEAM_INVARIANT(team, thr) \
4097 KMP_DEBUG_ASSERT( \
4098 __kmp_tasking_mode != tskm_task_teams || team->t.t_nproc == 1 || \
4099 thr->th.th_task_team == team->t.t_task_team[thr->th.th_task_state])
4100#else
4101#define KMP_DEBUG_ASSERT_TASKTEAM_INVARIANT(team, thr) /* Nothing */
4102#endif
4103
4104extern int __kmp_is_address_mapped(void *addr);
4105extern kmp_uint64 __kmp_hardware_timestamp(void);
4106
4107#if KMP_OS_UNIX
4108extern int __kmp_read_from_file(char const *path, char const *format, ...);
4109#endif
4110
4111/* ------------------------------------------------------------------------ */
4112//
4113// Assembly routines that have no compiler intrinsic replacement
4114//
4115
4116extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
4117 void *argv[]
4118#if OMPT_SUPPORT
4119 ,
4120 void **exit_frame_ptr
4121#endif
4122);
4123
4124/* ------------------------------------------------------------------------ */
4125
4126KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
4127KMP_EXPORT void __kmpc_end(ident_t *);
4128
4129KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
4130 kmpc_ctor_vec ctor,
4131 kmpc_cctor_vec cctor,
4132 kmpc_dtor_vec dtor,
4133 size_t vector_length);
4134KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
4135 kmpc_ctor ctor, kmpc_cctor cctor,
4136 kmpc_dtor dtor);
4137KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
4138 void *data, size_t size);
4139
4140KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
4141KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
4142KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
4143KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
4144
4145KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
4146KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
4147 kmpc_micro microtask, ...);
4148KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs,
4149 kmpc_micro microtask, kmp_int32 cond,
4150 void *args);
4151
4152KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
4153KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
4154
4155// Values of the order and scope arguments of __kmpc_flush_explicit. They are
4156// the compiler's __ATOMIC_* and __MEMORY_SCOPE_* values.
4157enum kmp_flush_order {
4158 kmp_flush_order_relaxed = 0,
4159 kmp_flush_order_consume = 1, // Completes the encoding; never emitted.
4160 kmp_flush_order_acquire = 2,
4161 kmp_flush_order_release = 3,
4162 kmp_flush_order_acq_rel = 4,
4163 kmp_flush_order_seq_cst = 5
4164};
4165enum kmp_flush_scope {
4166 kmp_flush_scope_all = 0, // memscope(all)
4167 kmp_flush_scope_device = 1, // memscope(device)
4168 kmp_flush_scope_cgroup = 2 // memscope(cgroup)
4169};
4170#ifdef __ATOMIC_SEQ_CST
4171static_assert(kmp_flush_order_relaxed == __ATOMIC_RELAXED &&
4172 kmp_flush_order_consume == __ATOMIC_CONSUME &&
4173 kmp_flush_order_acquire == __ATOMIC_ACQUIRE &&
4174 kmp_flush_order_release == __ATOMIC_RELEASE &&
4175 kmp_flush_order_acq_rel == __ATOMIC_ACQ_REL &&
4176 kmp_flush_order_seq_cst == __ATOMIC_SEQ_CST,
4177 "kmp_flush_order must match the __ATOMIC_* values");
4178#endif
4179#ifdef __MEMORY_SCOPE_SYSTEM
4180static_assert(kmp_flush_scope_all == __MEMORY_SCOPE_SYSTEM &&
4181 kmp_flush_scope_device == __MEMORY_SCOPE_DEVICE &&
4182 kmp_flush_scope_cgroup == __MEMORY_SCOPE_WRKGRP,
4183 "kmp_flush_scope must match the __MEMORY_SCOPE_* values");
4184#endif
4185
4186KMP_EXPORT void __kmpc_flush(ident_t *);
4187KMP_EXPORT void __kmpc_flush_explicit(ident_t *, kmp_int32 order,
4188 kmp_int32 scope);
4189KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
4190KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
4191KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
4192KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
4193 kmp_int32 filter);
4194KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
4195KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
4196KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
4197KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
4198 kmp_critical_name *);
4199KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
4200 kmp_critical_name *);
4201KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
4202 kmp_critical_name *, uint32_t hint);
4203
4204KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
4205KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
4206
4207KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
4208 kmp_int32 global_tid);
4209
4210KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
4211KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
4212
4213KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid);
4214KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid,
4215 kmp_int32 numberOfSections);
4216KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid);
4217
4218KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
4219 kmp_int32 schedtype, kmp_int32 *plastiter,
4220 kmp_int *plower, kmp_int *pupper,
4221 kmp_int *pstride, kmp_int incr,
4222 kmp_int chunk);
4223
4224KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
4225
4226KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
4227 size_t cpy_size, void *cpy_data,
4228 void (*cpy_func)(void *, void *),
4229 kmp_int32 didit);
4230
4231KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid,
4232 void *cpy_data);
4233
4234extern void KMPC_SET_NUM_THREADS(int arg);
4235extern void KMPC_SET_DYNAMIC(int flag);
4236extern void KMPC_SET_NESTED(int flag);
4237
4238/* OMP 3.0 tasking interface routines */
4239KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
4240 kmp_task_t *new_task);
4241KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4242 kmp_int32 flags,
4243 size_t sizeof_kmp_task_t,
4244 size_t sizeof_shareds,
4245 kmp_routine_entry_t task_entry);
4246KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
4247 ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
4248 size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
4249KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
4250 kmp_task_t *task);
4251KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
4252 kmp_task_t *task);
4253KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
4254 kmp_task_t *new_task);
4255KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
4256KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
4257 int end_part);
4258
4259#if TASK_UNUSED
4260void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
4261void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
4262 kmp_task_t *task);
4263#endif // TASK_UNUSED
4264
4265/* ------------------------------------------------------------------------ */
4266
4267KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
4268KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
4269
4270KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
4271 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
4272 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
4273 kmp_depend_info_t *noalias_dep_list);
4274
4275KMP_EXPORT kmp_base_depnode_t *__kmpc_task_get_depnode(kmp_task_t *task);
4276
4277KMP_EXPORT kmp_depnode_list_t *__kmpc_task_get_successors(kmp_task_t *task);
4278
4279KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
4280 kmp_int32 ndeps,
4281 kmp_depend_info_t *dep_list,
4282 kmp_int32 ndeps_noalias,
4283 kmp_depend_info_t *noalias_dep_list);
4284/* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause.
4285 * Placeholder for taskwait with nowait clause.*/
4286KMP_EXPORT void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid,
4287 kmp_int32 ndeps,
4288 kmp_depend_info_t *dep_list,
4289 kmp_int32 ndeps_noalias,
4290 kmp_depend_info_t *noalias_dep_list,
4291 kmp_int32 has_no_wait);
4292
4293extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
4294 bool serialize_immediate);
4295
4296KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
4297 kmp_int32 cncl_kind);
4298KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
4299 kmp_int32 cncl_kind);
4300KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
4301KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
4302
4303KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
4304KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
4305KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
4306 kmp_int32 if_val, kmp_uint64 *lb,
4307 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
4308 kmp_int32 sched, kmp_uint64 grainsize,
4309 void *task_dup);
4310KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
4311 kmp_task_t *task, kmp_int32 if_val,
4312 kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4313 kmp_int32 nogroup, kmp_int32 sched,
4314 kmp_uint64 grainsize, kmp_int32 modifier,
4315 void *task_dup);
4316KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
4317KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
4318KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
4319KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
4320 int is_ws, int num,
4321 void *data);
4322KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
4323 int num, void *data);
4324KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
4325 int is_ws);
4326KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
4327 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4328 kmp_task_affinity_info_t *affin_list);
4329KMP_EXPORT void __kmp_set_num_teams(int num_teams);
4330KMP_EXPORT int __kmp_get_max_teams(void);
4331KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
4332KMP_EXPORT int __kmp_get_teams_thread_limit(void);
4333
4334/* Interface target task integration */
4335KMP_EXPORT void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid);
4336KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid);
4337
4338/* Lock interface routines (fast versions with gtid passed in) */
4339KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
4340 void **user_lock);
4341KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
4342 void **user_lock);
4343KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
4344 void **user_lock);
4345KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
4346 void **user_lock);
4347KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4348KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
4349 void **user_lock);
4350KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
4351 void **user_lock);
4352KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
4353 void **user_lock);
4354KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4355KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
4356 void **user_lock);
4357
4358KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4359 void **user_lock, uintptr_t hint);
4360KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4361 void **user_lock,
4362 uintptr_t hint);
4363#if OMP_TASKGRAPH_EXPERIMENTAL
4364#endif
4365/* Interface to fast scalable reduce methods routines */
4366
4367KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
4368 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4369 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4370 kmp_critical_name *lck);
4371KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4372 kmp_critical_name *lck);
4373KMP_EXPORT kmp_int32 __kmpc_reduce(
4374 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4375 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4376 kmp_critical_name *lck);
4377KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4378 kmp_critical_name *lck);
4379
4380/* Internal fast reduction routines */
4381
4382extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4383 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4384 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4385 kmp_critical_name *lck);
4386
4387// this function is for testing set/get/determine reduce method
4388KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4389
4390KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4391KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4392
4393// C++ port
4394// missing 'extern "C"' declarations
4395
4396KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4397KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4398KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4399 kmp_int32 num_threads);
4400KMP_EXPORT void __kmpc_push_num_threads_strict(ident_t *loc,
4401 kmp_int32 global_tid,
4402 kmp_int32 num_threads,
4403 int severity,
4404 const char *message);
4405
4406KMP_EXPORT void __kmpc_push_num_threads_list(ident_t *loc, kmp_int32 global_tid,
4407 kmp_uint32 list_length,
4408 kmp_int32 *num_threads_list);
4409KMP_EXPORT void __kmpc_push_num_threads_list_strict(
4410 ident_t *loc, kmp_int32 global_tid, kmp_uint32 list_length,
4411 kmp_int32 *num_threads_list, int severity, const char *message);
4412
4413KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4414 int proc_bind);
4415KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4416 kmp_int32 num_teams,
4417 kmp_int32 num_threads);
4418KMP_EXPORT void __kmpc_set_thread_limit(ident_t *loc, kmp_int32 global_tid,
4419 kmp_int32 thread_limit);
4420/* Function for OpenMP 5.1 num_teams clause */
4421KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4422 kmp_int32 num_teams_lb,
4423 kmp_int32 num_teams_ub,
4424 kmp_int32 num_threads);
4425KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4426 kmpc_micro microtask, ...);
4427struct kmp_dim { // loop bounds info casted to kmp_int64
4428 kmp_int64 lo; // lower
4429 kmp_int64 up; // upper
4430 kmp_int64 st; // stride
4431};
4432KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4433 kmp_int32 num_dims,
4434 const struct kmp_dim *dims);
4435KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4436 const kmp_int64 *vec);
4437KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4438 const kmp_int64 *vec);
4439KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4440
4441KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4442 void *data, size_t size,
4443 void ***cache);
4444
4445// The routines below are not exported.
4446// Consider making them 'static' in corresponding source files.
4447void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4448 void *data_addr, size_t pc_size);
4449struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4450 void *data_addr,
4451 size_t pc_size);
4452void __kmp_threadprivate_resize_cache(int newCapacity);
4453void __kmp_cleanup_threadprivate_caches();
4454
4455// ompc_, kmpc_ entries moved from omp.h.
4456#if KMP_OS_WINDOWS
4457#define KMPC_CONVENTION __cdecl
4458#else
4459#define KMPC_CONVENTION
4460#endif
4461
4462#ifndef __OMP_H
4463typedef enum omp_sched_t {
4464 omp_sched_static = 1,
4465 omp_sched_dynamic = 2,
4466 omp_sched_guided = 3,
4467 omp_sched_auto = 4
4468} omp_sched_t;
4469typedef void *kmp_affinity_mask_t;
4470#endif
4471
4472KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4473KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4474KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4475KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4476KMP_EXPORT int KMPC_CONVENTION
4477kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4478KMP_EXPORT int KMPC_CONVENTION
4479kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4480KMP_EXPORT int KMPC_CONVENTION
4481kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4482
4483KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4484KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4485KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4486KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4487KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4488void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4489size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4490void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4491size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4492 char const *format);
4493
4494enum kmp_target_offload_kind {
4495 tgt_disabled = 0,
4496 tgt_default = 1,
4497 tgt_mandatory = 2
4498};
4499typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4500// Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4501extern kmp_target_offload_kind_t __kmp_target_offload;
4502extern int __kmpc_get_target_offload();
4503
4504// Constants used in libomptarget
4505#define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4506#define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4507
4508// OMP Pause Resource
4509
4510// The following enum is used both to set the status in __kmp_pause_status, and
4511// as the internal equivalent of the externally-visible omp_pause_resource_t.
4512typedef enum kmp_pause_status_t {
4513 kmp_not_paused = 0, // status is not paused, or, requesting resume
4514 kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4515 kmp_hard_paused = 2, // status is hard-paused, or, requesting hard pause
4516 kmp_stop_tool_paused = 3 // requesting stop_tool pause
4517} kmp_pause_status_t;
4518
4519// This stores the pause state of the runtime
4520extern kmp_pause_status_t __kmp_pause_status;
4521extern int __kmpc_pause_resource(kmp_pause_status_t level);
4522extern int __kmp_pause_resource(kmp_pause_status_t level);
4523// Soft resume sets __kmp_pause_status, and wakes up all threads.
4524extern void __kmp_resume_if_soft_paused();
4525// Hard resume simply resets the status to not paused. Library will appear to
4526// be uninitialized after hard pause. Let OMP constructs trigger required
4527// initializations.
4528static inline void __kmp_resume_if_hard_paused() {
4529 if (__kmp_pause_status == kmp_hard_paused) {
4530 __kmp_pause_status = kmp_not_paused;
4531 }
4532}
4533
4534extern void __kmp_omp_display_env(int verbose);
4535
4536// 1: it is initializing hidden helper team
4537extern volatile int __kmp_init_hidden_helper;
4538// 1: the hidden helper team is done
4539extern volatile int __kmp_hidden_helper_team_done;
4540// 1: enable hidden helper task
4541extern kmp_int32 __kmp_enable_hidden_helper;
4542// Main thread of hidden helper team
4543extern kmp_info_t *__kmp_hidden_helper_main_thread;
4544// Descriptors for the hidden helper threads
4545extern kmp_info_t **__kmp_hidden_helper_threads;
4546// Number of hidden helper threads
4547extern kmp_int32 __kmp_hidden_helper_threads_num;
4548// Number of hidden helper tasks that have not been executed yet
4549extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4550
4551extern void __kmp_hidden_helper_initialize();
4552extern void __kmp_hidden_helper_threads_initz_routine();
4553extern void __kmp_do_initialize_hidden_helper_threads();
4554extern void __kmp_hidden_helper_threads_initz_wait();
4555extern void __kmp_hidden_helper_initz_release();
4556extern void __kmp_hidden_helper_threads_deinitz_wait();
4557extern void __kmp_hidden_helper_threads_deinitz_release();
4558extern void __kmp_hidden_helper_main_thread_wait();
4559extern void __kmp_hidden_helper_worker_thread_wait();
4560extern void __kmp_hidden_helper_worker_thread_signal();
4561extern void __kmp_hidden_helper_main_thread_release();
4562
4563// Check whether a given thread is a hidden helper thread
4564#define KMP_HIDDEN_HELPER_THREAD(gtid) \
4565 ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4566
4567#define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4568 ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4569
4570#define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid) \
4571 ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4572
4573#define KMP_HIDDEN_HELPER_TEAM(team) \
4574 (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4575
4576// Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4577// main thread, is skipped.
4578#define KMP_GTID_TO_SHADOW_GTID(gtid) \
4579 ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4580
4581// Return the adjusted gtid value by subtracting from gtid the number
4582// of hidden helper threads. This adjusted value is the gtid the thread would
4583// have received if there were no hidden helper threads.
4584static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4585 int adjusted_gtid = gtid;
4586 if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4587 gtid - __kmp_hidden_helper_threads_num >= 0) {
4588 adjusted_gtid -= __kmp_hidden_helper_threads_num;
4589 }
4590 return adjusted_gtid;
4591}
4592
4593#if ENABLE_LIBOMPTARGET
4594// Pointers to callbacks registered by the offload library to be notified of
4595// task progress.
4596extern void (*kmp_target_sync_cb)(ident_t *loc_ref, int gtid,
4597 void *current_task, void *event);
4598#endif // ENABLE_LIBOMPTARGET
4599
4600// Support for error directive
4601typedef enum kmp_severity_t {
4602 severity_warning = 1,
4603 severity_fatal = 2
4604} kmp_severity_t;
4605extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4606
4607// Support for scope directive
4608KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4609KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4610
4611#ifdef __cplusplus
4612}
4613#endif
4614
4615template <bool C, bool S>
4616extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4617template <bool C, bool S>
4618extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4619template <bool C, bool S>
4620extern void __kmp_atomic_suspend_64(int th_gtid,
4621 kmp_atomic_flag_64<C, S> *flag);
4622extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4623#if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4624template <bool C, bool S>
4625extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4626template <bool C, bool S>
4627extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4628template <bool C, bool S>
4629extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4630extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4631#endif
4632template <bool C, bool S>
4633extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4634template <bool C, bool S>
4635extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4636template <bool C, bool S>
4637extern void __kmp_atomic_resume_64(int target_gtid,
4638 kmp_atomic_flag_64<C, S> *flag);
4639extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4640
4641template <bool C, bool S>
4642int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4643 kmp_flag_32<C, S> *flag, int final_spin,
4644 int *thread_finished,
4645#if USE_ITT_BUILD
4646 void *itt_sync_obj,
4647#endif /* USE_ITT_BUILD */
4648 kmp_int32 is_constrained);
4649template <bool C, bool S>
4650int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4651 kmp_flag_64<C, S> *flag, int final_spin,
4652 int *thread_finished,
4653#if USE_ITT_BUILD
4654 void *itt_sync_obj,
4655#endif /* USE_ITT_BUILD */
4656 kmp_int32 is_constrained);
4657template <bool C, bool S>
4658int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4659 kmp_atomic_flag_64<C, S> *flag,
4660 int final_spin, int *thread_finished,
4661#if USE_ITT_BUILD
4662 void *itt_sync_obj,
4663#endif /* USE_ITT_BUILD */
4664 kmp_int32 is_constrained);
4665int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4666 kmp_flag_oncore *flag, int final_spin,
4667 int *thread_finished,
4668#if USE_ITT_BUILD
4669 void *itt_sync_obj,
4670#endif /* USE_ITT_BUILD */
4671 kmp_int32 is_constrained);
4672
4673extern int __kmp_nesting_mode;
4674extern int __kmp_nesting_mode_nlevels;
4675extern int *__kmp_nesting_nth_level;
4676extern void __kmp_init_nesting_mode();
4677extern void __kmp_set_nesting_mode_threads();
4678
4685class kmp_safe_raii_file_t {
4686 FILE *f;
4687
4688 void close() {
4689 if (f && f != stdout && f != stderr) {
4690 fclose(f);
4691 f = nullptr;
4692 }
4693 }
4694
4695public:
4696 kmp_safe_raii_file_t() : f(nullptr) {}
4697 kmp_safe_raii_file_t(const char *filename, const char *mode,
4698 const char *env_var = nullptr)
4699 : f(nullptr) {
4700 open(filename, mode, env_var);
4701 }
4702 kmp_safe_raii_file_t(const kmp_safe_raii_file_t &other) = delete;
4703 kmp_safe_raii_file_t &operator=(const kmp_safe_raii_file_t &other) = delete;
4704 ~kmp_safe_raii_file_t() { close(); }
4705
4709 void open(const char *filename, const char *mode,
4710 const char *env_var = nullptr) {
4711 KMP_ASSERT(!f);
4712 f = fopen(filename, mode);
4713 if (!f) {
4714 int code = errno;
4715 if (env_var) {
4716 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4717 KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4718 } else {
4719 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4720 __kmp_msg_null);
4721 }
4722 }
4723 }
4724
4726 int try_open(const char *filename, const char *mode) {
4727 KMP_ASSERT(!f);
4728 f = fopen(filename, mode);
4729 if (!f)
4730 return errno;
4731 return 0;
4732 }
4733
4735 void set_stdout() {
4736 KMP_ASSERT(!f);
4737 f = stdout;
4738 }
4739
4741 void set_stderr() {
4742 KMP_ASSERT(!f);
4743 f = stderr;
4744 }
4745 operator bool() { return bool(f); }
4746 operator FILE *() { return f; }
4747};
4748
4749template <typename SourceType, typename TargetType,
4750 bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4751 bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4752 bool isSourceSigned = std::is_signed<SourceType>::value,
4753 bool isTargetSigned = std::is_signed<TargetType>::value>
4754struct kmp_convert {};
4755
4756// Both types are signed; Source smaller
4757template <typename SourceType, typename TargetType>
4758struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4759 static TargetType to(SourceType src) { return (TargetType)src; }
4760};
4761// Source equal
4762template <typename SourceType, typename TargetType>
4763struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4764 static TargetType to(SourceType src) { return src; }
4765};
4766// Source bigger
4767template <typename SourceType, typename TargetType>
4768struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4769 static TargetType to(SourceType src) {
4770 KMP_ASSERT(src <= static_cast<SourceType>(
4771 (std::numeric_limits<TargetType>::max)()));
4772 KMP_ASSERT(src >= static_cast<SourceType>(
4773 (std::numeric_limits<TargetType>::min)()));
4774 return (TargetType)src;
4775 }
4776};
4777
4778// Source signed, Target unsigned
4779// Source smaller
4780template <typename SourceType, typename TargetType>
4781struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4782 static TargetType to(SourceType src) {
4783 KMP_ASSERT(src >= 0);
4784 return (TargetType)src;
4785 }
4786};
4787// Source equal
4788template <typename SourceType, typename TargetType>
4789struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4790 static TargetType to(SourceType src) {
4791 KMP_ASSERT(src >= 0);
4792 return (TargetType)src;
4793 }
4794};
4795// Source bigger
4796template <typename SourceType, typename TargetType>
4797struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4798 static TargetType to(SourceType src) {
4799 KMP_ASSERT(src >= 0);
4800 KMP_ASSERT(src <= static_cast<SourceType>(
4801 (std::numeric_limits<TargetType>::max)()));
4802 return (TargetType)src;
4803 }
4804};
4805
4806// Source unsigned, Target signed
4807// Source smaller
4808template <typename SourceType, typename TargetType>
4809struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4810 static TargetType to(SourceType src) { return (TargetType)src; }
4811};
4812// Source equal
4813template <typename SourceType, typename TargetType>
4814struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4815 static TargetType to(SourceType src) {
4816 KMP_ASSERT(src <= static_cast<SourceType>(
4817 (std::numeric_limits<TargetType>::max)()));
4818 return (TargetType)src;
4819 }
4820};
4821// Source bigger
4822template <typename SourceType, typename TargetType>
4823struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4824 static TargetType to(SourceType src) {
4825 KMP_ASSERT(src <= static_cast<SourceType>(
4826 (std::numeric_limits<TargetType>::max)()));
4827 return (TargetType)src;
4828 }
4829};
4830
4831// Source unsigned, Target unsigned
4832// Source smaller
4833template <typename SourceType, typename TargetType>
4834struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4835 static TargetType to(SourceType src) { return (TargetType)src; }
4836};
4837// Source equal
4838template <typename SourceType, typename TargetType>
4839struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4840 static TargetType to(SourceType src) { return src; }
4841};
4842// Source bigger
4843template <typename SourceType, typename TargetType>
4844struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4845 static TargetType to(SourceType src) {
4846 KMP_ASSERT(src <= static_cast<SourceType>(
4847 (std::numeric_limits<TargetType>::max)()));
4848 return (TargetType)src;
4849 }
4850};
4851
4852template <typename T1, typename T2>
4853static inline void __kmp_type_convert(T1 src, T2 *dest) {
4854 *dest = kmp_convert<T1, T2>::to(src);
4855}
4856
4857#endif /* KMP_H */
void set_stdout()
Definition kmp.h:4735
void set_stderr()
Definition kmp.h:4741
int try_open(const char *filename, const char *mode)
Definition kmp.h:4726
void open(const char *filename, const char *mode, const char *env_var=nullptr)
Definition kmp.h:4709
struct ident ident_t
@ KMP_IDENT_KMPC
Definition kmp.h:179
@ KMP_IDENT_IMB
Definition kmp.h:177
@ KMP_IDENT_WORK_LOOP
Definition kmp.h:197
@ KMP_IDENT_BARRIER_IMPL
Definition kmp.h:188
@ KMP_IDENT_WORK_SECTIONS
Definition kmp.h:199
@ KMP_IDENT_AUTOPAR
Definition kmp.h:182
@ KMP_IDENT_ATOMIC_HINT_MASK
Definition kmp.h:206
@ KMP_IDENT_WORK_DISTRIBUTE
Definition kmp.h:201
@ KMP_IDENT_BARRIER_EXPL
Definition kmp.h:186
@ KMP_IDENT_ATOMIC_REDUCE
Definition kmp.h:184
KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *)
KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro microtask,...)
KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs, kmpc_micro microtask, kmp_int32 cond, void *args)
KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_set_thread_limit(ident_t *loc, kmp_int32 global_tid, kmp_int32 thread_limit)
KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_push_num_threads_list(ident_t *loc, kmp_int32 global_tid, kmp_uint32 list_length, kmp_int32 *num_threads_list)
KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs, kmpc_micro microtask,...)
KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid)
void(* kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
Definition kmp.h:1777
KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams_lb, kmp_int32 num_teams_ub, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags)
KMP_EXPORT void __kmpc_end(ident_t *)
KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
KMP_EXPORT kmp_int32 __kmpc_reduce(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void(*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck)
KMP_EXPORT void __kmpc_flush_explicit(ident_t *, kmp_int32 order, kmp_int32 scope)
KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_flush(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void(*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck)
KMP_EXPORT void * __kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d)
KMP_EXPORT void * __kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
KMP_EXPORT void * __kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid)
KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask)
KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, int is_ws)
KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins, kmp_task_affinity_info_t *affin_list)
KMP_EXPORT void * __kmpc_task_reduction_init(int gtid, int num_data, void *data)
KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask)
KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
KMP_EXPORT void * __kmpc_taskred_init(int gtid, int num_data, void *data)
KMP_EXPORT void ** __kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid)
void(* kmpc_dtor)(void *)
Definition kmp.h:1801
void *(* kmpc_cctor)(void *, void *)
Definition kmp.h:1808
KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data, kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor)
KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid, size_t cpy_size, void *cpy_data, void(*cpy_func)(void *, void *), kmp_int32 didit)
void *(* kmpc_cctor_vec)(void *, void *, size_t)
Definition kmp.h:1830
void *(* kmpc_ctor)(void *)
Definition kmp.h:1795
KMP_EXPORT void * __kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid, void *cpy_data)
void *(* kmpc_ctor_vec)(void *, size_t)
Definition kmp.h:1818
KMP_EXPORT void * __kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid, void *data, size_t size, void ***cache)
void(* kmpc_dtor_vec)(void *, size_t)
Definition kmp.h:1824
KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data, kmpc_ctor_vec ctor, kmpc_cctor_vec cctor, kmpc_dtor_vec dtor, size_t vector_length)
KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc)
KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *)
KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid)
sched_type
Definition kmp.h:340
KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid, kmp_int32 numberOfSections)
KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid, kmp_int32 filter)
void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_int32 lb, kmp_int32 ub, kmp_int32 st, kmp_int32 chunk)
KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
@ kmp_nm_guided_chunked
Definition kmp.h:391
@ kmp_sch_runtime_simd
Definition kmp.h:362
@ kmp_nm_ord_auto
Definition kmp.h:410
@ kmp_sch_auto
Definition kmp.h:347
@ kmp_nm_auto
Definition kmp.h:393
@ kmp_distribute_static_chunked
Definition kmp.h:378
@ kmp_sch_static
Definition kmp.h:343
@ kmp_sch_guided_simd
Definition kmp.h:361
@ kmp_sch_modifier_monotonic
Definition kmp.h:428
@ kmp_sch_default
Definition kmp.h:448
@ kmp_sch_modifier_nonmonotonic
Definition kmp.h:430
@ kmp_nm_ord_static
Definition kmp.h:406
@ kmp_distribute_static
Definition kmp.h:379
@ kmp_sch_guided_chunked
Definition kmp.h:345
@ kmp_nm_static
Definition kmp.h:389
@ kmp_sch_lower
Definition kmp.h:341
@ kmp_nm_upper
Definition kmp.h:412
@ kmp_ord_lower
Definition kmp.h:367
@ kmp_ord_static
Definition kmp.h:369
@ kmp_sch_upper
Definition kmp.h:365
@ kmp_ord_upper
Definition kmp.h:375
@ kmp_nm_lower
Definition kmp.h:385
@ kmp_ord_auto
Definition kmp.h:373
Definition kmp.h:217
kmp_int32 reserved_1
Definition kmp.h:218
char const * psource
Definition kmp.h:227
kmp_int32 reserved_2
Definition kmp.h:221
kmp_int32 reserved_3
Definition kmp.h:226
kmp_int32 flags
Definition kmp.h:219
Memory allocator information is shared with offload runtime.
Definition kmp.h:1117
Memory space informaition is shared with offload runtime.
Definition kmp.h:1109