LLVM OpenMP* Runtime Library
kmp.h
1
2/*
3 * kmp.h -- KPTS runtime header file.
4 */
5
6//===----------------------------------------------------------------------===//
7//
8// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
9// See https://llvm.org/LICENSE.txt for license information.
10// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
11//
12//===----------------------------------------------------------------------===//
13
14#ifndef KMP_H
15#define KMP_H
16
17#include "kmp_config.h"
18
19/* #define BUILD_PARALLEL_ORDERED 1 */
20
21/* This fix replaces gettimeofday with clock_gettime for better scalability on
22 the Altix. Requires user code to be linked with -lrt. */
23//#define FIX_SGI_CLOCK
24
25/* Defines for OpenMP 3.0 tasking and auto scheduling */
26
27#ifndef KMP_STATIC_STEAL_ENABLED
28#define KMP_STATIC_STEAL_ENABLED 1
29#endif
30
31#define TASK_CURRENT_NOT_QUEUED 0
32#define TASK_CURRENT_QUEUED 1
33
34#ifdef BUILD_TIED_TASK_STACK
35#define TASK_STACK_EMPTY 0 // entries when the stack is empty
36#define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
37// Number of entries in each task stack array
38#define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
39// Mask for determining index into stack block
40#define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
41#endif // BUILD_TIED_TASK_STACK
42
43#define TASK_NOT_PUSHED 1
44#define TASK_SUCCESSFULLY_PUSHED 0
45#define TASK_TIED 1
46#define TASK_UNTIED 0
47#define TASK_EXPLICIT 1
48#define TASK_IMPLICIT 0
49#define TASK_PROXY 1
50#define TASK_FULL 0
51#define TASK_DETACHABLE 1
52#define TASK_UNDETACHABLE 0
53
54#define KMP_CANCEL_THREADS
55#define KMP_THREAD_ATTR
56
57// Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being
58// built on Android
59#if defined(__ANDROID__)
60#undef KMP_CANCEL_THREADS
61#endif
62
63#include <signal.h>
64#include <stdarg.h>
65#include <stddef.h>
66#include <stdio.h>
67#include <stdlib.h>
68#include <string.h>
69#include <limits>
70#include <type_traits>
71/* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
72 Microsoft library. Some macros provided below to replace these functions */
73#ifndef __ABSOFT_WIN
74#include <sys/types.h>
75#endif
76#include <limits.h>
77#include <time.h>
78
79#include <errno.h>
80
81#include "kmp_os.h"
82
83#include "kmp_safe_c_api.h"
84
85#if KMP_STATS_ENABLED
86class kmp_stats_list;
87#endif
88
89#if KMP_USE_HIER_SCHED
90// Only include hierarchical scheduling if affinity is supported
91#undef KMP_USE_HIER_SCHED
92#define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
93#endif
94
95#if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
96#include "hwloc.h"
97#ifndef HWLOC_OBJ_NUMANODE
98#define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
99#endif
100#ifndef HWLOC_OBJ_PACKAGE
101#define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
102#endif
103#endif
104
105#if KMP_ARCH_X86 || KMP_ARCH_X86_64
106#include <xmmintrin.h>
107#endif
108
109// The below has to be defined before including "kmp_barrier.h".
110#define KMP_INTERNAL_MALLOC(sz) malloc(sz)
111#define KMP_INTERNAL_FREE(p) free(p)
112#define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
113#define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
114
115#include "kmp_debug.h"
116#include "kmp_lock.h"
117#include "kmp_version.h"
118#include "kmp_barrier.h"
119#if USE_DEBUGGER
120#include "kmp_debugger.h"
121#endif
122#include "kmp_i18n.h"
123
124#define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
125
126#include "kmp_wrapper_malloc.h"
127#if KMP_OS_UNIX
128#include <unistd.h>
129#if !defined NSIG && defined _NSIG
130#define NSIG _NSIG
131#endif
132#endif
133
134#if KMP_OS_LINUX
135#pragma weak clock_gettime
136#endif
137
138#if OMPT_SUPPORT
139#include "ompt-internal.h"
140#endif
141
142#if OMPD_SUPPORT
143#include "ompd-specific.h"
144#endif
145
146#ifndef UNLIKELY
147#define UNLIKELY(x) (x)
148#endif
149
150// Affinity format function
151#include "kmp_str.h"
152
153// 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
154// 3 - fast allocation using sync, non-sync free lists of any size, non-self
155// free lists of limited size.
156#ifndef USE_FAST_MEMORY
157#define USE_FAST_MEMORY 3
158#endif
159
160#ifndef KMP_NESTED_HOT_TEAMS
161#define KMP_NESTED_HOT_TEAMS 0
162#define USE_NESTED_HOT_ARG(x)
163#else
164#if KMP_NESTED_HOT_TEAMS
165#define USE_NESTED_HOT_ARG(x) , x
166#else
167#define USE_NESTED_HOT_ARG(x)
168#endif
169#endif
170
171// Assume using BGET compare_exchange instruction instead of lock by default.
172#ifndef USE_CMP_XCHG_FOR_BGET
173#define USE_CMP_XCHG_FOR_BGET 1
174#endif
175
176// Test to see if queuing lock is better than bootstrap lock for bget
177// #ifndef USE_QUEUING_LOCK_FOR_BGET
178// #define USE_QUEUING_LOCK_FOR_BGET
179// #endif
180
181#define KMP_NSEC_PER_SEC 1000000000L
182#define KMP_USEC_PER_SEC 1000000L
183
192enum {
197 /* 0x04 is no longer used */
206 KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
207 KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
208 KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
209
210 KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
211 KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
212
224 KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
225 KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
226 KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
227 KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
228 KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
229};
230
234typedef struct ident {
235 kmp_int32 reserved_1;
236 kmp_int32 flags;
238 kmp_int32 reserved_2;
239#if USE_ITT_BUILD
240/* but currently used for storing region-specific ITT */
241/* contextual information. */
242#endif /* USE_ITT_BUILD */
243 kmp_int32 reserved_3;
244 char const *psource;
248 // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
249 kmp_int32 get_openmp_version() {
250 return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
251 }
257// Some forward declarations.
258typedef union kmp_team kmp_team_t;
259typedef struct kmp_taskdata kmp_taskdata_t;
260typedef union kmp_task_team kmp_task_team_t;
261typedef union kmp_team kmp_team_p;
262typedef union kmp_info kmp_info_p;
263typedef union kmp_root kmp_root_p;
264
265template <bool C = false, bool S = true> class kmp_flag_32;
266template <bool C = false, bool S = true> class kmp_flag_64;
267template <bool C = false, bool S = true> class kmp_atomic_flag_64;
268class kmp_flag_oncore;
269
270#ifdef __cplusplus
271extern "C" {
272#endif
273
274/* ------------------------------------------------------------------------ */
275
276/* Pack two 32-bit signed integers into a 64-bit signed integer */
277/* ToDo: Fix word ordering for big-endian machines. */
278#define KMP_PACK_64(HIGH_32, LOW_32) \
279 ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
280
281// Generic string manipulation macros. Assume that _x is of type char *
282#define SKIP_WS(_x) \
283 { \
284 while (*(_x) == ' ' || *(_x) == '\t') \
285 (_x)++; \
286 }
287#define SKIP_DIGITS(_x) \
288 { \
289 while (*(_x) >= '0' && *(_x) <= '9') \
290 (_x)++; \
291 }
292#define SKIP_TOKEN(_x) \
293 { \
294 while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
295 (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
296 (_x)++; \
297 }
298#define SKIP_TO(_x, _c) \
299 { \
300 while (*(_x) != '\0' && *(_x) != (_c)) \
301 (_x)++; \
302 }
303
304/* ------------------------------------------------------------------------ */
305
306#define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
307#define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
308
309/* ------------------------------------------------------------------------ */
310/* Enumeration types */
311
312enum kmp_state_timer {
313 ts_stop,
314 ts_start,
315 ts_pause,
316
317 ts_last_state
318};
319
320enum dynamic_mode {
321 dynamic_default,
322#ifdef USE_LOAD_BALANCE
323 dynamic_load_balance,
324#endif /* USE_LOAD_BALANCE */
325 dynamic_random,
326 dynamic_thread_limit,
327 dynamic_max
328};
329
330/* external schedule constants, duplicate enum omp_sched in omp.h in order to
331 * not include it here */
332#ifndef KMP_SCHED_TYPE_DEFINED
333#define KMP_SCHED_TYPE_DEFINED
334typedef enum kmp_sched {
335 kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
336 // Note: need to adjust __kmp_sch_map global array in case enum is changed
337 kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
338 kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
339 kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
340 kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
341 kmp_sched_upper_std = 5, // upper bound for standard schedules
342 kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
343 kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
344#if KMP_STATIC_STEAL_ENABLED
345 kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
346#endif
347 kmp_sched_upper,
348 kmp_sched_default = kmp_sched_static, // default scheduling
349 kmp_sched_monotonic = 0x80000000
350} kmp_sched_t;
351#endif
352
357enum sched_type : kmp_int32 {
359 kmp_sch_static_chunked = 33,
361 kmp_sch_dynamic_chunked = 35,
363 kmp_sch_runtime = 37,
365 kmp_sch_trapezoidal = 39,
366
367 /* accessible only through KMP_SCHEDULE environment variable */
368 kmp_sch_static_greedy = 40,
369 kmp_sch_static_balanced = 41,
370 /* accessible only through KMP_SCHEDULE environment variable */
371 kmp_sch_guided_iterative_chunked = 42,
372 kmp_sch_guided_analytical_chunked = 43,
373 /* accessible only through KMP_SCHEDULE environment variable */
374 kmp_sch_static_steal = 44,
375
376 /* static with chunk adjustment (e.g., simd) */
377 kmp_sch_static_balanced_chunked = 45,
381 /* accessible only through KMP_SCHEDULE environment variable */
385 kmp_ord_static_chunked = 65,
387 kmp_ord_dynamic_chunked = 67,
388 kmp_ord_guided_chunked = 68,
389 kmp_ord_runtime = 69,
391 kmp_ord_trapezoidal = 71,
394 /* Schedules for Distribute construct */
398 /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
399 single iteration/chunk, even if the loop is serialized. For the schedule
400 types listed above, the entire iteration vector is returned if the loop is
401 serialized. This doesn't work for gcc/gcomp sections. */
404 kmp_nm_static_chunked =
405 (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
407 kmp_nm_dynamic_chunked = 163,
409 kmp_nm_runtime = 165,
411 kmp_nm_trapezoidal = 167,
412
413 /* accessible only through KMP_SCHEDULE environment variable */
414 kmp_nm_static_greedy = 168,
415 kmp_nm_static_balanced = 169,
416 /* accessible only through KMP_SCHEDULE environment variable */
417 kmp_nm_guided_iterative_chunked = 170,
418 kmp_nm_guided_analytical_chunked = 171,
419 kmp_nm_static_steal =
420 172, /* accessible only through OMP_SCHEDULE environment variable */
421
422 kmp_nm_ord_static_chunked = 193,
424 kmp_nm_ord_dynamic_chunked = 195,
425 kmp_nm_ord_guided_chunked = 196,
426 kmp_nm_ord_runtime = 197,
428 kmp_nm_ord_trapezoidal = 199,
431 /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
432 we need to distinguish the three possible cases (no modifier, monotonic
433 modifier, nonmonotonic modifier), we need separate bits for each modifier.
434 The absence of monotonic does not imply nonmonotonic, especially since 4.5
435 says that the behaviour of the "no modifier" case is implementation defined
436 in 4.5, but will become "nonmonotonic" in 5.0.
437
438 Since we're passing a full 32 bit value, we can use a couple of high bits
439 for these flags; out of paranoia we avoid the sign bit.
440
441 These modifiers can be or-ed into non-static schedules by the compiler to
442 pass the additional information. They will be stripped early in the
443 processing in __kmp_dispatch_init when setting up schedules, so most of the
444 code won't ever see schedules with these bits set. */
446 (1 << 29),
448 (1 << 30),
450#define SCHEDULE_WITHOUT_MODIFIERS(s) \
451 (enum sched_type)( \
453#define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
454#define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
455#define SCHEDULE_HAS_NO_MODIFIERS(s) \
456 (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
457#define SCHEDULE_GET_MODIFIERS(s) \
458 ((enum sched_type)( \
459 (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
460#define SCHEDULE_SET_MODIFIERS(s, m) \
461 (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
462#define SCHEDULE_NONMONOTONIC 0
463#define SCHEDULE_MONOTONIC 1
464
467
468// Apply modifiers on internal kind to standard kind
469static inline void
470__kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
471 enum sched_type internal_kind) {
472 if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
473 *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
474 }
475}
476
477// Apply modifiers on standard kind to internal kind
478static inline void
479__kmp_sched_apply_mods_intkind(kmp_sched_t kind,
480 enum sched_type *internal_kind) {
481 if ((int)kind & (int)kmp_sched_monotonic) {
482 *internal_kind = (enum sched_type)((int)*internal_kind |
484 }
485}
486
487// Get standard schedule without modifiers
488static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
489 return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
490}
491
492/* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
493typedef union kmp_r_sched {
494 struct {
495 enum sched_type r_sched_type;
496 int chunk;
497 };
498 kmp_int64 sched;
499} kmp_r_sched_t;
500
501extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
502// internal schedule types
503
504enum library_type {
505 library_none,
506 library_serial,
507 library_turnaround,
508 library_throughput
509};
510
511#if KMP_OS_LINUX
512enum clock_function_type {
513 clock_function_gettimeofday,
514 clock_function_clock_gettime
515};
516#endif /* KMP_OS_LINUX */
517
518#if KMP_MIC_SUPPORTED
519enum mic_type { non_mic, mic1, mic2, mic3, dummy };
520#endif
521
522/* -- fast reduction stuff ------------------------------------------------ */
523
524#undef KMP_FAST_REDUCTION_BARRIER
525#define KMP_FAST_REDUCTION_BARRIER 1
526
527#undef KMP_FAST_REDUCTION_CORE_DUO
528#if KMP_ARCH_X86 || KMP_ARCH_X86_64
529#define KMP_FAST_REDUCTION_CORE_DUO 1
530#endif
531
532enum _reduction_method {
533 reduction_method_not_defined = 0,
534 critical_reduce_block = (1 << 8),
535 atomic_reduce_block = (2 << 8),
536 tree_reduce_block = (3 << 8),
537 empty_reduce_block = (4 << 8)
538};
539
540// Description of the packed_reduction_method variable:
541// The packed_reduction_method variable consists of two enum types variables
542// that are packed together into 0-th byte and 1-st byte:
543// 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
544// barrier that will be used in fast reduction: bs_plain_barrier or
545// bs_reduction_barrier
546// 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
547// be used in fast reduction;
548// Reduction method is of 'enum _reduction_method' type and it's defined the way
549// so that the bits of 0-th byte are empty, so no need to execute a shift
550// instruction while packing/unpacking
551
552#if KMP_FAST_REDUCTION_BARRIER
553#define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
554 ((reduction_method) | (barrier_type))
555
556#define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
557 ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
558
559#define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
560 ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
561#else
562#define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
563 (reduction_method)
564
565#define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
566 (packed_reduction_method)
567
568#define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
569#endif
570
571#define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
572 ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
573 (which_reduction_block))
574
575#if KMP_FAST_REDUCTION_BARRIER
576#define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
577 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
578
579#define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
580 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
581#endif
582
583typedef int PACKED_REDUCTION_METHOD_T;
584
585/* -- end of fast reduction stuff ----------------------------------------- */
586
587#if KMP_OS_WINDOWS
588#define USE_CBLKDATA
589#if KMP_MSVC_COMPAT
590#pragma warning(push)
591#pragma warning(disable : 271 310)
592#endif
593#include <windows.h>
594#if KMP_MSVC_COMPAT
595#pragma warning(pop)
596#endif
597#endif
598
599#if KMP_OS_UNIX
600#include <dlfcn.h>
601#include <pthread.h>
602#endif
603
604enum kmp_hw_t : int {
605 KMP_HW_UNKNOWN = -1,
606 KMP_HW_SOCKET = 0,
607 KMP_HW_PROC_GROUP,
608 KMP_HW_NUMA,
609 KMP_HW_DIE,
610 KMP_HW_LLC,
611 KMP_HW_L3,
612 KMP_HW_TILE,
613 KMP_HW_MODULE,
614 KMP_HW_L2,
615 KMP_HW_L1,
616 KMP_HW_CORE,
617 KMP_HW_THREAD,
618 KMP_HW_LAST
619};
620
621typedef enum kmp_hw_core_type_t {
622 KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
623#if KMP_ARCH_X86 || KMP_ARCH_X86_64
624 KMP_HW_CORE_TYPE_ATOM = 0x20,
625 KMP_HW_CORE_TYPE_CORE = 0x40,
626 KMP_HW_MAX_NUM_CORE_TYPES = 3,
627#else
628 KMP_HW_MAX_NUM_CORE_TYPES = 1,
629#endif
630} kmp_hw_core_type_t;
631
632#define KMP_HW_MAX_NUM_CORE_EFFS 8
633
634#define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \
635 KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
636#define KMP_ASSERT_VALID_HW_TYPE(type) \
637 KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
638
639#define KMP_FOREACH_HW_TYPE(type) \
640 for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \
641 type = (kmp_hw_t)((int)type + 1))
642
643const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
644const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
645const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
646
647/* Only Linux* OS and Windows* OS support thread affinity. */
648#if KMP_AFFINITY_SUPPORTED
649
650// GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
651#if KMP_OS_WINDOWS
652#if _MSC_VER < 1600 && KMP_MSVC_COMPAT
653typedef struct GROUP_AFFINITY {
654 KAFFINITY Mask;
655 WORD Group;
656 WORD Reserved[3];
657} GROUP_AFFINITY;
658#endif /* _MSC_VER < 1600 */
659#if KMP_GROUP_AFFINITY
660extern int __kmp_num_proc_groups;
661#else
662static const int __kmp_num_proc_groups = 1;
663#endif /* KMP_GROUP_AFFINITY */
664typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
665extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
666
667typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
668extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
669
670typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
671extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
672
673typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
674 GROUP_AFFINITY *);
675extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
676#endif /* KMP_OS_WINDOWS */
677
678#if KMP_USE_HWLOC
679extern hwloc_topology_t __kmp_hwloc_topology;
680extern int __kmp_hwloc_error;
681#endif
682
683extern size_t __kmp_affin_mask_size;
684#define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
685#define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
686#define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
687#define KMP_CPU_SET_ITERATE(i, mask) \
688 for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
689#define KMP_CPU_SET(i, mask) (mask)->set(i)
690#define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
691#define KMP_CPU_CLR(i, mask) (mask)->clear(i)
692#define KMP_CPU_ZERO(mask) (mask)->zero()
693#define KMP_CPU_ISEMPTY(mask) (mask)->empty()
694#define KMP_CPU_COPY(dest, src) (dest)->copy(src)
695#define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
696#define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
697#define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
698#define KMP_CPU_EQUAL(dest, src) (dest)->is_equal(src)
699#define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
700#define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
701#define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
702#define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
703#define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
704#define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
705#define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
706#define KMP_CPU_ALLOC_ARRAY(arr, n) \
707 (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
708#define KMP_CPU_FREE_ARRAY(arr, n) \
709 __kmp_affinity_dispatch->deallocate_mask_array(arr)
710#define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
711#define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
712#define __kmp_get_system_affinity(mask, abort_bool) \
713 (mask)->get_system_affinity(abort_bool)
714#define __kmp_set_system_affinity(mask, abort_bool) \
715 (mask)->set_system_affinity(abort_bool)
716#define __kmp_get_proc_group(mask) (mask)->get_proc_group()
717
718class KMPAffinity {
719public:
720 class Mask {
721 public:
722 void *operator new(size_t n);
723 void operator delete(void *p);
724 void *operator new[](size_t n);
725 void operator delete[](void *p);
726 virtual ~Mask() {}
727 // Set bit i to 1
728 virtual void set(int i) {}
729 // Return bit i
730 virtual bool is_set(int i) const { return false; }
731 // Set bit i to 0
732 virtual void clear(int i) {}
733 // Zero out entire mask
734 virtual void zero() {}
735 // Check whether mask is empty
736 virtual bool empty() const { return true; }
737 // Copy src into this mask
738 virtual void copy(const Mask *src) {}
739 // this &= rhs
740 virtual void bitwise_and(const Mask *rhs) {}
741 // this |= rhs
742 virtual void bitwise_or(const Mask *rhs) {}
743 // this = ~this
744 virtual void bitwise_not() {}
745 // this == rhs
746 virtual bool is_equal(const Mask *rhs) const { return false; }
747 // API for iterating over an affinity mask
748 // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
749 virtual int begin() const { return 0; }
750 virtual int end() const { return 0; }
751 virtual int next(int previous) const { return 0; }
752#if KMP_OS_WINDOWS
753 virtual int set_process_affinity(bool abort_on_error) const { return -1; }
754#endif
755 // Set the system's affinity to this affinity mask's value
756 virtual int set_system_affinity(bool abort_on_error) const { return -1; }
757 // Set this affinity mask to the current system affinity
758 virtual int get_system_affinity(bool abort_on_error) { return -1; }
759 // Only 1 DWORD in the mask should have any procs set.
760 // Return the appropriate index, or -1 for an invalid mask.
761 virtual int get_proc_group() const { return -1; }
762 int get_max_cpu() const {
763 int cpu;
764 int max_cpu = -1;
765 KMP_CPU_SET_ITERATE(cpu, this) {
766 if (cpu > max_cpu)
767 max_cpu = cpu;
768 }
769 return max_cpu;
770 }
771 };
772 void *operator new(size_t n);
773 void operator delete(void *p);
774 // Need virtual destructor
775 virtual ~KMPAffinity() = default;
776 // Determine if affinity is capable
777 virtual void determine_capable(const char *env_var) {}
778 // Bind the current thread to os proc
779 virtual void bind_thread(int proc) {}
780 // Factory functions to allocate/deallocate a mask
781 virtual Mask *allocate_mask() { return nullptr; }
782 virtual void deallocate_mask(Mask *m) {}
783 virtual Mask *allocate_mask_array(int num) { return nullptr; }
784 virtual void deallocate_mask_array(Mask *m) {}
785 virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
786 static void pick_api();
787 static void destroy_api();
788 enum api_type {
789 NATIVE_OS
790#if KMP_USE_HWLOC
791 ,
792 HWLOC
793#endif
794 };
795 virtual api_type get_api_type() const {
796 KMP_ASSERT(0);
797 return NATIVE_OS;
798 }
799
800private:
801 static bool picked_api;
802};
803
804typedef KMPAffinity::Mask kmp_affin_mask_t;
805extern KMPAffinity *__kmp_affinity_dispatch;
806
807class kmp_affinity_raii_t {
808 kmp_affin_mask_t *mask;
809 bool restored;
810
811public:
812 kmp_affinity_raii_t(const kmp_affin_mask_t *new_mask = nullptr)
813 : restored(false) {
814 if (KMP_AFFINITY_CAPABLE()) {
815 KMP_CPU_ALLOC(mask);
816 KMP_ASSERT(mask != NULL);
817 __kmp_get_system_affinity(mask, /*abort_on_error=*/true);
818 if (new_mask)
819 __kmp_set_system_affinity(new_mask, /*abort_on_error=*/true);
820 }
821 }
822 void restore() {
823 if (!restored && KMP_AFFINITY_CAPABLE()) {
824 __kmp_set_system_affinity(mask, /*abort_on_error=*/true);
825 KMP_CPU_FREE(mask);
826 }
827 restored = true;
828 }
829 ~kmp_affinity_raii_t() { restore(); }
830};
831
832// Declare local char buffers with this size for printing debug and info
833// messages, using __kmp_affinity_print_mask().
834#define KMP_AFFIN_MASK_PRINT_LEN 1024
835
836enum affinity_type {
837 affinity_none = 0,
838 affinity_physical,
839 affinity_logical,
840 affinity_compact,
841 affinity_scatter,
842 affinity_explicit,
843 affinity_balanced,
844 affinity_disabled, // not used outsize the env var parser
845 affinity_default
846};
847
848enum affinity_top_method {
849 affinity_top_method_all = 0, // try all (supported) methods, in order
850#if KMP_ARCH_X86 || KMP_ARCH_X86_64
851 affinity_top_method_apicid,
852 affinity_top_method_x2apicid,
853 affinity_top_method_x2apicid_1f,
854#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
855 affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
856#if KMP_GROUP_AFFINITY
857 affinity_top_method_group,
858#endif /* KMP_GROUP_AFFINITY */
859 affinity_top_method_flat,
860#if KMP_USE_HWLOC
861 affinity_top_method_hwloc,
862#endif
863 affinity_top_method_default
864};
865
866#define affinity_respect_mask_default (2)
867
868typedef struct kmp_affinity_flags_t {
869 unsigned dups : 1;
870 unsigned verbose : 1;
871 unsigned warnings : 1;
872 unsigned respect : 2;
873 unsigned reset : 1;
874 unsigned initialized : 1;
875 unsigned core_types_gran : 1;
876 unsigned core_effs_gran : 1;
877 unsigned omp_places : 1;
878 unsigned reserved : 22;
879} kmp_affinity_flags_t;
880KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4);
881
882typedef struct kmp_affinity_ids_t {
883 int ids[KMP_HW_LAST];
884 int operator[](size_t idx) const { return ids[idx]; }
885 int &operator[](size_t idx) { return ids[idx]; }
886 kmp_affinity_ids_t &operator=(const kmp_affinity_ids_t &rhs) {
887 for (int i = 0; i < KMP_HW_LAST; ++i)
888 ids[i] = rhs[i];
889 return *this;
890 }
891} kmp_affinity_ids_t;
892
893typedef struct kmp_affinity_attrs_t {
894 int core_type : 8;
895 int core_eff : 8;
896 unsigned valid : 1;
897 unsigned reserved : 15;
898} kmp_affinity_attrs_t;
899#define KMP_AFFINITY_ATTRS_UNKNOWN \
900 { KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 }
901
902typedef struct kmp_affinity_t {
903 char *proclist;
904 enum affinity_type type;
905 kmp_hw_t gran;
906 int gran_levels;
907 kmp_affinity_attrs_t core_attr_gran;
908 int compact;
909 int offset;
910 kmp_affinity_flags_t flags;
911 unsigned num_masks;
912 kmp_affin_mask_t *masks;
913 kmp_affinity_ids_t *ids;
914 kmp_affinity_attrs_t *attrs;
915 unsigned num_os_id_masks;
916 kmp_affin_mask_t *os_id_masks;
917 const char *env_var;
918} kmp_affinity_t;
919
920#define KMP_AFFINITY_INIT(env) \
921 { \
922 nullptr, affinity_default, KMP_HW_UNKNOWN, -1, KMP_AFFINITY_ATTRS_UNKNOWN, \
923 0, 0, \
924 {TRUE, FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE, \
925 FALSE, FALSE, FALSE}, \
926 0, nullptr, nullptr, nullptr, 0, nullptr, env \
927 }
928
929extern enum affinity_top_method __kmp_affinity_top_method;
930extern kmp_affinity_t __kmp_affinity;
931extern kmp_affinity_t __kmp_hh_affinity;
932extern kmp_affinity_t *__kmp_affinities[2];
933
934extern void __kmp_affinity_bind_thread(int which);
935
936extern kmp_affin_mask_t *__kmp_affin_fullMask;
937extern kmp_affin_mask_t *__kmp_affin_origMask;
938extern char *__kmp_cpuinfo_file;
939
940#endif /* KMP_AFFINITY_SUPPORTED */
941
942// This needs to be kept in sync with the values in omp.h !!!
943typedef enum kmp_proc_bind_t {
944 proc_bind_false = 0,
945 proc_bind_true,
946 proc_bind_primary,
947 proc_bind_close,
948 proc_bind_spread,
949 proc_bind_intel, // use KMP_AFFINITY interface
950 proc_bind_default
951} kmp_proc_bind_t;
952
953typedef struct kmp_nested_proc_bind_t {
954 kmp_proc_bind_t *bind_types;
955 int size;
956 int used;
957} kmp_nested_proc_bind_t;
958
959extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
960extern kmp_proc_bind_t __kmp_teams_proc_bind;
961
962extern int __kmp_display_affinity;
963extern char *__kmp_affinity_format;
964static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
965#if OMPT_SUPPORT
966extern int __kmp_tool;
967extern char *__kmp_tool_libraries;
968#endif // OMPT_SUPPORT
969
970#if KMP_AFFINITY_SUPPORTED
971#define KMP_PLACE_ALL (-1)
972#define KMP_PLACE_UNDEFINED (-2)
973// Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
974#define KMP_AFFINITY_NON_PROC_BIND \
975 ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
976 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
977 (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced))
978#endif /* KMP_AFFINITY_SUPPORTED */
979
980extern int __kmp_affinity_num_places;
981
982typedef enum kmp_cancel_kind_t {
983 cancel_noreq = 0,
984 cancel_parallel = 1,
985 cancel_loop = 2,
986 cancel_sections = 3,
987 cancel_taskgroup = 4
988} kmp_cancel_kind_t;
989
990// KMP_HW_SUBSET support:
991typedef struct kmp_hws_item {
992 int num;
993 int offset;
994} kmp_hws_item_t;
995
996extern kmp_hws_item_t __kmp_hws_socket;
997extern kmp_hws_item_t __kmp_hws_die;
998extern kmp_hws_item_t __kmp_hws_node;
999extern kmp_hws_item_t __kmp_hws_tile;
1000extern kmp_hws_item_t __kmp_hws_core;
1001extern kmp_hws_item_t __kmp_hws_proc;
1002extern int __kmp_hws_requested;
1003extern int __kmp_hws_abs_flag; // absolute or per-item number requested
1004
1005/* ------------------------------------------------------------------------ */
1006
1007#define KMP_PAD(type, sz) \
1008 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
1009
1010// We need to avoid using -1 as a GTID as +1 is added to the gtid
1011// when storing it in a lock, and the value 0 is reserved.
1012#define KMP_GTID_DNE (-2) /* Does not exist */
1013#define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
1014#define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
1015#define KMP_GTID_UNKNOWN (-5) /* Is not known */
1016#define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
1017
1018/* OpenMP 5.0 Memory Management support */
1019
1020#ifndef __OMP_H
1021// Duplicate type definitions from omp.h
1022typedef uintptr_t omp_uintptr_t;
1023
1024typedef enum {
1025 omp_atk_sync_hint = 1,
1026 omp_atk_alignment = 2,
1027 omp_atk_access = 3,
1028 omp_atk_pool_size = 4,
1029 omp_atk_fallback = 5,
1030 omp_atk_fb_data = 6,
1031 omp_atk_pinned = 7,
1032 omp_atk_partition = 8
1033} omp_alloctrait_key_t;
1034
1035typedef enum {
1036 omp_atv_false = 0,
1037 omp_atv_true = 1,
1038 omp_atv_contended = 3,
1039 omp_atv_uncontended = 4,
1040 omp_atv_serialized = 5,
1041 omp_atv_sequential = omp_atv_serialized, // (deprecated)
1042 omp_atv_private = 6,
1043 omp_atv_all = 7,
1044 omp_atv_thread = 8,
1045 omp_atv_pteam = 9,
1046 omp_atv_cgroup = 10,
1047 omp_atv_default_mem_fb = 11,
1048 omp_atv_null_fb = 12,
1049 omp_atv_abort_fb = 13,
1050 omp_atv_allocator_fb = 14,
1051 omp_atv_environment = 15,
1052 omp_atv_nearest = 16,
1053 omp_atv_blocked = 17,
1054 omp_atv_interleaved = 18
1055} omp_alloctrait_value_t;
1056#define omp_atv_default ((omp_uintptr_t)-1)
1057
1058typedef void *omp_memspace_handle_t;
1059extern omp_memspace_handle_t const omp_default_mem_space;
1060extern omp_memspace_handle_t const omp_large_cap_mem_space;
1061extern omp_memspace_handle_t const omp_const_mem_space;
1062extern omp_memspace_handle_t const omp_high_bw_mem_space;
1063extern omp_memspace_handle_t const omp_low_lat_mem_space;
1064extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
1065extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
1066extern omp_memspace_handle_t const llvm_omp_target_device_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
1089// end of duplicate type definitions from omp.h
1090#endif
1091
1092extern int __kmp_memkind_available;
1093
1094typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
1095
1096typedef struct kmp_allocator_t {
1097 omp_memspace_handle_t memspace;
1098 void **memkind; // pointer to memkind
1099 size_t alignment;
1100 omp_alloctrait_value_t fb;
1101 kmp_allocator_t *fb_data;
1102 kmp_uint64 pool_size;
1103 kmp_uint64 pool_used;
1104 bool pinned;
1105} kmp_allocator_t;
1106
1107extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1108 omp_memspace_handle_t,
1109 int ntraits,
1110 omp_alloctrait_t traits[]);
1111extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1112extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1113extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1114// external interfaces, may be used by compiler
1115extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1116extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1117 omp_allocator_handle_t al);
1118extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1119 omp_allocator_handle_t al);
1120extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1121 omp_allocator_handle_t al,
1122 omp_allocator_handle_t free_al);
1123extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1124// internal interfaces, contain real implementation
1125extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1126 omp_allocator_handle_t al);
1127extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1128 omp_allocator_handle_t al);
1129extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1130 omp_allocator_handle_t al,
1131 omp_allocator_handle_t free_al);
1132extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1133
1134extern void __kmp_init_memkind();
1135extern void __kmp_fini_memkind();
1136extern void __kmp_init_target_mem();
1137
1138/* ------------------------------------------------------------------------ */
1139
1140#if ENABLE_LIBOMPTARGET
1141extern void __kmp_init_target_task();
1142#endif
1143
1144/* ------------------------------------------------------------------------ */
1145
1146#define KMP_UINT64_MAX \
1147 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1148
1149#define KMP_MIN_NTH 1
1150
1151#ifndef KMP_MAX_NTH
1152#if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1153#define KMP_MAX_NTH PTHREAD_THREADS_MAX
1154#else
1155#define KMP_MAX_NTH INT_MAX
1156#endif
1157#endif /* KMP_MAX_NTH */
1158
1159#ifdef PTHREAD_STACK_MIN
1160#define KMP_MIN_STKSIZE ((size_t)PTHREAD_STACK_MIN)
1161#else
1162#define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1163#endif
1164
1165#define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1166
1167#if KMP_ARCH_X86
1168#define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1169#elif KMP_ARCH_X86_64
1170#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1171#define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1172#else
1173#define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1174#endif
1175
1176#define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1177#define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1178#define KMP_MAX_MALLOC_POOL_INCR \
1179 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1180
1181#define KMP_MIN_STKOFFSET (0)
1182#define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1183#if KMP_OS_DARWIN
1184#define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1185#else
1186#define KMP_DEFAULT_STKOFFSET CACHE_LINE
1187#endif
1188
1189#define KMP_MIN_STKPADDING (0)
1190#define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1191
1192#define KMP_BLOCKTIME_MULTIPLIER \
1193 (1000) /* number of blocktime units per second */
1194#define KMP_MIN_BLOCKTIME (0)
1195#define KMP_MAX_BLOCKTIME \
1196 (INT_MAX) /* Must be this for "infinite" setting the work */
1197
1198/* __kmp_blocktime is in milliseconds */
1199#define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200))
1200
1201#if KMP_USE_MONITOR
1202#define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1203#define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1204#define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1205
1206/* Calculate new number of monitor wakeups for a specific block time based on
1207 previous monitor_wakeups. Only allow increasing number of wakeups */
1208#define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1209 (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \
1210 : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \
1211 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1212 ? (monitor_wakeups) \
1213 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1214
1215/* Calculate number of intervals for a specific block time based on
1216 monitor_wakeups */
1217#define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1218 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1219 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1220#else
1221#define KMP_BLOCKTIME(team, tid) \
1222 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1223#if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1224// HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1225extern kmp_uint64 __kmp_ticks_per_msec;
1226#if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1227#define KMP_NOW() ((kmp_uint64)_rdtsc())
1228#else
1229#define KMP_NOW() __kmp_hardware_timestamp()
1230#endif
1231#define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1232#define KMP_BLOCKTIME_INTERVAL(team, tid) \
1233 (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1234#define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1235#else
1236// System time is retrieved sporadically while blocking.
1237extern kmp_uint64 __kmp_now_nsec();
1238#define KMP_NOW() __kmp_now_nsec()
1239#define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1240#define KMP_BLOCKTIME_INTERVAL(team, tid) \
1241 (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1242#define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1243#endif
1244#endif // KMP_USE_MONITOR
1245
1246#define KMP_MIN_STATSCOLS 40
1247#define KMP_MAX_STATSCOLS 4096
1248#define KMP_DEFAULT_STATSCOLS 80
1249
1250#define KMP_MIN_INTERVAL 0
1251#define KMP_MAX_INTERVAL (INT_MAX - 1)
1252#define KMP_DEFAULT_INTERVAL 0
1253
1254#define KMP_MIN_CHUNK 1
1255#define KMP_MAX_CHUNK (INT_MAX - 1)
1256#define KMP_DEFAULT_CHUNK 1
1257
1258#define KMP_MIN_DISP_NUM_BUFF 1
1259#define KMP_DFLT_DISP_NUM_BUFF 7
1260#define KMP_MAX_DISP_NUM_BUFF 4096
1261
1262#define KMP_MAX_ORDERED 8
1263
1264#define KMP_MAX_FIELDS 32
1265
1266#define KMP_MAX_BRANCH_BITS 31
1267
1268#define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1269
1270#define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1271
1272#define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1273
1274/* Minimum number of threads before switch to TLS gtid (experimentally
1275 determined) */
1276/* josh TODO: what about OS X* tuning? */
1277#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1278#define KMP_TLS_GTID_MIN 5
1279#else
1280#define KMP_TLS_GTID_MIN INT_MAX
1281#endif
1282
1283#define KMP_MASTER_TID(tid) (0 == (tid))
1284#define KMP_WORKER_TID(tid) (0 != (tid))
1285
1286#define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1287#define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1288#define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1289
1290#ifndef TRUE
1291#define FALSE 0
1292#define TRUE (!FALSE)
1293#endif
1294
1295/* NOTE: all of the following constants must be even */
1296
1297#if KMP_OS_WINDOWS
1298#define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1299#define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1300#elif KMP_OS_LINUX
1301#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1302#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1303#elif KMP_OS_DARWIN
1304/* TODO: tune for KMP_OS_DARWIN */
1305#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1306#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1307#elif KMP_OS_DRAGONFLY
1308/* TODO: tune for KMP_OS_DRAGONFLY */
1309#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1310#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1311#elif KMP_OS_FREEBSD
1312/* TODO: tune for KMP_OS_FREEBSD */
1313#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1314#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1315#elif KMP_OS_NETBSD
1316/* TODO: tune for KMP_OS_NETBSD */
1317#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1318#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1319#elif KMP_OS_HURD
1320/* TODO: tune for KMP_OS_HURD */
1321#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1322#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1323#elif KMP_OS_OPENBSD
1324/* TODO: tune for KMP_OS_OPENBSD */
1325#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1326#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1327#endif
1328
1329#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1330typedef struct kmp_cpuid {
1331 kmp_uint32 eax;
1332 kmp_uint32 ebx;
1333 kmp_uint32 ecx;
1334 kmp_uint32 edx;
1335} kmp_cpuid_t;
1336
1337typedef struct kmp_cpuinfo_flags_t {
1338 unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1339 unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1340 unsigned hybrid : 1;
1341 unsigned reserved : 29; // Ensure size of 32 bits
1342} kmp_cpuinfo_flags_t;
1343
1344typedef struct kmp_cpuinfo {
1345 int initialized; // If 0, other fields are not initialized.
1346 int signature; // CPUID(1).EAX
1347 int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1348 int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1349 // Model << 4 ) + Model)
1350 int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1351 kmp_cpuinfo_flags_t flags;
1352 int apic_id;
1353 int physical_id;
1354 int logical_id;
1355 kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1356 char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1357} kmp_cpuinfo_t;
1358
1359extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1360
1361#if KMP_OS_UNIX
1362// subleaf is only needed for cache and topology discovery and can be set to
1363// zero in most cases
1364static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1365 __asm__ __volatile__("cpuid"
1366 : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1367 : "a"(leaf), "c"(subleaf));
1368}
1369// Load p into FPU control word
1370static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1371 __asm__ __volatile__("fldcw %0" : : "m"(*p));
1372}
1373// Store FPU control word into p
1374static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1375 __asm__ __volatile__("fstcw %0" : "=m"(*p));
1376}
1377static inline void __kmp_clear_x87_fpu_status_word() {
1378#if KMP_MIC
1379 // 32-bit protected mode x87 FPU state
1380 struct x87_fpu_state {
1381 unsigned cw;
1382 unsigned sw;
1383 unsigned tw;
1384 unsigned fip;
1385 unsigned fips;
1386 unsigned fdp;
1387 unsigned fds;
1388 };
1389 struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1390 __asm__ __volatile__("fstenv %0\n\t" // store FP env
1391 "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1392 "fldenv %0\n\t" // load FP env back
1393 : "+m"(fpu_state), "+m"(fpu_state.sw));
1394#else
1395 __asm__ __volatile__("fnclex");
1396#endif // KMP_MIC
1397}
1398#if __SSE__
1399static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1400static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1401#else
1402static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1403static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1404#endif
1405#else
1406// Windows still has these as external functions in assembly file
1407extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1408extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1409extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1410extern void __kmp_clear_x87_fpu_status_word();
1411static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1412static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1413#endif // KMP_OS_UNIX
1414
1415#define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1416
1417// User-level Monitor/Mwait
1418#if KMP_HAVE_UMWAIT
1419// We always try for UMWAIT first
1420#if KMP_HAVE_WAITPKG_INTRINSICS
1421#if KMP_HAVE_IMMINTRIN_H
1422#include <immintrin.h>
1423#elif KMP_HAVE_INTRIN_H
1424#include <intrin.h>
1425#endif
1426#endif // KMP_HAVE_WAITPKG_INTRINSICS
1427
1428KMP_ATTRIBUTE_TARGET_WAITPKG
1429static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1430#if !KMP_HAVE_WAITPKG_INTRINSICS
1431 uint32_t timeHi = uint32_t(counter >> 32);
1432 uint32_t timeLo = uint32_t(counter & 0xffffffff);
1433 char flag;
1434 __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1435 "setb %0"
1436 // The "=q" restraint means any register accessible as rl
1437 // in 32-bit mode: a, b, c, and d;
1438 // in 64-bit mode: any integer register
1439 : "=q"(flag)
1440 : "a"(timeLo), "d"(timeHi), "c"(hint)
1441 :);
1442 return flag;
1443#else
1444 return _tpause(hint, counter);
1445#endif
1446}
1447KMP_ATTRIBUTE_TARGET_WAITPKG
1448static inline void __kmp_umonitor(void *cacheline) {
1449#if !KMP_HAVE_WAITPKG_INTRINSICS
1450 __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1451 :
1452 : "a"(cacheline)
1453 :);
1454#else
1455 _umonitor(cacheline);
1456#endif
1457}
1458KMP_ATTRIBUTE_TARGET_WAITPKG
1459static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1460#if !KMP_HAVE_WAITPKG_INTRINSICS
1461 uint32_t timeHi = uint32_t(counter >> 32);
1462 uint32_t timeLo = uint32_t(counter & 0xffffffff);
1463 char flag;
1464 __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1465 "setb %0"
1466 // The "=q" restraint means any register accessible as rl
1467 // in 32-bit mode: a, b, c, and d;
1468 // in 64-bit mode: any integer register
1469 : "=q"(flag)
1470 : "a"(timeLo), "d"(timeHi), "c"(hint)
1471 :);
1472 return flag;
1473#else
1474 return _umwait(hint, counter);
1475#endif
1476}
1477#elif KMP_HAVE_MWAIT
1478#if KMP_OS_UNIX
1479#include <pmmintrin.h>
1480#else
1481#include <intrin.h>
1482#endif
1483#if KMP_OS_UNIX
1484__attribute__((target("sse3")))
1485#endif
1486static inline void
1487__kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1488 _mm_monitor(cacheline, extensions, hints);
1489}
1490#if KMP_OS_UNIX
1491__attribute__((target("sse3")))
1492#endif
1493static inline void
1494__kmp_mm_mwait(unsigned extensions, unsigned hints) {
1495 _mm_mwait(extensions, hints);
1496}
1497#endif // KMP_HAVE_UMWAIT
1498
1499#if KMP_ARCH_X86
1500extern void __kmp_x86_pause(void);
1501#elif KMP_MIC
1502// Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1503// regression after removal of extra PAUSE from spin loops. Changing
1504// the delay from 100 to 300 showed even better performance than double PAUSE
1505// on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1506static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1507#else
1508static inline void __kmp_x86_pause(void) { _mm_pause(); }
1509#endif
1510#define KMP_CPU_PAUSE() __kmp_x86_pause()
1511#elif KMP_ARCH_PPC64
1512#define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1513#define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1514#define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1515#define KMP_CPU_PAUSE() \
1516 do { \
1517 KMP_PPC64_PRI_LOW(); \
1518 KMP_PPC64_PRI_MED(); \
1519 KMP_PPC64_PRI_LOC_MB(); \
1520 } while (0)
1521#else
1522#define KMP_CPU_PAUSE() /* nothing to do */
1523#endif
1524
1525#define KMP_INIT_YIELD(count) \
1526 { (count) = __kmp_yield_init; }
1527
1528#define KMP_INIT_BACKOFF(time) \
1529 { (time) = __kmp_pause_init; }
1530
1531#define KMP_OVERSUBSCRIBED \
1532 (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1533
1534#define KMP_TRY_YIELD \
1535 ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1536
1537#define KMP_TRY_YIELD_OVERSUB \
1538 ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1539
1540#define KMP_YIELD(cond) \
1541 { \
1542 KMP_CPU_PAUSE(); \
1543 if ((cond) && (KMP_TRY_YIELD)) \
1544 __kmp_yield(); \
1545 }
1546
1547#define KMP_YIELD_OVERSUB() \
1548 { \
1549 KMP_CPU_PAUSE(); \
1550 if ((KMP_TRY_YIELD_OVERSUB)) \
1551 __kmp_yield(); \
1552 }
1553
1554// Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1555// there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1556#define KMP_YIELD_SPIN(count) \
1557 { \
1558 KMP_CPU_PAUSE(); \
1559 if (KMP_TRY_YIELD) { \
1560 (count) -= 2; \
1561 if (!(count)) { \
1562 __kmp_yield(); \
1563 (count) = __kmp_yield_next; \
1564 } \
1565 } \
1566 }
1567
1568// If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1569// (C0.2) state, which improves performance of other SMT threads on the same
1570// core, otherwise, use the fast (C0.1) default state, or whatever the user has
1571// requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1572// available, fall back to the regular CPU pause and yield combination.
1573#if KMP_HAVE_UMWAIT
1574#define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF)
1575#define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1576 { \
1577 if (__kmp_tpause_enabled) { \
1578 if (KMP_OVERSUBSCRIBED) { \
1579 __kmp_tpause(0, (time)); \
1580 } else { \
1581 __kmp_tpause(__kmp_tpause_hint, (time)); \
1582 } \
1583 (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK; \
1584 } else { \
1585 KMP_CPU_PAUSE(); \
1586 if ((KMP_TRY_YIELD_OVERSUB)) { \
1587 __kmp_yield(); \
1588 } else if (__kmp_use_yield == 1) { \
1589 (count) -= 2; \
1590 if (!(count)) { \
1591 __kmp_yield(); \
1592 (count) = __kmp_yield_next; \
1593 } \
1594 } \
1595 } \
1596 }
1597#else
1598#define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1599 { \
1600 KMP_CPU_PAUSE(); \
1601 if ((KMP_TRY_YIELD_OVERSUB)) \
1602 __kmp_yield(); \
1603 else if (__kmp_use_yield == 1) { \
1604 (count) -= 2; \
1605 if (!(count)) { \
1606 __kmp_yield(); \
1607 (count) = __kmp_yield_next; \
1608 } \
1609 } \
1610 }
1611#endif // KMP_HAVE_UMWAIT
1612
1613/* ------------------------------------------------------------------------ */
1614/* Support datatypes for the orphaned construct nesting checks. */
1615/* ------------------------------------------------------------------------ */
1616
1617/* When adding to this enum, add its corresponding string in cons_text_c[]
1618 * array in kmp_error.cpp */
1619enum cons_type {
1620 ct_none,
1621 ct_parallel,
1622 ct_pdo,
1623 ct_pdo_ordered,
1624 ct_psections,
1625 ct_psingle,
1626 ct_critical,
1627 ct_ordered_in_parallel,
1628 ct_ordered_in_pdo,
1629 ct_master,
1630 ct_reduce,
1631 ct_barrier,
1632 ct_masked
1633};
1634
1635#define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1636
1637struct cons_data {
1638 ident_t const *ident;
1639 enum cons_type type;
1640 int prev;
1641 kmp_user_lock_p
1642 name; /* address exclusively for critical section name comparison */
1643};
1644
1645struct cons_header {
1646 int p_top, w_top, s_top;
1647 int stack_size, stack_top;
1648 struct cons_data *stack_data;
1649};
1650
1651struct kmp_region_info {
1652 char *text;
1653 int offset[KMP_MAX_FIELDS];
1654 int length[KMP_MAX_FIELDS];
1655};
1656
1657/* ---------------------------------------------------------------------- */
1658/* ---------------------------------------------------------------------- */
1659
1660#if KMP_OS_WINDOWS
1661typedef HANDLE kmp_thread_t;
1662typedef DWORD kmp_key_t;
1663#endif /* KMP_OS_WINDOWS */
1664
1665#if KMP_OS_UNIX
1666typedef pthread_t kmp_thread_t;
1667typedef pthread_key_t kmp_key_t;
1668#endif
1669
1670extern kmp_key_t __kmp_gtid_threadprivate_key;
1671
1672typedef struct kmp_sys_info {
1673 long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1674 long minflt; /* the number of page faults serviced without any I/O */
1675 long majflt; /* the number of page faults serviced that required I/O */
1676 long nswap; /* the number of times a process was "swapped" out of memory */
1677 long inblock; /* the number of times the file system had to perform input */
1678 long oublock; /* the number of times the file system had to perform output */
1679 long nvcsw; /* the number of times a context switch was voluntarily */
1680 long nivcsw; /* the number of times a context switch was forced */
1681} kmp_sys_info_t;
1682
1683#if USE_ITT_BUILD
1684// We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1685// required type here. Later we will check the type meets requirements.
1686typedef int kmp_itt_mark_t;
1687#define KMP_ITT_DEBUG 0
1688#endif /* USE_ITT_BUILD */
1689
1690typedef kmp_int32 kmp_critical_name[8];
1691
1701typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1702typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1703 ...);
1704
1709/* ---------------------------------------------------------------------------
1710 */
1711/* Threadprivate initialization/finalization function declarations */
1712
1713/* for non-array objects: __kmpc_threadprivate_register() */
1714
1719typedef void *(*kmpc_ctor)(void *);
1720
1725typedef void (*kmpc_dtor)(
1726 void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1727 compiler */
1732typedef void *(*kmpc_cctor)(void *, void *);
1733
1734/* for array objects: __kmpc_threadprivate_register_vec() */
1735/* First arg: "this" pointer */
1736/* Last arg: number of array elements */
1742typedef void *(*kmpc_ctor_vec)(void *, size_t);
1748typedef void (*kmpc_dtor_vec)(void *, size_t);
1754typedef void *(*kmpc_cctor_vec)(void *, void *,
1755 size_t); /* function unused by compiler */
1756
1761/* keeps tracked of threadprivate cache allocations for cleanup later */
1762typedef struct kmp_cached_addr {
1763 void **addr; /* address of allocated cache */
1764 void ***compiler_cache; /* pointer to compiler's cache */
1765 void *data; /* pointer to global data */
1766 struct kmp_cached_addr *next; /* pointer to next cached address */
1767} kmp_cached_addr_t;
1768
1769struct private_data {
1770 struct private_data *next; /* The next descriptor in the list */
1771 void *data; /* The data buffer for this descriptor */
1772 int more; /* The repeat count for this descriptor */
1773 size_t size; /* The data size for this descriptor */
1774};
1775
1776struct private_common {
1777 struct private_common *next;
1778 struct private_common *link;
1779 void *gbl_addr;
1780 void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1781 size_t cmn_size;
1782};
1783
1784struct shared_common {
1785 struct shared_common *next;
1786 struct private_data *pod_init;
1787 void *obj_init;
1788 void *gbl_addr;
1789 union {
1790 kmpc_ctor ctor;
1791 kmpc_ctor_vec ctorv;
1792 } ct;
1793 union {
1794 kmpc_cctor cctor;
1795 kmpc_cctor_vec cctorv;
1796 } cct;
1797 union {
1798 kmpc_dtor dtor;
1799 kmpc_dtor_vec dtorv;
1800 } dt;
1801 size_t vec_len;
1802 int is_vec;
1803 size_t cmn_size;
1804};
1805
1806#define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1807#define KMP_HASH_TABLE_SIZE \
1808 (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1809#define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1810#define KMP_HASH(x) \
1811 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1812
1813struct common_table {
1814 struct private_common *data[KMP_HASH_TABLE_SIZE];
1815};
1816
1817struct shared_table {
1818 struct shared_common *data[KMP_HASH_TABLE_SIZE];
1819};
1820
1821/* ------------------------------------------------------------------------ */
1822
1823#if KMP_USE_HIER_SCHED
1824// Shared barrier data that exists inside a single unit of the scheduling
1825// hierarchy
1826typedef struct kmp_hier_private_bdata_t {
1827 kmp_int32 num_active;
1828 kmp_uint64 index;
1829 kmp_uint64 wait_val[2];
1830} kmp_hier_private_bdata_t;
1831#endif
1832
1833typedef struct kmp_sched_flags {
1834 unsigned ordered : 1;
1835 unsigned nomerge : 1;
1836 unsigned contains_last : 1;
1837#if KMP_USE_HIER_SCHED
1838 unsigned use_hier : 1;
1839 unsigned unused : 28;
1840#else
1841 unsigned unused : 29;
1842#endif
1843} kmp_sched_flags_t;
1844
1845KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1846
1847#if KMP_STATIC_STEAL_ENABLED
1848typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1849 kmp_int32 count;
1850 kmp_int32 ub;
1851 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1852 kmp_int32 lb;
1853 kmp_int32 st;
1854 kmp_int32 tc;
1855 kmp_lock_t *steal_lock; // lock used for chunk stealing
1856 // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1857 // a) parm3 is properly aligned and
1858 // b) all parm1-4 are on the same cache line.
1859 // Because of parm1-4 are used together, performance seems to be better
1860 // if they are on the same cache line (not measured though).
1861
1862 struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1863 kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1864 kmp_int32 parm2; // make no real change at least while padding is off.
1865 kmp_int32 parm3;
1866 kmp_int32 parm4;
1867 };
1868
1869 kmp_uint32 ordered_lower;
1870 kmp_uint32 ordered_upper;
1871#if KMP_OS_WINDOWS
1872 kmp_int32 last_upper;
1873#endif /* KMP_OS_WINDOWS */
1874} dispatch_private_info32_t;
1875
1876typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1877 kmp_int64 count; // current chunk number for static & static-steal scheduling
1878 kmp_int64 ub; /* upper-bound */
1879 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1880 kmp_int64 lb; /* lower-bound */
1881 kmp_int64 st; /* stride */
1882 kmp_int64 tc; /* trip count (number of iterations) */
1883 kmp_lock_t *steal_lock; // lock used for chunk stealing
1884 /* parm[1-4] are used in different ways by different scheduling algorithms */
1885
1886 // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1887 // a) parm3 is properly aligned and
1888 // b) all parm1-4 are in the same cache line.
1889 // Because of parm1-4 are used together, performance seems to be better
1890 // if they are in the same line (not measured though).
1891
1892 struct KMP_ALIGN(32) {
1893 kmp_int64 parm1;
1894 kmp_int64 parm2;
1895 kmp_int64 parm3;
1896 kmp_int64 parm4;
1897 };
1898
1899 kmp_uint64 ordered_lower;
1900 kmp_uint64 ordered_upper;
1901#if KMP_OS_WINDOWS
1902 kmp_int64 last_upper;
1903#endif /* KMP_OS_WINDOWS */
1904} dispatch_private_info64_t;
1905#else /* KMP_STATIC_STEAL_ENABLED */
1906typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1907 kmp_int32 lb;
1908 kmp_int32 ub;
1909 kmp_int32 st;
1910 kmp_int32 tc;
1911
1912 kmp_int32 parm1;
1913 kmp_int32 parm2;
1914 kmp_int32 parm3;
1915 kmp_int32 parm4;
1916
1917 kmp_int32 count;
1918
1919 kmp_uint32 ordered_lower;
1920 kmp_uint32 ordered_upper;
1921#if KMP_OS_WINDOWS
1922 kmp_int32 last_upper;
1923#endif /* KMP_OS_WINDOWS */
1924} dispatch_private_info32_t;
1925
1926typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1927 kmp_int64 lb; /* lower-bound */
1928 kmp_int64 ub; /* upper-bound */
1929 kmp_int64 st; /* stride */
1930 kmp_int64 tc; /* trip count (number of iterations) */
1931
1932 /* parm[1-4] are used in different ways by different scheduling algorithms */
1933 kmp_int64 parm1;
1934 kmp_int64 parm2;
1935 kmp_int64 parm3;
1936 kmp_int64 parm4;
1937
1938 kmp_int64 count; /* current chunk number for static scheduling */
1939
1940 kmp_uint64 ordered_lower;
1941 kmp_uint64 ordered_upper;
1942#if KMP_OS_WINDOWS
1943 kmp_int64 last_upper;
1944#endif /* KMP_OS_WINDOWS */
1945} dispatch_private_info64_t;
1946#endif /* KMP_STATIC_STEAL_ENABLED */
1947
1948typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1949 union private_info {
1950 dispatch_private_info32_t p32;
1951 dispatch_private_info64_t p64;
1952 } u;
1953 enum sched_type schedule; /* scheduling algorithm */
1954 kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1955 std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
1956 kmp_int32 ordered_bumped;
1957 // Stack of buffers for nest of serial regions
1958 struct dispatch_private_info *next;
1959 kmp_int32 type_size; /* the size of types in private_info */
1960#if KMP_USE_HIER_SCHED
1961 kmp_int32 hier_id;
1962 void *parent; /* hierarchical scheduling parent pointer */
1963#endif
1964 enum cons_type pushed_ws;
1965} dispatch_private_info_t;
1966
1967typedef struct dispatch_shared_info32 {
1968 /* chunk index under dynamic, number of idle threads under static-steal;
1969 iteration index otherwise */
1970 volatile kmp_uint32 iteration;
1971 volatile kmp_int32 num_done;
1972 volatile kmp_uint32 ordered_iteration;
1973 // Dummy to retain the structure size after making ordered_iteration scalar
1974 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1975} dispatch_shared_info32_t;
1976
1977typedef struct dispatch_shared_info64 {
1978 /* chunk index under dynamic, number of idle threads under static-steal;
1979 iteration index otherwise */
1980 volatile kmp_uint64 iteration;
1981 volatile kmp_int64 num_done;
1982 volatile kmp_uint64 ordered_iteration;
1983 // Dummy to retain the structure size after making ordered_iteration scalar
1984 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1985} dispatch_shared_info64_t;
1986
1987typedef struct dispatch_shared_info {
1988 union shared_info {
1989 dispatch_shared_info32_t s32;
1990 dispatch_shared_info64_t s64;
1991 } u;
1992 volatile kmp_uint32 buffer_index;
1993 volatile kmp_int32 doacross_buf_idx; // teamwise index
1994 volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1995 kmp_int32 doacross_num_done; // count finished threads
1996#if KMP_USE_HIER_SCHED
1997 void *hier;
1998#endif
1999#if KMP_USE_HWLOC
2000 // When linking with libhwloc, the ORDERED EPCC test slows down on big
2001 // machines (> 48 cores). Performance analysis showed that a cache thrash
2002 // was occurring and this padding helps alleviate the problem.
2003 char padding[64];
2004#endif
2005} dispatch_shared_info_t;
2006
2007typedef struct kmp_disp {
2008 /* Vector for ORDERED SECTION */
2009 void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
2010 /* Vector for END ORDERED SECTION */
2011 void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
2012
2013 dispatch_shared_info_t *th_dispatch_sh_current;
2014 dispatch_private_info_t *th_dispatch_pr_current;
2015
2016 dispatch_private_info_t *th_disp_buffer;
2017 kmp_uint32 th_disp_index;
2018 kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
2019 volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
2020 kmp_int64 *th_doacross_info; // info on loop bounds
2021#if KMP_USE_INTERNODE_ALIGNMENT
2022 char more_padding[INTERNODE_CACHE_LINE];
2023#endif
2024} kmp_disp_t;
2025
2026/* ------------------------------------------------------------------------ */
2027/* Barrier stuff */
2028
2029/* constants for barrier state update */
2030#define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
2031#define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
2032#define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
2033#define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
2034
2035#define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
2036#define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
2037#define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
2038
2039#if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
2040#error "Barrier sleep bit must be smaller than barrier bump bit"
2041#endif
2042#if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
2043#error "Barrier unused bit must be smaller than barrier bump bit"
2044#endif
2045
2046// Constants for release barrier wait state: currently, hierarchical only
2047#define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
2048#define KMP_BARRIER_OWN_FLAG \
2049 1 // Normal state; worker waiting on own b_go flag in release
2050#define KMP_BARRIER_PARENT_FLAG \
2051 2 // Special state; worker waiting on parent's b_go flag in release
2052#define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
2053 3 // Special state; tells worker to shift from parent to own b_go
2054#define KMP_BARRIER_SWITCHING \
2055 4 // Special state; worker resets appropriate flag on wake-up
2056
2057#define KMP_NOT_SAFE_TO_REAP \
2058 0 // Thread th_reap_state: not safe to reap (tasking)
2059#define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
2060
2061// The flag_type describes the storage used for the flag.
2062enum flag_type {
2063 flag32,
2064 flag64,
2065 atomic_flag64,
2066 flag_oncore,
2067 flag_unset
2068};
2069
2070enum barrier_type {
2071 bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
2072 barriers if enabled) */
2073 bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
2074#if KMP_FAST_REDUCTION_BARRIER
2075 bs_reduction_barrier, /* 2, All barriers that are used in reduction */
2076#endif // KMP_FAST_REDUCTION_BARRIER
2077 bs_last_barrier /* Just a placeholder to mark the end */
2078};
2079
2080// to work with reduction barriers just like with plain barriers
2081#if !KMP_FAST_REDUCTION_BARRIER
2082#define bs_reduction_barrier bs_plain_barrier
2083#endif // KMP_FAST_REDUCTION_BARRIER
2084
2085typedef enum kmp_bar_pat { /* Barrier communication patterns */
2086 bp_linear_bar =
2087 0, /* Single level (degenerate) tree */
2088 bp_tree_bar =
2089 1, /* Balanced tree with branching factor 2^n */
2090 bp_hyper_bar = 2, /* Hypercube-embedded tree with min
2091 branching factor 2^n */
2092 bp_hierarchical_bar = 3, /* Machine hierarchy tree */
2093 bp_dist_bar = 4, /* Distributed barrier */
2094 bp_last_bar /* Placeholder to mark the end */
2095} kmp_bar_pat_e;
2096
2097#define KMP_BARRIER_ICV_PUSH 1
2098
2099/* Record for holding the values of the internal controls stack records */
2100typedef struct kmp_internal_control {
2101 int serial_nesting_level; /* corresponds to the value of the
2102 th_team_serialized field */
2103 kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
2104 thread) */
2105 kmp_int8
2106 bt_set; /* internal control for whether blocktime is explicitly set */
2107 int blocktime; /* internal control for blocktime */
2108#if KMP_USE_MONITOR
2109 int bt_intervals; /* internal control for blocktime intervals */
2110#endif
2111 int nproc; /* internal control for #threads for next parallel region (per
2112 thread) */
2113 int thread_limit; /* internal control for thread-limit-var */
2114 int max_active_levels; /* internal control for max_active_levels */
2115 kmp_r_sched_t
2116 sched; /* internal control for runtime schedule {sched,chunk} pair */
2117 kmp_proc_bind_t proc_bind; /* internal control for affinity */
2118 kmp_int32 default_device; /* internal control for default device */
2119 struct kmp_internal_control *next;
2120} kmp_internal_control_t;
2121
2122static inline void copy_icvs(kmp_internal_control_t *dst,
2123 kmp_internal_control_t *src) {
2124 *dst = *src;
2125}
2126
2127/* Thread barrier needs volatile barrier fields */
2128typedef struct KMP_ALIGN_CACHE kmp_bstate {
2129 // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2130 // uses of it). It is not explicitly aligned below, because we *don't* want
2131 // it to be padded -- instead, we fit b_go into the same cache line with
2132 // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2133 kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2134 // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2135 // same NGO store
2136 volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2137 KMP_ALIGN_CACHE volatile kmp_uint64
2138 b_arrived; // STATE => task reached synch point.
2139 kmp_uint32 *skip_per_level;
2140 kmp_uint32 my_level;
2141 kmp_int32 parent_tid;
2142 kmp_int32 old_tid;
2143 kmp_uint32 depth;
2144 struct kmp_bstate *parent_bar;
2145 kmp_team_t *team;
2146 kmp_uint64 leaf_state;
2147 kmp_uint32 nproc;
2148 kmp_uint8 base_leaf_kids;
2149 kmp_uint8 leaf_kids;
2150 kmp_uint8 offset;
2151 kmp_uint8 wait_flag;
2152 kmp_uint8 use_oncore_barrier;
2153#if USE_DEBUGGER
2154 // The following field is intended for the debugger solely. Only the worker
2155 // thread itself accesses this field: the worker increases it by 1 when it
2156 // arrives to a barrier.
2157 KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2158#endif /* USE_DEBUGGER */
2159} kmp_bstate_t;
2160
2161union KMP_ALIGN_CACHE kmp_barrier_union {
2162 double b_align; /* use worst case alignment */
2163 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2164 kmp_bstate_t bb;
2165};
2166
2167typedef union kmp_barrier_union kmp_balign_t;
2168
2169/* Team barrier needs only non-volatile arrived counter */
2170union KMP_ALIGN_CACHE kmp_barrier_team_union {
2171 double b_align; /* use worst case alignment */
2172 char b_pad[CACHE_LINE];
2173 struct {
2174 kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2175#if USE_DEBUGGER
2176 // The following two fields are indended for the debugger solely. Only
2177 // primary thread of the team accesses these fields: the first one is
2178 // increased by 1 when the primary thread arrives to a barrier, the second
2179 // one is increased by one when all the threads arrived.
2180 kmp_uint b_master_arrived;
2181 kmp_uint b_team_arrived;
2182#endif
2183 };
2184};
2185
2186typedef union kmp_barrier_team_union kmp_balign_team_t;
2187
2188/* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2189 threads when a condition changes. This is to workaround an NPTL bug where
2190 padding was added to pthread_cond_t which caused the initialization routine
2191 to write outside of the structure if compiled on pre-NPTL threads. */
2192#if KMP_OS_WINDOWS
2193typedef struct kmp_win32_mutex {
2194 /* The Lock */
2195 CRITICAL_SECTION cs;
2196} kmp_win32_mutex_t;
2197
2198typedef struct kmp_win32_cond {
2199 /* Count of the number of waiters. */
2200 int waiters_count_;
2201
2202 /* Serialize access to <waiters_count_> */
2203 kmp_win32_mutex_t waiters_count_lock_;
2204
2205 /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2206 int release_count_;
2207
2208 /* Keeps track of the current "generation" so that we don't allow */
2209 /* one thread to steal all the "releases" from the broadcast. */
2210 int wait_generation_count_;
2211
2212 /* A manual-reset event that's used to block and release waiting threads. */
2213 HANDLE event_;
2214} kmp_win32_cond_t;
2215#endif
2216
2217#if KMP_OS_UNIX
2218
2219union KMP_ALIGN_CACHE kmp_cond_union {
2220 double c_align;
2221 char c_pad[CACHE_LINE];
2222 pthread_cond_t c_cond;
2223};
2224
2225typedef union kmp_cond_union kmp_cond_align_t;
2226
2227union KMP_ALIGN_CACHE kmp_mutex_union {
2228 double m_align;
2229 char m_pad[CACHE_LINE];
2230 pthread_mutex_t m_mutex;
2231};
2232
2233typedef union kmp_mutex_union kmp_mutex_align_t;
2234
2235#endif /* KMP_OS_UNIX */
2236
2237typedef struct kmp_desc_base {
2238 void *ds_stackbase;
2239 size_t ds_stacksize;
2240 int ds_stackgrow;
2241 kmp_thread_t ds_thread;
2242 volatile int ds_tid;
2243 int ds_gtid;
2244#if KMP_OS_WINDOWS
2245 volatile int ds_alive;
2246 DWORD ds_thread_id;
2247/* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2248 However, debugger support (libomp_db) cannot work with handles, because they
2249 uncomparable. For example, debugger requests info about thread with handle h.
2250 h is valid within debugger process, and meaningless within debugee process.
2251 Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2252 within debugee process, but it is a *new* handle which does *not* equal to
2253 any other handle in debugee... The only way to compare handles is convert
2254 them to system-wide ids. GetThreadId() function is available only in
2255 Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2256 on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2257 thread id by call to GetCurrentThreadId() from within the thread and save it
2258 to let libomp_db identify threads. */
2259#endif /* KMP_OS_WINDOWS */
2260} kmp_desc_base_t;
2261
2262typedef union KMP_ALIGN_CACHE kmp_desc {
2263 double ds_align; /* use worst case alignment */
2264 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2265 kmp_desc_base_t ds;
2266} kmp_desc_t;
2267
2268typedef struct kmp_local {
2269 volatile int this_construct; /* count of single's encountered by thread */
2270 void *reduce_data;
2271#if KMP_USE_BGET
2272 void *bget_data;
2273 void *bget_list;
2274#if !USE_CMP_XCHG_FOR_BGET
2275#ifdef USE_QUEUING_LOCK_FOR_BGET
2276 kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2277#else
2278 kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2279// bootstrap lock so we can use it at library
2280// shutdown.
2281#endif /* USE_LOCK_FOR_BGET */
2282#endif /* ! USE_CMP_XCHG_FOR_BGET */
2283#endif /* KMP_USE_BGET */
2284
2285 PACKED_REDUCTION_METHOD_T
2286 packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2287 __kmpc_end_reduce*() */
2288
2289} kmp_local_t;
2290
2291#define KMP_CHECK_UPDATE(a, b) \
2292 if ((a) != (b)) \
2293 (a) = (b)
2294#define KMP_CHECK_UPDATE_SYNC(a, b) \
2295 if ((a) != (b)) \
2296 TCW_SYNC_PTR((a), (b))
2297
2298#define get__blocktime(xteam, xtid) \
2299 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2300#define get__bt_set(xteam, xtid) \
2301 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2302#if KMP_USE_MONITOR
2303#define get__bt_intervals(xteam, xtid) \
2304 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2305#endif
2306
2307#define get__dynamic_2(xteam, xtid) \
2308 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2309#define get__nproc_2(xteam, xtid) \
2310 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2311#define get__sched_2(xteam, xtid) \
2312 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2313
2314#define set__blocktime_team(xteam, xtid, xval) \
2315 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2316 (xval))
2317
2318#if KMP_USE_MONITOR
2319#define set__bt_intervals_team(xteam, xtid, xval) \
2320 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2321 (xval))
2322#endif
2323
2324#define set__bt_set_team(xteam, xtid, xval) \
2325 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2326
2327#define set__dynamic(xthread, xval) \
2328 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2329#define get__dynamic(xthread) \
2330 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2331
2332#define set__nproc(xthread, xval) \
2333 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2334
2335#define set__thread_limit(xthread, xval) \
2336 (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2337
2338#define set__max_active_levels(xthread, xval) \
2339 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2340
2341#define get__max_active_levels(xthread) \
2342 ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2343
2344#define set__sched(xthread, xval) \
2345 (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2346
2347#define set__proc_bind(xthread, xval) \
2348 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2349#define get__proc_bind(xthread) \
2350 ((xthread)->th.th_current_task->td_icvs.proc_bind)
2351
2352// OpenMP tasking data structures
2353
2354typedef enum kmp_tasking_mode {
2355 tskm_immediate_exec = 0,
2356 tskm_extra_barrier = 1,
2357 tskm_task_teams = 2,
2358 tskm_max = 2
2359} kmp_tasking_mode_t;
2360
2361extern kmp_tasking_mode_t
2362 __kmp_tasking_mode; /* determines how/when to execute tasks */
2363extern int __kmp_task_stealing_constraint;
2364extern int __kmp_enable_task_throttling;
2365extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2366// specified, defaults to 0 otherwise
2367// Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2368extern kmp_int32 __kmp_max_task_priority;
2369// Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2370extern kmp_uint64 __kmp_taskloop_min_tasks;
2371
2372/* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2373 taskdata first */
2374#define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2375#define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2376
2377// The tt_found_tasks flag is a signal to all threads in the team that tasks
2378// were spawned and queued since the previous barrier release.
2379#define KMP_TASKING_ENABLED(task_team) \
2380 (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2388typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2389
2390typedef union kmp_cmplrdata {
2391 kmp_int32 priority;
2392 kmp_routine_entry_t
2393 destructors; /* pointer to function to invoke deconstructors of
2394 firstprivate C++ objects */
2395 /* future data */
2396} kmp_cmplrdata_t;
2397
2398/* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2401typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2402 void *shareds;
2403 kmp_routine_entry_t
2404 routine;
2405 kmp_int32 part_id;
2406 kmp_cmplrdata_t
2407 data1; /* Two known optional additions: destructors and priority */
2408 kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2409 /* future data */
2410 /* private vars */
2411} kmp_task_t;
2412
2417typedef struct kmp_taskgroup {
2418 std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2419 std::atomic<kmp_int32>
2420 cancel_request; // request for cancellation of this taskgroup
2421 struct kmp_taskgroup *parent; // parent taskgroup
2422 // Block of data to perform task reduction
2423 void *reduce_data; // reduction related info
2424 kmp_int32 reduce_num_data; // number of data items to reduce
2425 uintptr_t *gomp_data; // gomp reduction data
2426} kmp_taskgroup_t;
2427
2428// forward declarations
2429typedef union kmp_depnode kmp_depnode_t;
2430typedef struct kmp_depnode_list kmp_depnode_list_t;
2431typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2432
2433// macros for checking dep flag as an integer
2434#define KMP_DEP_IN 0x1
2435#define KMP_DEP_OUT 0x2
2436#define KMP_DEP_INOUT 0x3
2437#define KMP_DEP_MTX 0x4
2438#define KMP_DEP_SET 0x8
2439#define KMP_DEP_ALL 0x80
2440// Compiler sends us this info:
2441typedef struct kmp_depend_info {
2442 kmp_intptr_t base_addr;
2443 size_t len;
2444 union {
2445 kmp_uint8 flag; // flag as an unsigned char
2446 struct { // flag as a set of 8 bits
2447 unsigned in : 1;
2448 unsigned out : 1;
2449 unsigned mtx : 1;
2450 unsigned set : 1;
2451 unsigned unused : 3;
2452 unsigned all : 1;
2453 } flags;
2454 };
2455} kmp_depend_info_t;
2456
2457// Internal structures to work with task dependencies:
2458struct kmp_depnode_list {
2459 kmp_depnode_t *node;
2460 kmp_depnode_list_t *next;
2461};
2462
2463// Max number of mutexinoutset dependencies per node
2464#define MAX_MTX_DEPS 4
2465
2466typedef struct kmp_base_depnode {
2467 kmp_depnode_list_t *successors; /* used under lock */
2468 kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2469 kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2470 kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2471 kmp_lock_t lock; /* guards shared fields: task, successors */
2472#if KMP_SUPPORT_GRAPH_OUTPUT
2473 kmp_uint32 id;
2474#endif
2475 std::atomic<kmp_int32> npredecessors;
2476 std::atomic<kmp_int32> nrefs;
2477} kmp_base_depnode_t;
2478
2479union KMP_ALIGN_CACHE kmp_depnode {
2480 double dn_align; /* use worst case alignment */
2481 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2482 kmp_base_depnode_t dn;
2483};
2484
2485struct kmp_dephash_entry {
2486 kmp_intptr_t addr;
2487 kmp_depnode_t *last_out;
2488 kmp_depnode_list_t *last_set;
2489 kmp_depnode_list_t *prev_set;
2490 kmp_uint8 last_flag;
2491 kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2492 kmp_dephash_entry_t *next_in_bucket;
2493};
2494
2495typedef struct kmp_dephash {
2496 kmp_dephash_entry_t **buckets;
2497 size_t size;
2498 kmp_depnode_t *last_all;
2499 size_t generation;
2500 kmp_uint32 nelements;
2501 kmp_uint32 nconflicts;
2502} kmp_dephash_t;
2503
2504typedef struct kmp_task_affinity_info {
2505 kmp_intptr_t base_addr;
2506 size_t len;
2507 struct {
2508 bool flag1 : 1;
2509 bool flag2 : 1;
2510 kmp_int32 reserved : 30;
2511 } flags;
2512} kmp_task_affinity_info_t;
2513
2514typedef enum kmp_event_type_t {
2515 KMP_EVENT_UNINITIALIZED = 0,
2516 KMP_EVENT_ALLOW_COMPLETION = 1
2517} kmp_event_type_t;
2518
2519typedef struct {
2520 kmp_event_type_t type;
2521 kmp_tas_lock_t lock;
2522 union {
2523 kmp_task_t *task;
2524 } ed;
2525} kmp_event_t;
2526
2527#if OMPX_TASKGRAPH
2528// Initial number of allocated nodes while recording
2529#define INIT_MAPSIZE 50
2530
2531typedef struct kmp_taskgraph_flags { /*This needs to be exactly 32 bits */
2532 unsigned nowait : 1;
2533 unsigned re_record : 1;
2534 unsigned reserved : 30;
2535} kmp_taskgraph_flags_t;
2536
2538typedef struct kmp_node_info {
2539 kmp_task_t *task; // Pointer to the actual task
2540 kmp_int32 *successors; // Array of the succesors ids
2541 kmp_int32 nsuccessors; // Number of succesors of the node
2542 std::atomic<kmp_int32>
2543 npredecessors_counter; // Number of predessors on the fly
2544 kmp_int32 npredecessors; // Total number of predecessors
2545 kmp_int32 successors_size; // Number of allocated succesors ids
2546 kmp_taskdata_t *parent_task; // Parent implicit task
2547} kmp_node_info_t;
2548
2550typedef enum kmp_tdg_status {
2551 KMP_TDG_NONE = 0,
2552 KMP_TDG_RECORDING = 1,
2553 KMP_TDG_READY = 2
2554} kmp_tdg_status_t;
2555
2557typedef struct kmp_tdg_info {
2558 kmp_int32 tdg_id; // Unique idenfifier of the TDG
2559 kmp_taskgraph_flags_t tdg_flags; // Flags related to a TDG
2560 kmp_int32 map_size; // Number of allocated TDG nodes
2561 kmp_int32 num_roots; // Number of roots tasks int the TDG
2562 kmp_int32 *root_tasks; // Array of tasks identifiers that are roots
2563 kmp_node_info_t *record_map; // Array of TDG nodes
2564 kmp_tdg_status_t tdg_status =
2565 KMP_TDG_NONE; // Status of the TDG (recording, ready...)
2566 std::atomic<kmp_int32> num_tasks; // Number of TDG nodes
2567 kmp_bootstrap_lock_t
2568 graph_lock; // Protect graph attributes when updated via taskloop_recur
2569 // Taskloop reduction related
2570 void *rec_taskred_data; // Data to pass to __kmpc_task_reduction_init or
2571 // __kmpc_taskred_init
2572 kmp_int32 rec_num_taskred;
2573} kmp_tdg_info_t;
2574
2575extern int __kmp_tdg_dot;
2576extern kmp_int32 __kmp_max_tdgs;
2577extern kmp_tdg_info_t **__kmp_global_tdgs;
2578extern kmp_int32 __kmp_curr_tdg_idx;
2579extern kmp_int32 __kmp_successors_size;
2580extern std::atomic<kmp_int32> __kmp_tdg_task_id;
2581extern kmp_int32 __kmp_num_tdg;
2582#endif
2583
2584#ifdef BUILD_TIED_TASK_STACK
2585
2586/* Tied Task stack definitions */
2587typedef struct kmp_stack_block {
2588 kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2589 struct kmp_stack_block *sb_next;
2590 struct kmp_stack_block *sb_prev;
2591} kmp_stack_block_t;
2592
2593typedef struct kmp_task_stack {
2594 kmp_stack_block_t ts_first_block; // first block of stack entries
2595 kmp_taskdata_t **ts_top; // pointer to the top of stack
2596 kmp_int32 ts_entries; // number of entries on the stack
2597} kmp_task_stack_t;
2598
2599#endif // BUILD_TIED_TASK_STACK
2600
2601typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2602 /* Compiler flags */ /* Total compiler flags must be 16 bits */
2603 unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2604 unsigned final : 1; /* task is final(1) so execute immediately */
2605 unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2606 code path */
2607 unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2608 invoke destructors from the runtime */
2609 unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2610 context of the RTL) */
2611 unsigned priority_specified : 1; /* set if the compiler provides priority
2612 setting for the task */
2613 unsigned detachable : 1; /* 1 == can detach */
2614 unsigned hidden_helper : 1; /* 1 == hidden helper task */
2615 unsigned reserved : 8; /* reserved for compiler use */
2616
2617 /* Library flags */ /* Total library flags must be 16 bits */
2618 unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2619 unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2620 unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2621 // (1) or may be deferred (0)
2622 unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2623 // (0) [>= 2 threads]
2624 /* If either team_serial or tasking_ser is set, task team may be NULL */
2625 /* Task State Flags: */
2626 unsigned started : 1; /* 1==started, 0==not started */
2627 unsigned executing : 1; /* 1==executing, 0==not executing */
2628 unsigned complete : 1; /* 1==complete, 0==not complete */
2629 unsigned freed : 1; /* 1==freed, 0==allocated */
2630 unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2631#if OMPX_TASKGRAPH
2632 unsigned onced : 1; /* 1==ran once already, 0==never ran, record & replay purposes */
2633 unsigned reserved31 : 6; /* reserved for library use */
2634#else
2635 unsigned reserved31 : 7; /* reserved for library use */
2636#endif
2637
2638} kmp_tasking_flags_t;
2639
2640typedef struct kmp_target_data {
2641 void *async_handle; // libomptarget async handle for task completion query
2642} kmp_target_data_t;
2643
2644struct kmp_taskdata { /* aligned during dynamic allocation */
2645 kmp_int32 td_task_id; /* id, assigned by debugger */
2646 kmp_tasking_flags_t td_flags; /* task flags */
2647 kmp_team_t *td_team; /* team for this task */
2648 kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2649 /* Currently not used except for perhaps IDB */
2650 kmp_taskdata_t *td_parent; /* parent task */
2651 kmp_int32 td_level; /* task nesting level */
2652 std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2653 ident_t *td_ident; /* task identifier */
2654 // Taskwait data.
2655 ident_t *td_taskwait_ident;
2656 kmp_uint32 td_taskwait_counter;
2657 kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2658 KMP_ALIGN_CACHE kmp_internal_control_t
2659 td_icvs; /* Internal control variables for the task */
2660 KMP_ALIGN_CACHE std::atomic<kmp_int32>
2661 td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2662 deallocated */
2663 std::atomic<kmp_int32>
2664 td_incomplete_child_tasks; /* Child tasks not yet complete */
2665 kmp_taskgroup_t
2666 *td_taskgroup; // Each task keeps pointer to its current taskgroup
2667 kmp_dephash_t
2668 *td_dephash; // Dependencies for children tasks are tracked from here
2669 kmp_depnode_t
2670 *td_depnode; // Pointer to graph node if this task has dependencies
2671 kmp_task_team_t *td_task_team;
2672 size_t td_size_alloc; // Size of task structure, including shareds etc.
2673#if defined(KMP_GOMP_COMPAT)
2674 // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2675 kmp_int32 td_size_loop_bounds;
2676#endif
2677 kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2678#if defined(KMP_GOMP_COMPAT)
2679 // GOMP sends in a copy function for copy constructors
2680 void (*td_copy_func)(void *, void *);
2681#endif
2682 kmp_event_t td_allow_completion_event;
2683#if OMPT_SUPPORT
2684 ompt_task_info_t ompt_task_info;
2685#endif
2686#if OMPX_TASKGRAPH
2687 bool is_taskgraph = 0; // whether the task is within a TDG
2688 kmp_tdg_info_t *tdg; // used to associate task with a TDG
2689#endif
2690 kmp_target_data_t td_target_data;
2691}; // struct kmp_taskdata
2692
2693// Make sure padding above worked
2694KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2695
2696// Data for task team but per thread
2697typedef struct kmp_base_thread_data {
2698 kmp_info_p *td_thr; // Pointer back to thread info
2699 // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2700 // queued?
2701 kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2702 kmp_taskdata_t *
2703 *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2704 kmp_int32 td_deque_size; // Size of deck
2705 kmp_uint32 td_deque_head; // Head of deque (will wrap)
2706 kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2707 kmp_int32 td_deque_ntasks; // Number of tasks in deque
2708 // GEH: shouldn't this be volatile since used in while-spin?
2709 kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2710#ifdef BUILD_TIED_TASK_STACK
2711 kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2712// scheduling constraint
2713#endif // BUILD_TIED_TASK_STACK
2714} kmp_base_thread_data_t;
2715
2716#define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2717#define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2718
2719#define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2720#define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2721
2722typedef union KMP_ALIGN_CACHE kmp_thread_data {
2723 kmp_base_thread_data_t td;
2724 double td_align; /* use worst case alignment */
2725 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2726} kmp_thread_data_t;
2727
2728typedef struct kmp_task_pri {
2729 kmp_thread_data_t td;
2730 kmp_int32 priority;
2731 kmp_task_pri *next;
2732} kmp_task_pri_t;
2733
2734// Data for task teams which are used when tasking is enabled for the team
2735typedef struct kmp_base_task_team {
2736 kmp_bootstrap_lock_t
2737 tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2738 /* must be bootstrap lock since used at library shutdown*/
2739
2740 // TODO: check performance vs kmp_tas_lock_t
2741 kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */
2742 kmp_task_pri_t *tt_task_pri_list;
2743
2744 kmp_task_team_t *tt_next; /* For linking the task team free list */
2745 kmp_thread_data_t
2746 *tt_threads_data; /* Array of per-thread structures for task team */
2747 /* Data survives task team deallocation */
2748 kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2749 executing this team? */
2750 /* TRUE means tt_threads_data is set up and initialized */
2751 kmp_int32 tt_nproc; /* #threads in team */
2752 kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2753 kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2754 kmp_int32 tt_untied_task_encountered;
2755 std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued
2756 // There is hidden helper thread encountered in this task team so that we must
2757 // wait when waiting on task team
2758 kmp_int32 tt_hidden_helper_task_encountered;
2759
2760 KMP_ALIGN_CACHE
2761 std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2762
2763 KMP_ALIGN_CACHE
2764 volatile kmp_uint32
2765 tt_active; /* is the team still actively executing tasks */
2766} kmp_base_task_team_t;
2767
2768union KMP_ALIGN_CACHE kmp_task_team {
2769 kmp_base_task_team_t tt;
2770 double tt_align; /* use worst case alignment */
2771 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2772};
2773
2774#if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2775// Free lists keep same-size free memory slots for fast memory allocation
2776// routines
2777typedef struct kmp_free_list {
2778 void *th_free_list_self; // Self-allocated tasks free list
2779 void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2780 // threads
2781 void *th_free_list_other; // Non-self free list (to be returned to owner's
2782 // sync list)
2783} kmp_free_list_t;
2784#endif
2785#if KMP_NESTED_HOT_TEAMS
2786// Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2787// are not put in teams pool, and they don't put threads in threads pool.
2788typedef struct kmp_hot_team_ptr {
2789 kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2790 kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2791} kmp_hot_team_ptr_t;
2792#endif
2793typedef struct kmp_teams_size {
2794 kmp_int32 nteams; // number of teams in a league
2795 kmp_int32 nth; // number of threads in each team of the league
2796} kmp_teams_size_t;
2797
2798// This struct stores a thread that acts as a "root" for a contention
2799// group. Contention groups are rooted at kmp_root threads, but also at
2800// each primary thread of each team created in the teams construct.
2801// This struct therefore also stores a thread_limit associated with
2802// that contention group, and a counter to track the number of threads
2803// active in that contention group. Each thread has a list of these: CG
2804// root threads have an entry in their list in which cg_root refers to
2805// the thread itself, whereas other workers in the CG will have a
2806// single entry where cg_root is same as the entry containing their CG
2807// root. When a thread encounters a teams construct, it will add a new
2808// entry to the front of its list, because it now roots a new CG.
2809typedef struct kmp_cg_root {
2810 kmp_info_p *cg_root; // "root" thread for a contention group
2811 // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2812 // thread_limit clause for teams primary threads
2813 kmp_int32 cg_thread_limit;
2814 kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2815 struct kmp_cg_root *up; // pointer to higher level CG root in list
2816} kmp_cg_root_t;
2817
2818// OpenMP thread data structures
2819
2820typedef struct KMP_ALIGN_CACHE kmp_base_info {
2821 /* Start with the readonly data which is cache aligned and padded. This is
2822 written before the thread starts working by the primary thread. Uber
2823 masters may update themselves later. Usage does not consider serialized
2824 regions. */
2825 kmp_desc_t th_info;
2826 kmp_team_p *th_team; /* team we belong to */
2827 kmp_root_p *th_root; /* pointer to root of task hierarchy */
2828 kmp_info_p *th_next_pool; /* next available thread in the pool */
2829 kmp_disp_t *th_dispatch; /* thread's dispatch data */
2830 int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2831
2832 /* The following are cached from the team info structure */
2833 /* TODO use these in more places as determined to be needed via profiling */
2834 int th_team_nproc; /* number of threads in a team */
2835 kmp_info_p *th_team_master; /* the team's primary thread */
2836 int th_team_serialized; /* team is serialized */
2837 microtask_t th_teams_microtask; /* save entry address for teams construct */
2838 int th_teams_level; /* save initial level of teams construct */
2839/* it is 0 on device but may be any on host */
2840
2841/* The blocktime info is copied from the team struct to the thread struct */
2842/* at the start of a barrier, and the values stored in the team are used */
2843/* at points in the code where the team struct is no longer guaranteed */
2844/* to exist (from the POV of worker threads). */
2845#if KMP_USE_MONITOR
2846 int th_team_bt_intervals;
2847 int th_team_bt_set;
2848#else
2849 kmp_uint64 th_team_bt_intervals;
2850#endif
2851
2852#if KMP_AFFINITY_SUPPORTED
2853 kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2854 kmp_affinity_ids_t th_topology_ids; /* thread's current topology ids */
2855 kmp_affinity_attrs_t th_topology_attrs; /* thread's current topology attrs */
2856#endif
2857 omp_allocator_handle_t th_def_allocator; /* default allocator */
2858 /* The data set by the primary thread at reinit, then R/W by the worker */
2859 KMP_ALIGN_CACHE int
2860 th_set_nproc; /* if > 0, then only use this request for the next fork */
2861#if KMP_NESTED_HOT_TEAMS
2862 kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2863#endif
2864 kmp_proc_bind_t
2865 th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2866 kmp_teams_size_t
2867 th_teams_size; /* number of teams/threads in teams construct */
2868#if KMP_AFFINITY_SUPPORTED
2869 int th_current_place; /* place currently bound to */
2870 int th_new_place; /* place to bind to in par reg */
2871 int th_first_place; /* first place in partition */
2872 int th_last_place; /* last place in partition */
2873#endif
2874 int th_prev_level; /* previous level for affinity format */
2875 int th_prev_num_threads; /* previous num_threads for affinity format */
2876#if USE_ITT_BUILD
2877 kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2878 kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2879 kmp_uint64 th_frame_time; /* frame timestamp */
2880#endif /* USE_ITT_BUILD */
2881 kmp_local_t th_local;
2882 struct private_common *th_pri_head;
2883
2884 /* Now the data only used by the worker (after initial allocation) */
2885 /* TODO the first serial team should actually be stored in the info_t
2886 structure. this will help reduce initial allocation overhead */
2887 KMP_ALIGN_CACHE kmp_team_p
2888 *th_serial_team; /*serialized team held in reserve*/
2889
2890#if OMPT_SUPPORT
2891 ompt_thread_info_t ompt_thread_info;
2892#endif
2893
2894 /* The following are also read by the primary thread during reinit */
2895 struct common_table *th_pri_common;
2896
2897 volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2898 /* while awaiting queuing lock acquire */
2899
2900 volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2901 flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2902
2903 ident_t *th_ident;
2904 unsigned th_x; // Random number generator data
2905 unsigned th_a; // Random number generator data
2906
2907 /* Tasking-related data for the thread */
2908 kmp_task_team_t *th_task_team; // Task team struct
2909 kmp_taskdata_t *th_current_task; // Innermost Task being executed
2910 kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2911 kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2912 // at nested levels
2913 kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2914 kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2915 kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2916 // tasking, thus safe to reap
2917
2918 /* More stuff for keeping track of active/sleeping threads (this part is
2919 written by the worker thread) */
2920 kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2921 int th_active; // ! sleeping; 32 bits for TCR/TCW
2922 std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
2923 // 0 = not used in team; 1 = used in team;
2924 // 2 = transitioning to not used in team; 3 = transitioning to used in team
2925 struct cons_header *th_cons; // used for consistency check
2926#if KMP_USE_HIER_SCHED
2927 // used for hierarchical scheduling
2928 kmp_hier_private_bdata_t *th_hier_bar_data;
2929#endif
2930
2931 /* Add the syncronizing data which is cache aligned and padded. */
2932 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2933
2934 KMP_ALIGN_CACHE volatile kmp_int32
2935 th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2936
2937#if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2938#define NUM_LISTS 4
2939 kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2940// allocation routines
2941#endif
2942
2943#if KMP_OS_WINDOWS
2944 kmp_win32_cond_t th_suspend_cv;
2945 kmp_win32_mutex_t th_suspend_mx;
2946 std::atomic<int> th_suspend_init;
2947#endif
2948#if KMP_OS_UNIX
2949 kmp_cond_align_t th_suspend_cv;
2950 kmp_mutex_align_t th_suspend_mx;
2951 std::atomic<int> th_suspend_init_count;
2952#endif
2953
2954#if USE_ITT_BUILD
2955 kmp_itt_mark_t th_itt_mark_single;
2956// alignment ???
2957#endif /* USE_ITT_BUILD */
2958#if KMP_STATS_ENABLED
2959 kmp_stats_list *th_stats;
2960#endif
2961#if KMP_OS_UNIX
2962 std::atomic<bool> th_blocking;
2963#endif
2964 kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2965} kmp_base_info_t;
2966
2967typedef union KMP_ALIGN_CACHE kmp_info {
2968 double th_align; /* use worst case alignment */
2969 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2970 kmp_base_info_t th;
2971} kmp_info_t;
2972
2973// OpenMP thread team data structures
2974
2975typedef struct kmp_base_data {
2976 volatile kmp_uint32 t_value;
2977} kmp_base_data_t;
2978
2979typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2980 double dt_align; /* use worst case alignment */
2981 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2982 kmp_base_data_t dt;
2983} kmp_sleep_team_t;
2984
2985typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2986 double dt_align; /* use worst case alignment */
2987 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2988 kmp_base_data_t dt;
2989} kmp_ordered_team_t;
2990
2991typedef int (*launch_t)(int gtid);
2992
2993/* Minimum number of ARGV entries to malloc if necessary */
2994#define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2995
2996// Set up how many argv pointers will fit in cache lines containing
2997// t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2998// larger value for more space between the primary write/worker read section and
2999// read/write by all section seems to buy more performance on EPCC PARALLEL.
3000#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3001#define KMP_INLINE_ARGV_BYTES \
3002 (4 * CACHE_LINE - \
3003 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
3004 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
3005 CACHE_LINE))
3006#else
3007#define KMP_INLINE_ARGV_BYTES \
3008 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
3009#endif
3010#define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
3011
3012typedef struct KMP_ALIGN_CACHE kmp_base_team {
3013 // Synchronization Data
3014 // ---------------------------------------------------------------------------
3015 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
3016 kmp_balign_team_t t_bar[bs_last_barrier];
3017 std::atomic<int> t_construct; // count of single directive encountered by team
3018 char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
3019
3020 // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
3021 std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
3022 std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
3023
3024 // Primary thread only
3025 // ---------------------------------------------------------------------------
3026 KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
3027 int t_master_this_cons; // "this_construct" single counter of primary thread
3028 // in parent team
3029 ident_t *t_ident; // if volatile, have to change too much other crud to
3030 // volatile too
3031 kmp_team_p *t_parent; // parent team
3032 kmp_team_p *t_next_pool; // next free team in the team pool
3033 kmp_disp_t *t_dispatch; // thread's dispatch data
3034 kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
3035 kmp_proc_bind_t t_proc_bind; // bind type for par region
3036#if USE_ITT_BUILD
3037 kmp_uint64 t_region_time; // region begin timestamp
3038#endif /* USE_ITT_BUILD */
3039
3040 // Primary thread write, workers read
3041 // --------------------------------------------------------------------------
3042 KMP_ALIGN_CACHE void **t_argv;
3043 int t_argc;
3044 int t_nproc; // number of threads in team
3045 microtask_t t_pkfn;
3046 launch_t t_invoke; // procedure to launch the microtask
3047
3048#if OMPT_SUPPORT
3049 ompt_team_info_t ompt_team_info;
3050 ompt_lw_taskteam_t *ompt_serialized_team_info;
3051#endif
3052
3053#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3054 kmp_int8 t_fp_control_saved;
3055 kmp_int8 t_pad2b;
3056 kmp_int16 t_x87_fpu_control_word; // FP control regs
3057 kmp_uint32 t_mxcsr;
3058#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3059
3060 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
3061
3062 KMP_ALIGN_CACHE kmp_info_t **t_threads;
3063 kmp_taskdata_t
3064 *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
3065 int t_level; // nested parallel level
3066
3067 KMP_ALIGN_CACHE int t_max_argc;
3068 int t_max_nproc; // max threads this team can handle (dynamically expandable)
3069 int t_serialized; // levels deep of serialized teams
3070 dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
3071 int t_id; // team's id, assigned by debugger.
3072 int t_active_level; // nested active parallel level
3073 kmp_r_sched_t t_sched; // run-time schedule for the team
3074#if KMP_AFFINITY_SUPPORTED
3075 int t_first_place; // first & last place in parent thread's partition.
3076 int t_last_place; // Restore these values to primary thread after par region.
3077#endif // KMP_AFFINITY_SUPPORTED
3078 int t_display_affinity;
3079 int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
3080 // omp_set_num_threads() call
3081 omp_allocator_handle_t t_def_allocator; /* default allocator */
3082
3083// Read/write by workers as well
3084#if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
3085 // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
3086 // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
3087 // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
3088 // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
3089 char dummy_padding[1024];
3090#endif
3091 // Internal control stack for additional nested teams.
3092 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
3093 // for SERIALIZED teams nested 2 or more levels deep
3094 // typed flag to store request state of cancellation
3095 std::atomic<kmp_int32> t_cancel_request;
3096 int t_master_active; // save on fork, restore on join
3097 void *t_copypriv_data; // team specific pointer to copyprivate data array
3098#if KMP_OS_WINDOWS
3099 std::atomic<kmp_uint32> t_copyin_counter;
3100#endif
3101#if USE_ITT_BUILD
3102 void *t_stack_id; // team specific stack stitching id (for ittnotify)
3103#endif /* USE_ITT_BUILD */
3104 distributedBarrier *b; // Distributed barrier data associated with team
3105} kmp_base_team_t;
3106
3107union KMP_ALIGN_CACHE kmp_team {
3108 kmp_base_team_t t;
3109 double t_align; /* use worst case alignment */
3110 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
3111};
3112
3113typedef union KMP_ALIGN_CACHE kmp_time_global {
3114 double dt_align; /* use worst case alignment */
3115 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3116 kmp_base_data_t dt;
3117} kmp_time_global_t;
3118
3119typedef struct kmp_base_global {
3120 /* cache-aligned */
3121 kmp_time_global_t g_time;
3122
3123 /* non cache-aligned */
3124 volatile int g_abort;
3125 volatile int g_done;
3126
3127 int g_dynamic;
3128 enum dynamic_mode g_dynamic_mode;
3129} kmp_base_global_t;
3130
3131typedef union KMP_ALIGN_CACHE kmp_global {
3132 kmp_base_global_t g;
3133 double g_align; /* use worst case alignment */
3134 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
3135} kmp_global_t;
3136
3137typedef struct kmp_base_root {
3138 // TODO: GEH - combine r_active with r_in_parallel then r_active ==
3139 // (r_in_parallel>= 0)
3140 // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
3141 // the synch overhead or keeping r_active
3142 volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
3143 // keeps a count of active parallel regions per root
3144 std::atomic<int> r_in_parallel;
3145 // GEH: This is misnamed, should be r_active_levels
3146 kmp_team_t *r_root_team;
3147 kmp_team_t *r_hot_team;
3148 kmp_info_t *r_uber_thread;
3149 kmp_lock_t r_begin_lock;
3150 volatile int r_begin;
3151 int r_blocktime; /* blocktime for this root and descendants */
3152#if KMP_AFFINITY_SUPPORTED
3153 int r_affinity_assigned;
3154#endif // KMP_AFFINITY_SUPPORTED
3155} kmp_base_root_t;
3156
3157typedef union KMP_ALIGN_CACHE kmp_root {
3158 kmp_base_root_t r;
3159 double r_align; /* use worst case alignment */
3160 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
3161} kmp_root_t;
3162
3163struct fortran_inx_info {
3164 kmp_int32 data;
3165};
3166
3167// This list type exists to hold old __kmp_threads arrays so that
3168// old references to them may complete while reallocation takes place when
3169// expanding the array. The items in this list are kept alive until library
3170// shutdown.
3171typedef struct kmp_old_threads_list_t {
3172 kmp_info_t **threads;
3173 struct kmp_old_threads_list_t *next;
3174} kmp_old_threads_list_t;
3175
3176/* ------------------------------------------------------------------------ */
3177
3178extern int __kmp_settings;
3179extern int __kmp_duplicate_library_ok;
3180#if USE_ITT_BUILD
3181extern int __kmp_forkjoin_frames;
3182extern int __kmp_forkjoin_frames_mode;
3183#endif
3184extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3185extern int __kmp_determ_red;
3186
3187#ifdef KMP_DEBUG
3188extern int kmp_a_debug;
3189extern int kmp_b_debug;
3190extern int kmp_c_debug;
3191extern int kmp_d_debug;
3192extern int kmp_e_debug;
3193extern int kmp_f_debug;
3194#endif /* KMP_DEBUG */
3195
3196/* For debug information logging using rotating buffer */
3197#define KMP_DEBUG_BUF_LINES_INIT 512
3198#define KMP_DEBUG_BUF_LINES_MIN 1
3199
3200#define KMP_DEBUG_BUF_CHARS_INIT 128
3201#define KMP_DEBUG_BUF_CHARS_MIN 2
3202
3203extern int
3204 __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3205extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3206extern int
3207 __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3208extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3209 entry pointer */
3210
3211extern char *__kmp_debug_buffer; /* Debug buffer itself */
3212extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3213 printed in buffer so far */
3214extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3215 recommended in warnings */
3216/* end rotating debug buffer */
3217
3218#ifdef KMP_DEBUG
3219extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3220
3221#define KMP_PAR_RANGE_ROUTINE_LEN 1024
3222extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3223#define KMP_PAR_RANGE_FILENAME_LEN 1024
3224extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3225extern int __kmp_par_range_lb;
3226extern int __kmp_par_range_ub;
3227#endif
3228
3229/* For printing out dynamic storage map for threads and teams */
3230extern int
3231 __kmp_storage_map; /* True means print storage map for threads and teams */
3232extern int __kmp_storage_map_verbose; /* True means storage map includes
3233 placement info */
3234extern int __kmp_storage_map_verbose_specified;
3235
3236#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3237extern kmp_cpuinfo_t __kmp_cpuinfo;
3238static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3239#elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3240static inline bool __kmp_is_hybrid_cpu() { return true; }
3241#else
3242static inline bool __kmp_is_hybrid_cpu() { return false; }
3243#endif
3244
3245extern volatile int __kmp_init_serial;
3246extern volatile int __kmp_init_gtid;
3247extern volatile int __kmp_init_common;
3248extern volatile int __kmp_need_register_serial;
3249extern volatile int __kmp_init_middle;
3250extern volatile int __kmp_init_parallel;
3251#if KMP_USE_MONITOR
3252extern volatile int __kmp_init_monitor;
3253#endif
3254extern volatile int __kmp_init_user_locks;
3255extern volatile int __kmp_init_hidden_helper_threads;
3256extern int __kmp_init_counter;
3257extern int __kmp_root_counter;
3258extern int __kmp_version;
3259
3260/* list of address of allocated caches for commons */
3261extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3262
3263/* Barrier algorithm types and options */
3264extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3265extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3266extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3267extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3268extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3269extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3270extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3271extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3272extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3273extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3274extern char const *__kmp_barrier_type_name[bs_last_barrier];
3275extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3276
3277/* Global Locks */
3278extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3279extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3280extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3281extern kmp_bootstrap_lock_t
3282 __kmp_exit_lock; /* exit() is not always thread-safe */
3283#if KMP_USE_MONITOR
3284extern kmp_bootstrap_lock_t
3285 __kmp_monitor_lock; /* control monitor thread creation */
3286#endif
3287extern kmp_bootstrap_lock_t
3288 __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3289 __kmp_threads expansion to co-exist */
3290
3291extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
3292extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
3293extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3294
3295extern enum library_type __kmp_library;
3296
3297extern enum sched_type __kmp_sched; /* default runtime scheduling */
3298extern enum sched_type __kmp_static; /* default static scheduling method */
3299extern enum sched_type __kmp_guided; /* default guided scheduling method */
3300extern enum sched_type __kmp_auto; /* default auto scheduling method */
3301extern int __kmp_chunk; /* default runtime chunk size */
3302extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3303
3304extern size_t __kmp_stksize; /* stack size per thread */
3305#if KMP_USE_MONITOR
3306extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3307#endif
3308extern size_t __kmp_stkoffset; /* stack offset per thread */
3309extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3310
3311extern size_t
3312 __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3313extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3314extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3315extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
3316extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3317extern int __kmp_generate_warnings; /* should we issue warnings? */
3318extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3319
3320#ifdef DEBUG_SUSPEND
3321extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3322#endif
3323
3324extern kmp_int32 __kmp_use_yield;
3325extern kmp_int32 __kmp_use_yield_exp_set;
3326extern kmp_uint32 __kmp_yield_init;
3327extern kmp_uint32 __kmp_yield_next;
3328extern kmp_uint64 __kmp_pause_init;
3329
3330/* ------------------------------------------------------------------------- */
3331extern int __kmp_allThreadsSpecified;
3332
3333extern size_t __kmp_align_alloc;
3334/* following data protected by initialization routines */
3335extern int __kmp_xproc; /* number of processors in the system */
3336extern int __kmp_avail_proc; /* number of processors available to the process */
3337extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3338extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3339// maximum total number of concurrently-existing threads on device
3340extern int __kmp_max_nth;
3341// maximum total number of concurrently-existing threads in a contention group
3342extern int __kmp_cg_max_nth;
3343extern int __kmp_teams_max_nth; // max threads used in a teams construct
3344extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3345 __kmp_root */
3346extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3347 region a la OMP_NUM_THREADS */
3348extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3349 initialization */
3350extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3351 used (fixed) */
3352extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3353 (__kmpc_threadprivate_cached()) */
3354extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
3355 blocking (env setting) */
3356extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */
3357#if KMP_USE_MONITOR
3358extern int
3359 __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3360extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3361 blocking */
3362#endif
3363#ifdef KMP_ADJUST_BLOCKTIME
3364extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3365#endif /* KMP_ADJUST_BLOCKTIME */
3366#ifdef KMP_DFLT_NTH_CORES
3367extern int __kmp_ncores; /* Total number of cores for threads placement */
3368#endif
3369/* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3370extern int __kmp_abort_delay;
3371
3372extern int __kmp_need_register_atfork_specified;
3373extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3374 to install fork handler */
3375extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3376 0 - not set, will be set at runtime
3377 1 - using stack search
3378 2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3379 X*) or TlsGetValue(Windows* OS))
3380 3 - static TLS (__declspec(thread) __kmp_gtid),
3381 Linux* OS .so only. */
3382extern int
3383 __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3384#ifdef KMP_TDATA_GTID
3385extern KMP_THREAD_LOCAL int __kmp_gtid;
3386#endif
3387extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3388extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3389#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3390extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3391extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3392extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3393#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3394
3395// max_active_levels for nested parallelism enabled by default via
3396// OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3397extern int __kmp_dflt_max_active_levels;
3398// Indicates whether value of __kmp_dflt_max_active_levels was already
3399// explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3400extern bool __kmp_dflt_max_active_levels_set;
3401extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3402 concurrent execution per team */
3403#if KMP_NESTED_HOT_TEAMS
3404extern int __kmp_hot_teams_mode;
3405extern int __kmp_hot_teams_max_level;
3406#endif
3407
3408#if KMP_OS_LINUX
3409extern enum clock_function_type __kmp_clock_function;
3410extern int __kmp_clock_function_param;
3411#endif /* KMP_OS_LINUX */
3412
3413#if KMP_MIC_SUPPORTED
3414extern enum mic_type __kmp_mic_type;
3415#endif
3416
3417#ifdef USE_LOAD_BALANCE
3418extern double __kmp_load_balance_interval; // load balance algorithm interval
3419#endif /* USE_LOAD_BALANCE */
3420
3421// OpenMP 3.1 - Nested num threads array
3422typedef struct kmp_nested_nthreads_t {
3423 int *nth;
3424 int size;
3425 int used;
3426} kmp_nested_nthreads_t;
3427
3428extern kmp_nested_nthreads_t __kmp_nested_nth;
3429
3430#if KMP_USE_ADAPTIVE_LOCKS
3431
3432// Parameters for the speculative lock backoff system.
3433struct kmp_adaptive_backoff_params_t {
3434 // Number of soft retries before it counts as a hard retry.
3435 kmp_uint32 max_soft_retries;
3436 // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3437 // the right
3438 kmp_uint32 max_badness;
3439};
3440
3441extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3442
3443#if KMP_DEBUG_ADAPTIVE_LOCKS
3444extern const char *__kmp_speculative_statsfile;
3445#endif
3446
3447#endif // KMP_USE_ADAPTIVE_LOCKS
3448
3449extern int __kmp_display_env; /* TRUE or FALSE */
3450extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3451extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3452extern int __kmp_nteams;
3453extern int __kmp_teams_thread_limit;
3454
3455/* ------------------------------------------------------------------------- */
3456
3457/* the following are protected by the fork/join lock */
3458/* write: lock read: anytime */
3459extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3460/* Holds old arrays of __kmp_threads until library shutdown */
3461extern kmp_old_threads_list_t *__kmp_old_threads_list;
3462/* read/write: lock */
3463extern volatile kmp_team_t *__kmp_team_pool;
3464extern volatile kmp_info_t *__kmp_thread_pool;
3465extern kmp_info_t *__kmp_thread_pool_insert_pt;
3466
3467// total num threads reachable from some root thread including all root threads
3468extern volatile int __kmp_nth;
3469/* total number of threads reachable from some root thread including all root
3470 threads, and those in the thread pool */
3471extern volatile int __kmp_all_nth;
3472extern std::atomic<int> __kmp_thread_pool_active_nth;
3473
3474extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3475/* end data protected by fork/join lock */
3476/* ------------------------------------------------------------------------- */
3477
3478#define __kmp_get_gtid() __kmp_get_global_thread_id()
3479#define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3480#define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3481#define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3482#define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3483
3484// AT: Which way is correct?
3485// AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3486// AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3487#define __kmp_get_team_num_threads(gtid) \
3488 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3489
3490static inline bool KMP_UBER_GTID(int gtid) {
3491 KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3492 KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3493 return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3494 __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3495}
3496
3497static inline int __kmp_tid_from_gtid(int gtid) {
3498 KMP_DEBUG_ASSERT(gtid >= 0);
3499 return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3500}
3501
3502static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3503 KMP_DEBUG_ASSERT(tid >= 0 && team);
3504 return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3505}
3506
3507static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3508 KMP_DEBUG_ASSERT(thr);
3509 return thr->th.th_info.ds.ds_gtid;
3510}
3511
3512static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3513 KMP_DEBUG_ASSERT(gtid >= 0);
3514 return __kmp_threads[gtid];
3515}
3516
3517static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3518 KMP_DEBUG_ASSERT(gtid >= 0);
3519 return __kmp_threads[gtid]->th.th_team;
3520}
3521
3522static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3523 if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3524 KMP_FATAL(ThreadIdentInvalid);
3525}
3526
3527#if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3528extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3529extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3530extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3531extern int __kmp_mwait_hints; // Hints to pass in to mwait
3532#endif
3533
3534#if KMP_HAVE_UMWAIT
3535extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3536extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3537extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3538extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3539#endif
3540
3541/* ------------------------------------------------------------------------- */
3542
3543extern kmp_global_t __kmp_global; /* global status */
3544
3545extern kmp_info_t __kmp_monitor;
3546// For Debugging Support Library
3547extern std::atomic<kmp_int32> __kmp_team_counter;
3548// For Debugging Support Library
3549extern std::atomic<kmp_int32> __kmp_task_counter;
3550
3551#if USE_DEBUGGER
3552#define _KMP_GEN_ID(counter) \
3553 (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3554#else
3555#define _KMP_GEN_ID(counter) (~0)
3556#endif /* USE_DEBUGGER */
3557
3558#define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3559#define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3560
3561/* ------------------------------------------------------------------------ */
3562
3563extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3564 size_t size, char const *format, ...);
3565
3566extern void __kmp_serial_initialize(void);
3567extern void __kmp_middle_initialize(void);
3568extern void __kmp_parallel_initialize(void);
3569
3570extern void __kmp_internal_begin(void);
3571extern void __kmp_internal_end_library(int gtid);
3572extern void __kmp_internal_end_thread(int gtid);
3573extern void __kmp_internal_end_atexit(void);
3574extern void __kmp_internal_end_dtor(void);
3575extern void __kmp_internal_end_dest(void *);
3576
3577extern int __kmp_register_root(int initial_thread);
3578extern void __kmp_unregister_root(int gtid);
3579extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3580
3581extern int __kmp_ignore_mppbeg(void);
3582extern int __kmp_ignore_mppend(void);
3583
3584extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3585extern void __kmp_exit_single(int gtid);
3586
3587extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3588extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3589
3590#ifdef USE_LOAD_BALANCE
3591extern int __kmp_get_load_balance(int);
3592#endif
3593
3594extern int __kmp_get_global_thread_id(void);
3595extern int __kmp_get_global_thread_id_reg(void);
3596extern void __kmp_exit_thread(int exit_status);
3597extern void __kmp_abort(char const *format, ...);
3598extern void __kmp_abort_thread(void);
3599KMP_NORETURN extern void __kmp_abort_process(void);
3600extern void __kmp_warn(char const *format, ...);
3601
3602extern void __kmp_set_num_threads(int new_nth, int gtid);
3603
3604// Returns current thread (pointer to kmp_info_t). Current thread *must* be
3605// registered.
3606static inline kmp_info_t *__kmp_entry_thread() {
3607 int gtid = __kmp_entry_gtid();
3608
3609 return __kmp_threads[gtid];
3610}
3611
3612extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3613extern int __kmp_get_max_active_levels(int gtid);
3614extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3615extern int __kmp_get_team_size(int gtid, int level);
3616extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3617extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3618
3619extern unsigned short __kmp_get_random(kmp_info_t *thread);
3620extern void __kmp_init_random(kmp_info_t *thread);
3621
3622extern kmp_r_sched_t __kmp_get_schedule_global(void);
3623extern void __kmp_adjust_num_threads(int new_nproc);
3624extern void __kmp_check_stksize(size_t *val);
3625
3626extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3627extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3628extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3629#define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3630#define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3631#define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3632
3633#if USE_FAST_MEMORY
3634extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3635 size_t size KMP_SRC_LOC_DECL);
3636extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3637extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3638extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3639#define __kmp_fast_allocate(this_thr, size) \
3640 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3641#define __kmp_fast_free(this_thr, ptr) \
3642 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3643#endif
3644
3645extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3646extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3647 size_t elsize KMP_SRC_LOC_DECL);
3648extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3649 size_t size KMP_SRC_LOC_DECL);
3650extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3651#define __kmp_thread_malloc(th, size) \
3652 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3653#define __kmp_thread_calloc(th, nelem, elsize) \
3654 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3655#define __kmp_thread_realloc(th, ptr, size) \
3656 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3657#define __kmp_thread_free(th, ptr) \
3658 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3659
3660extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3661
3662extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3663 kmp_proc_bind_t proc_bind);
3664extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3665 int num_threads);
3666extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3667 int num_teams_ub, int num_threads);
3668
3669extern void __kmp_yield();
3670
3671extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3672 enum sched_type schedule, kmp_int32 lb,
3673 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3674extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3675 enum sched_type schedule, kmp_uint32 lb,
3676 kmp_uint32 ub, kmp_int32 st,
3677 kmp_int32 chunk);
3678extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3679 enum sched_type schedule, kmp_int64 lb,
3680 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3681extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3682 enum sched_type schedule, kmp_uint64 lb,
3683 kmp_uint64 ub, kmp_int64 st,
3684 kmp_int64 chunk);
3685
3686extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3687 kmp_int32 *p_last, kmp_int32 *p_lb,
3688 kmp_int32 *p_ub, kmp_int32 *p_st);
3689extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3690 kmp_int32 *p_last, kmp_uint32 *p_lb,
3691 kmp_uint32 *p_ub, kmp_int32 *p_st);
3692extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3693 kmp_int32 *p_last, kmp_int64 *p_lb,
3694 kmp_int64 *p_ub, kmp_int64 *p_st);
3695extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3696 kmp_int32 *p_last, kmp_uint64 *p_lb,
3697 kmp_uint64 *p_ub, kmp_int64 *p_st);
3698
3699extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3700extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3701extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3702extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3703
3704#ifdef KMP_GOMP_COMPAT
3705
3706extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3707 enum sched_type schedule, kmp_int32 lb,
3708 kmp_int32 ub, kmp_int32 st,
3709 kmp_int32 chunk, int push_ws);
3710extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3711 enum sched_type schedule, kmp_uint32 lb,
3712 kmp_uint32 ub, kmp_int32 st,
3713 kmp_int32 chunk, int push_ws);
3714extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3715 enum sched_type schedule, kmp_int64 lb,
3716 kmp_int64 ub, kmp_int64 st,
3717 kmp_int64 chunk, int push_ws);
3718extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3719 enum sched_type schedule, kmp_uint64 lb,
3720 kmp_uint64 ub, kmp_int64 st,
3721 kmp_int64 chunk, int push_ws);
3722extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3723extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3724extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3725extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3726
3727#endif /* KMP_GOMP_COMPAT */
3728
3729extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3730extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3731extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3732extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3733extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3734extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3735 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3736 void *obj);
3737extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3738 kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3739
3740extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3741 int final_spin
3742#if USE_ITT_BUILD
3743 ,
3744 void *itt_sync_obj
3745#endif
3746);
3747extern void __kmp_release_64(kmp_flag_64<> *flag);
3748
3749extern void __kmp_infinite_loop(void);
3750
3751extern void __kmp_cleanup(void);
3752
3753#if KMP_HANDLE_SIGNALS
3754extern int __kmp_handle_signals;
3755extern void __kmp_install_signals(int parallel_init);
3756extern void __kmp_remove_signals(void);
3757#endif
3758
3759extern void __kmp_clear_system_time(void);
3760extern void __kmp_read_system_time(double *delta);
3761
3762extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3763
3764extern void __kmp_expand_host_name(char *buffer, size_t size);
3765extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3766
3767#if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM))
3768extern void
3769__kmp_initialize_system_tick(void); /* Initialize timer tick value */
3770#endif
3771
3772extern void
3773__kmp_runtime_initialize(void); /* machine specific initialization */
3774extern void __kmp_runtime_destroy(void);
3775
3776#if KMP_AFFINITY_SUPPORTED
3777extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3778 kmp_affin_mask_t *mask);
3779extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3780 kmp_affin_mask_t *mask);
3781extern void __kmp_affinity_initialize(kmp_affinity_t &affinity);
3782extern void __kmp_affinity_uninitialize(void);
3783extern void __kmp_affinity_set_init_mask(
3784 int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3785extern void __kmp_affinity_set_place(int gtid);
3786extern void __kmp_affinity_determine_capable(const char *env_var);
3787extern int __kmp_aux_set_affinity(void **mask);
3788extern int __kmp_aux_get_affinity(void **mask);
3789extern int __kmp_aux_get_affinity_max_proc();
3790extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3791extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3792extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3793extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3794#if KMP_OS_LINUX || KMP_OS_FREEBSD
3795extern int kmp_set_thread_affinity_mask_initial(void);
3796#endif
3797static inline void __kmp_assign_root_init_mask() {
3798 int gtid = __kmp_entry_gtid();
3799 kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3800 if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3801 __kmp_affinity_set_init_mask(gtid, TRUE);
3802 r->r.r_affinity_assigned = TRUE;
3803 }
3804}
3805static inline void __kmp_reset_root_init_mask(int gtid) {
3806 if (!KMP_AFFINITY_CAPABLE())
3807 return;
3808 kmp_info_t *th = __kmp_threads[gtid];
3809 kmp_root_t *r = th->th.th_root;
3810 if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3811 __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3812 KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3813 r->r.r_affinity_assigned = FALSE;
3814 }
3815}
3816#else /* KMP_AFFINITY_SUPPORTED */
3817#define __kmp_assign_root_init_mask() /* Nothing */
3818static inline void __kmp_reset_root_init_mask(int gtid) {}
3819#endif /* KMP_AFFINITY_SUPPORTED */
3820// No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3821// format string is for affinity, so platforms that do not support
3822// affinity can still use the other fields, e.g., %n for num_threads
3823extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3824 kmp_str_buf_t *buffer);
3825extern void __kmp_aux_display_affinity(int gtid, const char *format);
3826
3827extern void __kmp_cleanup_hierarchy();
3828extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3829
3830#if KMP_USE_FUTEX
3831
3832extern int __kmp_futex_determine_capable(void);
3833
3834#endif // KMP_USE_FUTEX
3835
3836extern void __kmp_gtid_set_specific(int gtid);
3837extern int __kmp_gtid_get_specific(void);
3838
3839extern double __kmp_read_cpu_time(void);
3840
3841extern int __kmp_read_system_info(struct kmp_sys_info *info);
3842
3843#if KMP_USE_MONITOR
3844extern void __kmp_create_monitor(kmp_info_t *th);
3845#endif
3846
3847extern void *__kmp_launch_thread(kmp_info_t *thr);
3848
3849extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3850
3851#if KMP_OS_WINDOWS
3852extern int __kmp_still_running(kmp_info_t *th);
3853extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3854extern void __kmp_free_handle(kmp_thread_t tHandle);
3855#endif
3856
3857#if KMP_USE_MONITOR
3858extern void __kmp_reap_monitor(kmp_info_t *th);
3859#endif
3860extern void __kmp_reap_worker(kmp_info_t *th);
3861extern void __kmp_terminate_thread(int gtid);
3862
3863extern int __kmp_try_suspend_mx(kmp_info_t *th);
3864extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3865extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3866
3867extern void __kmp_elapsed(double *);
3868extern void __kmp_elapsed_tick(double *);
3869
3870extern void __kmp_enable(int old_state);
3871extern void __kmp_disable(int *old_state);
3872
3873extern void __kmp_thread_sleep(int millis);
3874
3875extern void __kmp_common_initialize(void);
3876extern void __kmp_common_destroy(void);
3877extern void __kmp_common_destroy_gtid(int gtid);
3878
3879#if KMP_OS_UNIX
3880extern void __kmp_register_atfork(void);
3881#endif
3882extern void __kmp_suspend_initialize(void);
3883extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3884extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3885
3886extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3887 int tid);
3888extern kmp_team_t *
3889__kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3890#if OMPT_SUPPORT
3891 ompt_data_t ompt_parallel_data,
3892#endif
3893 kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3894 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3895extern void __kmp_free_thread(kmp_info_t *);
3896extern void __kmp_free_team(kmp_root_t *,
3897 kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3898extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3899
3900/* ------------------------------------------------------------------------ */
3901
3902extern void __kmp_initialize_bget(kmp_info_t *th);
3903extern void __kmp_finalize_bget(kmp_info_t *th);
3904
3905KMP_EXPORT void *kmpc_malloc(size_t size);
3906KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3907KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3908KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3909KMP_EXPORT void kmpc_free(void *ptr);
3910
3911/* declarations for internal use */
3912
3913extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3914 size_t reduce_size, void *reduce_data,
3915 void (*reduce)(void *, void *));
3916extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3917extern int __kmp_barrier_gomp_cancel(int gtid);
3918
3923enum fork_context_e {
3924 fork_context_gnu,
3926 fork_context_intel,
3927 fork_context_last
3928};
3929extern int __kmp_fork_call(ident_t *loc, int gtid,
3930 enum fork_context_e fork_context, kmp_int32 argc,
3931 microtask_t microtask, launch_t invoker,
3932 kmp_va_list ap);
3933
3934extern void __kmp_join_call(ident_t *loc, int gtid
3935#if OMPT_SUPPORT
3936 ,
3937 enum fork_context_e fork_context
3938#endif
3939 ,
3940 int exit_teams = 0);
3941
3942extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3943extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3944extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3945extern int __kmp_invoke_task_func(int gtid);
3946extern void __kmp_run_before_invoked_task(int gtid, int tid,
3947 kmp_info_t *this_thr,
3948 kmp_team_t *team);
3949extern void __kmp_run_after_invoked_task(int gtid, int tid,
3950 kmp_info_t *this_thr,
3951 kmp_team_t *team);
3952
3953// should never have been exported
3954KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3955extern int __kmp_invoke_teams_master(int gtid);
3956extern void __kmp_teams_master(int gtid);
3957extern int __kmp_aux_get_team_num();
3958extern int __kmp_aux_get_num_teams();
3959extern void __kmp_save_internal_controls(kmp_info_t *thread);
3960extern void __kmp_user_set_library(enum library_type arg);
3961extern void __kmp_aux_set_library(enum library_type arg);
3962extern void __kmp_aux_set_stacksize(size_t arg);
3963extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3964extern void __kmp_aux_set_defaults(char const *str, size_t len);
3965
3966/* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3967void kmpc_set_blocktime(int arg);
3968void ompc_set_nested(int flag);
3969void ompc_set_dynamic(int flag);
3970void ompc_set_num_threads(int arg);
3971
3972extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3973 kmp_team_t *team, int tid);
3974extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3975extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3976 kmp_tasking_flags_t *flags,
3977 size_t sizeof_kmp_task_t,
3978 size_t sizeof_shareds,
3979 kmp_routine_entry_t task_entry);
3980extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3981 kmp_team_t *team, int tid,
3982 int set_curr_task);
3983extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3984extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3985
3986extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3987 int gtid,
3988 kmp_task_t *task);
3989extern void __kmp_fulfill_event(kmp_event_t *event);
3990
3991extern void __kmp_free_task_team(kmp_info_t *thread,
3992 kmp_task_team_t *task_team);
3993extern void __kmp_reap_task_teams(void);
3994extern void __kmp_wait_to_unref_task_teams(void);
3995extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3996 int always);
3997extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3998extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3999#if USE_ITT_BUILD
4000 ,
4001 void *itt_sync_obj
4002#endif /* USE_ITT_BUILD */
4003 ,
4004 int wait = 1);
4005extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
4006 int gtid);
4007
4008extern int __kmp_is_address_mapped(void *addr);
4009extern kmp_uint64 __kmp_hardware_timestamp(void);
4010
4011#if KMP_OS_UNIX
4012extern int __kmp_read_from_file(char const *path, char const *format, ...);
4013#endif
4014
4015/* ------------------------------------------------------------------------ */
4016//
4017// Assembly routines that have no compiler intrinsic replacement
4018//
4019
4020extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
4021 void *argv[]
4022#if OMPT_SUPPORT
4023 ,
4024 void **exit_frame_ptr
4025#endif
4026);
4027
4028/* ------------------------------------------------------------------------ */
4029
4030KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
4031KMP_EXPORT void __kmpc_end(ident_t *);
4032
4033KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
4034 kmpc_ctor_vec ctor,
4035 kmpc_cctor_vec cctor,
4036 kmpc_dtor_vec dtor,
4037 size_t vector_length);
4038KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
4039 kmpc_ctor ctor, kmpc_cctor cctor,
4040 kmpc_dtor dtor);
4041KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
4042 void *data, size_t size);
4043
4044KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
4045KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
4046KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
4047KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
4048
4049KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
4050KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
4051 kmpc_micro microtask, ...);
4052KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs,
4053 kmpc_micro microtask, kmp_int32 cond,
4054 void *args);
4055
4056KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
4057KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
4058
4059KMP_EXPORT void __kmpc_flush(ident_t *);
4060KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
4061KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
4062KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
4063KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
4064 kmp_int32 filter);
4065KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
4066KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
4067KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
4068KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
4069 kmp_critical_name *);
4070KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
4071 kmp_critical_name *);
4072KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
4073 kmp_critical_name *, uint32_t hint);
4074
4075KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
4076KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
4077
4078KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
4079 kmp_int32 global_tid);
4080
4081KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
4082KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
4083
4084KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid);
4085KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid,
4086 kmp_int32 numberOfSections);
4087KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid);
4088
4089KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
4090 kmp_int32 schedtype, kmp_int32 *plastiter,
4091 kmp_int *plower, kmp_int *pupper,
4092 kmp_int *pstride, kmp_int incr,
4093 kmp_int chunk);
4094
4095KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
4096
4097KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
4098 size_t cpy_size, void *cpy_data,
4099 void (*cpy_func)(void *, void *),
4100 kmp_int32 didit);
4101
4102KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid,
4103 void *cpy_data);
4104
4105extern void KMPC_SET_NUM_THREADS(int arg);
4106extern void KMPC_SET_DYNAMIC(int flag);
4107extern void KMPC_SET_NESTED(int flag);
4108
4109/* OMP 3.0 tasking interface routines */
4110KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
4111 kmp_task_t *new_task);
4112KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4113 kmp_int32 flags,
4114 size_t sizeof_kmp_task_t,
4115 size_t sizeof_shareds,
4116 kmp_routine_entry_t task_entry);
4117KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
4118 ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
4119 size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
4120KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
4121 kmp_task_t *task);
4122KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
4123 kmp_task_t *task);
4124KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
4125 kmp_task_t *new_task);
4126KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
4127KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
4128 int end_part);
4129
4130#if TASK_UNUSED
4131void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
4132void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
4133 kmp_task_t *task);
4134#endif // TASK_UNUSED
4135
4136/* ------------------------------------------------------------------------ */
4137
4138KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
4139KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
4140
4141KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
4142 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
4143 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
4144 kmp_depend_info_t *noalias_dep_list);
4145KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
4146 kmp_int32 ndeps,
4147 kmp_depend_info_t *dep_list,
4148 kmp_int32 ndeps_noalias,
4149 kmp_depend_info_t *noalias_dep_list);
4150/* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause.
4151 * Placeholder for taskwait with nowait clause.*/
4152KMP_EXPORT void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid,
4153 kmp_int32 ndeps,
4154 kmp_depend_info_t *dep_list,
4155 kmp_int32 ndeps_noalias,
4156 kmp_depend_info_t *noalias_dep_list,
4157 kmp_int32 has_no_wait);
4158
4159extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
4160 bool serialize_immediate);
4161
4162KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
4163 kmp_int32 cncl_kind);
4164KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
4165 kmp_int32 cncl_kind);
4166KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
4167KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
4168
4169KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
4170KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
4171KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
4172 kmp_int32 if_val, kmp_uint64 *lb,
4173 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
4174 kmp_int32 sched, kmp_uint64 grainsize,
4175 void *task_dup);
4176KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
4177 kmp_task_t *task, kmp_int32 if_val,
4178 kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4179 kmp_int32 nogroup, kmp_int32 sched,
4180 kmp_uint64 grainsize, kmp_int32 modifier,
4181 void *task_dup);
4182KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
4183KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
4184KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
4185KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
4186 int is_ws, int num,
4187 void *data);
4188KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
4189 int num, void *data);
4190KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
4191 int is_ws);
4192KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
4193 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4194 kmp_task_affinity_info_t *affin_list);
4195KMP_EXPORT void __kmp_set_num_teams(int num_teams);
4196KMP_EXPORT int __kmp_get_max_teams(void);
4197KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
4198KMP_EXPORT int __kmp_get_teams_thread_limit(void);
4199
4200/* Interface target task integration */
4201KMP_EXPORT void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid);
4202KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid);
4203
4204/* Lock interface routines (fast versions with gtid passed in) */
4205KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
4206 void **user_lock);
4207KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
4208 void **user_lock);
4209KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
4210 void **user_lock);
4211KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
4212 void **user_lock);
4213KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4214KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
4215 void **user_lock);
4216KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
4217 void **user_lock);
4218KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
4219 void **user_lock);
4220KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4221KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
4222 void **user_lock);
4223
4224KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4225 void **user_lock, uintptr_t hint);
4226KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4227 void **user_lock,
4228 uintptr_t hint);
4229
4230#if OMPX_TASKGRAPH
4231// Taskgraph's Record & Replay mechanism
4232// __kmp_tdg_is_recording: check whether a given TDG is recording
4233// status: the tdg's current status
4234static inline bool __kmp_tdg_is_recording(kmp_tdg_status_t status) {
4235 return status == KMP_TDG_RECORDING;
4236}
4237
4238KMP_EXPORT kmp_int32 __kmpc_start_record_task(ident_t *loc, kmp_int32 gtid,
4239 kmp_int32 input_flags,
4240 kmp_int32 tdg_id);
4241KMP_EXPORT void __kmpc_end_record_task(ident_t *loc, kmp_int32 gtid,
4242 kmp_int32 input_flags, kmp_int32 tdg_id);
4243#endif
4244/* Interface to fast scalable reduce methods routines */
4245
4246KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
4247 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4248 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4249 kmp_critical_name *lck);
4250KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4251 kmp_critical_name *lck);
4252KMP_EXPORT kmp_int32 __kmpc_reduce(
4253 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4254 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4255 kmp_critical_name *lck);
4256KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4257 kmp_critical_name *lck);
4258
4259/* Internal fast reduction routines */
4260
4261extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4262 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4263 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4264 kmp_critical_name *lck);
4265
4266// this function is for testing set/get/determine reduce method
4267KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4268
4269KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4270KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4271
4272// C++ port
4273// missing 'extern "C"' declarations
4274
4275KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4276KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4277KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4278 kmp_int32 num_threads);
4279
4280KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4281 int proc_bind);
4282KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4283 kmp_int32 num_teams,
4284 kmp_int32 num_threads);
4285/* Function for OpenMP 5.1 num_teams clause */
4286KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4287 kmp_int32 num_teams_lb,
4288 kmp_int32 num_teams_ub,
4289 kmp_int32 num_threads);
4290KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4291 kmpc_micro microtask, ...);
4292struct kmp_dim { // loop bounds info casted to kmp_int64
4293 kmp_int64 lo; // lower
4294 kmp_int64 up; // upper
4295 kmp_int64 st; // stride
4296};
4297KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4298 kmp_int32 num_dims,
4299 const struct kmp_dim *dims);
4300KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4301 const kmp_int64 *vec);
4302KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4303 const kmp_int64 *vec);
4304KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4305
4306KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4307 void *data, size_t size,
4308 void ***cache);
4309
4310// The routines below are not exported.
4311// Consider making them 'static' in corresponding source files.
4312void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4313 void *data_addr, size_t pc_size);
4314struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4315 void *data_addr,
4316 size_t pc_size);
4317void __kmp_threadprivate_resize_cache(int newCapacity);
4318void __kmp_cleanup_threadprivate_caches();
4319
4320// ompc_, kmpc_ entries moved from omp.h.
4321#if KMP_OS_WINDOWS
4322#define KMPC_CONVENTION __cdecl
4323#else
4324#define KMPC_CONVENTION
4325#endif
4326
4327#ifndef __OMP_H
4328typedef enum omp_sched_t {
4329 omp_sched_static = 1,
4330 omp_sched_dynamic = 2,
4331 omp_sched_guided = 3,
4332 omp_sched_auto = 4
4333} omp_sched_t;
4334typedef void *kmp_affinity_mask_t;
4335#endif
4336
4337KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4338KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4339KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4340KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4341KMP_EXPORT int KMPC_CONVENTION
4342kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4343KMP_EXPORT int KMPC_CONVENTION
4344kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4345KMP_EXPORT int KMPC_CONVENTION
4346kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4347
4348KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4349KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4350KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4351KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4352KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4353void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4354size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4355void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4356size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4357 char const *format);
4358
4359enum kmp_target_offload_kind {
4360 tgt_disabled = 0,
4361 tgt_default = 1,
4362 tgt_mandatory = 2
4363};
4364typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4365// Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4366extern kmp_target_offload_kind_t __kmp_target_offload;
4367extern int __kmpc_get_target_offload();
4368
4369// Constants used in libomptarget
4370#define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4371#define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4372
4373// OMP Pause Resource
4374
4375// The following enum is used both to set the status in __kmp_pause_status, and
4376// as the internal equivalent of the externally-visible omp_pause_resource_t.
4377typedef enum kmp_pause_status_t {
4378 kmp_not_paused = 0, // status is not paused, or, requesting resume
4379 kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4380 kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4381} kmp_pause_status_t;
4382
4383// This stores the pause state of the runtime
4384extern kmp_pause_status_t __kmp_pause_status;
4385extern int __kmpc_pause_resource(kmp_pause_status_t level);
4386extern int __kmp_pause_resource(kmp_pause_status_t level);
4387// Soft resume sets __kmp_pause_status, and wakes up all threads.
4388extern void __kmp_resume_if_soft_paused();
4389// Hard resume simply resets the status to not paused. Library will appear to
4390// be uninitialized after hard pause. Let OMP constructs trigger required
4391// initializations.
4392static inline void __kmp_resume_if_hard_paused() {
4393 if (__kmp_pause_status == kmp_hard_paused) {
4394 __kmp_pause_status = kmp_not_paused;
4395 }
4396}
4397
4398extern void __kmp_omp_display_env(int verbose);
4399
4400// 1: it is initializing hidden helper team
4401extern volatile int __kmp_init_hidden_helper;
4402// 1: the hidden helper team is done
4403extern volatile int __kmp_hidden_helper_team_done;
4404// 1: enable hidden helper task
4405extern kmp_int32 __kmp_enable_hidden_helper;
4406// Main thread of hidden helper team
4407extern kmp_info_t *__kmp_hidden_helper_main_thread;
4408// Descriptors for the hidden helper threads
4409extern kmp_info_t **__kmp_hidden_helper_threads;
4410// Number of hidden helper threads
4411extern kmp_int32 __kmp_hidden_helper_threads_num;
4412// Number of hidden helper tasks that have not been executed yet
4413extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4414
4415extern void __kmp_hidden_helper_initialize();
4416extern void __kmp_hidden_helper_threads_initz_routine();
4417extern void __kmp_do_initialize_hidden_helper_threads();
4418extern void __kmp_hidden_helper_threads_initz_wait();
4419extern void __kmp_hidden_helper_initz_release();
4420extern void __kmp_hidden_helper_threads_deinitz_wait();
4421extern void __kmp_hidden_helper_threads_deinitz_release();
4422extern void __kmp_hidden_helper_main_thread_wait();
4423extern void __kmp_hidden_helper_worker_thread_wait();
4424extern void __kmp_hidden_helper_worker_thread_signal();
4425extern void __kmp_hidden_helper_main_thread_release();
4426
4427// Check whether a given thread is a hidden helper thread
4428#define KMP_HIDDEN_HELPER_THREAD(gtid) \
4429 ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4430
4431#define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4432 ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4433
4434#define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid) \
4435 ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4436
4437#define KMP_HIDDEN_HELPER_TEAM(team) \
4438 (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4439
4440// Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4441// main thread, is skipped.
4442#define KMP_GTID_TO_SHADOW_GTID(gtid) \
4443 ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4444
4445// Return the adjusted gtid value by subtracting from gtid the number
4446// of hidden helper threads. This adjusted value is the gtid the thread would
4447// have received if there were no hidden helper threads.
4448static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4449 int adjusted_gtid = gtid;
4450 if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4451 gtid - __kmp_hidden_helper_threads_num >= 0) {
4452 adjusted_gtid -= __kmp_hidden_helper_threads_num;
4453 }
4454 return adjusted_gtid;
4455}
4456
4457// Support for error directive
4458typedef enum kmp_severity_t {
4459 severity_warning = 1,
4460 severity_fatal = 2
4461} kmp_severity_t;
4462extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4463
4464// Support for scope directive
4465KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4466KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4467
4468#ifdef __cplusplus
4469}
4470#endif
4471
4472template <bool C, bool S>
4473extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4474template <bool C, bool S>
4475extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4476template <bool C, bool S>
4477extern void __kmp_atomic_suspend_64(int th_gtid,
4478 kmp_atomic_flag_64<C, S> *flag);
4479extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4480#if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4481template <bool C, bool S>
4482extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4483template <bool C, bool S>
4484extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4485template <bool C, bool S>
4486extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4487extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4488#endif
4489template <bool C, bool S>
4490extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4491template <bool C, bool S>
4492extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4493template <bool C, bool S>
4494extern void __kmp_atomic_resume_64(int target_gtid,
4495 kmp_atomic_flag_64<C, S> *flag);
4496extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4497
4498template <bool C, bool S>
4499int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4500 kmp_flag_32<C, S> *flag, int final_spin,
4501 int *thread_finished,
4502#if USE_ITT_BUILD
4503 void *itt_sync_obj,
4504#endif /* USE_ITT_BUILD */
4505 kmp_int32 is_constrained);
4506template <bool C, bool S>
4507int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4508 kmp_flag_64<C, S> *flag, int final_spin,
4509 int *thread_finished,
4510#if USE_ITT_BUILD
4511 void *itt_sync_obj,
4512#endif /* USE_ITT_BUILD */
4513 kmp_int32 is_constrained);
4514template <bool C, bool S>
4515int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4516 kmp_atomic_flag_64<C, S> *flag,
4517 int final_spin, int *thread_finished,
4518#if USE_ITT_BUILD
4519 void *itt_sync_obj,
4520#endif /* USE_ITT_BUILD */
4521 kmp_int32 is_constrained);
4522int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4523 kmp_flag_oncore *flag, int final_spin,
4524 int *thread_finished,
4525#if USE_ITT_BUILD
4526 void *itt_sync_obj,
4527#endif /* USE_ITT_BUILD */
4528 kmp_int32 is_constrained);
4529
4530extern int __kmp_nesting_mode;
4531extern int __kmp_nesting_mode_nlevels;
4532extern int *__kmp_nesting_nth_level;
4533extern void __kmp_init_nesting_mode();
4534extern void __kmp_set_nesting_mode_threads();
4535
4543 FILE *f;
4544
4545 void close() {
4546 if (f && f != stdout && f != stderr) {
4547 fclose(f);
4548 f = nullptr;
4549 }
4550 }
4551
4552public:
4553 kmp_safe_raii_file_t() : f(nullptr) {}
4554 kmp_safe_raii_file_t(const char *filename, const char *mode,
4555 const char *env_var = nullptr)
4556 : f(nullptr) {
4557 open(filename, mode, env_var);
4558 }
4559 ~kmp_safe_raii_file_t() { close(); }
4560
4564 void open(const char *filename, const char *mode,
4565 const char *env_var = nullptr) {
4566 KMP_ASSERT(!f);
4567 f = fopen(filename, mode);
4568 if (!f) {
4569 int code = errno;
4570 if (env_var) {
4571 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4572 KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4573 } else {
4574 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4575 __kmp_msg_null);
4576 }
4577 }
4578 }
4581 int try_open(const char *filename, const char *mode) {
4582 KMP_ASSERT(!f);
4583 f = fopen(filename, mode);
4584 if (!f)
4585 return errno;
4586 return 0;
4587 }
4590 void set_stdout() {
4591 KMP_ASSERT(!f);
4592 f = stdout;
4593 }
4596 void set_stderr() {
4597 KMP_ASSERT(!f);
4598 f = stderr;
4599 }
4600 operator bool() { return bool(f); }
4601 operator FILE *() { return f; }
4602};
4603
4604template <typename SourceType, typename TargetType,
4605 bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4606 bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4607 bool isSourceSigned = std::is_signed<SourceType>::value,
4608 bool isTargetSigned = std::is_signed<TargetType>::value>
4609struct kmp_convert {};
4610
4611// Both types are signed; Source smaller
4612template <typename SourceType, typename TargetType>
4613struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4614 static TargetType to(SourceType src) { return (TargetType)src; }
4615};
4616// Source equal
4617template <typename SourceType, typename TargetType>
4618struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4619 static TargetType to(SourceType src) { return src; }
4620};
4621// Source bigger
4622template <typename SourceType, typename TargetType>
4623struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4624 static TargetType to(SourceType src) {
4625 KMP_ASSERT(src <= static_cast<SourceType>(
4626 (std::numeric_limits<TargetType>::max)()));
4627 KMP_ASSERT(src >= static_cast<SourceType>(
4628 (std::numeric_limits<TargetType>::min)()));
4629 return (TargetType)src;
4630 }
4631};
4632
4633// Source signed, Target unsigned
4634// Source smaller
4635template <typename SourceType, typename TargetType>
4636struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4637 static TargetType to(SourceType src) {
4638 KMP_ASSERT(src >= 0);
4639 return (TargetType)src;
4640 }
4641};
4642// Source equal
4643template <typename SourceType, typename TargetType>
4644struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4645 static TargetType to(SourceType src) {
4646 KMP_ASSERT(src >= 0);
4647 return (TargetType)src;
4648 }
4649};
4650// Source bigger
4651template <typename SourceType, typename TargetType>
4652struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4653 static TargetType to(SourceType src) {
4654 KMP_ASSERT(src >= 0);
4655 KMP_ASSERT(src <= static_cast<SourceType>(
4656 (std::numeric_limits<TargetType>::max)()));
4657 return (TargetType)src;
4658 }
4659};
4660
4661// Source unsigned, Target signed
4662// Source smaller
4663template <typename SourceType, typename TargetType>
4664struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4665 static TargetType to(SourceType src) { return (TargetType)src; }
4666};
4667// Source equal
4668template <typename SourceType, typename TargetType>
4669struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4670 static TargetType to(SourceType src) {
4671 KMP_ASSERT(src <= static_cast<SourceType>(
4672 (std::numeric_limits<TargetType>::max)()));
4673 return (TargetType)src;
4674 }
4675};
4676// Source bigger
4677template <typename SourceType, typename TargetType>
4678struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4679 static TargetType to(SourceType src) {
4680 KMP_ASSERT(src <= static_cast<SourceType>(
4681 (std::numeric_limits<TargetType>::max)()));
4682 return (TargetType)src;
4683 }
4684};
4685
4686// Source unsigned, Target unsigned
4687// Source smaller
4688template <typename SourceType, typename TargetType>
4689struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4690 static TargetType to(SourceType src) { return (TargetType)src; }
4691};
4692// Source equal
4693template <typename SourceType, typename TargetType>
4694struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4695 static TargetType to(SourceType src) { return src; }
4696};
4697// Source bigger
4698template <typename SourceType, typename TargetType>
4699struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4700 static TargetType to(SourceType src) {
4701 KMP_ASSERT(src <= static_cast<SourceType>(
4702 (std::numeric_limits<TargetType>::max)()));
4703 return (TargetType)src;
4704 }
4705};
4706
4707template <typename T1, typename T2>
4708static inline void __kmp_type_convert(T1 src, T2 *dest) {
4709 *dest = kmp_convert<T1, T2>::to(src);
4710}
4711
4712#endif /* KMP_H */
void set_stdout()
Definition: kmp.h:4590
void set_stderr()
Definition: kmp.h:4596
int try_open(const char *filename, const char *mode)
Definition: kmp.h:4581
void open(const char *filename, const char *mode, const char *env_var=nullptr)
Definition: kmp.h:4564
struct ident ident_t
@ KMP_IDENT_KMPC
Definition: kmp.h:196
@ KMP_IDENT_IMB
Definition: kmp.h:194
@ KMP_IDENT_WORK_LOOP
Definition: kmp.h:214
@ KMP_IDENT_BARRIER_IMPL
Definition: kmp.h:205
@ KMP_IDENT_WORK_SECTIONS
Definition: kmp.h:216
@ KMP_IDENT_AUTOPAR
Definition: kmp.h:199
@ KMP_IDENT_ATOMIC_HINT_MASK
Definition: kmp.h:223
@ KMP_IDENT_WORK_DISTRIBUTE
Definition: kmp.h:218
@ KMP_IDENT_BARRIER_EXPL
Definition: kmp.h:203
@ KMP_IDENT_ATOMIC_REDUCE
Definition: kmp.h:201
KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *)
KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro microtask,...)
KMP_EXPORT void __kmpc_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_serialized_parallel(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs, kmpc_micro microtask,...)
KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid)
void(* kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
Definition: kmp.h:1701
KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams_lb, kmp_int32 num_teams_ub, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags)
KMP_EXPORT void __kmpc_end(ident_t *)
KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
KMP_EXPORT kmp_int32 __kmpc_reduce(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void(*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck)
KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_flush(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void(*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck)
KMP_EXPORT void * __kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d)
void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int sched, kmp_uint64 grainsize, void *task_dup)
KMP_EXPORT void * __kmpc_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)
void __kmpc_taskloop_5(ident_t *loc, int gtid, kmp_task_t *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int sched, kmp_uint64 grainsize, int modifier, void *task_dup)
KMP_EXPORT void * __kmpc_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:1725
void *(* kmpc_cctor)(void *, void *)
Definition: kmp.h:1732
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:1754
void *(* kmpc_ctor)(void *)
Definition: kmp.h:1719
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:1742
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:1748
KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data, kmpc_ctor_vec ctor, kmpc_cctor_vec cctor, kmpc_dtor_vec dtor, size_t vector_length)
KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc)
KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *)
KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid)
int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_int32 *p_lb, kmp_int32 *p_ub, kmp_int32 *p_st)
sched_type
Definition: kmp.h:357
KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid)
void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid, kmp_critical_name *, uint32_t hint)
KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid, kmp_int32 numberOfSections)
void __kmpc_doacross_init(ident_t *loc, int gtid, int num_dims, const struct kmp_dim *dims)
int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_uint32 *p_lb, kmp_uint32 *p_ub, kmp_int32 *p_st)
KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid)
int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_uint64 *p_lb, kmp_uint64 *p_ub, kmp_int64 *p_st)
void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid)
int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_int64 *p_lb, kmp_int64 *p_ub, kmp_int64 *p_st)
void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid, kmp_int32 filter)
void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_int32 lb, kmp_int32 ub, kmp_int32 st, kmp_int32 chunk)
void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_uint32 lb, kmp_uint32 ub, kmp_int32 st, kmp_int32 chunk)
void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_uint64 lb, kmp_uint64 ub, kmp_int64 st, kmp_int64 chunk)
void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid)
void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_int64 lb, kmp_int64 ub, kmp_int64 st, kmp_int64 chunk)
KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
@ kmp_nm_guided_chunked
Definition: kmp.h:408
@ kmp_sch_runtime_simd
Definition: kmp.h:379
@ kmp_nm_ord_auto
Definition: kmp.h:427
@ kmp_sch_auto
Definition: kmp.h:364
@ kmp_nm_auto
Definition: kmp.h:410
@ kmp_distribute_static_chunked
Definition: kmp.h:395
@ kmp_sch_static
Definition: kmp.h:360
@ kmp_sch_guided_simd
Definition: kmp.h:378
@ kmp_sch_modifier_monotonic
Definition: kmp.h:445
@ kmp_sch_default
Definition: kmp.h:465
@ kmp_sch_modifier_nonmonotonic
Definition: kmp.h:447
@ kmp_nm_ord_static
Definition: kmp.h:423
@ kmp_distribute_static
Definition: kmp.h:396
@ kmp_sch_guided_chunked
Definition: kmp.h:362
@ kmp_nm_static
Definition: kmp.h:406
@ kmp_sch_lower
Definition: kmp.h:358
@ kmp_nm_upper
Definition: kmp.h:429
@ kmp_ord_lower
Definition: kmp.h:384
@ kmp_ord_static
Definition: kmp.h:386
@ kmp_sch_upper
Definition: kmp.h:382
@ kmp_ord_upper
Definition: kmp.h:392
@ kmp_nm_lower
Definition: kmp.h:402
@ kmp_ord_auto
Definition: kmp.h:390
Definition: kmp.h:234
kmp_int32 reserved_1
Definition: kmp.h:235
char const * psource
Definition: kmp.h:244
kmp_int32 reserved_2
Definition: kmp.h:238
kmp_int32 reserved_3
Definition: kmp.h:243
kmp_int32 flags
Definition: kmp.h:236