LLVM OpenMP* Runtime Library
Loading...
Searching...
No Matches
kmp.h
1
2/*
3 * kmp.h -- KPTS runtime header file.
4 */
5
6//===----------------------------------------------------------------------===//
7//
8// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
9// See https://llvm.org/LICENSE.txt for license information.
10// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
11//
12//===----------------------------------------------------------------------===//
13
14#ifndef KMP_H
15#define KMP_H
16
17#include "kmp_config.h"
18
19/* #define BUILD_PARALLEL_ORDERED 1 */
20
21/* This fix replaces gettimeofday with clock_gettime for better scalability on
22 the Altix. Requires user code to be linked with -lrt. */
23//#define FIX_SGI_CLOCK
24
25/* Defines for OpenMP 3.0 tasking and auto scheduling */
26
27#ifndef KMP_STATIC_STEAL_ENABLED
28#define KMP_STATIC_STEAL_ENABLED 1
29#endif
30#define KMP_WEIGHTED_ITERATIONS_SUPPORTED \
31 (KMP_AFFINITY_SUPPORTED && KMP_STATIC_STEAL_ENABLED && \
32 (KMP_ARCH_X86 || KMP_ARCH_X86_64))
33
34#define TASK_CURRENT_NOT_QUEUED 0
35#define TASK_CURRENT_QUEUED 1
36
37#ifdef BUILD_TIED_TASK_STACK
38#define TASK_STACK_EMPTY 0 // entries when the stack is empty
39#define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
40// Number of entries in each task stack array
41#define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
42// Mask for determining index into stack block
43#define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
44#endif // BUILD_TIED_TASK_STACK
45
46#define TASK_NOT_PUSHED 1
47#define TASK_SUCCESSFULLY_PUSHED 0
48#define TASK_TIED 1
49#define TASK_UNTIED 0
50#define TASK_EXPLICIT 1
51#define TASK_IMPLICIT 0
52#define TASK_PROXY 1
53#define TASK_FULL 0
54#define TASK_DETACHABLE 1
55#define TASK_UNDETACHABLE 0
56
57#define KMP_CANCEL_THREADS
58#define KMP_THREAD_ATTR
59
60// Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being
61// built on Android
62#if defined(__ANDROID__)
63#undef KMP_CANCEL_THREADS
64#endif
65
66// Some WASI targets (e.g., wasm32-wasi-threads) do not support thread
67// cancellation.
68#if KMP_OS_WASI
69#undef KMP_CANCEL_THREADS
70#endif
71
72#if !KMP_OS_WASI
73#include <signal.h>
74#endif
75#include <stdarg.h>
76#include <stddef.h>
77#include <stdio.h>
78#include <stdlib.h>
79#include <string.h>
80#include <limits>
81#include <type_traits>
82/* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
83 Microsoft library. Some macros provided below to replace these functions */
84#ifndef __ABSOFT_WIN
85#include <sys/types.h>
86#endif
87#include <limits.h>
88#include <time.h>
89
90#include <errno.h>
91
92#include "kmp_os.h"
93
94#include "kmp_safe_c_api.h"
95
96#if KMP_STATS_ENABLED
97class kmp_stats_list;
98#endif
99
100#if KMP_USE_HIER_SCHED
101// Only include hierarchical scheduling if affinity is supported
102#undef KMP_USE_HIER_SCHED
103#define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
104#endif
105
106#if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
107#include "hwloc.h"
108#ifndef HWLOC_OBJ_NUMANODE
109#define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
110#endif
111#ifndef HWLOC_OBJ_PACKAGE
112#define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
113#endif
114#endif
115
116#if KMP_ARCH_X86 || KMP_ARCH_X86_64
117#include <xmmintrin.h>
118#endif
119
120// The below has to be defined before including "kmp_barrier.h".
121#define KMP_INTERNAL_MALLOC(sz) malloc(sz)
122#define KMP_INTERNAL_FREE(p) free(p)
123#define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
124#define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
125
126#include "kmp_debug.h"
127#include "kmp_lock.h"
128#include "kmp_version.h"
129#include "kmp_barrier.h"
130#if USE_DEBUGGER
131#include "kmp_debugger.h"
132#endif
133#include "kmp_i18n.h"
134
135#define KMP_HANDLE_SIGNALS ((KMP_OS_UNIX && !KMP_OS_WASI) || KMP_OS_WINDOWS)
136
137#include "kmp_wrapper_malloc.h"
138#if KMP_OS_UNIX
139#include <unistd.h>
140#if !defined NSIG && defined _NSIG
141#define NSIG _NSIG
142#endif
143#endif
144
145#if KMP_OS_LINUX
146#pragma weak clock_gettime
147#endif
148
149#if OMPT_SUPPORT
150#include "ompt-internal.h"
151#endif
152
153#if OMPD_SUPPORT
154#include "ompd-specific.h"
155#endif
156
157#ifndef UNLIKELY
158#define UNLIKELY(x) (x)
159#endif
160
161// Affinity format function
162#include "kmp_str.h"
163
164// 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
165// 3 - fast allocation using sync, non-sync free lists of any size, non-self
166// free lists of limited size.
167#ifndef USE_FAST_MEMORY
168#define USE_FAST_MEMORY 3
169#endif
170
171#ifndef KMP_NESTED_HOT_TEAMS
172#define KMP_NESTED_HOT_TEAMS 0
173#define USE_NESTED_HOT_ARG(x)
174#else
175#if KMP_NESTED_HOT_TEAMS
176#define USE_NESTED_HOT_ARG(x) , x
177#else
178#define USE_NESTED_HOT_ARG(x)
179#endif
180#endif
181
182// Assume using BGET compare_exchange instruction instead of lock by default.
183#ifndef USE_CMP_XCHG_FOR_BGET
184#define USE_CMP_XCHG_FOR_BGET 1
185#endif
186
187// Test to see if queuing lock is better than bootstrap lock for bget
188// #ifndef USE_QUEUING_LOCK_FOR_BGET
189// #define USE_QUEUING_LOCK_FOR_BGET
190// #endif
191
192#define KMP_NSEC_PER_SEC 1000000000L
193#define KMP_USEC_PER_SEC 1000000L
194#define KMP_NSEC_PER_USEC 1000L
195
204enum {
209 /* 0x04 is no longer used */
218 KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
219 KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
220 KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
221
222 KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
223 KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
224
236 KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
237 KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
238 KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
239 KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
240 KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
241};
242
246typedef struct ident {
247 kmp_int32 reserved_1;
248 kmp_int32 flags;
250 kmp_int32 reserved_2;
251#if USE_ITT_BUILD
252/* but currently used for storing region-specific ITT */
253/* contextual information. */
254#endif /* USE_ITT_BUILD */
255 kmp_int32 reserved_3;
256 char const *psource;
260 // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
261 kmp_int32 get_openmp_version() {
262 return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
263 }
269// Some forward declarations.
270typedef union kmp_team kmp_team_t;
271typedef struct kmp_taskdata kmp_taskdata_t;
272typedef union kmp_task_team kmp_task_team_t;
273typedef union kmp_team kmp_team_p;
274typedef union kmp_info kmp_info_p;
275typedef union kmp_root kmp_root_p;
276
277template <bool C = false, bool S = true> class kmp_flag_32;
278template <bool C = false, bool S = true> class kmp_flag_64;
279template <bool C = false, bool S = true> class kmp_atomic_flag_64;
280class kmp_flag_oncore;
281
282#ifdef __cplusplus
283extern "C" {
284#endif
285
286/* ------------------------------------------------------------------------ */
287
288/* Pack two 32-bit signed integers into a 64-bit signed integer */
289/* ToDo: Fix word ordering for big-endian machines. */
290#define KMP_PACK_64(HIGH_32, LOW_32) \
291 ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
292
293// Generic string manipulation macros. Assume that _x is of type char *
294#define SKIP_WS(_x) \
295 { \
296 while (*(_x) == ' ' || *(_x) == '\t') \
297 (_x)++; \
298 }
299#define SKIP_DIGITS(_x) \
300 { \
301 while (*(_x) >= '0' && *(_x) <= '9') \
302 (_x)++; \
303 }
304#define SKIP_TOKEN(_x) \
305 { \
306 while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
307 (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
308 (_x)++; \
309 }
310#define SKIP_TO(_x, _c) \
311 { \
312 while (*(_x) != '\0' && *(_x) != (_c)) \
313 (_x)++; \
314 }
315
316/* ------------------------------------------------------------------------ */
317
318#define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
319#define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
320
321/* ------------------------------------------------------------------------ */
322/* Enumeration types */
323
324enum kmp_state_timer {
325 ts_stop,
326 ts_start,
327 ts_pause,
328
329 ts_last_state
330};
331
332enum dynamic_mode {
333 dynamic_default,
334#ifdef USE_LOAD_BALANCE
335 dynamic_load_balance,
336#endif /* USE_LOAD_BALANCE */
337 dynamic_random,
338 dynamic_thread_limit,
339 dynamic_max
340};
341
342/* external schedule constants, duplicate enum omp_sched in omp.h in order to
343 * not include it here */
344#ifndef KMP_SCHED_TYPE_DEFINED
345#define KMP_SCHED_TYPE_DEFINED
346typedef enum kmp_sched {
347 kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
348 // Note: need to adjust __kmp_sch_map global array in case enum is changed
349 kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
350 kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
351 kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
352 kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
353 kmp_sched_upper_std = 5, // upper bound for standard schedules
354 kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
355 kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
356#if KMP_STATIC_STEAL_ENABLED
357 kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
358#endif
359 kmp_sched_upper,
360 kmp_sched_default = kmp_sched_static, // default scheduling
361 kmp_sched_monotonic = 0x80000000
362} kmp_sched_t;
363#endif
364
369enum sched_type : kmp_int32 {
371 kmp_sch_static_chunked = 33,
373 kmp_sch_dynamic_chunked = 35,
375 kmp_sch_runtime = 37,
377 kmp_sch_trapezoidal = 39,
378
379 /* accessible only through KMP_SCHEDULE environment variable */
380 kmp_sch_static_greedy = 40,
381 kmp_sch_static_balanced = 41,
382 /* accessible only through KMP_SCHEDULE environment variable */
383 kmp_sch_guided_iterative_chunked = 42,
384 kmp_sch_guided_analytical_chunked = 43,
385 /* accessible only through KMP_SCHEDULE environment variable */
386 kmp_sch_static_steal = 44,
387
388 /* static with chunk adjustment (e.g., simd) */
389 kmp_sch_static_balanced_chunked = 45,
393 /* accessible only through KMP_SCHEDULE environment variable */
397 kmp_ord_static_chunked = 65,
399 kmp_ord_dynamic_chunked = 67,
400 kmp_ord_guided_chunked = 68,
401 kmp_ord_runtime = 69,
403 kmp_ord_trapezoidal = 71,
406 /* Schedules for Distribute construct */
410 /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
411 single iteration/chunk, even if the loop is serialized. For the schedule
412 types listed above, the entire iteration vector is returned if the loop is
413 serialized. This doesn't work for gcc/gcomp sections. */
416 kmp_nm_static_chunked =
417 (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
419 kmp_nm_dynamic_chunked = 163,
421 kmp_nm_runtime = 165,
423 kmp_nm_trapezoidal = 167,
424
425 /* accessible only through KMP_SCHEDULE environment variable */
426 kmp_nm_static_greedy = 168,
427 kmp_nm_static_balanced = 169,
428 /* accessible only through KMP_SCHEDULE environment variable */
429 kmp_nm_guided_iterative_chunked = 170,
430 kmp_nm_guided_analytical_chunked = 171,
431 kmp_nm_static_steal =
432 172, /* accessible only through OMP_SCHEDULE environment variable */
433
434 kmp_nm_ord_static_chunked = 193,
436 kmp_nm_ord_dynamic_chunked = 195,
437 kmp_nm_ord_guided_chunked = 196,
438 kmp_nm_ord_runtime = 197,
440 kmp_nm_ord_trapezoidal = 199,
443 /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
444 we need to distinguish the three possible cases (no modifier, monotonic
445 modifier, nonmonotonic modifier), we need separate bits for each modifier.
446 The absence of monotonic does not imply nonmonotonic, especially since 4.5
447 says that the behaviour of the "no modifier" case is implementation defined
448 in 4.5, but will become "nonmonotonic" in 5.0.
449
450 Since we're passing a full 32 bit value, we can use a couple of high bits
451 for these flags; out of paranoia we avoid the sign bit.
452
453 These modifiers can be or-ed into non-static schedules by the compiler to
454 pass the additional information. They will be stripped early in the
455 processing in __kmp_dispatch_init when setting up schedules, so most of the
456 code won't ever see schedules with these bits set. */
458 (1 << 29),
460 (1 << 30),
462#define SCHEDULE_WITHOUT_MODIFIERS(s) \
463 (enum sched_type)( \
465#define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
466#define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
467#define SCHEDULE_HAS_NO_MODIFIERS(s) \
468 (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
469#define SCHEDULE_GET_MODIFIERS(s) \
470 ((enum sched_type)( \
471 (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
472#define SCHEDULE_SET_MODIFIERS(s, m) \
473 (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
474#define SCHEDULE_NONMONOTONIC 0
475#define SCHEDULE_MONOTONIC 1
476
479
480// Apply modifiers on internal kind to standard kind
481static inline void
482__kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
483 enum sched_type internal_kind) {
484 if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
485 *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
486 }
487}
488
489// Apply modifiers on standard kind to internal kind
490static inline void
491__kmp_sched_apply_mods_intkind(kmp_sched_t kind,
492 enum sched_type *internal_kind) {
493 if ((int)kind & (int)kmp_sched_monotonic) {
494 *internal_kind = (enum sched_type)((int)*internal_kind |
496 }
497}
498
499// Get standard schedule without modifiers
500static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
501 return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
502}
503
504/* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
505typedef union kmp_r_sched {
506 struct {
507 enum sched_type r_sched_type;
508 int chunk;
509 };
510 kmp_int64 sched;
511} kmp_r_sched_t;
512
513extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
514// internal schedule types
515
516enum library_type {
517 library_none,
518 library_serial,
519 library_turnaround,
520 library_throughput
521};
522
523#if KMP_OS_LINUX
524enum clock_function_type {
525 clock_function_gettimeofday,
526 clock_function_clock_gettime
527};
528#endif /* KMP_OS_LINUX */
529
530#if KMP_MIC_SUPPORTED
531enum mic_type { non_mic, mic1, mic2, mic3, dummy };
532#endif
533
534/* -- fast reduction stuff ------------------------------------------------ */
535
536#undef KMP_FAST_REDUCTION_BARRIER
537#define KMP_FAST_REDUCTION_BARRIER 1
538
539#undef KMP_FAST_REDUCTION_CORE_DUO
540#if KMP_ARCH_X86 || KMP_ARCH_X86_64
541#define KMP_FAST_REDUCTION_CORE_DUO 1
542#endif
543
544enum _reduction_method {
545 reduction_method_not_defined = 0,
546 critical_reduce_block = (1 << 8),
547 atomic_reduce_block = (2 << 8),
548 tree_reduce_block = (3 << 8),
549 empty_reduce_block = (4 << 8)
550};
551
552// Description of the packed_reduction_method variable:
553// The packed_reduction_method variable consists of two enum types variables
554// that are packed together into 0-th byte and 1-st byte:
555// 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
556// barrier that will be used in fast reduction: bs_plain_barrier or
557// bs_reduction_barrier
558// 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
559// be used in fast reduction;
560// Reduction method is of 'enum _reduction_method' type and it's defined the way
561// so that the bits of 0-th byte are empty, so no need to execute a shift
562// instruction while packing/unpacking
563
564#if KMP_FAST_REDUCTION_BARRIER
565#define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
566 ((reduction_method) | (barrier_type))
567
568#define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
569 ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
570
571#define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
572 ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
573#else
574#define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
575 (reduction_method)
576
577#define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
578 (packed_reduction_method)
579
580#define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
581#endif
582
583#define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
584 ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
585 (which_reduction_block))
586
587#if KMP_FAST_REDUCTION_BARRIER
588#define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
589 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
590
591#define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
592 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
593#endif
594
595typedef int PACKED_REDUCTION_METHOD_T;
596
597/* -- end of fast reduction stuff ----------------------------------------- */
598
599#if KMP_OS_WINDOWS
600#define USE_CBLKDATA
601#if KMP_MSVC_COMPAT
602#pragma warning(push)
603#pragma warning(disable : 271 310)
604#endif
605#include <windows.h>
606#if KMP_MSVC_COMPAT
607#pragma warning(pop)
608#endif
609#endif
610
611#if KMP_OS_UNIX
612#if !KMP_OS_WASI
613#include <dlfcn.h>
614#endif
615#include <pthread.h>
616#endif
617
618enum kmp_hw_t : int {
619 KMP_HW_UNKNOWN = -1,
620 KMP_HW_SOCKET = 0,
621 KMP_HW_PROC_GROUP,
622 KMP_HW_NUMA,
623 KMP_HW_DIE,
624 KMP_HW_LLC,
625 KMP_HW_L3,
626 KMP_HW_TILE,
627 KMP_HW_MODULE,
628 KMP_HW_L2,
629 KMP_HW_L1,
630 KMP_HW_CORE,
631 KMP_HW_THREAD,
632 KMP_HW_LAST
633};
634
635typedef enum kmp_hw_core_type_t {
636 KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
637#if KMP_ARCH_X86 || KMP_ARCH_X86_64
638 KMP_HW_CORE_TYPE_ATOM = 0x20,
639 KMP_HW_CORE_TYPE_CORE = 0x40,
640 KMP_HW_MAX_NUM_CORE_TYPES = 3,
641#else
642 KMP_HW_MAX_NUM_CORE_TYPES = 1,
643#endif
644} kmp_hw_core_type_t;
645
646#define KMP_HW_MAX_NUM_CORE_EFFS 8
647
648#define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \
649 KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
650#define KMP_ASSERT_VALID_HW_TYPE(type) \
651 KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
652
653#define KMP_FOREACH_HW_TYPE(type) \
654 for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \
655 type = (kmp_hw_t)((int)type + 1))
656
657const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
658const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
659const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
660
661/* Only Linux* OS and Windows* OS support thread affinity. */
662#if KMP_AFFINITY_SUPPORTED
663
664// GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
665#if KMP_OS_WINDOWS
666#if _MSC_VER < 1600 && KMP_MSVC_COMPAT
667typedef struct GROUP_AFFINITY {
668 KAFFINITY Mask;
669 WORD Group;
670 WORD Reserved[3];
671} GROUP_AFFINITY;
672#endif /* _MSC_VER < 1600 */
673#if KMP_GROUP_AFFINITY
674extern int __kmp_num_proc_groups;
675#else
676static const int __kmp_num_proc_groups = 1;
677#endif /* KMP_GROUP_AFFINITY */
678typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
679extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
680
681typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
682extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
683
684typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
685extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
686
687typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
688 GROUP_AFFINITY *);
689extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
690#endif /* KMP_OS_WINDOWS */
691
692#if KMP_USE_HWLOC
693extern hwloc_topology_t __kmp_hwloc_topology;
694extern int __kmp_hwloc_error;
695#endif
696
697extern size_t __kmp_affin_mask_size;
698#define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
699#define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
700#define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
701#define KMP_CPU_SET_ITERATE(i, mask) \
702 for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
703#define KMP_CPU_SET(i, mask) (mask)->set(i)
704#define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
705#define KMP_CPU_CLR(i, mask) (mask)->clear(i)
706#define KMP_CPU_ZERO(mask) (mask)->zero()
707#define KMP_CPU_ISEMPTY(mask) (mask)->empty()
708#define KMP_CPU_COPY(dest, src) (dest)->copy(src)
709#define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
710#define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
711#define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
712#define KMP_CPU_EQUAL(dest, src) (dest)->is_equal(src)
713#define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
714#define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
715#define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
716#define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
717#define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
718#define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
719#define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
720#define KMP_CPU_ALLOC_ARRAY(arr, n) \
721 (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
722#define KMP_CPU_FREE_ARRAY(arr, n) \
723 __kmp_affinity_dispatch->deallocate_mask_array(arr)
724#define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
725#define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
726#define __kmp_get_system_affinity(mask, abort_bool) \
727 (mask)->get_system_affinity(abort_bool)
728#define __kmp_set_system_affinity(mask, abort_bool) \
729 (mask)->set_system_affinity(abort_bool)
730#define __kmp_get_proc_group(mask) (mask)->get_proc_group()
731
732class KMPAffinity {
733public:
734 class Mask {
735 public:
736 void *operator new(size_t n);
737 void operator delete(void *p);
738 void *operator new[](size_t n);
739 void operator delete[](void *p);
740 virtual ~Mask() {}
741 // Set bit i to 1
742 virtual void set(int i) {}
743 // Return bit i
744 virtual bool is_set(int i) const { return false; }
745 // Set bit i to 0
746 virtual void clear(int i) {}
747 // Zero out entire mask
748 virtual void zero() {}
749 // Check whether mask is empty
750 virtual bool empty() const { return true; }
751 // Copy src into this mask
752 virtual void copy(const Mask *src) {}
753 // this &= rhs
754 virtual void bitwise_and(const Mask *rhs) {}
755 // this |= rhs
756 virtual void bitwise_or(const Mask *rhs) {}
757 // this = ~this
758 virtual void bitwise_not() {}
759 // this == rhs
760 virtual bool is_equal(const Mask *rhs) const { return false; }
761 // API for iterating over an affinity mask
762 // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
763 virtual int begin() const { return 0; }
764 virtual int end() const { return 0; }
765 virtual int next(int previous) const { return 0; }
766#if KMP_OS_WINDOWS
767 virtual int set_process_affinity(bool abort_on_error) const { return -1; }
768#endif
769 // Set the system's affinity to this affinity mask's value
770 virtual int set_system_affinity(bool abort_on_error) const { return -1; }
771 // Set this affinity mask to the current system affinity
772 virtual int get_system_affinity(bool abort_on_error) { return -1; }
773 // Only 1 DWORD in the mask should have any procs set.
774 // Return the appropriate index, or -1 for an invalid mask.
775 virtual int get_proc_group() const { return -1; }
776 int get_max_cpu() const {
777 int cpu;
778 int max_cpu = -1;
779 KMP_CPU_SET_ITERATE(cpu, this) {
780 if (cpu > max_cpu)
781 max_cpu = cpu;
782 }
783 return max_cpu;
784 }
785 };
786 void *operator new(size_t n);
787 void operator delete(void *p);
788 // Need virtual destructor
789 virtual ~KMPAffinity() = default;
790 // Determine if affinity is capable
791 virtual void determine_capable(const char *env_var) {}
792 // Bind the current thread to os proc
793 virtual void bind_thread(int proc) {}
794 // Factory functions to allocate/deallocate a mask
795 virtual Mask *allocate_mask() { return nullptr; }
796 virtual void deallocate_mask(Mask *m) {}
797 virtual Mask *allocate_mask_array(int num) { return nullptr; }
798 virtual void deallocate_mask_array(Mask *m) {}
799 virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
800 static void pick_api();
801 static void destroy_api();
802 enum api_type {
803 NATIVE_OS
804#if KMP_USE_HWLOC
805 ,
806 HWLOC
807#endif
808 };
809 virtual api_type get_api_type() const {
810 KMP_ASSERT(0);
811 return NATIVE_OS;
812 }
813
814private:
815 static bool picked_api;
816};
817
818typedef KMPAffinity::Mask kmp_affin_mask_t;
819extern KMPAffinity *__kmp_affinity_dispatch;
820
821class kmp_affinity_raii_t {
822 kmp_affin_mask_t *mask;
823 bool restored;
824
825public:
826 kmp_affinity_raii_t(const kmp_affin_mask_t *new_mask = nullptr)
827 : restored(false) {
828 if (KMP_AFFINITY_CAPABLE()) {
829 KMP_CPU_ALLOC(mask);
830 KMP_ASSERT(mask != NULL);
831 __kmp_get_system_affinity(mask, /*abort_on_error=*/true);
832 if (new_mask)
833 __kmp_set_system_affinity(new_mask, /*abort_on_error=*/true);
834 }
835 }
836 void restore() {
837 if (!restored && KMP_AFFINITY_CAPABLE()) {
838 __kmp_set_system_affinity(mask, /*abort_on_error=*/true);
839 KMP_CPU_FREE(mask);
840 }
841 restored = true;
842 }
843 ~kmp_affinity_raii_t() { restore(); }
844};
845
846// Declare local char buffers with this size for printing debug and info
847// messages, using __kmp_affinity_print_mask().
848#define KMP_AFFIN_MASK_PRINT_LEN 1024
849
850enum affinity_type {
851 affinity_none = 0,
852 affinity_physical,
853 affinity_logical,
854 affinity_compact,
855 affinity_scatter,
856 affinity_explicit,
857 affinity_balanced,
858 affinity_disabled, // not used outsize the env var parser
859 affinity_default
860};
861
862enum affinity_top_method {
863 affinity_top_method_all = 0, // try all (supported) methods, in order
864#if KMP_ARCH_X86 || KMP_ARCH_X86_64
865 affinity_top_method_apicid,
866 affinity_top_method_x2apicid,
867 affinity_top_method_x2apicid_1f,
868#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
869 affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
870#if KMP_GROUP_AFFINITY
871 affinity_top_method_group,
872#endif /* KMP_GROUP_AFFINITY */
873 affinity_top_method_flat,
874#if KMP_USE_HWLOC
875 affinity_top_method_hwloc,
876#endif
877 affinity_top_method_default
878};
879
880#define affinity_respect_mask_default (2)
881
882typedef struct kmp_affinity_flags_t {
883 unsigned dups : 1;
884 unsigned verbose : 1;
885 unsigned warnings : 1;
886 unsigned respect : 2;
887 unsigned reset : 1;
888 unsigned initialized : 1;
889 unsigned core_types_gran : 1;
890 unsigned core_effs_gran : 1;
891 unsigned omp_places : 1;
892 unsigned reserved : 22;
893} kmp_affinity_flags_t;
894KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4);
895
896typedef struct kmp_affinity_ids_t {
897 int os_id;
898 int ids[KMP_HW_LAST];
899} kmp_affinity_ids_t;
900
901typedef struct kmp_affinity_attrs_t {
902 int core_type : 8;
903 int core_eff : 8;
904 unsigned valid : 1;
905 unsigned reserved : 15;
906} kmp_affinity_attrs_t;
907#define KMP_AFFINITY_ATTRS_UNKNOWN \
908 { KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 }
909
910typedef struct kmp_affinity_t {
911 char *proclist;
912 enum affinity_type type;
913 kmp_hw_t gran;
914 int gran_levels;
915 kmp_affinity_attrs_t core_attr_gran;
916 int compact;
917 int offset;
918 kmp_affinity_flags_t flags;
919 unsigned num_masks;
920 kmp_affin_mask_t *masks;
921 kmp_affinity_ids_t *ids;
922 kmp_affinity_attrs_t *attrs;
923 unsigned num_os_id_masks;
924 kmp_affin_mask_t *os_id_masks;
925 const char *env_var;
926} kmp_affinity_t;
927
928#define KMP_AFFINITY_INIT(env) \
929 { \
930 nullptr, affinity_default, KMP_HW_UNKNOWN, -1, KMP_AFFINITY_ATTRS_UNKNOWN, \
931 0, 0, \
932 {TRUE, FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE, \
933 FALSE, FALSE, FALSE}, \
934 0, nullptr, nullptr, nullptr, 0, nullptr, env \
935 }
936
937extern enum affinity_top_method __kmp_affinity_top_method;
938extern kmp_affinity_t __kmp_affinity;
939extern kmp_affinity_t __kmp_hh_affinity;
940extern kmp_affinity_t *__kmp_affinities[2];
941
942extern void __kmp_affinity_bind_thread(int which);
943
944extern kmp_affin_mask_t *__kmp_affin_fullMask;
945extern kmp_affin_mask_t *__kmp_affin_origMask;
946extern char *__kmp_cpuinfo_file;
947
948#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
949extern int __kmp_first_osid_with_ecore;
950#endif
951
952#endif /* KMP_AFFINITY_SUPPORTED */
953
954// This needs to be kept in sync with the values in omp.h !!!
955typedef enum kmp_proc_bind_t {
956 proc_bind_false = 0,
957 proc_bind_true,
958 proc_bind_primary,
959 proc_bind_close,
960 proc_bind_spread,
961 proc_bind_intel, // use KMP_AFFINITY interface
962 proc_bind_default
963} kmp_proc_bind_t;
964
965typedef struct kmp_nested_proc_bind_t {
966 kmp_proc_bind_t *bind_types;
967 int size;
968 int used;
969} kmp_nested_proc_bind_t;
970
971extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
972extern kmp_proc_bind_t __kmp_teams_proc_bind;
973
974extern int __kmp_display_affinity;
975extern char *__kmp_affinity_format;
976static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
977#if OMPT_SUPPORT
978extern int __kmp_tool;
979extern char *__kmp_tool_libraries;
980#endif // OMPT_SUPPORT
981
982#if KMP_AFFINITY_SUPPORTED
983#define KMP_PLACE_ALL (-1)
984#define KMP_PLACE_UNDEFINED (-2)
985// Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
986#define KMP_AFFINITY_NON_PROC_BIND \
987 ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
988 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
989 (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced))
990#endif /* KMP_AFFINITY_SUPPORTED */
991
992extern int __kmp_affinity_num_places;
993
994typedef enum kmp_cancel_kind_t {
995 cancel_noreq = 0,
996 cancel_parallel = 1,
997 cancel_loop = 2,
998 cancel_sections = 3,
999 cancel_taskgroup = 4
1000} kmp_cancel_kind_t;
1001
1002// KMP_HW_SUBSET support:
1003typedef struct kmp_hws_item {
1004 int num;
1005 int offset;
1006} kmp_hws_item_t;
1007
1008extern kmp_hws_item_t __kmp_hws_socket;
1009extern kmp_hws_item_t __kmp_hws_die;
1010extern kmp_hws_item_t __kmp_hws_node;
1011extern kmp_hws_item_t __kmp_hws_tile;
1012extern kmp_hws_item_t __kmp_hws_core;
1013extern kmp_hws_item_t __kmp_hws_proc;
1014extern int __kmp_hws_requested;
1015extern int __kmp_hws_abs_flag; // absolute or per-item number requested
1016
1017/* ------------------------------------------------------------------------ */
1018
1019#define KMP_PAD(type, sz) \
1020 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
1021
1022// We need to avoid using -1 as a GTID as +1 is added to the gtid
1023// when storing it in a lock, and the value 0 is reserved.
1024#define KMP_GTID_DNE (-2) /* Does not exist */
1025#define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
1026#define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
1027#define KMP_GTID_UNKNOWN (-5) /* Is not known */
1028#define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
1029
1030/* OpenMP 5.0 Memory Management support */
1031
1032#ifndef __OMP_H
1033// Duplicate type definitions from omp.h
1034typedef uintptr_t omp_uintptr_t;
1035
1036typedef enum {
1037 omp_atk_sync_hint = 1,
1038 omp_atk_alignment = 2,
1039 omp_atk_access = 3,
1040 omp_atk_pool_size = 4,
1041 omp_atk_fallback = 5,
1042 omp_atk_fb_data = 6,
1043 omp_atk_pinned = 7,
1044 omp_atk_partition = 8
1045} omp_alloctrait_key_t;
1046
1047typedef enum {
1048 omp_atv_false = 0,
1049 omp_atv_true = 1,
1050 omp_atv_contended = 3,
1051 omp_atv_uncontended = 4,
1052 omp_atv_serialized = 5,
1053 omp_atv_sequential = omp_atv_serialized, // (deprecated)
1054 omp_atv_private = 6,
1055 omp_atv_all = 7,
1056 omp_atv_thread = 8,
1057 omp_atv_pteam = 9,
1058 omp_atv_cgroup = 10,
1059 omp_atv_default_mem_fb = 11,
1060 omp_atv_null_fb = 12,
1061 omp_atv_abort_fb = 13,
1062 omp_atv_allocator_fb = 14,
1063 omp_atv_environment = 15,
1064 omp_atv_nearest = 16,
1065 omp_atv_blocked = 17,
1066 omp_atv_interleaved = 18
1067} omp_alloctrait_value_t;
1068#define omp_atv_default ((omp_uintptr_t)-1)
1069
1070typedef void *omp_memspace_handle_t;
1071extern omp_memspace_handle_t const omp_default_mem_space;
1072extern omp_memspace_handle_t const omp_large_cap_mem_space;
1073extern omp_memspace_handle_t const omp_const_mem_space;
1074extern omp_memspace_handle_t const omp_high_bw_mem_space;
1075extern omp_memspace_handle_t const omp_low_lat_mem_space;
1076extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
1077extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
1078extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
1079
1080typedef struct {
1081 omp_alloctrait_key_t key;
1082 omp_uintptr_t value;
1083} omp_alloctrait_t;
1084
1085typedef void *omp_allocator_handle_t;
1086extern omp_allocator_handle_t const omp_null_allocator;
1087extern omp_allocator_handle_t const omp_default_mem_alloc;
1088extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
1089extern omp_allocator_handle_t const omp_const_mem_alloc;
1090extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
1091extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
1092extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
1093extern omp_allocator_handle_t const omp_pteam_mem_alloc;
1094extern omp_allocator_handle_t const omp_thread_mem_alloc;
1095extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
1096extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
1097extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
1098extern omp_allocator_handle_t const kmp_max_mem_alloc;
1099extern omp_allocator_handle_t __kmp_def_allocator;
1100
1101// end of duplicate type definitions from omp.h
1102#endif
1103
1104extern int __kmp_memkind_available;
1105
1106typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
1107
1108typedef struct kmp_allocator_t {
1109 omp_memspace_handle_t memspace;
1110 void **memkind; // pointer to memkind
1111 size_t alignment;
1112 omp_alloctrait_value_t fb;
1113 kmp_allocator_t *fb_data;
1114 kmp_uint64 pool_size;
1115 kmp_uint64 pool_used;
1116 bool pinned;
1117} kmp_allocator_t;
1118
1119extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1120 omp_memspace_handle_t,
1121 int ntraits,
1122 omp_alloctrait_t traits[]);
1123extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1124extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1125extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1126// external interfaces, may be used by compiler
1127extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1128extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1129 omp_allocator_handle_t al);
1130extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1131 omp_allocator_handle_t al);
1132extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1133 omp_allocator_handle_t al,
1134 omp_allocator_handle_t free_al);
1135extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1136// internal interfaces, contain real implementation
1137extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1138 omp_allocator_handle_t al);
1139extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1140 omp_allocator_handle_t al);
1141extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1142 omp_allocator_handle_t al,
1143 omp_allocator_handle_t free_al);
1144extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1145
1146extern void __kmp_init_memkind();
1147extern void __kmp_fini_memkind();
1148extern void __kmp_init_target_mem();
1149
1150/* ------------------------------------------------------------------------ */
1151
1152#if ENABLE_LIBOMPTARGET
1153extern void __kmp_init_target_task();
1154#endif
1155
1156/* ------------------------------------------------------------------------ */
1157
1158#define KMP_UINT64_MAX \
1159 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1160
1161#define KMP_MIN_NTH 1
1162
1163#ifndef KMP_MAX_NTH
1164#if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1165#define KMP_MAX_NTH PTHREAD_THREADS_MAX
1166#else
1167#ifdef __ve__
1168// VE's pthread supports only up to 64 threads per a VE process.
1169// Please check p. 14 of following documentation for more details.
1170// https://sxauroratsubasa.sakura.ne.jp/documents/veos/en/VEOS_high_level_design.pdf
1171#define KMP_MAX_NTH 64
1172#else
1173#define KMP_MAX_NTH INT_MAX
1174#endif
1175#endif
1176#endif /* KMP_MAX_NTH */
1177
1178#ifdef PTHREAD_STACK_MIN
1179#define KMP_MIN_STKSIZE ((size_t)PTHREAD_STACK_MIN)
1180#else
1181#define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1182#endif
1183
1184#if KMP_OS_AIX && KMP_ARCH_PPC
1185#define KMP_MAX_STKSIZE 0x10000000 /* 256Mb max size on 32-bit AIX */
1186#else
1187#define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1188#endif
1189
1190#if KMP_ARCH_X86
1191#define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1192#elif KMP_ARCH_X86_64
1193#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1194#define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1195#elif KMP_ARCH_VE
1196// Minimum stack size for pthread for VE is 4MB.
1197// https://www.hpc.nec/documents/veos/en/glibc/Difference_Points_glibc.htm
1198#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1199#elif KMP_OS_AIX
1200// The default stack size for worker threads on AIX is 4MB.
1201#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1202#else
1203#define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1204#endif
1205
1206#define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1207#define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1208#define KMP_MAX_MALLOC_POOL_INCR \
1209 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1210
1211#define KMP_MIN_STKOFFSET (0)
1212#define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1213#if KMP_OS_DARWIN
1214#define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1215#else
1216#define KMP_DEFAULT_STKOFFSET CACHE_LINE
1217#endif
1218
1219#define KMP_MIN_STKPADDING (0)
1220#define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1221
1222#define KMP_BLOCKTIME_MULTIPLIER \
1223 (1000000) /* number of blocktime units per second */
1224#define KMP_MIN_BLOCKTIME (0)
1225#define KMP_MAX_BLOCKTIME \
1226 (INT_MAX) /* Must be this for "infinite" setting the work */
1227
1228/* __kmp_blocktime is in microseconds */
1229#define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200000))
1230
1231#if KMP_USE_MONITOR
1232#define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1233#define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1234#define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1235
1236/* Calculate new number of monitor wakeups for a specific block time based on
1237 previous monitor_wakeups. Only allow increasing number of wakeups */
1238#define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1239 (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \
1240 : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \
1241 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1242 ? (monitor_wakeups) \
1243 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1244
1245/* Calculate number of intervals for a specific block time based on
1246 monitor_wakeups */
1247#define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1248 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1249 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1250#else
1251#define KMP_BLOCKTIME(team, tid) \
1252 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1253#if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1254// HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1255extern kmp_uint64 __kmp_ticks_per_msec;
1256extern kmp_uint64 __kmp_ticks_per_usec;
1257#if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1258#define KMP_NOW() ((kmp_uint64)_rdtsc())
1259#else
1260#define KMP_NOW() __kmp_hardware_timestamp()
1261#endif
1262#define KMP_BLOCKTIME_INTERVAL(team, tid) \
1263 ((kmp_uint64)KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_usec)
1264#define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1265#else
1266// System time is retrieved sporadically while blocking.
1267extern kmp_uint64 __kmp_now_nsec();
1268#define KMP_NOW() __kmp_now_nsec()
1269#define KMP_BLOCKTIME_INTERVAL(team, tid) \
1270 ((kmp_uint64)KMP_BLOCKTIME(team, tid) * (kmp_uint64)KMP_NSEC_PER_USEC)
1271#define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1272#endif
1273#endif // KMP_USE_MONITOR
1274
1275#define KMP_MIN_STATSCOLS 40
1276#define KMP_MAX_STATSCOLS 4096
1277#define KMP_DEFAULT_STATSCOLS 80
1278
1279#define KMP_MIN_INTERVAL 0
1280#define KMP_MAX_INTERVAL (INT_MAX - 1)
1281#define KMP_DEFAULT_INTERVAL 0
1282
1283#define KMP_MIN_CHUNK 1
1284#define KMP_MAX_CHUNK (INT_MAX - 1)
1285#define KMP_DEFAULT_CHUNK 1
1286
1287#define KMP_MIN_DISP_NUM_BUFF 1
1288#define KMP_DFLT_DISP_NUM_BUFF 7
1289#define KMP_MAX_DISP_NUM_BUFF 4096
1290
1291#define KMP_MAX_ORDERED 8
1292
1293#define KMP_MAX_FIELDS 32
1294
1295#define KMP_MAX_BRANCH_BITS 31
1296
1297#define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1298
1299#define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1300
1301#define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1302
1303/* Minimum number of threads before switch to TLS gtid (experimentally
1304 determined) */
1305/* josh TODO: what about OS X* tuning? */
1306#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1307#define KMP_TLS_GTID_MIN 5
1308#else
1309#define KMP_TLS_GTID_MIN INT_MAX
1310#endif
1311
1312#define KMP_MASTER_TID(tid) (0 == (tid))
1313#define KMP_WORKER_TID(tid) (0 != (tid))
1314
1315#define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1316#define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1317#define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1318
1319#ifndef TRUE
1320#define FALSE 0
1321#define TRUE (!FALSE)
1322#endif
1323
1324/* NOTE: all of the following constants must be even */
1325
1326#if KMP_OS_WINDOWS
1327#define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1328#define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1329#elif KMP_OS_LINUX
1330#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1331#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1332#elif KMP_OS_DARWIN
1333/* TODO: tune for KMP_OS_DARWIN */
1334#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1335#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1336#elif KMP_OS_DRAGONFLY
1337/* TODO: tune for KMP_OS_DRAGONFLY */
1338#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1339#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1340#elif KMP_OS_FREEBSD
1341/* TODO: tune for KMP_OS_FREEBSD */
1342#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1343#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1344#elif KMP_OS_NETBSD
1345/* TODO: tune for KMP_OS_NETBSD */
1346#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1347#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1348#elif KMP_OS_OPENBSD
1349/* TODO: tune for KMP_OS_OPENBSD */
1350#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1351#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1352#elif KMP_OS_HURD
1353/* TODO: tune for KMP_OS_HURD */
1354#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1355#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1356#elif KMP_OS_SOLARIS
1357/* TODO: tune for KMP_OS_SOLARIS */
1358#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1359#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1360#elif KMP_OS_WASI
1361/* TODO: tune for KMP_OS_WASI */
1362#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1363#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1364#elif KMP_OS_AIX
1365/* TODO: tune for KMP_OS_AIX */
1366#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1367#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1368#endif
1369
1370#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1371typedef struct kmp_cpuid {
1372 kmp_uint32 eax;
1373 kmp_uint32 ebx;
1374 kmp_uint32 ecx;
1375 kmp_uint32 edx;
1376} kmp_cpuid_t;
1377
1378typedef struct kmp_cpuinfo_flags_t {
1379 unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1380 unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1381 unsigned hybrid : 1;
1382 unsigned reserved : 29; // Ensure size of 32 bits
1383} kmp_cpuinfo_flags_t;
1384
1385typedef struct kmp_cpuinfo {
1386 int initialized; // If 0, other fields are not initialized.
1387 int signature; // CPUID(1).EAX
1388 int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1389 int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1390 // Model << 4 ) + Model)
1391 int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1392 kmp_cpuinfo_flags_t flags;
1393 int apic_id;
1394 int physical_id;
1395 int logical_id;
1396 kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1397 char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1398} kmp_cpuinfo_t;
1399
1400extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1401
1402#if KMP_OS_UNIX
1403// subleaf is only needed for cache and topology discovery and can be set to
1404// zero in most cases
1405static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1406 __asm__ __volatile__("cpuid"
1407 : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1408 : "a"(leaf), "c"(subleaf));
1409}
1410// Load p into FPU control word
1411static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1412 __asm__ __volatile__("fldcw %0" : : "m"(*p));
1413}
1414// Store FPU control word into p
1415static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1416 __asm__ __volatile__("fstcw %0" : "=m"(*p));
1417}
1418static inline void __kmp_clear_x87_fpu_status_word() {
1419#if KMP_MIC
1420 // 32-bit protected mode x87 FPU state
1421 struct x87_fpu_state {
1422 unsigned cw;
1423 unsigned sw;
1424 unsigned tw;
1425 unsigned fip;
1426 unsigned fips;
1427 unsigned fdp;
1428 unsigned fds;
1429 };
1430 struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1431 __asm__ __volatile__("fstenv %0\n\t" // store FP env
1432 "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1433 "fldenv %0\n\t" // load FP env back
1434 : "+m"(fpu_state), "+m"(fpu_state.sw));
1435#else
1436 __asm__ __volatile__("fnclex");
1437#endif // KMP_MIC
1438}
1439#if __SSE__
1440static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1441static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1442#else
1443static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1444static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1445#endif
1446#else
1447// Windows still has these as external functions in assembly file
1448extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1449extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1450extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1451extern void __kmp_clear_x87_fpu_status_word();
1452static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1453static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1454#endif // KMP_OS_UNIX
1455
1456#define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1457
1458// User-level Monitor/Mwait
1459#if KMP_HAVE_UMWAIT
1460// We always try for UMWAIT first
1461#if KMP_HAVE_WAITPKG_INTRINSICS
1462#if KMP_HAVE_IMMINTRIN_H
1463#include <immintrin.h>
1464#elif KMP_HAVE_INTRIN_H
1465#include <intrin.h>
1466#endif
1467#endif // KMP_HAVE_WAITPKG_INTRINSICS
1468
1469KMP_ATTRIBUTE_TARGET_WAITPKG
1470static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1471#if !KMP_HAVE_WAITPKG_INTRINSICS
1472 uint32_t timeHi = uint32_t(counter >> 32);
1473 uint32_t timeLo = uint32_t(counter & 0xffffffff);
1474 char flag;
1475 __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1476 "setb %0"
1477 // The "=q" restraint means any register accessible as rl
1478 // in 32-bit mode: a, b, c, and d;
1479 // in 64-bit mode: any integer register
1480 : "=q"(flag)
1481 : "a"(timeLo), "d"(timeHi), "c"(hint)
1482 :);
1483 return flag;
1484#else
1485 return _tpause(hint, counter);
1486#endif
1487}
1488KMP_ATTRIBUTE_TARGET_WAITPKG
1489static inline void __kmp_umonitor(void *cacheline) {
1490#if !KMP_HAVE_WAITPKG_INTRINSICS
1491 __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1492 :
1493 : "a"(cacheline)
1494 :);
1495#else
1496 _umonitor(cacheline);
1497#endif
1498}
1499KMP_ATTRIBUTE_TARGET_WAITPKG
1500static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1501#if !KMP_HAVE_WAITPKG_INTRINSICS
1502 uint32_t timeHi = uint32_t(counter >> 32);
1503 uint32_t timeLo = uint32_t(counter & 0xffffffff);
1504 char flag;
1505 __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1506 "setb %0"
1507 // The "=q" restraint means any register accessible as rl
1508 // in 32-bit mode: a, b, c, and d;
1509 // in 64-bit mode: any integer register
1510 : "=q"(flag)
1511 : "a"(timeLo), "d"(timeHi), "c"(hint)
1512 :);
1513 return flag;
1514#else
1515 return _umwait(hint, counter);
1516#endif
1517}
1518#elif KMP_HAVE_MWAIT
1519#if KMP_OS_UNIX
1520#include <pmmintrin.h>
1521#else
1522#include <intrin.h>
1523#endif
1524#if KMP_OS_UNIX
1525__attribute__((target("sse3")))
1526#endif
1527static inline void
1528__kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1529 _mm_monitor(cacheline, extensions, hints);
1530}
1531#if KMP_OS_UNIX
1532__attribute__((target("sse3")))
1533#endif
1534static inline void
1535__kmp_mm_mwait(unsigned extensions, unsigned hints) {
1536 _mm_mwait(extensions, hints);
1537}
1538#endif // KMP_HAVE_UMWAIT
1539
1540#if KMP_ARCH_X86
1541extern void __kmp_x86_pause(void);
1542#elif KMP_MIC
1543// Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1544// regression after removal of extra PAUSE from spin loops. Changing
1545// the delay from 100 to 300 showed even better performance than double PAUSE
1546// on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1547static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1548#else
1549static inline void __kmp_x86_pause(void) { _mm_pause(); }
1550#endif
1551#define KMP_CPU_PAUSE() __kmp_x86_pause()
1552#elif KMP_ARCH_PPC64
1553#define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1554#define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1555#define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1556#define KMP_CPU_PAUSE() \
1557 do { \
1558 KMP_PPC64_PRI_LOW(); \
1559 KMP_PPC64_PRI_MED(); \
1560 KMP_PPC64_PRI_LOC_MB(); \
1561 } while (0)
1562#else
1563#define KMP_CPU_PAUSE() /* nothing to do */
1564#endif
1565
1566#define KMP_INIT_YIELD(count) \
1567 { (count) = __kmp_yield_init; }
1568
1569#define KMP_INIT_BACKOFF(time) \
1570 { (time) = __kmp_pause_init; }
1571
1572#define KMP_OVERSUBSCRIBED \
1573 (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1574
1575#define KMP_TRY_YIELD \
1576 ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1577
1578#define KMP_TRY_YIELD_OVERSUB \
1579 ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1580
1581#define KMP_YIELD(cond) \
1582 { \
1583 KMP_CPU_PAUSE(); \
1584 if ((cond) && (KMP_TRY_YIELD)) \
1585 __kmp_yield(); \
1586 }
1587
1588#define KMP_YIELD_OVERSUB() \
1589 { \
1590 KMP_CPU_PAUSE(); \
1591 if ((KMP_TRY_YIELD_OVERSUB)) \
1592 __kmp_yield(); \
1593 }
1594
1595// Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1596// there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1597#define KMP_YIELD_SPIN(count) \
1598 { \
1599 KMP_CPU_PAUSE(); \
1600 if (KMP_TRY_YIELD) { \
1601 (count) -= 2; \
1602 if (!(count)) { \
1603 __kmp_yield(); \
1604 (count) = __kmp_yield_next; \
1605 } \
1606 } \
1607 }
1608
1609// If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1610// (C0.2) state, which improves performance of other SMT threads on the same
1611// core, otherwise, use the fast (C0.1) default state, or whatever the user has
1612// requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1613// available, fall back to the regular CPU pause and yield combination.
1614#if KMP_HAVE_UMWAIT
1615#define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF)
1616#define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1617 { \
1618 if (__kmp_tpause_enabled) { \
1619 if (KMP_OVERSUBSCRIBED) { \
1620 __kmp_tpause(0, (time)); \
1621 } else { \
1622 __kmp_tpause(__kmp_tpause_hint, (time)); \
1623 } \
1624 (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK; \
1625 } else { \
1626 KMP_CPU_PAUSE(); \
1627 if ((KMP_TRY_YIELD_OVERSUB)) { \
1628 __kmp_yield(); \
1629 } else if (__kmp_use_yield == 1) { \
1630 (count) -= 2; \
1631 if (!(count)) { \
1632 __kmp_yield(); \
1633 (count) = __kmp_yield_next; \
1634 } \
1635 } \
1636 } \
1637 }
1638#else
1639#define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1640 { \
1641 KMP_CPU_PAUSE(); \
1642 if ((KMP_TRY_YIELD_OVERSUB)) \
1643 __kmp_yield(); \
1644 else if (__kmp_use_yield == 1) { \
1645 (count) -= 2; \
1646 if (!(count)) { \
1647 __kmp_yield(); \
1648 (count) = __kmp_yield_next; \
1649 } \
1650 } \
1651 }
1652#endif // KMP_HAVE_UMWAIT
1653
1654/* ------------------------------------------------------------------------ */
1655/* Support datatypes for the orphaned construct nesting checks. */
1656/* ------------------------------------------------------------------------ */
1657
1658/* When adding to this enum, add its corresponding string in cons_text_c[]
1659 * array in kmp_error.cpp */
1660enum cons_type {
1661 ct_none,
1662 ct_parallel,
1663 ct_pdo,
1664 ct_pdo_ordered,
1665 ct_psections,
1666 ct_psingle,
1667 ct_critical,
1668 ct_ordered_in_parallel,
1669 ct_ordered_in_pdo,
1670 ct_master,
1671 ct_reduce,
1672 ct_barrier,
1673 ct_masked
1674};
1675
1676#define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1677
1678struct cons_data {
1679 ident_t const *ident;
1680 enum cons_type type;
1681 int prev;
1682 kmp_user_lock_p
1683 name; /* address exclusively for critical section name comparison */
1684};
1685
1686struct cons_header {
1687 int p_top, w_top, s_top;
1688 int stack_size, stack_top;
1689 struct cons_data *stack_data;
1690};
1691
1692struct kmp_region_info {
1693 char *text;
1694 int offset[KMP_MAX_FIELDS];
1695 int length[KMP_MAX_FIELDS];
1696};
1697
1698/* ---------------------------------------------------------------------- */
1699/* ---------------------------------------------------------------------- */
1700
1701#if KMP_OS_WINDOWS
1702typedef HANDLE kmp_thread_t;
1703typedef DWORD kmp_key_t;
1704#endif /* KMP_OS_WINDOWS */
1705
1706#if KMP_OS_UNIX
1707typedef pthread_t kmp_thread_t;
1708typedef pthread_key_t kmp_key_t;
1709#endif
1710
1711extern kmp_key_t __kmp_gtid_threadprivate_key;
1712
1713typedef struct kmp_sys_info {
1714 long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1715 long minflt; /* the number of page faults serviced without any I/O */
1716 long majflt; /* the number of page faults serviced that required I/O */
1717 long nswap; /* the number of times a process was "swapped" out of memory */
1718 long inblock; /* the number of times the file system had to perform input */
1719 long oublock; /* the number of times the file system had to perform output */
1720 long nvcsw; /* the number of times a context switch was voluntarily */
1721 long nivcsw; /* the number of times a context switch was forced */
1722} kmp_sys_info_t;
1723
1724#if USE_ITT_BUILD
1725// We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1726// required type here. Later we will check the type meets requirements.
1727typedef int kmp_itt_mark_t;
1728#define KMP_ITT_DEBUG 0
1729#endif /* USE_ITT_BUILD */
1730
1731typedef kmp_int32 kmp_critical_name[8];
1732
1742typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1743typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1744 ...);
1745
1750/* ---------------------------------------------------------------------------
1751 */
1752/* Threadprivate initialization/finalization function declarations */
1753
1754/* for non-array objects: __kmpc_threadprivate_register() */
1755
1760typedef void *(*kmpc_ctor)(void *);
1761
1766typedef void (*kmpc_dtor)(
1767 void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1768 compiler */
1773typedef void *(*kmpc_cctor)(void *, void *);
1774
1775/* for array objects: __kmpc_threadprivate_register_vec() */
1776/* First arg: "this" pointer */
1777/* Last arg: number of array elements */
1783typedef void *(*kmpc_ctor_vec)(void *, size_t);
1789typedef void (*kmpc_dtor_vec)(void *, size_t);
1795typedef void *(*kmpc_cctor_vec)(void *, void *,
1796 size_t); /* function unused by compiler */
1797
1802/* keeps tracked of threadprivate cache allocations for cleanup later */
1803typedef struct kmp_cached_addr {
1804 void **addr; /* address of allocated cache */
1805 void ***compiler_cache; /* pointer to compiler's cache */
1806 void *data; /* pointer to global data */
1807 struct kmp_cached_addr *next; /* pointer to next cached address */
1808} kmp_cached_addr_t;
1809
1810struct private_data {
1811 struct private_data *next; /* The next descriptor in the list */
1812 void *data; /* The data buffer for this descriptor */
1813 int more; /* The repeat count for this descriptor */
1814 size_t size; /* The data size for this descriptor */
1815};
1816
1817struct private_common {
1818 struct private_common *next;
1819 struct private_common *link;
1820 void *gbl_addr;
1821 void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1822 size_t cmn_size;
1823};
1824
1825struct shared_common {
1826 struct shared_common *next;
1827 struct private_data *pod_init;
1828 void *obj_init;
1829 void *gbl_addr;
1830 union {
1831 kmpc_ctor ctor;
1832 kmpc_ctor_vec ctorv;
1833 } ct;
1834 union {
1835 kmpc_cctor cctor;
1836 kmpc_cctor_vec cctorv;
1837 } cct;
1838 union {
1839 kmpc_dtor dtor;
1840 kmpc_dtor_vec dtorv;
1841 } dt;
1842 size_t vec_len;
1843 int is_vec;
1844 size_t cmn_size;
1845};
1846
1847#define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1848#define KMP_HASH_TABLE_SIZE \
1849 (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1850#define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1851#define KMP_HASH(x) \
1852 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1853
1854struct common_table {
1855 struct private_common *data[KMP_HASH_TABLE_SIZE];
1856};
1857
1858struct shared_table {
1859 struct shared_common *data[KMP_HASH_TABLE_SIZE];
1860};
1861
1862/* ------------------------------------------------------------------------ */
1863
1864#if KMP_USE_HIER_SCHED
1865// Shared barrier data that exists inside a single unit of the scheduling
1866// hierarchy
1867typedef struct kmp_hier_private_bdata_t {
1868 kmp_int32 num_active;
1869 kmp_uint64 index;
1870 kmp_uint64 wait_val[2];
1871} kmp_hier_private_bdata_t;
1872#endif
1873
1874typedef struct kmp_sched_flags {
1875 unsigned ordered : 1;
1876 unsigned nomerge : 1;
1877 unsigned contains_last : 1;
1878 unsigned use_hier : 1; // Used in KMP_USE_HIER_SCHED code
1879 unsigned use_hybrid : 1; // Used in KMP_WEIGHTED_ITERATIONS_SUPPORTED code
1880 unsigned unused : 27;
1881} kmp_sched_flags_t;
1882
1883KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1884
1885#if KMP_STATIC_STEAL_ENABLED
1886typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1887 kmp_int32 count;
1888 kmp_int32 ub;
1889 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1890 kmp_int32 lb;
1891 kmp_int32 st;
1892 kmp_int32 tc;
1893 kmp_lock_t *steal_lock; // lock used for chunk stealing
1894
1895 kmp_uint32 ordered_lower;
1896 kmp_uint32 ordered_upper;
1897
1898 // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1899 // a) parm3 is properly aligned and
1900 // b) all parm1-4 are on the same cache line.
1901 // Because of parm1-4 are used together, performance seems to be better
1902 // if they are on the same cache line (not measured though).
1903
1904 struct KMP_ALIGN(32) {
1905 kmp_int32 parm1;
1906 kmp_int32 parm2;
1907 kmp_int32 parm3;
1908 kmp_int32 parm4;
1909 };
1910
1911#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
1912 kmp_uint32 pchunks;
1913 kmp_uint32 num_procs_with_pcore;
1914 kmp_int32 first_thread_with_ecore;
1915#endif
1916#if KMP_OS_WINDOWS
1917 kmp_int32 last_upper;
1918#endif /* KMP_OS_WINDOWS */
1919} dispatch_private_info32_t;
1920
1921#if CACHE_LINE <= 128
1922KMP_BUILD_ASSERT(sizeof(dispatch_private_info32_t) <= 128);
1923#endif
1924
1925typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1926 kmp_int64 count; // current chunk number for static & static-steal scheduling
1927 kmp_int64 ub; /* upper-bound */
1928 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1929 kmp_int64 lb; /* lower-bound */
1930 kmp_int64 st; /* stride */
1931 kmp_int64 tc; /* trip count (number of iterations) */
1932 kmp_lock_t *steal_lock; // lock used for chunk stealing
1933
1934 kmp_uint64 ordered_lower;
1935 kmp_uint64 ordered_upper;
1936 /* parm[1-4] are used in different ways by different scheduling algorithms */
1937
1938 // KMP_ALIGN(32) ensures ( if the KMP_ALIGN macro is turned on )
1939 // a) parm3 is properly aligned and
1940 // b) all parm1-4 are in the same cache line.
1941 // Because of parm1-4 are used together, performance seems to be better
1942 // if they are in the same line (not measured though).
1943 struct KMP_ALIGN(32) {
1944 kmp_int64 parm1;
1945 kmp_int64 parm2;
1946 kmp_int64 parm3;
1947 kmp_int64 parm4;
1948 };
1949
1950#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
1951 kmp_uint64 pchunks;
1952 kmp_uint64 num_procs_with_pcore;
1953 kmp_int64 first_thread_with_ecore;
1954#endif
1955
1956#if KMP_OS_WINDOWS
1957 kmp_int64 last_upper;
1958#endif /* KMP_OS_WINDOWS */
1959} dispatch_private_info64_t;
1960
1961#if CACHE_LINE <= 128
1962KMP_BUILD_ASSERT(sizeof(dispatch_private_info64_t) <= 128);
1963#endif
1964
1965#else /* KMP_STATIC_STEAL_ENABLED */
1966typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1967 kmp_int32 lb;
1968 kmp_int32 ub;
1969 kmp_int32 st;
1970 kmp_int32 tc;
1971
1972 kmp_int32 parm1;
1973 kmp_int32 parm2;
1974 kmp_int32 parm3;
1975 kmp_int32 parm4;
1976
1977 kmp_int32 count;
1978
1979 kmp_uint32 ordered_lower;
1980 kmp_uint32 ordered_upper;
1981#if KMP_OS_WINDOWS
1982 kmp_int32 last_upper;
1983#endif /* KMP_OS_WINDOWS */
1984} dispatch_private_info32_t;
1985
1986typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1987 kmp_int64 lb; /* lower-bound */
1988 kmp_int64 ub; /* upper-bound */
1989 kmp_int64 st; /* stride */
1990 kmp_int64 tc; /* trip count (number of iterations) */
1991
1992 /* parm[1-4] are used in different ways by different scheduling algorithms */
1993 kmp_int64 parm1;
1994 kmp_int64 parm2;
1995 kmp_int64 parm3;
1996 kmp_int64 parm4;
1997
1998 kmp_int64 count; /* current chunk number for static scheduling */
1999
2000 kmp_uint64 ordered_lower;
2001 kmp_uint64 ordered_upper;
2002#if KMP_OS_WINDOWS
2003 kmp_int64 last_upper;
2004#endif /* KMP_OS_WINDOWS */
2005} dispatch_private_info64_t;
2006#endif /* KMP_STATIC_STEAL_ENABLED */
2007
2008typedef struct KMP_ALIGN_CACHE dispatch_private_info {
2009 union private_info {
2010 dispatch_private_info32_t p32;
2011 dispatch_private_info64_t p64;
2012 } u;
2013 enum sched_type schedule; /* scheduling algorithm */
2014 kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
2015 std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
2016 kmp_int32 ordered_bumped;
2017 // Stack of buffers for nest of serial regions
2018 struct dispatch_private_info *next;
2019 kmp_int32 type_size; /* the size of types in private_info */
2020#if KMP_USE_HIER_SCHED
2021 kmp_int32 hier_id;
2022 void *parent; /* hierarchical scheduling parent pointer */
2023#endif
2024 enum cons_type pushed_ws;
2025} dispatch_private_info_t;
2026
2027typedef struct dispatch_shared_info32 {
2028 /* chunk index under dynamic, number of idle threads under static-steal;
2029 iteration index otherwise */
2030 volatile kmp_uint32 iteration;
2031 volatile kmp_int32 num_done;
2032 volatile kmp_uint32 ordered_iteration;
2033 // Dummy to retain the structure size after making ordered_iteration scalar
2034 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
2035} dispatch_shared_info32_t;
2036
2037typedef struct dispatch_shared_info64 {
2038 /* chunk index under dynamic, number of idle threads under static-steal;
2039 iteration index otherwise */
2040 volatile kmp_uint64 iteration;
2041 volatile kmp_int64 num_done;
2042 volatile kmp_uint64 ordered_iteration;
2043 // Dummy to retain the structure size after making ordered_iteration scalar
2044 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
2045} dispatch_shared_info64_t;
2046
2047typedef struct dispatch_shared_info {
2048 union shared_info {
2049 dispatch_shared_info32_t s32;
2050 dispatch_shared_info64_t s64;
2051 } u;
2052 volatile kmp_uint32 buffer_index;
2053 volatile kmp_int32 doacross_buf_idx; // teamwise index
2054 volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
2055 kmp_int32 doacross_num_done; // count finished threads
2056#if KMP_USE_HIER_SCHED
2057 void *hier;
2058#endif
2059#if KMP_USE_HWLOC
2060 // When linking with libhwloc, the ORDERED EPCC test slows down on big
2061 // machines (> 48 cores). Performance analysis showed that a cache thrash
2062 // was occurring and this padding helps alleviate the problem.
2063 char padding[64];
2064#endif
2065} dispatch_shared_info_t;
2066
2067typedef struct kmp_disp {
2068 /* Vector for ORDERED SECTION */
2069 void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
2070 /* Vector for END ORDERED SECTION */
2071 void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
2072
2073 dispatch_shared_info_t *th_dispatch_sh_current;
2074 dispatch_private_info_t *th_dispatch_pr_current;
2075
2076 dispatch_private_info_t *th_disp_buffer;
2077 kmp_uint32 th_disp_index;
2078 kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
2079 volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
2080 kmp_int64 *th_doacross_info; // info on loop bounds
2081#if KMP_USE_INTERNODE_ALIGNMENT
2082 char more_padding[INTERNODE_CACHE_LINE];
2083#endif
2084} kmp_disp_t;
2085
2086/* ------------------------------------------------------------------------ */
2087/* Barrier stuff */
2088
2089/* constants for barrier state update */
2090#define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
2091#define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
2092#define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
2093#define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
2094
2095#define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
2096#define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
2097#define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
2098
2099#if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
2100#error "Barrier sleep bit must be smaller than barrier bump bit"
2101#endif
2102#if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
2103#error "Barrier unused bit must be smaller than barrier bump bit"
2104#endif
2105
2106// Constants for release barrier wait state: currently, hierarchical only
2107#define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
2108#define KMP_BARRIER_OWN_FLAG \
2109 1 // Normal state; worker waiting on own b_go flag in release
2110#define KMP_BARRIER_PARENT_FLAG \
2111 2 // Special state; worker waiting on parent's b_go flag in release
2112#define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
2113 3 // Special state; tells worker to shift from parent to own b_go
2114#define KMP_BARRIER_SWITCHING \
2115 4 // Special state; worker resets appropriate flag on wake-up
2116
2117#define KMP_NOT_SAFE_TO_REAP \
2118 0 // Thread th_reap_state: not safe to reap (tasking)
2119#define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
2120
2121// The flag_type describes the storage used for the flag.
2122enum flag_type {
2123 flag32,
2124 flag64,
2125 atomic_flag64,
2126 flag_oncore,
2127 flag_unset
2128};
2129
2130enum barrier_type {
2131 bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
2132 barriers if enabled) */
2133 bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
2134#if KMP_FAST_REDUCTION_BARRIER
2135 bs_reduction_barrier, /* 2, All barriers that are used in reduction */
2136#endif // KMP_FAST_REDUCTION_BARRIER
2137 bs_last_barrier /* Just a placeholder to mark the end */
2138};
2139
2140// to work with reduction barriers just like with plain barriers
2141#if !KMP_FAST_REDUCTION_BARRIER
2142#define bs_reduction_barrier bs_plain_barrier
2143#endif // KMP_FAST_REDUCTION_BARRIER
2144
2145typedef enum kmp_bar_pat { /* Barrier communication patterns */
2146 bp_linear_bar =
2147 0, /* Single level (degenerate) tree */
2148 bp_tree_bar =
2149 1, /* Balanced tree with branching factor 2^n */
2150 bp_hyper_bar = 2, /* Hypercube-embedded tree with min
2151 branching factor 2^n */
2152 bp_hierarchical_bar = 3, /* Machine hierarchy tree */
2153 bp_dist_bar = 4, /* Distributed barrier */
2154 bp_last_bar /* Placeholder to mark the end */
2155} kmp_bar_pat_e;
2156
2157#define KMP_BARRIER_ICV_PUSH 1
2158
2159/* Record for holding the values of the internal controls stack records */
2160typedef struct kmp_internal_control {
2161 int serial_nesting_level; /* corresponds to the value of the
2162 th_team_serialized field */
2163 kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
2164 thread) */
2165 kmp_int8
2166 bt_set; /* internal control for whether blocktime is explicitly set */
2167 int blocktime; /* internal control for blocktime */
2168#if KMP_USE_MONITOR
2169 int bt_intervals; /* internal control for blocktime intervals */
2170#endif
2171 int nproc; /* internal control for #threads for next parallel region (per
2172 thread) */
2173 int thread_limit; /* internal control for thread-limit-var */
2174 int task_thread_limit; /* internal control for thread-limit-var of a task*/
2175 int max_active_levels; /* internal control for max_active_levels */
2176 kmp_r_sched_t
2177 sched; /* internal control for runtime schedule {sched,chunk} pair */
2178 kmp_proc_bind_t proc_bind; /* internal control for affinity */
2179 kmp_int32 default_device; /* internal control for default device */
2180 struct kmp_internal_control *next;
2181} kmp_internal_control_t;
2182
2183static inline void copy_icvs(kmp_internal_control_t *dst,
2184 kmp_internal_control_t *src) {
2185 *dst = *src;
2186}
2187
2188/* Thread barrier needs volatile barrier fields */
2189typedef struct KMP_ALIGN_CACHE kmp_bstate {
2190 // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2191 // uses of it). It is not explicitly aligned below, because we *don't* want
2192 // it to be padded -- instead, we fit b_go into the same cache line with
2193 // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2194 kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2195 // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2196 // same NGO store
2197 volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2198 KMP_ALIGN_CACHE volatile kmp_uint64
2199 b_arrived; // STATE => task reached synch point.
2200 kmp_uint32 *skip_per_level;
2201 kmp_uint32 my_level;
2202 kmp_int32 parent_tid;
2203 kmp_int32 old_tid;
2204 kmp_uint32 depth;
2205 struct kmp_bstate *parent_bar;
2206 kmp_team_t *team;
2207 kmp_uint64 leaf_state;
2208 kmp_uint32 nproc;
2209 kmp_uint8 base_leaf_kids;
2210 kmp_uint8 leaf_kids;
2211 kmp_uint8 offset;
2212 kmp_uint8 wait_flag;
2213 kmp_uint8 use_oncore_barrier;
2214#if USE_DEBUGGER
2215 // The following field is intended for the debugger solely. Only the worker
2216 // thread itself accesses this field: the worker increases it by 1 when it
2217 // arrives to a barrier.
2218 KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2219#endif /* USE_DEBUGGER */
2220} kmp_bstate_t;
2221
2222union KMP_ALIGN_CACHE kmp_barrier_union {
2223 double b_align; /* use worst case alignment */
2224 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2225 kmp_bstate_t bb;
2226};
2227
2228typedef union kmp_barrier_union kmp_balign_t;
2229
2230/* Team barrier needs only non-volatile arrived counter */
2231union KMP_ALIGN_CACHE kmp_barrier_team_union {
2232 double b_align; /* use worst case alignment */
2233 char b_pad[CACHE_LINE];
2234 struct {
2235 kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2236#if USE_DEBUGGER
2237 // The following two fields are indended for the debugger solely. Only
2238 // primary thread of the team accesses these fields: the first one is
2239 // increased by 1 when the primary thread arrives to a barrier, the second
2240 // one is increased by one when all the threads arrived.
2241 kmp_uint b_master_arrived;
2242 kmp_uint b_team_arrived;
2243#endif
2244 };
2245};
2246
2247typedef union kmp_barrier_team_union kmp_balign_team_t;
2248
2249/* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2250 threads when a condition changes. This is to workaround an NPTL bug where
2251 padding was added to pthread_cond_t which caused the initialization routine
2252 to write outside of the structure if compiled on pre-NPTL threads. */
2253#if KMP_OS_WINDOWS
2254typedef struct kmp_win32_mutex {
2255 /* The Lock */
2256 CRITICAL_SECTION cs;
2257} kmp_win32_mutex_t;
2258
2259typedef struct kmp_win32_cond {
2260 /* Count of the number of waiters. */
2261 int waiters_count_;
2262
2263 /* Serialize access to <waiters_count_> */
2264 kmp_win32_mutex_t waiters_count_lock_;
2265
2266 /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2267 int release_count_;
2268
2269 /* Keeps track of the current "generation" so that we don't allow */
2270 /* one thread to steal all the "releases" from the broadcast. */
2271 int wait_generation_count_;
2272
2273 /* A manual-reset event that's used to block and release waiting threads. */
2274 HANDLE event_;
2275} kmp_win32_cond_t;
2276#endif
2277
2278#if KMP_OS_UNIX
2279
2280union KMP_ALIGN_CACHE kmp_cond_union {
2281 double c_align;
2282 char c_pad[CACHE_LINE];
2283 pthread_cond_t c_cond;
2284};
2285
2286typedef union kmp_cond_union kmp_cond_align_t;
2287
2288union KMP_ALIGN_CACHE kmp_mutex_union {
2289 double m_align;
2290 char m_pad[CACHE_LINE];
2291 pthread_mutex_t m_mutex;
2292};
2293
2294typedef union kmp_mutex_union kmp_mutex_align_t;
2295
2296#endif /* KMP_OS_UNIX */
2297
2298typedef struct kmp_desc_base {
2299 void *ds_stackbase;
2300 size_t ds_stacksize;
2301 int ds_stackgrow;
2302 kmp_thread_t ds_thread;
2303 volatile int ds_tid;
2304 int ds_gtid;
2305#if KMP_OS_WINDOWS
2306 volatile int ds_alive;
2307 DWORD ds_thread_id;
2308/* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2309 However, debugger support (libomp_db) cannot work with handles, because they
2310 uncomparable. For example, debugger requests info about thread with handle h.
2311 h is valid within debugger process, and meaningless within debugee process.
2312 Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2313 within debugee process, but it is a *new* handle which does *not* equal to
2314 any other handle in debugee... The only way to compare handles is convert
2315 them to system-wide ids. GetThreadId() function is available only in
2316 Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2317 on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2318 thread id by call to GetCurrentThreadId() from within the thread and save it
2319 to let libomp_db identify threads. */
2320#endif /* KMP_OS_WINDOWS */
2321} kmp_desc_base_t;
2322
2323typedef union KMP_ALIGN_CACHE kmp_desc {
2324 double ds_align; /* use worst case alignment */
2325 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2326 kmp_desc_base_t ds;
2327} kmp_desc_t;
2328
2329typedef struct kmp_local {
2330 volatile int this_construct; /* count of single's encountered by thread */
2331 void *reduce_data;
2332#if KMP_USE_BGET
2333 void *bget_data;
2334 void *bget_list;
2335#if !USE_CMP_XCHG_FOR_BGET
2336#ifdef USE_QUEUING_LOCK_FOR_BGET
2337 kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2338#else
2339 kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2340// bootstrap lock so we can use it at library
2341// shutdown.
2342#endif /* USE_LOCK_FOR_BGET */
2343#endif /* ! USE_CMP_XCHG_FOR_BGET */
2344#endif /* KMP_USE_BGET */
2345
2346 PACKED_REDUCTION_METHOD_T
2347 packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2348 __kmpc_end_reduce*() */
2349
2350} kmp_local_t;
2351
2352#define KMP_CHECK_UPDATE(a, b) \
2353 if ((a) != (b)) \
2354 (a) = (b)
2355#define KMP_CHECK_UPDATE_SYNC(a, b) \
2356 if ((a) != (b)) \
2357 TCW_SYNC_PTR((a), (b))
2358
2359#define get__blocktime(xteam, xtid) \
2360 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2361#define get__bt_set(xteam, xtid) \
2362 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2363#if KMP_USE_MONITOR
2364#define get__bt_intervals(xteam, xtid) \
2365 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2366#endif
2367
2368#define get__dynamic_2(xteam, xtid) \
2369 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2370#define get__nproc_2(xteam, xtid) \
2371 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2372#define get__sched_2(xteam, xtid) \
2373 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2374
2375#define set__blocktime_team(xteam, xtid, xval) \
2376 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2377 (xval))
2378
2379#if KMP_USE_MONITOR
2380#define set__bt_intervals_team(xteam, xtid, xval) \
2381 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2382 (xval))
2383#endif
2384
2385#define set__bt_set_team(xteam, xtid, xval) \
2386 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2387
2388#define set__dynamic(xthread, xval) \
2389 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2390#define get__dynamic(xthread) \
2391 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2392
2393#define set__nproc(xthread, xval) \
2394 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2395
2396#define set__thread_limit(xthread, xval) \
2397 (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2398
2399#define set__max_active_levels(xthread, xval) \
2400 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2401
2402#define get__max_active_levels(xthread) \
2403 ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2404
2405#define set__sched(xthread, xval) \
2406 (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2407
2408#define set__proc_bind(xthread, xval) \
2409 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2410#define get__proc_bind(xthread) \
2411 ((xthread)->th.th_current_task->td_icvs.proc_bind)
2412
2413// OpenMP tasking data structures
2414
2415typedef enum kmp_tasking_mode {
2416 tskm_immediate_exec = 0,
2417 tskm_extra_barrier = 1,
2418 tskm_task_teams = 2,
2419 tskm_max = 2
2420} kmp_tasking_mode_t;
2421
2422extern kmp_tasking_mode_t
2423 __kmp_tasking_mode; /* determines how/when to execute tasks */
2424extern int __kmp_task_stealing_constraint;
2425extern int __kmp_enable_task_throttling;
2426extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2427// specified, defaults to 0 otherwise
2428// Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2429extern kmp_int32 __kmp_max_task_priority;
2430// Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2431extern kmp_uint64 __kmp_taskloop_min_tasks;
2432
2433/* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2434 taskdata first */
2435#define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2436#define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2437
2438// The tt_found_tasks flag is a signal to all threads in the team that tasks
2439// were spawned and queued since the previous barrier release.
2440#define KMP_TASKING_ENABLED(task_team) \
2441 (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2449typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2450
2451typedef union kmp_cmplrdata {
2452 kmp_int32 priority;
2453 kmp_routine_entry_t
2454 destructors; /* pointer to function to invoke deconstructors of
2455 firstprivate C++ objects */
2456 /* future data */
2457} kmp_cmplrdata_t;
2458
2459/* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2462typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2463 void *shareds;
2464 kmp_routine_entry_t
2465 routine;
2466 kmp_int32 part_id;
2467 kmp_cmplrdata_t
2468 data1; /* Two known optional additions: destructors and priority */
2469 kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2470 /* future data */
2471 /* private vars */
2472} kmp_task_t;
2473
2478typedef struct kmp_taskgroup {
2479 std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2480 std::atomic<kmp_int32>
2481 cancel_request; // request for cancellation of this taskgroup
2482 struct kmp_taskgroup *parent; // parent taskgroup
2483 // Block of data to perform task reduction
2484 void *reduce_data; // reduction related info
2485 kmp_int32 reduce_num_data; // number of data items to reduce
2486 uintptr_t *gomp_data; // gomp reduction data
2487} kmp_taskgroup_t;
2488
2489// forward declarations
2490typedef union kmp_depnode kmp_depnode_t;
2491typedef struct kmp_depnode_list kmp_depnode_list_t;
2492typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2493
2494// macros for checking dep flag as an integer
2495#define KMP_DEP_IN 0x1
2496#define KMP_DEP_OUT 0x2
2497#define KMP_DEP_INOUT 0x3
2498#define KMP_DEP_MTX 0x4
2499#define KMP_DEP_SET 0x8
2500#define KMP_DEP_ALL 0x80
2501// Compiler sends us this info. Note: some test cases contain an explicit copy
2502// of this struct and should be in sync with any changes here.
2503typedef struct kmp_depend_info {
2504 kmp_intptr_t base_addr;
2505 size_t len;
2506 union {
2507 kmp_uint8 flag; // flag as an unsigned char
2508 struct { // flag as a set of 8 bits
2509#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
2510 /* Same fields as in the #else branch, but in reverse order */
2511 unsigned all : 1;
2512 unsigned unused : 3;
2513 unsigned set : 1;
2514 unsigned mtx : 1;
2515 unsigned out : 1;
2516 unsigned in : 1;
2517#else
2518 unsigned in : 1;
2519 unsigned out : 1;
2520 unsigned mtx : 1;
2521 unsigned set : 1;
2522 unsigned unused : 3;
2523 unsigned all : 1;
2524#endif
2525 } flags;
2526 };
2527} kmp_depend_info_t;
2528
2529// Internal structures to work with task dependencies:
2530struct kmp_depnode_list {
2531 kmp_depnode_t *node;
2532 kmp_depnode_list_t *next;
2533};
2534
2535// Max number of mutexinoutset dependencies per node
2536#define MAX_MTX_DEPS 4
2537
2538typedef struct kmp_base_depnode {
2539 kmp_depnode_list_t *successors; /* used under lock */
2540 kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2541 kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2542 kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2543 kmp_lock_t lock; /* guards shared fields: task, successors */
2544#if KMP_SUPPORT_GRAPH_OUTPUT
2545 kmp_uint32 id;
2546#endif
2547 std::atomic<kmp_int32> npredecessors;
2548 std::atomic<kmp_int32> nrefs;
2549} kmp_base_depnode_t;
2550
2551union KMP_ALIGN_CACHE kmp_depnode {
2552 double dn_align; /* use worst case alignment */
2553 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2554 kmp_base_depnode_t dn;
2555};
2556
2557struct kmp_dephash_entry {
2558 kmp_intptr_t addr;
2559 kmp_depnode_t *last_out;
2560 kmp_depnode_list_t *last_set;
2561 kmp_depnode_list_t *prev_set;
2562 kmp_uint8 last_flag;
2563 kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2564 kmp_dephash_entry_t *next_in_bucket;
2565};
2566
2567typedef struct kmp_dephash {
2568 kmp_dephash_entry_t **buckets;
2569 size_t size;
2570 kmp_depnode_t *last_all;
2571 size_t generation;
2572 kmp_uint32 nelements;
2573 kmp_uint32 nconflicts;
2574} kmp_dephash_t;
2575
2576typedef struct kmp_task_affinity_info {
2577 kmp_intptr_t base_addr;
2578 size_t len;
2579 struct {
2580 bool flag1 : 1;
2581 bool flag2 : 1;
2582 kmp_int32 reserved : 30;
2583 } flags;
2584} kmp_task_affinity_info_t;
2585
2586typedef enum kmp_event_type_t {
2587 KMP_EVENT_UNINITIALIZED = 0,
2588 KMP_EVENT_ALLOW_COMPLETION = 1
2589} kmp_event_type_t;
2590
2591typedef struct {
2592 kmp_event_type_t type;
2593 kmp_tas_lock_t lock;
2594 union {
2595 kmp_task_t *task;
2596 } ed;
2597} kmp_event_t;
2598
2599#if OMPX_TASKGRAPH
2600// Initial number of allocated nodes while recording
2601#define INIT_MAPSIZE 50
2602
2603typedef struct kmp_taskgraph_flags { /*This needs to be exactly 32 bits */
2604 unsigned nowait : 1;
2605 unsigned re_record : 1;
2606 unsigned reserved : 30;
2607} kmp_taskgraph_flags_t;
2608
2610typedef struct kmp_node_info {
2611 kmp_task_t *task; // Pointer to the actual task
2612 kmp_int32 *successors; // Array of the succesors ids
2613 kmp_int32 nsuccessors; // Number of succesors of the node
2614 std::atomic<kmp_int32>
2615 npredecessors_counter; // Number of predessors on the fly
2616 kmp_int32 npredecessors; // Total number of predecessors
2617 kmp_int32 successors_size; // Number of allocated succesors ids
2618 kmp_taskdata_t *parent_task; // Parent implicit task
2619} kmp_node_info_t;
2620
2622typedef enum kmp_tdg_status {
2623 KMP_TDG_NONE = 0,
2624 KMP_TDG_RECORDING = 1,
2625 KMP_TDG_READY = 2
2626} kmp_tdg_status_t;
2627
2629typedef struct kmp_tdg_info {
2630 kmp_int32 tdg_id; // Unique idenfifier of the TDG
2631 kmp_taskgraph_flags_t tdg_flags; // Flags related to a TDG
2632 kmp_int32 map_size; // Number of allocated TDG nodes
2633 kmp_int32 num_roots; // Number of roots tasks int the TDG
2634 kmp_int32 *root_tasks; // Array of tasks identifiers that are roots
2635 kmp_node_info_t *record_map; // Array of TDG nodes
2636 kmp_tdg_status_t tdg_status =
2637 KMP_TDG_NONE; // Status of the TDG (recording, ready...)
2638 std::atomic<kmp_int32> num_tasks; // Number of TDG nodes
2639 kmp_bootstrap_lock_t
2640 graph_lock; // Protect graph attributes when updated via taskloop_recur
2641 // Taskloop reduction related
2642 void *rec_taskred_data; // Data to pass to __kmpc_task_reduction_init or
2643 // __kmpc_taskred_init
2644 kmp_int32 rec_num_taskred;
2645} kmp_tdg_info_t;
2646
2647extern int __kmp_tdg_dot;
2648extern kmp_int32 __kmp_max_tdgs;
2649extern kmp_tdg_info_t **__kmp_global_tdgs;
2650extern kmp_int32 __kmp_curr_tdg_idx;
2651extern kmp_int32 __kmp_successors_size;
2652extern std::atomic<kmp_int32> __kmp_tdg_task_id;
2653extern kmp_int32 __kmp_num_tdg;
2654#endif
2655
2656#ifdef BUILD_TIED_TASK_STACK
2657
2658/* Tied Task stack definitions */
2659typedef struct kmp_stack_block {
2660 kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2661 struct kmp_stack_block *sb_next;
2662 struct kmp_stack_block *sb_prev;
2663} kmp_stack_block_t;
2664
2665typedef struct kmp_task_stack {
2666 kmp_stack_block_t ts_first_block; // first block of stack entries
2667 kmp_taskdata_t **ts_top; // pointer to the top of stack
2668 kmp_int32 ts_entries; // number of entries on the stack
2669} kmp_task_stack_t;
2670
2671#endif // BUILD_TIED_TASK_STACK
2672
2673typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2674#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
2675 /* Same fields as in the #else branch, but in reverse order */
2676#if OMPX_TASKGRAPH
2677 unsigned reserved31 : 6;
2678 unsigned onced : 1;
2679#else
2680 unsigned reserved31 : 7;
2681#endif
2682 unsigned native : 1;
2683 unsigned freed : 1;
2684 unsigned complete : 1;
2685 unsigned executing : 1;
2686 unsigned started : 1;
2687 unsigned team_serial : 1;
2688 unsigned tasking_ser : 1;
2689 unsigned task_serial : 1;
2690 unsigned tasktype : 1;
2691 unsigned reserved : 8;
2692 unsigned hidden_helper : 1;
2693 unsigned detachable : 1;
2694 unsigned priority_specified : 1;
2695 unsigned proxy : 1;
2696 unsigned destructors_thunk : 1;
2697 unsigned merged_if0 : 1;
2698 unsigned final : 1;
2699 unsigned tiedness : 1;
2700#else
2701 /* Compiler flags */ /* Total compiler flags must be 16 bits */
2702 unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2703 unsigned final : 1; /* task is final(1) so execute immediately */
2704 unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2705 code path */
2706 unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2707 invoke destructors from the runtime */
2708 unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2709 context of the RTL) */
2710 unsigned priority_specified : 1; /* set if the compiler provides priority
2711 setting for the task */
2712 unsigned detachable : 1; /* 1 == can detach */
2713 unsigned hidden_helper : 1; /* 1 == hidden helper task */
2714 unsigned reserved : 8; /* reserved for compiler use */
2715
2716 /* Library flags */ /* Total library flags must be 16 bits */
2717 unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2718 unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2719 unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2720 // (1) or may be deferred (0)
2721 unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2722 // (0) [>= 2 threads]
2723 /* If either team_serial or tasking_ser is set, task team may be NULL */
2724 /* Task State Flags: */
2725 unsigned started : 1; /* 1==started, 0==not started */
2726 unsigned executing : 1; /* 1==executing, 0==not executing */
2727 unsigned complete : 1; /* 1==complete, 0==not complete */
2728 unsigned freed : 1; /* 1==freed, 0==allocated */
2729 unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2730#if OMPX_TASKGRAPH
2731 unsigned onced : 1; /* 1==ran once already, 0==never ran, record & replay purposes */
2732 unsigned reserved31 : 6; /* reserved for library use */
2733#else
2734 unsigned reserved31 : 7; /* reserved for library use */
2735#endif
2736#endif
2737} kmp_tasking_flags_t;
2738
2739typedef struct kmp_target_data {
2740 void *async_handle; // libomptarget async handle for task completion query
2741} kmp_target_data_t;
2742
2743struct kmp_taskdata { /* aligned during dynamic allocation */
2744 kmp_int32 td_task_id; /* id, assigned by debugger */
2745 kmp_tasking_flags_t td_flags; /* task flags */
2746 kmp_team_t *td_team; /* team for this task */
2747 kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2748 /* Currently not used except for perhaps IDB */
2749 kmp_taskdata_t *td_parent; /* parent task */
2750 kmp_int32 td_level; /* task nesting level */
2751 std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2752 ident_t *td_ident; /* task identifier */
2753 // Taskwait data.
2754 ident_t *td_taskwait_ident;
2755 kmp_uint32 td_taskwait_counter;
2756 kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2757 KMP_ALIGN_CACHE kmp_internal_control_t
2758 td_icvs; /* Internal control variables for the task */
2759 KMP_ALIGN_CACHE std::atomic<kmp_int32>
2760 td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2761 deallocated */
2762 std::atomic<kmp_int32>
2763 td_incomplete_child_tasks; /* Child tasks not yet complete */
2764 kmp_taskgroup_t
2765 *td_taskgroup; // Each task keeps pointer to its current taskgroup
2766 kmp_dephash_t
2767 *td_dephash; // Dependencies for children tasks are tracked from here
2768 kmp_depnode_t
2769 *td_depnode; // Pointer to graph node if this task has dependencies
2770 kmp_task_team_t *td_task_team;
2771 size_t td_size_alloc; // Size of task structure, including shareds etc.
2772#if defined(KMP_GOMP_COMPAT)
2773 // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2774 kmp_int32 td_size_loop_bounds;
2775#endif
2776 kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2777#if defined(KMP_GOMP_COMPAT)
2778 // GOMP sends in a copy function for copy constructors
2779 void (*td_copy_func)(void *, void *);
2780#endif
2781 kmp_event_t td_allow_completion_event;
2782#if OMPT_SUPPORT
2783 ompt_task_info_t ompt_task_info;
2784#endif
2785#if OMPX_TASKGRAPH
2786 bool is_taskgraph = 0; // whether the task is within a TDG
2787 kmp_tdg_info_t *tdg; // used to associate task with a TDG
2788#endif
2789 kmp_target_data_t td_target_data;
2790}; // struct kmp_taskdata
2791
2792// Make sure padding above worked
2793KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2794
2795// Data for task team but per thread
2796typedef struct kmp_base_thread_data {
2797 kmp_info_p *td_thr; // Pointer back to thread info
2798 // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2799 // queued?
2800 kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2801 kmp_taskdata_t *
2802 *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2803 kmp_int32 td_deque_size; // Size of deck
2804 kmp_uint32 td_deque_head; // Head of deque (will wrap)
2805 kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2806 kmp_int32 td_deque_ntasks; // Number of tasks in deque
2807 // GEH: shouldn't this be volatile since used in while-spin?
2808 kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2809#ifdef BUILD_TIED_TASK_STACK
2810 kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2811// scheduling constraint
2812#endif // BUILD_TIED_TASK_STACK
2813} kmp_base_thread_data_t;
2814
2815#define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2816#define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2817
2818#define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2819#define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2820
2821typedef union KMP_ALIGN_CACHE kmp_thread_data {
2822 kmp_base_thread_data_t td;
2823 double td_align; /* use worst case alignment */
2824 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2825} kmp_thread_data_t;
2826
2827typedef struct kmp_task_pri {
2828 kmp_thread_data_t td;
2829 kmp_int32 priority;
2830 kmp_task_pri *next;
2831} kmp_task_pri_t;
2832
2833// Data for task teams which are used when tasking is enabled for the team
2834typedef struct kmp_base_task_team {
2835 kmp_bootstrap_lock_t
2836 tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2837 /* must be bootstrap lock since used at library shutdown*/
2838
2839 // TODO: check performance vs kmp_tas_lock_t
2840 kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */
2841 kmp_task_pri_t *tt_task_pri_list;
2842
2843 kmp_task_team_t *tt_next; /* For linking the task team free list */
2844 kmp_thread_data_t
2845 *tt_threads_data; /* Array of per-thread structures for task team */
2846 /* Data survives task team deallocation */
2847 kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2848 executing this team? */
2849 /* TRUE means tt_threads_data is set up and initialized */
2850 kmp_int32 tt_nproc; /* #threads in team */
2851 kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2852 kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2853 kmp_int32 tt_untied_task_encountered;
2854 std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued
2855 // There is hidden helper thread encountered in this task team so that we must
2856 // wait when waiting on task team
2857 kmp_int32 tt_hidden_helper_task_encountered;
2858
2859 KMP_ALIGN_CACHE
2860 std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2861
2862 KMP_ALIGN_CACHE
2863 volatile kmp_uint32
2864 tt_active; /* is the team still actively executing tasks */
2865} kmp_base_task_team_t;
2866
2867union KMP_ALIGN_CACHE kmp_task_team {
2868 kmp_base_task_team_t tt;
2869 double tt_align; /* use worst case alignment */
2870 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2871};
2872
2873#if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2874// Free lists keep same-size free memory slots for fast memory allocation
2875// routines
2876typedef struct kmp_free_list {
2877 void *th_free_list_self; // Self-allocated tasks free list
2878 void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2879 // threads
2880 void *th_free_list_other; // Non-self free list (to be returned to owner's
2881 // sync list)
2882} kmp_free_list_t;
2883#endif
2884#if KMP_NESTED_HOT_TEAMS
2885// Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2886// are not put in teams pool, and they don't put threads in threads pool.
2887typedef struct kmp_hot_team_ptr {
2888 kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2889 kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2890} kmp_hot_team_ptr_t;
2891#endif
2892typedef struct kmp_teams_size {
2893 kmp_int32 nteams; // number of teams in a league
2894 kmp_int32 nth; // number of threads in each team of the league
2895} kmp_teams_size_t;
2896
2897// This struct stores a thread that acts as a "root" for a contention
2898// group. Contention groups are rooted at kmp_root threads, but also at
2899// each primary thread of each team created in the teams construct.
2900// This struct therefore also stores a thread_limit associated with
2901// that contention group, and a counter to track the number of threads
2902// active in that contention group. Each thread has a list of these: CG
2903// root threads have an entry in their list in which cg_root refers to
2904// the thread itself, whereas other workers in the CG will have a
2905// single entry where cg_root is same as the entry containing their CG
2906// root. When a thread encounters a teams construct, it will add a new
2907// entry to the front of its list, because it now roots a new CG.
2908typedef struct kmp_cg_root {
2909 kmp_info_p *cg_root; // "root" thread for a contention group
2910 // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2911 // thread_limit clause for teams primary threads
2912 kmp_int32 cg_thread_limit;
2913 kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2914 struct kmp_cg_root *up; // pointer to higher level CG root in list
2915} kmp_cg_root_t;
2916
2917// OpenMP thread data structures
2918
2919typedef struct KMP_ALIGN_CACHE kmp_base_info {
2920 /* Start with the readonly data which is cache aligned and padded. This is
2921 written before the thread starts working by the primary thread. Uber
2922 masters may update themselves later. Usage does not consider serialized
2923 regions. */
2924 kmp_desc_t th_info;
2925 kmp_team_p *th_team; /* team we belong to */
2926 kmp_root_p *th_root; /* pointer to root of task hierarchy */
2927 kmp_info_p *th_next_pool; /* next available thread in the pool */
2928 kmp_disp_t *th_dispatch; /* thread's dispatch data */
2929 int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2930
2931 /* The following are cached from the team info structure */
2932 /* TODO use these in more places as determined to be needed via profiling */
2933 int th_team_nproc; /* number of threads in a team */
2934 kmp_info_p *th_team_master; /* the team's primary thread */
2935 int th_team_serialized; /* team is serialized */
2936 microtask_t th_teams_microtask; /* save entry address for teams construct */
2937 int th_teams_level; /* save initial level of teams construct */
2938/* it is 0 on device but may be any on host */
2939
2940/* The blocktime info is copied from the team struct to the thread struct */
2941/* at the start of a barrier, and the values stored in the team are used */
2942/* at points in the code where the team struct is no longer guaranteed */
2943/* to exist (from the POV of worker threads). */
2944#if KMP_USE_MONITOR
2945 int th_team_bt_intervals;
2946 int th_team_bt_set;
2947#else
2948 kmp_uint64 th_team_bt_intervals;
2949#endif
2950
2951#if KMP_AFFINITY_SUPPORTED
2952 kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2953 kmp_affinity_ids_t th_topology_ids; /* thread's current topology ids */
2954 kmp_affinity_attrs_t th_topology_attrs; /* thread's current topology attrs */
2955#endif
2956 omp_allocator_handle_t th_def_allocator; /* default allocator */
2957 /* The data set by the primary thread at reinit, then R/W by the worker */
2958 KMP_ALIGN_CACHE int
2959 th_set_nproc; /* if > 0, then only use this request for the next fork */
2960#if KMP_NESTED_HOT_TEAMS
2961 kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2962#endif
2963 kmp_proc_bind_t
2964 th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2965 kmp_teams_size_t
2966 th_teams_size; /* number of teams/threads in teams construct */
2967#if KMP_AFFINITY_SUPPORTED
2968 int th_current_place; /* place currently bound to */
2969 int th_new_place; /* place to bind to in par reg */
2970 int th_first_place; /* first place in partition */
2971 int th_last_place; /* last place in partition */
2972#endif
2973 int th_prev_level; /* previous level for affinity format */
2974 int th_prev_num_threads; /* previous num_threads for affinity format */
2975#if USE_ITT_BUILD
2976 kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2977 kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2978 kmp_uint64 th_frame_time; /* frame timestamp */
2979#endif /* USE_ITT_BUILD */
2980 kmp_local_t th_local;
2981 struct private_common *th_pri_head;
2982
2983 /* Now the data only used by the worker (after initial allocation) */
2984 /* TODO the first serial team should actually be stored in the info_t
2985 structure. this will help reduce initial allocation overhead */
2986 KMP_ALIGN_CACHE kmp_team_p
2987 *th_serial_team; /*serialized team held in reserve*/
2988
2989#if OMPT_SUPPORT
2990 ompt_thread_info_t ompt_thread_info;
2991#endif
2992
2993 /* The following are also read by the primary thread during reinit */
2994 struct common_table *th_pri_common;
2995
2996 volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2997 /* while awaiting queuing lock acquire */
2998
2999 volatile void *th_sleep_loc; // this points at a kmp_flag<T>
3000 flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
3001
3002 ident_t *th_ident;
3003 unsigned th_x; // Random number generator data
3004 unsigned th_a; // Random number generator data
3005
3006 /* Tasking-related data for the thread */
3007 kmp_task_team_t *th_task_team; // Task team struct
3008 kmp_taskdata_t *th_current_task; // Innermost Task being executed
3009 kmp_uint8 th_task_state; // alternating 0/1 for task team identification
3010 kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
3011 // at nested levels
3012 kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
3013 kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
3014 kmp_uint32 th_reap_state; // Non-zero indicates thread is not
3015 // tasking, thus safe to reap
3016
3017 /* More stuff for keeping track of active/sleeping threads (this part is
3018 written by the worker thread) */
3019 kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
3020 int th_active; // ! sleeping; 32 bits for TCR/TCW
3021 std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
3022 // 0 = not used in team; 1 = used in team;
3023 // 2 = transitioning to not used in team; 3 = transitioning to used in team
3024 struct cons_header *th_cons; // used for consistency check
3025#if KMP_USE_HIER_SCHED
3026 // used for hierarchical scheduling
3027 kmp_hier_private_bdata_t *th_hier_bar_data;
3028#endif
3029
3030 /* Add the syncronizing data which is cache aligned and padded. */
3031 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
3032
3033 KMP_ALIGN_CACHE volatile kmp_int32
3034 th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
3035
3036#if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
3037#define NUM_LISTS 4
3038 kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
3039// allocation routines
3040#endif
3041
3042#if KMP_OS_WINDOWS
3043 kmp_win32_cond_t th_suspend_cv;
3044 kmp_win32_mutex_t th_suspend_mx;
3045 std::atomic<int> th_suspend_init;
3046#endif
3047#if KMP_OS_UNIX
3048 kmp_cond_align_t th_suspend_cv;
3049 kmp_mutex_align_t th_suspend_mx;
3050 std::atomic<int> th_suspend_init_count;
3051#endif
3052
3053#if USE_ITT_BUILD
3054 kmp_itt_mark_t th_itt_mark_single;
3055// alignment ???
3056#endif /* USE_ITT_BUILD */
3057#if KMP_STATS_ENABLED
3058 kmp_stats_list *th_stats;
3059#endif
3060#if KMP_OS_UNIX
3061 std::atomic<bool> th_blocking;
3062#endif
3063 kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
3064} kmp_base_info_t;
3065
3066typedef union KMP_ALIGN_CACHE kmp_info {
3067 double th_align; /* use worst case alignment */
3068 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
3069 kmp_base_info_t th;
3070} kmp_info_t;
3071
3072// OpenMP thread team data structures
3073
3074typedef struct kmp_base_data {
3075 volatile kmp_uint32 t_value;
3076} kmp_base_data_t;
3077
3078typedef union KMP_ALIGN_CACHE kmp_sleep_team {
3079 double dt_align; /* use worst case alignment */
3080 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3081 kmp_base_data_t dt;
3082} kmp_sleep_team_t;
3083
3084typedef union KMP_ALIGN_CACHE kmp_ordered_team {
3085 double dt_align; /* use worst case alignment */
3086 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3087 kmp_base_data_t dt;
3088} kmp_ordered_team_t;
3089
3090typedef int (*launch_t)(int gtid);
3091
3092/* Minimum number of ARGV entries to malloc if necessary */
3093#define KMP_MIN_MALLOC_ARGV_ENTRIES 100
3094
3095// Set up how many argv pointers will fit in cache lines containing
3096// t_inline_argv. Historically, we have supported at least 96 bytes. Using a
3097// larger value for more space between the primary write/worker read section and
3098// read/write by all section seems to buy more performance on EPCC PARALLEL.
3099#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3100#define KMP_INLINE_ARGV_BYTES \
3101 (4 * CACHE_LINE - \
3102 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
3103 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
3104 CACHE_LINE))
3105#else
3106#define KMP_INLINE_ARGV_BYTES \
3107 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
3108#endif
3109#define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
3110
3111typedef struct KMP_ALIGN_CACHE kmp_base_team {
3112 // Synchronization Data
3113 // ---------------------------------------------------------------------------
3114 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
3115 kmp_balign_team_t t_bar[bs_last_barrier];
3116 std::atomic<int> t_construct; // count of single directive encountered by team
3117 char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
3118
3119 // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
3120 std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
3121 std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
3122
3123 // Primary thread only
3124 // ---------------------------------------------------------------------------
3125 KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
3126 int t_master_this_cons; // "this_construct" single counter of primary thread
3127 // in parent team
3128 ident_t *t_ident; // if volatile, have to change too much other crud to
3129 // volatile too
3130 kmp_team_p *t_parent; // parent team
3131 kmp_team_p *t_next_pool; // next free team in the team pool
3132 kmp_disp_t *t_dispatch; // thread's dispatch data
3133 kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
3134 kmp_proc_bind_t t_proc_bind; // bind type for par region
3135#if USE_ITT_BUILD
3136 kmp_uint64 t_region_time; // region begin timestamp
3137#endif /* USE_ITT_BUILD */
3138
3139 // Primary thread write, workers read
3140 // --------------------------------------------------------------------------
3141 KMP_ALIGN_CACHE void **t_argv;
3142 int t_argc;
3143 int t_nproc; // number of threads in team
3144 microtask_t t_pkfn;
3145 launch_t t_invoke; // procedure to launch the microtask
3146
3147#if OMPT_SUPPORT
3148 ompt_team_info_t ompt_team_info;
3149 ompt_lw_taskteam_t *ompt_serialized_team_info;
3150#endif
3151
3152#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3153 kmp_int8 t_fp_control_saved;
3154 kmp_int8 t_pad2b;
3155 kmp_int16 t_x87_fpu_control_word; // FP control regs
3156 kmp_uint32 t_mxcsr;
3157#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3158
3159 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
3160
3161 KMP_ALIGN_CACHE kmp_info_t **t_threads;
3162 kmp_taskdata_t
3163 *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
3164 int t_level; // nested parallel level
3165
3166 KMP_ALIGN_CACHE int t_max_argc;
3167 int t_max_nproc; // max threads this team can handle (dynamically expandable)
3168 int t_serialized; // levels deep of serialized teams
3169 dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
3170 int t_id; // team's id, assigned by debugger.
3171 int t_active_level; // nested active parallel level
3172 kmp_r_sched_t t_sched; // run-time schedule for the team
3173#if KMP_AFFINITY_SUPPORTED
3174 int t_first_place; // first & last place in parent thread's partition.
3175 int t_last_place; // Restore these values to primary thread after par region.
3176#endif // KMP_AFFINITY_SUPPORTED
3177 int t_display_affinity;
3178 int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
3179 // omp_set_num_threads() call
3180 omp_allocator_handle_t t_def_allocator; /* default allocator */
3181
3182// Read/write by workers as well
3183#if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
3184 // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
3185 // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
3186 // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
3187 // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
3188 char dummy_padding[1024];
3189#endif
3190 // Internal control stack for additional nested teams.
3191 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
3192 // for SERIALIZED teams nested 2 or more levels deep
3193 // typed flag to store request state of cancellation
3194 std::atomic<kmp_int32> t_cancel_request;
3195 int t_master_active; // save on fork, restore on join
3196 void *t_copypriv_data; // team specific pointer to copyprivate data array
3197#if KMP_OS_WINDOWS
3198 std::atomic<kmp_uint32> t_copyin_counter;
3199#endif
3200#if USE_ITT_BUILD
3201 void *t_stack_id; // team specific stack stitching id (for ittnotify)
3202#endif /* USE_ITT_BUILD */
3203 distributedBarrier *b; // Distributed barrier data associated with team
3204} kmp_base_team_t;
3205
3206union KMP_ALIGN_CACHE kmp_team {
3207 kmp_base_team_t t;
3208 double t_align; /* use worst case alignment */
3209 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
3210};
3211
3212typedef union KMP_ALIGN_CACHE kmp_time_global {
3213 double dt_align; /* use worst case alignment */
3214 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3215 kmp_base_data_t dt;
3216} kmp_time_global_t;
3217
3218typedef struct kmp_base_global {
3219 /* cache-aligned */
3220 kmp_time_global_t g_time;
3221
3222 /* non cache-aligned */
3223 volatile int g_abort;
3224 volatile int g_done;
3225
3226 int g_dynamic;
3227 enum dynamic_mode g_dynamic_mode;
3228} kmp_base_global_t;
3229
3230typedef union KMP_ALIGN_CACHE kmp_global {
3231 kmp_base_global_t g;
3232 double g_align; /* use worst case alignment */
3233 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
3234} kmp_global_t;
3235
3236typedef struct kmp_base_root {
3237 // TODO: GEH - combine r_active with r_in_parallel then r_active ==
3238 // (r_in_parallel>= 0)
3239 // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
3240 // the synch overhead or keeping r_active
3241 volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
3242 // keeps a count of active parallel regions per root
3243 std::atomic<int> r_in_parallel;
3244 // GEH: This is misnamed, should be r_active_levels
3245 kmp_team_t *r_root_team;
3246 kmp_team_t *r_hot_team;
3247 kmp_info_t *r_uber_thread;
3248 kmp_lock_t r_begin_lock;
3249 volatile int r_begin;
3250 int r_blocktime; /* blocktime for this root and descendants */
3251#if KMP_AFFINITY_SUPPORTED
3252 int r_affinity_assigned;
3253#endif // KMP_AFFINITY_SUPPORTED
3254} kmp_base_root_t;
3255
3256typedef union KMP_ALIGN_CACHE kmp_root {
3257 kmp_base_root_t r;
3258 double r_align; /* use worst case alignment */
3259 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
3260} kmp_root_t;
3261
3262struct fortran_inx_info {
3263 kmp_int32 data;
3264};
3265
3266// This list type exists to hold old __kmp_threads arrays so that
3267// old references to them may complete while reallocation takes place when
3268// expanding the array. The items in this list are kept alive until library
3269// shutdown.
3270typedef struct kmp_old_threads_list_t {
3271 kmp_info_t **threads;
3272 struct kmp_old_threads_list_t *next;
3273} kmp_old_threads_list_t;
3274
3275/* ------------------------------------------------------------------------ */
3276
3277extern int __kmp_settings;
3278extern int __kmp_duplicate_library_ok;
3279#if USE_ITT_BUILD
3280extern int __kmp_forkjoin_frames;
3281extern int __kmp_forkjoin_frames_mode;
3282#endif
3283extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3284extern int __kmp_determ_red;
3285
3286#ifdef KMP_DEBUG
3287extern int kmp_a_debug;
3288extern int kmp_b_debug;
3289extern int kmp_c_debug;
3290extern int kmp_d_debug;
3291extern int kmp_e_debug;
3292extern int kmp_f_debug;
3293#endif /* KMP_DEBUG */
3294
3295/* For debug information logging using rotating buffer */
3296#define KMP_DEBUG_BUF_LINES_INIT 512
3297#define KMP_DEBUG_BUF_LINES_MIN 1
3298
3299#define KMP_DEBUG_BUF_CHARS_INIT 128
3300#define KMP_DEBUG_BUF_CHARS_MIN 2
3301
3302extern int
3303 __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3304extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3305extern int
3306 __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3307extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3308 entry pointer */
3309
3310extern char *__kmp_debug_buffer; /* Debug buffer itself */
3311extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3312 printed in buffer so far */
3313extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3314 recommended in warnings */
3315/* end rotating debug buffer */
3316
3317#ifdef KMP_DEBUG
3318extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3319
3320#define KMP_PAR_RANGE_ROUTINE_LEN 1024
3321extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3322#define KMP_PAR_RANGE_FILENAME_LEN 1024
3323extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3324extern int __kmp_par_range_lb;
3325extern int __kmp_par_range_ub;
3326#endif
3327
3328/* For printing out dynamic storage map for threads and teams */
3329extern int
3330 __kmp_storage_map; /* True means print storage map for threads and teams */
3331extern int __kmp_storage_map_verbose; /* True means storage map includes
3332 placement info */
3333extern int __kmp_storage_map_verbose_specified;
3334
3335#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3336extern kmp_cpuinfo_t __kmp_cpuinfo;
3337static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3338#elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3339static inline bool __kmp_is_hybrid_cpu() { return true; }
3340#else
3341static inline bool __kmp_is_hybrid_cpu() { return false; }
3342#endif
3343
3344extern volatile int __kmp_init_serial;
3345extern volatile int __kmp_init_gtid;
3346extern volatile int __kmp_init_common;
3347extern volatile int __kmp_need_register_serial;
3348extern volatile int __kmp_init_middle;
3349extern volatile int __kmp_init_parallel;
3350#if KMP_USE_MONITOR
3351extern volatile int __kmp_init_monitor;
3352#endif
3353extern volatile int __kmp_init_user_locks;
3354extern volatile int __kmp_init_hidden_helper_threads;
3355extern int __kmp_init_counter;
3356extern int __kmp_root_counter;
3357extern int __kmp_version;
3358
3359/* list of address of allocated caches for commons */
3360extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3361
3362/* Barrier algorithm types and options */
3363extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3364extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3365extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3366extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3367extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3368extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3369extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3370extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3371extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3372extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3373extern char const *__kmp_barrier_type_name[bs_last_barrier];
3374extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3375
3376/* Global Locks */
3377extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3378extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3379extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3380extern kmp_bootstrap_lock_t
3381 __kmp_exit_lock; /* exit() is not always thread-safe */
3382#if KMP_USE_MONITOR
3383extern kmp_bootstrap_lock_t
3384 __kmp_monitor_lock; /* control monitor thread creation */
3385#endif
3386extern kmp_bootstrap_lock_t
3387 __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3388 __kmp_threads expansion to co-exist */
3389
3390extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
3391extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
3392extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3393
3394extern enum library_type __kmp_library;
3395
3396extern enum sched_type __kmp_sched; /* default runtime scheduling */
3397extern enum sched_type __kmp_static; /* default static scheduling method */
3398extern enum sched_type __kmp_guided; /* default guided scheduling method */
3399extern enum sched_type __kmp_auto; /* default auto scheduling method */
3400extern int __kmp_chunk; /* default runtime chunk size */
3401extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3402
3403extern size_t __kmp_stksize; /* stack size per thread */
3404#if KMP_USE_MONITOR
3405extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3406#endif
3407extern size_t __kmp_stkoffset; /* stack offset per thread */
3408extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3409
3410extern size_t
3411 __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3412extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3413extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3414extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
3415extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3416extern int __kmp_generate_warnings; /* should we issue warnings? */
3417extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3418
3419#ifdef DEBUG_SUSPEND
3420extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3421#endif
3422
3423extern kmp_int32 __kmp_use_yield;
3424extern kmp_int32 __kmp_use_yield_exp_set;
3425extern kmp_uint32 __kmp_yield_init;
3426extern kmp_uint32 __kmp_yield_next;
3427extern kmp_uint64 __kmp_pause_init;
3428
3429/* ------------------------------------------------------------------------- */
3430extern int __kmp_allThreadsSpecified;
3431
3432extern size_t __kmp_align_alloc;
3433/* following data protected by initialization routines */
3434extern int __kmp_xproc; /* number of processors in the system */
3435extern int __kmp_avail_proc; /* number of processors available to the process */
3436extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3437extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3438// maximum total number of concurrently-existing threads on device
3439extern int __kmp_max_nth;
3440// maximum total number of concurrently-existing threads in a contention group
3441extern int __kmp_cg_max_nth;
3442extern int __kmp_task_max_nth; // max threads used in a task
3443extern int __kmp_teams_max_nth; // max threads used in a teams construct
3444extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3445 __kmp_root */
3446extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3447 region a la OMP_NUM_THREADS */
3448extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3449 initialization */
3450extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3451 used (fixed) */
3452extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3453 (__kmpc_threadprivate_cached()) */
3454extern int __kmp_dflt_blocktime; /* number of microseconds to wait before
3455 blocking (env setting) */
3456extern char __kmp_blocktime_units; /* 'm' or 'u' to note units specified */
3457extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */
3458
3459// Convert raw blocktime from ms to us if needed.
3460static inline void __kmp_aux_convert_blocktime(int *bt) {
3461 if (__kmp_blocktime_units == 'm') {
3462 if (*bt > INT_MAX / 1000) {
3463 *bt = INT_MAX / 1000;
3464 KMP_INFORM(MaxValueUsing, "kmp_set_blocktime(ms)", bt);
3465 }
3466 *bt = *bt * 1000;
3467 }
3468}
3469
3470#if KMP_USE_MONITOR
3471extern int
3472 __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3473extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3474 blocking */
3475#endif
3476#ifdef KMP_ADJUST_BLOCKTIME
3477extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3478#endif /* KMP_ADJUST_BLOCKTIME */
3479#ifdef KMP_DFLT_NTH_CORES
3480extern int __kmp_ncores; /* Total number of cores for threads placement */
3481#endif
3482/* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3483extern int __kmp_abort_delay;
3484
3485extern int __kmp_need_register_atfork_specified;
3486extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3487 to install fork handler */
3488extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3489 0 - not set, will be set at runtime
3490 1 - using stack search
3491 2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3492 X*) or TlsGetValue(Windows* OS))
3493 3 - static TLS (__declspec(thread) __kmp_gtid),
3494 Linux* OS .so only. */
3495extern int
3496 __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3497#ifdef KMP_TDATA_GTID
3498extern KMP_THREAD_LOCAL int __kmp_gtid;
3499#endif
3500extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3501extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3502#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3503extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3504extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3505extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3506#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3507
3508// max_active_levels for nested parallelism enabled by default via
3509// OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3510extern int __kmp_dflt_max_active_levels;
3511// Indicates whether value of __kmp_dflt_max_active_levels was already
3512// explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3513extern bool __kmp_dflt_max_active_levels_set;
3514extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3515 concurrent execution per team */
3516#if KMP_NESTED_HOT_TEAMS
3517extern int __kmp_hot_teams_mode;
3518extern int __kmp_hot_teams_max_level;
3519#endif
3520
3521#if KMP_OS_LINUX
3522extern enum clock_function_type __kmp_clock_function;
3523extern int __kmp_clock_function_param;
3524#endif /* KMP_OS_LINUX */
3525
3526#if KMP_MIC_SUPPORTED
3527extern enum mic_type __kmp_mic_type;
3528#endif
3529
3530#ifdef USE_LOAD_BALANCE
3531extern double __kmp_load_balance_interval; // load balance algorithm interval
3532#endif /* USE_LOAD_BALANCE */
3533
3534// OpenMP 3.1 - Nested num threads array
3535typedef struct kmp_nested_nthreads_t {
3536 int *nth;
3537 int size;
3538 int used;
3539} kmp_nested_nthreads_t;
3540
3541extern kmp_nested_nthreads_t __kmp_nested_nth;
3542
3543#if KMP_USE_ADAPTIVE_LOCKS
3544
3545// Parameters for the speculative lock backoff system.
3546struct kmp_adaptive_backoff_params_t {
3547 // Number of soft retries before it counts as a hard retry.
3548 kmp_uint32 max_soft_retries;
3549 // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3550 // the right
3551 kmp_uint32 max_badness;
3552};
3553
3554extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3555
3556#if KMP_DEBUG_ADAPTIVE_LOCKS
3557extern const char *__kmp_speculative_statsfile;
3558#endif
3559
3560#endif // KMP_USE_ADAPTIVE_LOCKS
3561
3562extern int __kmp_display_env; /* TRUE or FALSE */
3563extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3564extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3565extern int __kmp_nteams;
3566extern int __kmp_teams_thread_limit;
3567
3568/* ------------------------------------------------------------------------- */
3569
3570/* the following are protected by the fork/join lock */
3571/* write: lock read: anytime */
3572extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3573/* Holds old arrays of __kmp_threads until library shutdown */
3574extern kmp_old_threads_list_t *__kmp_old_threads_list;
3575/* read/write: lock */
3576extern volatile kmp_team_t *__kmp_team_pool;
3577extern volatile kmp_info_t *__kmp_thread_pool;
3578extern kmp_info_t *__kmp_thread_pool_insert_pt;
3579
3580// total num threads reachable from some root thread including all root threads
3581extern volatile int __kmp_nth;
3582/* total number of threads reachable from some root thread including all root
3583 threads, and those in the thread pool */
3584extern volatile int __kmp_all_nth;
3585extern std::atomic<int> __kmp_thread_pool_active_nth;
3586
3587extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3588/* end data protected by fork/join lock */
3589/* ------------------------------------------------------------------------- */
3590
3591#define __kmp_get_gtid() __kmp_get_global_thread_id()
3592#define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3593#define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3594#define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3595#define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3596
3597// AT: Which way is correct?
3598// AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3599// AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3600#define __kmp_get_team_num_threads(gtid) \
3601 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3602
3603static inline bool KMP_UBER_GTID(int gtid) {
3604 KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3605 KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3606 return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3607 __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3608}
3609
3610static inline int __kmp_tid_from_gtid(int gtid) {
3611 KMP_DEBUG_ASSERT(gtid >= 0);
3612 return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3613}
3614
3615static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3616 KMP_DEBUG_ASSERT(tid >= 0 && team);
3617 return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3618}
3619
3620static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3621 KMP_DEBUG_ASSERT(thr);
3622 return thr->th.th_info.ds.ds_gtid;
3623}
3624
3625static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3626 KMP_DEBUG_ASSERT(gtid >= 0);
3627 return __kmp_threads[gtid];
3628}
3629
3630static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3631 KMP_DEBUG_ASSERT(gtid >= 0);
3632 return __kmp_threads[gtid]->th.th_team;
3633}
3634
3635static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3636 if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3637 KMP_FATAL(ThreadIdentInvalid);
3638}
3639
3640#if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3641extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3642extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3643extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3644extern int __kmp_mwait_hints; // Hints to pass in to mwait
3645#endif
3646
3647#if KMP_HAVE_UMWAIT
3648extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3649extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3650extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3651extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3652#endif
3653
3654/* ------------------------------------------------------------------------- */
3655
3656extern kmp_global_t __kmp_global; /* global status */
3657
3658extern kmp_info_t __kmp_monitor;
3659// For Debugging Support Library
3660extern std::atomic<kmp_int32> __kmp_team_counter;
3661// For Debugging Support Library
3662extern std::atomic<kmp_int32> __kmp_task_counter;
3663
3664#if USE_DEBUGGER
3665#define _KMP_GEN_ID(counter) \
3666 (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3667#else
3668#define _KMP_GEN_ID(counter) (~0)
3669#endif /* USE_DEBUGGER */
3670
3671#define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3672#define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3673
3674/* ------------------------------------------------------------------------ */
3675
3676extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3677 size_t size, char const *format, ...);
3678
3679extern void __kmp_serial_initialize(void);
3680extern void __kmp_middle_initialize(void);
3681extern void __kmp_parallel_initialize(void);
3682
3683extern void __kmp_internal_begin(void);
3684extern void __kmp_internal_end_library(int gtid);
3685extern void __kmp_internal_end_thread(int gtid);
3686extern void __kmp_internal_end_atexit(void);
3687extern void __kmp_internal_end_dtor(void);
3688extern void __kmp_internal_end_dest(void *);
3689
3690extern int __kmp_register_root(int initial_thread);
3691extern void __kmp_unregister_root(int gtid);
3692extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3693
3694extern int __kmp_ignore_mppbeg(void);
3695extern int __kmp_ignore_mppend(void);
3696
3697extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3698extern void __kmp_exit_single(int gtid);
3699
3700extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3701extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3702
3703#ifdef USE_LOAD_BALANCE
3704extern int __kmp_get_load_balance(int);
3705#endif
3706
3707extern int __kmp_get_global_thread_id(void);
3708extern int __kmp_get_global_thread_id_reg(void);
3709extern void __kmp_exit_thread(int exit_status);
3710extern void __kmp_abort(char const *format, ...);
3711extern void __kmp_abort_thread(void);
3712KMP_NORETURN extern void __kmp_abort_process(void);
3713extern void __kmp_warn(char const *format, ...);
3714
3715extern void __kmp_set_num_threads(int new_nth, int gtid);
3716
3717extern bool __kmp_detect_shm();
3718extern bool __kmp_detect_tmp();
3719
3720// Returns current thread (pointer to kmp_info_t). Current thread *must* be
3721// registered.
3722static inline kmp_info_t *__kmp_entry_thread() {
3723 int gtid = __kmp_entry_gtid();
3724
3725 return __kmp_threads[gtid];
3726}
3727
3728extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3729extern int __kmp_get_max_active_levels(int gtid);
3730extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3731extern int __kmp_get_team_size(int gtid, int level);
3732extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3733extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3734
3735extern unsigned short __kmp_get_random(kmp_info_t *thread);
3736extern void __kmp_init_random(kmp_info_t *thread);
3737
3738extern kmp_r_sched_t __kmp_get_schedule_global(void);
3739extern void __kmp_adjust_num_threads(int new_nproc);
3740extern void __kmp_check_stksize(size_t *val);
3741
3742extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3743extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3744extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3745#define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3746#define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3747#define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3748
3749#if USE_FAST_MEMORY
3750extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3751 size_t size KMP_SRC_LOC_DECL);
3752extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3753extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3754extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3755#define __kmp_fast_allocate(this_thr, size) \
3756 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3757#define __kmp_fast_free(this_thr, ptr) \
3758 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3759#endif
3760
3761extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3762extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3763 size_t elsize KMP_SRC_LOC_DECL);
3764extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3765 size_t size KMP_SRC_LOC_DECL);
3766extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3767#define __kmp_thread_malloc(th, size) \
3768 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3769#define __kmp_thread_calloc(th, nelem, elsize) \
3770 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3771#define __kmp_thread_realloc(th, ptr, size) \
3772 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3773#define __kmp_thread_free(th, ptr) \
3774 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3775
3776extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3777
3778extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3779 kmp_proc_bind_t proc_bind);
3780extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3781 int num_threads);
3782extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3783 int num_teams_ub, int num_threads);
3784
3785extern void __kmp_yield();
3786
3787extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3788 enum sched_type schedule, kmp_int32 lb,
3789 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3790extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3791 enum sched_type schedule, kmp_uint32 lb,
3792 kmp_uint32 ub, kmp_int32 st,
3793 kmp_int32 chunk);
3794extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3795 enum sched_type schedule, kmp_int64 lb,
3796 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3797extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3798 enum sched_type schedule, kmp_uint64 lb,
3799 kmp_uint64 ub, kmp_int64 st,
3800 kmp_int64 chunk);
3801
3802extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3803 kmp_int32 *p_last, kmp_int32 *p_lb,
3804 kmp_int32 *p_ub, kmp_int32 *p_st);
3805extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3806 kmp_int32 *p_last, kmp_uint32 *p_lb,
3807 kmp_uint32 *p_ub, kmp_int32 *p_st);
3808extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3809 kmp_int32 *p_last, kmp_int64 *p_lb,
3810 kmp_int64 *p_ub, kmp_int64 *p_st);
3811extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3812 kmp_int32 *p_last, kmp_uint64 *p_lb,
3813 kmp_uint64 *p_ub, kmp_int64 *p_st);
3814
3815extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3816extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3817extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3818extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3819
3820#ifdef KMP_GOMP_COMPAT
3821
3822extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3823 enum sched_type schedule, kmp_int32 lb,
3824 kmp_int32 ub, kmp_int32 st,
3825 kmp_int32 chunk, int push_ws);
3826extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3827 enum sched_type schedule, kmp_uint32 lb,
3828 kmp_uint32 ub, kmp_int32 st,
3829 kmp_int32 chunk, int push_ws);
3830extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3831 enum sched_type schedule, kmp_int64 lb,
3832 kmp_int64 ub, kmp_int64 st,
3833 kmp_int64 chunk, int push_ws);
3834extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3835 enum sched_type schedule, kmp_uint64 lb,
3836 kmp_uint64 ub, kmp_int64 st,
3837 kmp_int64 chunk, int push_ws);
3838extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3839extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3840extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3841extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3842
3843#endif /* KMP_GOMP_COMPAT */
3844
3845extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3846extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3847extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3848extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3849extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3850extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3851 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3852 void *obj);
3853extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3854 kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3855
3856extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3857 int final_spin
3858#if USE_ITT_BUILD
3859 ,
3860 void *itt_sync_obj
3861#endif
3862);
3863extern void __kmp_release_64(kmp_flag_64<> *flag);
3864
3865extern void __kmp_infinite_loop(void);
3866
3867extern void __kmp_cleanup(void);
3868
3869#if KMP_HANDLE_SIGNALS
3870extern int __kmp_handle_signals;
3871extern void __kmp_install_signals(int parallel_init);
3872extern void __kmp_remove_signals(void);
3873#endif
3874
3875extern void __kmp_clear_system_time(void);
3876extern void __kmp_read_system_time(double *delta);
3877
3878extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3879
3880extern void __kmp_expand_host_name(char *buffer, size_t size);
3881extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3882
3883#if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM))
3884extern void
3885__kmp_initialize_system_tick(void); /* Initialize timer tick value */
3886#endif
3887
3888extern void
3889__kmp_runtime_initialize(void); /* machine specific initialization */
3890extern void __kmp_runtime_destroy(void);
3891
3892#if KMP_AFFINITY_SUPPORTED
3893extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3894 kmp_affin_mask_t *mask);
3895extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3896 kmp_affin_mask_t *mask);
3897extern void __kmp_affinity_initialize(kmp_affinity_t &affinity);
3898extern void __kmp_affinity_uninitialize(void);
3899extern void __kmp_affinity_set_init_mask(
3900 int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3901void __kmp_affinity_bind_init_mask(int gtid);
3902extern void __kmp_affinity_bind_place(int gtid);
3903extern void __kmp_affinity_determine_capable(const char *env_var);
3904extern int __kmp_aux_set_affinity(void **mask);
3905extern int __kmp_aux_get_affinity(void **mask);
3906extern int __kmp_aux_get_affinity_max_proc();
3907extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3908extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3909extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3910extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3911#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
3912extern int __kmp_get_first_osid_with_ecore(void);
3913#endif
3914#if KMP_OS_LINUX || KMP_OS_FREEBSD
3915extern int kmp_set_thread_affinity_mask_initial(void);
3916#endif
3917static inline void __kmp_assign_root_init_mask() {
3918 int gtid = __kmp_entry_gtid();
3919 kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3920 if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3921 __kmp_affinity_set_init_mask(gtid, /*isa_root=*/TRUE);
3922 __kmp_affinity_bind_init_mask(gtid);
3923 r->r.r_affinity_assigned = TRUE;
3924 }
3925}
3926static inline void __kmp_reset_root_init_mask(int gtid) {
3927 if (!KMP_AFFINITY_CAPABLE())
3928 return;
3929 kmp_info_t *th = __kmp_threads[gtid];
3930 kmp_root_t *r = th->th.th_root;
3931 if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3932 __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3933 KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3934 r->r.r_affinity_assigned = FALSE;
3935 }
3936}
3937#else /* KMP_AFFINITY_SUPPORTED */
3938#define __kmp_assign_root_init_mask() /* Nothing */
3939static inline void __kmp_reset_root_init_mask(int gtid) {}
3940#endif /* KMP_AFFINITY_SUPPORTED */
3941// No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3942// format string is for affinity, so platforms that do not support
3943// affinity can still use the other fields, e.g., %n for num_threads
3944extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3945 kmp_str_buf_t *buffer);
3946extern void __kmp_aux_display_affinity(int gtid, const char *format);
3947
3948extern void __kmp_cleanup_hierarchy();
3949extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3950
3951#if KMP_USE_FUTEX
3952
3953extern int __kmp_futex_determine_capable(void);
3954
3955#endif // KMP_USE_FUTEX
3956
3957extern void __kmp_gtid_set_specific(int gtid);
3958extern int __kmp_gtid_get_specific(void);
3959
3960extern double __kmp_read_cpu_time(void);
3961
3962extern int __kmp_read_system_info(struct kmp_sys_info *info);
3963
3964#if KMP_USE_MONITOR
3965extern void __kmp_create_monitor(kmp_info_t *th);
3966#endif
3967
3968extern void *__kmp_launch_thread(kmp_info_t *thr);
3969
3970extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3971
3972#if KMP_OS_WINDOWS
3973extern int __kmp_still_running(kmp_info_t *th);
3974extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3975extern void __kmp_free_handle(kmp_thread_t tHandle);
3976#endif
3977
3978#if KMP_USE_MONITOR
3979extern void __kmp_reap_monitor(kmp_info_t *th);
3980#endif
3981extern void __kmp_reap_worker(kmp_info_t *th);
3982extern void __kmp_terminate_thread(int gtid);
3983
3984extern int __kmp_try_suspend_mx(kmp_info_t *th);
3985extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3986extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3987
3988extern void __kmp_elapsed(double *);
3989extern void __kmp_elapsed_tick(double *);
3990
3991extern void __kmp_enable(int old_state);
3992extern void __kmp_disable(int *old_state);
3993
3994extern void __kmp_thread_sleep(int millis);
3995
3996extern void __kmp_common_initialize(void);
3997extern void __kmp_common_destroy(void);
3998extern void __kmp_common_destroy_gtid(int gtid);
3999
4000#if KMP_OS_UNIX
4001extern void __kmp_register_atfork(void);
4002#endif
4003extern void __kmp_suspend_initialize(void);
4004extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
4005extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
4006
4007extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4008 int tid);
4009extern kmp_team_t *
4010__kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4011#if OMPT_SUPPORT
4012 ompt_data_t ompt_parallel_data,
4013#endif
4014 kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
4015 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
4016extern void __kmp_free_thread(kmp_info_t *);
4017extern void __kmp_free_team(kmp_root_t *,
4018 kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
4019extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
4020
4021/* ------------------------------------------------------------------------ */
4022
4023extern void __kmp_initialize_bget(kmp_info_t *th);
4024extern void __kmp_finalize_bget(kmp_info_t *th);
4025
4026KMP_EXPORT void *kmpc_malloc(size_t size);
4027KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
4028KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
4029KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
4030KMP_EXPORT void kmpc_free(void *ptr);
4031
4032/* declarations for internal use */
4033
4034extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
4035 size_t reduce_size, void *reduce_data,
4036 void (*reduce)(void *, void *));
4037extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
4038extern int __kmp_barrier_gomp_cancel(int gtid);
4039
4044enum fork_context_e {
4045 fork_context_gnu,
4047 fork_context_intel,
4048 fork_context_last
4049};
4050extern int __kmp_fork_call(ident_t *loc, int gtid,
4051 enum fork_context_e fork_context, kmp_int32 argc,
4052 microtask_t microtask, launch_t invoker,
4053 kmp_va_list ap);
4054
4055extern void __kmp_join_call(ident_t *loc, int gtid
4056#if OMPT_SUPPORT
4057 ,
4058 enum fork_context_e fork_context
4059#endif
4060 ,
4061 int exit_teams = 0);
4062
4063extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
4064extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
4065extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
4066extern int __kmp_invoke_task_func(int gtid);
4067extern void __kmp_run_before_invoked_task(int gtid, int tid,
4068 kmp_info_t *this_thr,
4069 kmp_team_t *team);
4070extern void __kmp_run_after_invoked_task(int gtid, int tid,
4071 kmp_info_t *this_thr,
4072 kmp_team_t *team);
4073
4074// should never have been exported
4075KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
4076extern int __kmp_invoke_teams_master(int gtid);
4077extern void __kmp_teams_master(int gtid);
4078extern int __kmp_aux_get_team_num();
4079extern int __kmp_aux_get_num_teams();
4080extern void __kmp_save_internal_controls(kmp_info_t *thread);
4081extern void __kmp_user_set_library(enum library_type arg);
4082extern void __kmp_aux_set_library(enum library_type arg);
4083extern void __kmp_aux_set_stacksize(size_t arg);
4084extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
4085extern void __kmp_aux_set_defaults(char const *str, size_t len);
4086
4087/* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
4088void kmpc_set_blocktime(int arg);
4089void ompc_set_nested(int flag);
4090void ompc_set_dynamic(int flag);
4091void ompc_set_num_threads(int arg);
4092
4093extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
4094 kmp_team_t *team, int tid);
4095extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
4096extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4097 kmp_tasking_flags_t *flags,
4098 size_t sizeof_kmp_task_t,
4099 size_t sizeof_shareds,
4100 kmp_routine_entry_t task_entry);
4101extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
4102 kmp_team_t *team, int tid,
4103 int set_curr_task);
4104extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
4105extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
4106
4107extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
4108 int gtid,
4109 kmp_task_t *task);
4110extern void __kmp_fulfill_event(kmp_event_t *event);
4111
4112extern void __kmp_free_task_team(kmp_info_t *thread,
4113 kmp_task_team_t *task_team);
4114extern void __kmp_reap_task_teams(void);
4115extern void __kmp_wait_to_unref_task_teams(void);
4116extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
4117 int always);
4118extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
4119extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
4120#if USE_ITT_BUILD
4121 ,
4122 void *itt_sync_obj
4123#endif /* USE_ITT_BUILD */
4124 ,
4125 int wait = 1);
4126extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
4127 int gtid);
4128
4129extern int __kmp_is_address_mapped(void *addr);
4130extern kmp_uint64 __kmp_hardware_timestamp(void);
4131
4132#if KMP_OS_UNIX
4133extern int __kmp_read_from_file(char const *path, char const *format, ...);
4134#endif
4135
4136/* ------------------------------------------------------------------------ */
4137//
4138// Assembly routines that have no compiler intrinsic replacement
4139//
4140
4141extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
4142 void *argv[]
4143#if OMPT_SUPPORT
4144 ,
4145 void **exit_frame_ptr
4146#endif
4147);
4148
4149/* ------------------------------------------------------------------------ */
4150
4151KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
4152KMP_EXPORT void __kmpc_end(ident_t *);
4153
4154KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
4155 kmpc_ctor_vec ctor,
4156 kmpc_cctor_vec cctor,
4157 kmpc_dtor_vec dtor,
4158 size_t vector_length);
4159KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
4160 kmpc_ctor ctor, kmpc_cctor cctor,
4161 kmpc_dtor dtor);
4162KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
4163 void *data, size_t size);
4164
4165KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
4166KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
4167KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
4168KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
4169
4170KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
4171KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
4172 kmpc_micro microtask, ...);
4173KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs,
4174 kmpc_micro microtask, kmp_int32 cond,
4175 void *args);
4176
4177KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
4178KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
4179
4180KMP_EXPORT void __kmpc_flush(ident_t *);
4181KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
4182KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
4183KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
4184KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
4185 kmp_int32 filter);
4186KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
4187KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
4188KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
4189KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
4190 kmp_critical_name *);
4191KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
4192 kmp_critical_name *);
4193KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
4194 kmp_critical_name *, uint32_t hint);
4195
4196KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
4197KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
4198
4199KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
4200 kmp_int32 global_tid);
4201
4202KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
4203KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
4204
4205KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid);
4206KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid,
4207 kmp_int32 numberOfSections);
4208KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid);
4209
4210KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
4211 kmp_int32 schedtype, kmp_int32 *plastiter,
4212 kmp_int *plower, kmp_int *pupper,
4213 kmp_int *pstride, kmp_int incr,
4214 kmp_int chunk);
4215
4216KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
4217
4218KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
4219 size_t cpy_size, void *cpy_data,
4220 void (*cpy_func)(void *, void *),
4221 kmp_int32 didit);
4222
4223KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid,
4224 void *cpy_data);
4225
4226extern void KMPC_SET_NUM_THREADS(int arg);
4227extern void KMPC_SET_DYNAMIC(int flag);
4228extern void KMPC_SET_NESTED(int flag);
4229
4230/* OMP 3.0 tasking interface routines */
4231KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
4232 kmp_task_t *new_task);
4233KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4234 kmp_int32 flags,
4235 size_t sizeof_kmp_task_t,
4236 size_t sizeof_shareds,
4237 kmp_routine_entry_t task_entry);
4238KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
4239 ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
4240 size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
4241KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
4242 kmp_task_t *task);
4243KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
4244 kmp_task_t *task);
4245KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
4246 kmp_task_t *new_task);
4247KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
4248KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
4249 int end_part);
4250
4251#if TASK_UNUSED
4252void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
4253void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
4254 kmp_task_t *task);
4255#endif // TASK_UNUSED
4256
4257/* ------------------------------------------------------------------------ */
4258
4259KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
4260KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
4261
4262KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
4263 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
4264 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
4265 kmp_depend_info_t *noalias_dep_list);
4266
4267KMP_EXPORT kmp_base_depnode_t *__kmpc_task_get_depnode(kmp_task_t *task);
4268
4269KMP_EXPORT kmp_depnode_list_t *__kmpc_task_get_successors(kmp_task_t *task);
4270
4271KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
4272 kmp_int32 ndeps,
4273 kmp_depend_info_t *dep_list,
4274 kmp_int32 ndeps_noalias,
4275 kmp_depend_info_t *noalias_dep_list);
4276/* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause.
4277 * Placeholder for taskwait with nowait clause.*/
4278KMP_EXPORT void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid,
4279 kmp_int32 ndeps,
4280 kmp_depend_info_t *dep_list,
4281 kmp_int32 ndeps_noalias,
4282 kmp_depend_info_t *noalias_dep_list,
4283 kmp_int32 has_no_wait);
4284
4285extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
4286 bool serialize_immediate);
4287
4288KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
4289 kmp_int32 cncl_kind);
4290KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
4291 kmp_int32 cncl_kind);
4292KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
4293KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
4294
4295KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
4296KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
4297KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
4298 kmp_int32 if_val, kmp_uint64 *lb,
4299 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
4300 kmp_int32 sched, kmp_uint64 grainsize,
4301 void *task_dup);
4302KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
4303 kmp_task_t *task, kmp_int32 if_val,
4304 kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4305 kmp_int32 nogroup, kmp_int32 sched,
4306 kmp_uint64 grainsize, kmp_int32 modifier,
4307 void *task_dup);
4308KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
4309KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
4310KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
4311KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
4312 int is_ws, int num,
4313 void *data);
4314KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
4315 int num, void *data);
4316KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
4317 int is_ws);
4318KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
4319 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4320 kmp_task_affinity_info_t *affin_list);
4321KMP_EXPORT void __kmp_set_num_teams(int num_teams);
4322KMP_EXPORT int __kmp_get_max_teams(void);
4323KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
4324KMP_EXPORT int __kmp_get_teams_thread_limit(void);
4325
4326/* Interface target task integration */
4327KMP_EXPORT void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid);
4328KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid);
4329
4330/* Lock interface routines (fast versions with gtid passed in) */
4331KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
4332 void **user_lock);
4333KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
4334 void **user_lock);
4335KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
4336 void **user_lock);
4337KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
4338 void **user_lock);
4339KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4340KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
4341 void **user_lock);
4342KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
4343 void **user_lock);
4344KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
4345 void **user_lock);
4346KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4347KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
4348 void **user_lock);
4349
4350KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4351 void **user_lock, uintptr_t hint);
4352KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4353 void **user_lock,
4354 uintptr_t hint);
4355
4356#if OMPX_TASKGRAPH
4357// Taskgraph's Record & Replay mechanism
4358// __kmp_tdg_is_recording: check whether a given TDG is recording
4359// status: the tdg's current status
4360static inline bool __kmp_tdg_is_recording(kmp_tdg_status_t status) {
4361 return status == KMP_TDG_RECORDING;
4362}
4363
4364KMP_EXPORT kmp_int32 __kmpc_start_record_task(ident_t *loc, kmp_int32 gtid,
4365 kmp_int32 input_flags,
4366 kmp_int32 tdg_id);
4367KMP_EXPORT void __kmpc_end_record_task(ident_t *loc, kmp_int32 gtid,
4368 kmp_int32 input_flags, kmp_int32 tdg_id);
4369#endif
4370/* Interface to fast scalable reduce methods routines */
4371
4372KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
4373 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4374 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4375 kmp_critical_name *lck);
4376KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4377 kmp_critical_name *lck);
4378KMP_EXPORT kmp_int32 __kmpc_reduce(
4379 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4380 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4381 kmp_critical_name *lck);
4382KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4383 kmp_critical_name *lck);
4384
4385/* Internal fast reduction routines */
4386
4387extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4388 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4389 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4390 kmp_critical_name *lck);
4391
4392// this function is for testing set/get/determine reduce method
4393KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4394
4395KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4396KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4397
4398// C++ port
4399// missing 'extern "C"' declarations
4400
4401KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4402KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4403KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4404 kmp_int32 num_threads);
4405
4406KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4407 int proc_bind);
4408KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4409 kmp_int32 num_teams,
4410 kmp_int32 num_threads);
4411KMP_EXPORT void __kmpc_set_thread_limit(ident_t *loc, kmp_int32 global_tid,
4412 kmp_int32 thread_limit);
4413/* Function for OpenMP 5.1 num_teams clause */
4414KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4415 kmp_int32 num_teams_lb,
4416 kmp_int32 num_teams_ub,
4417 kmp_int32 num_threads);
4418KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4419 kmpc_micro microtask, ...);
4420struct kmp_dim { // loop bounds info casted to kmp_int64
4421 kmp_int64 lo; // lower
4422 kmp_int64 up; // upper
4423 kmp_int64 st; // stride
4424};
4425KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4426 kmp_int32 num_dims,
4427 const struct kmp_dim *dims);
4428KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4429 const kmp_int64 *vec);
4430KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4431 const kmp_int64 *vec);
4432KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4433
4434KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4435 void *data, size_t size,
4436 void ***cache);
4437
4438// The routines below are not exported.
4439// Consider making them 'static' in corresponding source files.
4440void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4441 void *data_addr, size_t pc_size);
4442struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4443 void *data_addr,
4444 size_t pc_size);
4445void __kmp_threadprivate_resize_cache(int newCapacity);
4446void __kmp_cleanup_threadprivate_caches();
4447
4448// ompc_, kmpc_ entries moved from omp.h.
4449#if KMP_OS_WINDOWS
4450#define KMPC_CONVENTION __cdecl
4451#else
4452#define KMPC_CONVENTION
4453#endif
4454
4455#ifndef __OMP_H
4456typedef enum omp_sched_t {
4457 omp_sched_static = 1,
4458 omp_sched_dynamic = 2,
4459 omp_sched_guided = 3,
4460 omp_sched_auto = 4
4461} omp_sched_t;
4462typedef void *kmp_affinity_mask_t;
4463#endif
4464
4465KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4466KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4467KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4468KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4469KMP_EXPORT int KMPC_CONVENTION
4470kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4471KMP_EXPORT int KMPC_CONVENTION
4472kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4473KMP_EXPORT int KMPC_CONVENTION
4474kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4475
4476KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4477KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4478KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4479KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4480KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4481void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4482size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4483void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4484size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4485 char const *format);
4486
4487enum kmp_target_offload_kind {
4488 tgt_disabled = 0,
4489 tgt_default = 1,
4490 tgt_mandatory = 2
4491};
4492typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4493// Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4494extern kmp_target_offload_kind_t __kmp_target_offload;
4495extern int __kmpc_get_target_offload();
4496
4497// Constants used in libomptarget
4498#define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4499#define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4500
4501// OMP Pause Resource
4502
4503// The following enum is used both to set the status in __kmp_pause_status, and
4504// as the internal equivalent of the externally-visible omp_pause_resource_t.
4505typedef enum kmp_pause_status_t {
4506 kmp_not_paused = 0, // status is not paused, or, requesting resume
4507 kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4508 kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4509} kmp_pause_status_t;
4510
4511// This stores the pause state of the runtime
4512extern kmp_pause_status_t __kmp_pause_status;
4513extern int __kmpc_pause_resource(kmp_pause_status_t level);
4514extern int __kmp_pause_resource(kmp_pause_status_t level);
4515// Soft resume sets __kmp_pause_status, and wakes up all threads.
4516extern void __kmp_resume_if_soft_paused();
4517// Hard resume simply resets the status to not paused. Library will appear to
4518// be uninitialized after hard pause. Let OMP constructs trigger required
4519// initializations.
4520static inline void __kmp_resume_if_hard_paused() {
4521 if (__kmp_pause_status == kmp_hard_paused) {
4522 __kmp_pause_status = kmp_not_paused;
4523 }
4524}
4525
4526extern void __kmp_omp_display_env(int verbose);
4527
4528// 1: it is initializing hidden helper team
4529extern volatile int __kmp_init_hidden_helper;
4530// 1: the hidden helper team is done
4531extern volatile int __kmp_hidden_helper_team_done;
4532// 1: enable hidden helper task
4533extern kmp_int32 __kmp_enable_hidden_helper;
4534// Main thread of hidden helper team
4535extern kmp_info_t *__kmp_hidden_helper_main_thread;
4536// Descriptors for the hidden helper threads
4537extern kmp_info_t **__kmp_hidden_helper_threads;
4538// Number of hidden helper threads
4539extern kmp_int32 __kmp_hidden_helper_threads_num;
4540// Number of hidden helper tasks that have not been executed yet
4541extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4542
4543extern void __kmp_hidden_helper_initialize();
4544extern void __kmp_hidden_helper_threads_initz_routine();
4545extern void __kmp_do_initialize_hidden_helper_threads();
4546extern void __kmp_hidden_helper_threads_initz_wait();
4547extern void __kmp_hidden_helper_initz_release();
4548extern void __kmp_hidden_helper_threads_deinitz_wait();
4549extern void __kmp_hidden_helper_threads_deinitz_release();
4550extern void __kmp_hidden_helper_main_thread_wait();
4551extern void __kmp_hidden_helper_worker_thread_wait();
4552extern void __kmp_hidden_helper_worker_thread_signal();
4553extern void __kmp_hidden_helper_main_thread_release();
4554
4555// Check whether a given thread is a hidden helper thread
4556#define KMP_HIDDEN_HELPER_THREAD(gtid) \
4557 ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4558
4559#define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4560 ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4561
4562#define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid) \
4563 ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4564
4565#define KMP_HIDDEN_HELPER_TEAM(team) \
4566 (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4567
4568// Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4569// main thread, is skipped.
4570#define KMP_GTID_TO_SHADOW_GTID(gtid) \
4571 ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4572
4573// Return the adjusted gtid value by subtracting from gtid the number
4574// of hidden helper threads. This adjusted value is the gtid the thread would
4575// have received if there were no hidden helper threads.
4576static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4577 int adjusted_gtid = gtid;
4578 if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4579 gtid - __kmp_hidden_helper_threads_num >= 0) {
4580 adjusted_gtid -= __kmp_hidden_helper_threads_num;
4581 }
4582 return adjusted_gtid;
4583}
4584
4585// Support for error directive
4586typedef enum kmp_severity_t {
4587 severity_warning = 1,
4588 severity_fatal = 2
4589} kmp_severity_t;
4590extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4591
4592// Support for scope directive
4593KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4594KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4595
4596#ifdef __cplusplus
4597}
4598#endif
4599
4600template <bool C, bool S>
4601extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4602template <bool C, bool S>
4603extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4604template <bool C, bool S>
4605extern void __kmp_atomic_suspend_64(int th_gtid,
4606 kmp_atomic_flag_64<C, S> *flag);
4607extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4608#if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4609template <bool C, bool S>
4610extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4611template <bool C, bool S>
4612extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4613template <bool C, bool S>
4614extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4615extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4616#endif
4617template <bool C, bool S>
4618extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4619template <bool C, bool S>
4620extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4621template <bool C, bool S>
4622extern void __kmp_atomic_resume_64(int target_gtid,
4623 kmp_atomic_flag_64<C, S> *flag);
4624extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4625
4626template <bool C, bool S>
4627int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4628 kmp_flag_32<C, S> *flag, int final_spin,
4629 int *thread_finished,
4630#if USE_ITT_BUILD
4631 void *itt_sync_obj,
4632#endif /* USE_ITT_BUILD */
4633 kmp_int32 is_constrained);
4634template <bool C, bool S>
4635int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4636 kmp_flag_64<C, S> *flag, int final_spin,
4637 int *thread_finished,
4638#if USE_ITT_BUILD
4639 void *itt_sync_obj,
4640#endif /* USE_ITT_BUILD */
4641 kmp_int32 is_constrained);
4642template <bool C, bool S>
4643int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4644 kmp_atomic_flag_64<C, S> *flag,
4645 int final_spin, int *thread_finished,
4646#if USE_ITT_BUILD
4647 void *itt_sync_obj,
4648#endif /* USE_ITT_BUILD */
4649 kmp_int32 is_constrained);
4650int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4651 kmp_flag_oncore *flag, int final_spin,
4652 int *thread_finished,
4653#if USE_ITT_BUILD
4654 void *itt_sync_obj,
4655#endif /* USE_ITT_BUILD */
4656 kmp_int32 is_constrained);
4657
4658extern int __kmp_nesting_mode;
4659extern int __kmp_nesting_mode_nlevels;
4660extern int *__kmp_nesting_nth_level;
4661extern void __kmp_init_nesting_mode();
4662extern void __kmp_set_nesting_mode_threads();
4663
4671 FILE *f;
4672
4673 void close() {
4674 if (f && f != stdout && f != stderr) {
4675 fclose(f);
4676 f = nullptr;
4677 }
4678 }
4679
4680public:
4681 kmp_safe_raii_file_t() : f(nullptr) {}
4682 kmp_safe_raii_file_t(const char *filename, const char *mode,
4683 const char *env_var = nullptr)
4684 : f(nullptr) {
4685 open(filename, mode, env_var);
4686 }
4687 ~kmp_safe_raii_file_t() { close(); }
4688
4692 void open(const char *filename, const char *mode,
4693 const char *env_var = nullptr) {
4694 KMP_ASSERT(!f);
4695 f = fopen(filename, mode);
4696 if (!f) {
4697 int code = errno;
4698 if (env_var) {
4699 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4700 KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4701 } else {
4702 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4703 __kmp_msg_null);
4704 }
4705 }
4706 }
4709 int try_open(const char *filename, const char *mode) {
4710 KMP_ASSERT(!f);
4711 f = fopen(filename, mode);
4712 if (!f)
4713 return errno;
4714 return 0;
4715 }
4718 void set_stdout() {
4719 KMP_ASSERT(!f);
4720 f = stdout;
4721 }
4724 void set_stderr() {
4725 KMP_ASSERT(!f);
4726 f = stderr;
4727 }
4728 operator bool() { return bool(f); }
4729 operator FILE *() { return f; }
4730};
4731
4732template <typename SourceType, typename TargetType,
4733 bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4734 bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4735 bool isSourceSigned = std::is_signed<SourceType>::value,
4736 bool isTargetSigned = std::is_signed<TargetType>::value>
4737struct kmp_convert {};
4738
4739// Both types are signed; Source smaller
4740template <typename SourceType, typename TargetType>
4741struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4742 static TargetType to(SourceType src) { return (TargetType)src; }
4743};
4744// Source equal
4745template <typename SourceType, typename TargetType>
4746struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4747 static TargetType to(SourceType src) { return src; }
4748};
4749// Source bigger
4750template <typename SourceType, typename TargetType>
4751struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4752 static TargetType to(SourceType src) {
4753 KMP_ASSERT(src <= static_cast<SourceType>(
4754 (std::numeric_limits<TargetType>::max)()));
4755 KMP_ASSERT(src >= static_cast<SourceType>(
4756 (std::numeric_limits<TargetType>::min)()));
4757 return (TargetType)src;
4758 }
4759};
4760
4761// Source signed, Target unsigned
4762// Source smaller
4763template <typename SourceType, typename TargetType>
4764struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4765 static TargetType to(SourceType src) {
4766 KMP_ASSERT(src >= 0);
4767 return (TargetType)src;
4768 }
4769};
4770// Source equal
4771template <typename SourceType, typename TargetType>
4772struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4773 static TargetType to(SourceType src) {
4774 KMP_ASSERT(src >= 0);
4775 return (TargetType)src;
4776 }
4777};
4778// Source bigger
4779template <typename SourceType, typename TargetType>
4780struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4781 static TargetType to(SourceType src) {
4782 KMP_ASSERT(src >= 0);
4783 KMP_ASSERT(src <= static_cast<SourceType>(
4784 (std::numeric_limits<TargetType>::max)()));
4785 return (TargetType)src;
4786 }
4787};
4788
4789// Source unsigned, Target signed
4790// Source smaller
4791template <typename SourceType, typename TargetType>
4792struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4793 static TargetType to(SourceType src) { return (TargetType)src; }
4794};
4795// Source equal
4796template <typename SourceType, typename TargetType>
4797struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4798 static TargetType to(SourceType src) {
4799 KMP_ASSERT(src <= static_cast<SourceType>(
4800 (std::numeric_limits<TargetType>::max)()));
4801 return (TargetType)src;
4802 }
4803};
4804// Source bigger
4805template <typename SourceType, typename TargetType>
4806struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4807 static TargetType to(SourceType src) {
4808 KMP_ASSERT(src <= static_cast<SourceType>(
4809 (std::numeric_limits<TargetType>::max)()));
4810 return (TargetType)src;
4811 }
4812};
4813
4814// Source unsigned, Target unsigned
4815// Source smaller
4816template <typename SourceType, typename TargetType>
4817struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4818 static TargetType to(SourceType src) { return (TargetType)src; }
4819};
4820// Source equal
4821template <typename SourceType, typename TargetType>
4822struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4823 static TargetType to(SourceType src) { return src; }
4824};
4825// Source bigger
4826template <typename SourceType, typename TargetType>
4827struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4828 static TargetType to(SourceType src) {
4829 KMP_ASSERT(src <= static_cast<SourceType>(
4830 (std::numeric_limits<TargetType>::max)()));
4831 return (TargetType)src;
4832 }
4833};
4834
4835template <typename T1, typename T2>
4836static inline void __kmp_type_convert(T1 src, T2 *dest) {
4837 *dest = kmp_convert<T1, T2>::to(src);
4838}
4839
4840#endif /* KMP_H */
void set_stdout()
Definition kmp.h:4718
void set_stderr()
Definition kmp.h:4724
int try_open(const char *filename, const char *mode)
Definition kmp.h:4709
void open(const char *filename, const char *mode, const char *env_var=nullptr)
Definition kmp.h:4692
struct ident ident_t
@ KMP_IDENT_KMPC
Definition kmp.h:208
@ KMP_IDENT_IMB
Definition kmp.h:206
@ KMP_IDENT_WORK_LOOP
Definition kmp.h:226
@ KMP_IDENT_BARRIER_IMPL
Definition kmp.h:217
@ KMP_IDENT_WORK_SECTIONS
Definition kmp.h:228
@ KMP_IDENT_AUTOPAR
Definition kmp.h:211
@ KMP_IDENT_ATOMIC_HINT_MASK
Definition kmp.h:235
@ KMP_IDENT_WORK_DISTRIBUTE
Definition kmp.h:230
@ KMP_IDENT_BARRIER_EXPL
Definition kmp.h:215
@ KMP_IDENT_ATOMIC_REDUCE
Definition kmp.h:213
KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *)
KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro microtask,...)
KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs, kmpc_micro microtask, kmp_int32 cond, void *args)
KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_set_thread_limit(ident_t *loc, kmp_int32 global_tid, kmp_int32 thread_limit)
KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_push_num_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:1742
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)
KMP_EXPORT void * __kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
KMP_EXPORT void * __kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid)
KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask)
KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, int is_ws)
KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins, kmp_task_affinity_info_t *affin_list)
KMP_EXPORT void * __kmpc_task_reduction_init(int gtid, int num_data, void *data)
KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask)
KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
KMP_EXPORT void * __kmpc_taskred_init(int gtid, int num_data, void *data)
KMP_EXPORT void ** __kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid)
void(* kmpc_dtor)(void *)
Definition kmp.h:1766
void *(* kmpc_cctor)(void *, void *)
Definition kmp.h:1773
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:1795
void *(* kmpc_ctor)(void *)
Definition kmp.h:1760
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:1783
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:1789
KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data, kmpc_ctor_vec ctor, kmpc_cctor_vec cctor, kmpc_dtor_vec dtor, size_t vector_length)
KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc)
KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *)
KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid)
sched_type
Definition kmp.h:369
KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid, kmp_int32 numberOfSections)
KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid, kmp_int32 filter)
void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_int32 lb, kmp_int32 ub, kmp_int32 st, kmp_int32 chunk)
KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
@ kmp_nm_guided_chunked
Definition kmp.h:420
@ kmp_sch_runtime_simd
Definition kmp.h:391
@ kmp_nm_ord_auto
Definition kmp.h:439
@ kmp_sch_auto
Definition kmp.h:376
@ kmp_nm_auto
Definition kmp.h:422
@ kmp_distribute_static_chunked
Definition kmp.h:407
@ kmp_sch_static
Definition kmp.h:372
@ kmp_sch_guided_simd
Definition kmp.h:390
@ kmp_sch_modifier_monotonic
Definition kmp.h:457
@ kmp_sch_default
Definition kmp.h:477
@ kmp_sch_modifier_nonmonotonic
Definition kmp.h:459
@ kmp_nm_ord_static
Definition kmp.h:435
@ kmp_distribute_static
Definition kmp.h:408
@ kmp_sch_guided_chunked
Definition kmp.h:374
@ kmp_nm_static
Definition kmp.h:418
@ kmp_sch_lower
Definition kmp.h:370
@ kmp_nm_upper
Definition kmp.h:441
@ kmp_ord_lower
Definition kmp.h:396
@ kmp_ord_static
Definition kmp.h:398
@ kmp_sch_upper
Definition kmp.h:394
@ kmp_ord_upper
Definition kmp.h:404
@ kmp_nm_lower
Definition kmp.h:414
@ kmp_ord_auto
Definition kmp.h:402
Definition kmp.h:246
kmp_int32 reserved_1
Definition kmp.h:247
char const * psource
Definition kmp.h:256
kmp_int32 reserved_2
Definition kmp.h:250
kmp_int32 reserved_3
Definition kmp.h:255
kmp_int32 flags
Definition kmp.h:248