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
31#define TASK_CURRENT_NOT_QUEUED 0
32#define TASK_CURRENT_QUEUED 1
33
34#ifdef BUILD_TIED_TASK_STACK
35#define TASK_STACK_EMPTY 0 // entries when the stack is empty
36#define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
37// Number of entries in each task stack array
38#define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
39// Mask for determining index into stack block
40#define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
41#endif // BUILD_TIED_TASK_STACK
42
43#define TASK_NOT_PUSHED 1
44#define TASK_SUCCESSFULLY_PUSHED 0
45#define TASK_TIED 1
46#define TASK_UNTIED 0
47#define TASK_EXPLICIT 1
48#define TASK_IMPLICIT 0
49#define TASK_PROXY 1
50#define TASK_FULL 0
51#define TASK_DETACHABLE 1
52#define TASK_UNDETACHABLE 0
53
54#define KMP_CANCEL_THREADS
55#define KMP_THREAD_ATTR
56
57// Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being
58// built on Android
59#if defined(__ANDROID__)
60#undef KMP_CANCEL_THREADS
61#endif
62
63#include <signal.h>
64#include <stdarg.h>
65#include <stddef.h>
66#include <stdio.h>
67#include <stdlib.h>
68#include <string.h>
69#include <limits>
70#include <type_traits>
71/* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
72 Microsoft library. Some macros provided below to replace these functions */
73#ifndef __ABSOFT_WIN
74#include <sys/types.h>
75#endif
76#include <limits.h>
77#include <time.h>
78
79#include <errno.h>
80
81#include "kmp_os.h"
82
83#include "kmp_safe_c_api.h"
84
85#if KMP_STATS_ENABLED
86class kmp_stats_list;
87#endif
88
89#if KMP_USE_HIER_SCHED
90// Only include hierarchical scheduling if affinity is supported
91#undef KMP_USE_HIER_SCHED
92#define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
93#endif
94
95#if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
96#include "hwloc.h"
97#ifndef HWLOC_OBJ_NUMANODE
98#define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
99#endif
100#ifndef HWLOC_OBJ_PACKAGE
101#define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
102#endif
103#endif
104
105#if KMP_ARCH_X86 || KMP_ARCH_X86_64
106#include <xmmintrin.h>
107#endif
108
109// The below has to be defined before including "kmp_barrier.h".
110#define KMP_INTERNAL_MALLOC(sz) malloc(sz)
111#define KMP_INTERNAL_FREE(p) free(p)
112#define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
113#define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
114
115#include "kmp_debug.h"
116#include "kmp_lock.h"
117#include "kmp_version.h"
118#include "kmp_barrier.h"
119#if USE_DEBUGGER
120#include "kmp_debugger.h"
121#endif
122#include "kmp_i18n.h"
123
124#define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
125
126#include "kmp_wrapper_malloc.h"
127#if KMP_OS_UNIX
128#include <unistd.h>
129#if !defined NSIG && defined _NSIG
130#define NSIG _NSIG
131#endif
132#endif
133
134#if KMP_OS_LINUX
135#pragma weak clock_gettime
136#endif
137
138#if OMPT_SUPPORT
139#include "ompt-internal.h"
140#endif
141
142#if OMPD_SUPPORT
143#include "ompd-specific.h"
144#endif
145
146#ifndef UNLIKELY
147#define UNLIKELY(x) (x)
148#endif
149
150// Affinity format function
151#include "kmp_str.h"
152
153// 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
154// 3 - fast allocation using sync, non-sync free lists of any size, non-self
155// free lists of limited size.
156#ifndef USE_FAST_MEMORY
157#define USE_FAST_MEMORY 3
158#endif
159
160#ifndef KMP_NESTED_HOT_TEAMS
161#define KMP_NESTED_HOT_TEAMS 0
162#define USE_NESTED_HOT_ARG(x)
163#else
164#if KMP_NESTED_HOT_TEAMS
165#define USE_NESTED_HOT_ARG(x) , x
166#else
167#define USE_NESTED_HOT_ARG(x)
168#endif
169#endif
170
171// Assume using BGET compare_exchange instruction instead of lock by default.
172#ifndef USE_CMP_XCHG_FOR_BGET
173#define USE_CMP_XCHG_FOR_BGET 1
174#endif
175
176// Test to see if queuing lock is better than bootstrap lock for bget
177// #ifndef USE_QUEUING_LOCK_FOR_BGET
178// #define USE_QUEUING_LOCK_FOR_BGET
179// #endif
180
181#define KMP_NSEC_PER_SEC 1000000000L
182#define KMP_USEC_PER_SEC 1000000L
183#define KMP_NSEC_PER_USEC 1000L
184
193enum {
198 /* 0x04 is no longer used */
207 KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
208 KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
209 KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
210
211 KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
212 KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
213
225 KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
226 KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
227 KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
228 KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
229 KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
230};
231
235typedef struct ident {
236 kmp_int32 reserved_1;
237 kmp_int32 flags;
239 kmp_int32 reserved_2;
240#if USE_ITT_BUILD
241/* but currently used for storing region-specific ITT */
242/* contextual information. */
243#endif /* USE_ITT_BUILD */
244 kmp_int32 reserved_3;
245 char const *psource;
249 // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
250 kmp_int32 get_openmp_version() {
251 return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
252 }
258// Some forward declarations.
259typedef union kmp_team kmp_team_t;
260typedef struct kmp_taskdata kmp_taskdata_t;
261typedef union kmp_task_team kmp_task_team_t;
262typedef union kmp_team kmp_team_p;
263typedef union kmp_info kmp_info_p;
264typedef union kmp_root kmp_root_p;
265
266template <bool C = false, bool S = true> class kmp_flag_32;
267template <bool C = false, bool S = true> class kmp_flag_64;
268template <bool C = false, bool S = true> class kmp_atomic_flag_64;
269class kmp_flag_oncore;
270
271#ifdef __cplusplus
272extern "C" {
273#endif
274
275/* ------------------------------------------------------------------------ */
276
277/* Pack two 32-bit signed integers into a 64-bit signed integer */
278/* ToDo: Fix word ordering for big-endian machines. */
279#define KMP_PACK_64(HIGH_32, LOW_32) \
280 ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
281
282// Generic string manipulation macros. Assume that _x is of type char *
283#define SKIP_WS(_x) \
284 { \
285 while (*(_x) == ' ' || *(_x) == '\t') \
286 (_x)++; \
287 }
288#define SKIP_DIGITS(_x) \
289 { \
290 while (*(_x) >= '0' && *(_x) <= '9') \
291 (_x)++; \
292 }
293#define SKIP_TOKEN(_x) \
294 { \
295 while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
296 (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
297 (_x)++; \
298 }
299#define SKIP_TO(_x, _c) \
300 { \
301 while (*(_x) != '\0' && *(_x) != (_c)) \
302 (_x)++; \
303 }
304
305/* ------------------------------------------------------------------------ */
306
307#define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
308#define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
309
310/* ------------------------------------------------------------------------ */
311/* Enumeration types */
312
313enum kmp_state_timer {
314 ts_stop,
315 ts_start,
316 ts_pause,
317
318 ts_last_state
319};
320
321enum dynamic_mode {
322 dynamic_default,
323#ifdef USE_LOAD_BALANCE
324 dynamic_load_balance,
325#endif /* USE_LOAD_BALANCE */
326 dynamic_random,
327 dynamic_thread_limit,
328 dynamic_max
329};
330
331/* external schedule constants, duplicate enum omp_sched in omp.h in order to
332 * not include it here */
333#ifndef KMP_SCHED_TYPE_DEFINED
334#define KMP_SCHED_TYPE_DEFINED
335typedef enum kmp_sched {
336 kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
337 // Note: need to adjust __kmp_sch_map global array in case enum is changed
338 kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
339 kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
340 kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
341 kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
342 kmp_sched_upper_std = 5, // upper bound for standard schedules
343 kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
344 kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
345#if KMP_STATIC_STEAL_ENABLED
346 kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
347#endif
348 kmp_sched_upper,
349 kmp_sched_default = kmp_sched_static, // default scheduling
350 kmp_sched_monotonic = 0x80000000
351} kmp_sched_t;
352#endif
353
358enum sched_type : kmp_int32 {
360 kmp_sch_static_chunked = 33,
362 kmp_sch_dynamic_chunked = 35,
364 kmp_sch_runtime = 37,
366 kmp_sch_trapezoidal = 39,
367
368 /* accessible only through KMP_SCHEDULE environment variable */
369 kmp_sch_static_greedy = 40,
370 kmp_sch_static_balanced = 41,
371 /* accessible only through KMP_SCHEDULE environment variable */
372 kmp_sch_guided_iterative_chunked = 42,
373 kmp_sch_guided_analytical_chunked = 43,
374 /* accessible only through KMP_SCHEDULE environment variable */
375 kmp_sch_static_steal = 44,
376
377 /* static with chunk adjustment (e.g., simd) */
378 kmp_sch_static_balanced_chunked = 45,
382 /* accessible only through KMP_SCHEDULE environment variable */
386 kmp_ord_static_chunked = 65,
388 kmp_ord_dynamic_chunked = 67,
389 kmp_ord_guided_chunked = 68,
390 kmp_ord_runtime = 69,
392 kmp_ord_trapezoidal = 71,
395 /* Schedules for Distribute construct */
399 /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
400 single iteration/chunk, even if the loop is serialized. For the schedule
401 types listed above, the entire iteration vector is returned if the loop is
402 serialized. This doesn't work for gcc/gcomp sections. */
405 kmp_nm_static_chunked =
406 (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
408 kmp_nm_dynamic_chunked = 163,
410 kmp_nm_runtime = 165,
412 kmp_nm_trapezoidal = 167,
413
414 /* accessible only through KMP_SCHEDULE environment variable */
415 kmp_nm_static_greedy = 168,
416 kmp_nm_static_balanced = 169,
417 /* accessible only through KMP_SCHEDULE environment variable */
418 kmp_nm_guided_iterative_chunked = 170,
419 kmp_nm_guided_analytical_chunked = 171,
420 kmp_nm_static_steal =
421 172, /* accessible only through OMP_SCHEDULE environment variable */
422
423 kmp_nm_ord_static_chunked = 193,
425 kmp_nm_ord_dynamic_chunked = 195,
426 kmp_nm_ord_guided_chunked = 196,
427 kmp_nm_ord_runtime = 197,
429 kmp_nm_ord_trapezoidal = 199,
432 /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
433 we need to distinguish the three possible cases (no modifier, monotonic
434 modifier, nonmonotonic modifier), we need separate bits for each modifier.
435 The absence of monotonic does not imply nonmonotonic, especially since 4.5
436 says that the behaviour of the "no modifier" case is implementation defined
437 in 4.5, but will become "nonmonotonic" in 5.0.
438
439 Since we're passing a full 32 bit value, we can use a couple of high bits
440 for these flags; out of paranoia we avoid the sign bit.
441
442 These modifiers can be or-ed into non-static schedules by the compiler to
443 pass the additional information. They will be stripped early in the
444 processing in __kmp_dispatch_init when setting up schedules, so most of the
445 code won't ever see schedules with these bits set. */
447 (1 << 29),
449 (1 << 30),
451#define SCHEDULE_WITHOUT_MODIFIERS(s) \
452 (enum sched_type)( \
454#define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
455#define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
456#define SCHEDULE_HAS_NO_MODIFIERS(s) \
457 (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
458#define SCHEDULE_GET_MODIFIERS(s) \
459 ((enum sched_type)( \
460 (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
461#define SCHEDULE_SET_MODIFIERS(s, m) \
462 (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
463#define SCHEDULE_NONMONOTONIC 0
464#define SCHEDULE_MONOTONIC 1
465
468
469// Apply modifiers on internal kind to standard kind
470static inline void
471__kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
472 enum sched_type internal_kind) {
473 if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
474 *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
475 }
476}
477
478// Apply modifiers on standard kind to internal kind
479static inline void
480__kmp_sched_apply_mods_intkind(kmp_sched_t kind,
481 enum sched_type *internal_kind) {
482 if ((int)kind & (int)kmp_sched_monotonic) {
483 *internal_kind = (enum sched_type)((int)*internal_kind |
485 }
486}
487
488// Get standard schedule without modifiers
489static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
490 return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
491}
492
493/* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
494typedef union kmp_r_sched {
495 struct {
496 enum sched_type r_sched_type;
497 int chunk;
498 };
499 kmp_int64 sched;
500} kmp_r_sched_t;
501
502extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
503// internal schedule types
504
505enum library_type {
506 library_none,
507 library_serial,
508 library_turnaround,
509 library_throughput
510};
511
512#if KMP_OS_LINUX
513enum clock_function_type {
514 clock_function_gettimeofday,
515 clock_function_clock_gettime
516};
517#endif /* KMP_OS_LINUX */
518
519#if KMP_MIC_SUPPORTED
520enum mic_type { non_mic, mic1, mic2, mic3, dummy };
521#endif
522
523/* -- fast reduction stuff ------------------------------------------------ */
524
525#undef KMP_FAST_REDUCTION_BARRIER
526#define KMP_FAST_REDUCTION_BARRIER 1
527
528#undef KMP_FAST_REDUCTION_CORE_DUO
529#if KMP_ARCH_X86 || KMP_ARCH_X86_64
530#define KMP_FAST_REDUCTION_CORE_DUO 1
531#endif
532
533enum _reduction_method {
534 reduction_method_not_defined = 0,
535 critical_reduce_block = (1 << 8),
536 atomic_reduce_block = (2 << 8),
537 tree_reduce_block = (3 << 8),
538 empty_reduce_block = (4 << 8)
539};
540
541// Description of the packed_reduction_method variable:
542// The packed_reduction_method variable consists of two enum types variables
543// that are packed together into 0-th byte and 1-st byte:
544// 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
545// barrier that will be used in fast reduction: bs_plain_barrier or
546// bs_reduction_barrier
547// 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
548// be used in fast reduction;
549// Reduction method is of 'enum _reduction_method' type and it's defined the way
550// so that the bits of 0-th byte are empty, so no need to execute a shift
551// instruction while packing/unpacking
552
553#if KMP_FAST_REDUCTION_BARRIER
554#define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
555 ((reduction_method) | (barrier_type))
556
557#define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
558 ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
559
560#define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
561 ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
562#else
563#define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
564 (reduction_method)
565
566#define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
567 (packed_reduction_method)
568
569#define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
570#endif
571
572#define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
573 ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
574 (which_reduction_block))
575
576#if KMP_FAST_REDUCTION_BARRIER
577#define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
578 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
579
580#define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
581 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
582#endif
583
584typedef int PACKED_REDUCTION_METHOD_T;
585
586/* -- end of fast reduction stuff ----------------------------------------- */
587
588#if KMP_OS_WINDOWS
589#define USE_CBLKDATA
590#if KMP_MSVC_COMPAT
591#pragma warning(push)
592#pragma warning(disable : 271 310)
593#endif
594#include <windows.h>
595#if KMP_MSVC_COMPAT
596#pragma warning(pop)
597#endif
598#endif
599
600#if KMP_OS_UNIX
601#include <dlfcn.h>
602#include <pthread.h>
603#endif
604
605enum kmp_hw_t : int {
606 KMP_HW_UNKNOWN = -1,
607 KMP_HW_SOCKET = 0,
608 KMP_HW_PROC_GROUP,
609 KMP_HW_NUMA,
610 KMP_HW_DIE,
611 KMP_HW_LLC,
612 KMP_HW_L3,
613 KMP_HW_TILE,
614 KMP_HW_MODULE,
615 KMP_HW_L2,
616 KMP_HW_L1,
617 KMP_HW_CORE,
618 KMP_HW_THREAD,
619 KMP_HW_LAST
620};
621
622typedef enum kmp_hw_core_type_t {
623 KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
624#if KMP_ARCH_X86 || KMP_ARCH_X86_64
625 KMP_HW_CORE_TYPE_ATOM = 0x20,
626 KMP_HW_CORE_TYPE_CORE = 0x40,
627 KMP_HW_MAX_NUM_CORE_TYPES = 3,
628#else
629 KMP_HW_MAX_NUM_CORE_TYPES = 1,
630#endif
631} kmp_hw_core_type_t;
632
633#define KMP_HW_MAX_NUM_CORE_EFFS 8
634
635#define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \
636 KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
637#define KMP_ASSERT_VALID_HW_TYPE(type) \
638 KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
639
640#define KMP_FOREACH_HW_TYPE(type) \
641 for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \
642 type = (kmp_hw_t)((int)type + 1))
643
644const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
645const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
646const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
647
648/* Only Linux* OS and Windows* OS support thread affinity. */
649#if KMP_AFFINITY_SUPPORTED
650
651// GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
652#if KMP_OS_WINDOWS
653#if _MSC_VER < 1600 && KMP_MSVC_COMPAT
654typedef struct GROUP_AFFINITY {
655 KAFFINITY Mask;
656 WORD Group;
657 WORD Reserved[3];
658} GROUP_AFFINITY;
659#endif /* _MSC_VER < 1600 */
660#if KMP_GROUP_AFFINITY
661extern int __kmp_num_proc_groups;
662#else
663static const int __kmp_num_proc_groups = 1;
664#endif /* KMP_GROUP_AFFINITY */
665typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
666extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
667
668typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
669extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
670
671typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
672extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
673
674typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
675 GROUP_AFFINITY *);
676extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
677#endif /* KMP_OS_WINDOWS */
678
679#if KMP_USE_HWLOC
680extern hwloc_topology_t __kmp_hwloc_topology;
681extern int __kmp_hwloc_error;
682#endif
683
684extern size_t __kmp_affin_mask_size;
685#define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
686#define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
687#define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
688#define KMP_CPU_SET_ITERATE(i, mask) \
689 for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
690#define KMP_CPU_SET(i, mask) (mask)->set(i)
691#define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
692#define KMP_CPU_CLR(i, mask) (mask)->clear(i)
693#define KMP_CPU_ZERO(mask) (mask)->zero()
694#define KMP_CPU_ISEMPTY(mask) (mask)->empty()
695#define KMP_CPU_COPY(dest, src) (dest)->copy(src)
696#define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
697#define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
698#define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
699#define KMP_CPU_EQUAL(dest, src) (dest)->is_equal(src)
700#define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
701#define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
702#define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
703#define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
704#define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
705#define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
706#define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
707#define KMP_CPU_ALLOC_ARRAY(arr, n) \
708 (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
709#define KMP_CPU_FREE_ARRAY(arr, n) \
710 __kmp_affinity_dispatch->deallocate_mask_array(arr)
711#define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
712#define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
713#define __kmp_get_system_affinity(mask, abort_bool) \
714 (mask)->get_system_affinity(abort_bool)
715#define __kmp_set_system_affinity(mask, abort_bool) \
716 (mask)->set_system_affinity(abort_bool)
717#define __kmp_get_proc_group(mask) (mask)->get_proc_group()
718
719class KMPAffinity {
720public:
721 class Mask {
722 public:
723 void *operator new(size_t n);
724 void operator delete(void *p);
725 void *operator new[](size_t n);
726 void operator delete[](void *p);
727 virtual ~Mask() {}
728 // Set bit i to 1
729 virtual void set(int i) {}
730 // Return bit i
731 virtual bool is_set(int i) const { return false; }
732 // Set bit i to 0
733 virtual void clear(int i) {}
734 // Zero out entire mask
735 virtual void zero() {}
736 // Check whether mask is empty
737 virtual bool empty() const { return true; }
738 // Copy src into this mask
739 virtual void copy(const Mask *src) {}
740 // this &= rhs
741 virtual void bitwise_and(const Mask *rhs) {}
742 // this |= rhs
743 virtual void bitwise_or(const Mask *rhs) {}
744 // this = ~this
745 virtual void bitwise_not() {}
746 // this == rhs
747 virtual bool is_equal(const Mask *rhs) const { return false; }
748 // API for iterating over an affinity mask
749 // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
750 virtual int begin() const { return 0; }
751 virtual int end() const { return 0; }
752 virtual int next(int previous) const { return 0; }
753#if KMP_OS_WINDOWS
754 virtual int set_process_affinity(bool abort_on_error) const { return -1; }
755#endif
756 // Set the system's affinity to this affinity mask's value
757 virtual int set_system_affinity(bool abort_on_error) const { return -1; }
758 // Set this affinity mask to the current system affinity
759 virtual int get_system_affinity(bool abort_on_error) { return -1; }
760 // Only 1 DWORD in the mask should have any procs set.
761 // Return the appropriate index, or -1 for an invalid mask.
762 virtual int get_proc_group() const { return -1; }
763 int get_max_cpu() const {
764 int cpu;
765 int max_cpu = -1;
766 KMP_CPU_SET_ITERATE(cpu, this) {
767 if (cpu > max_cpu)
768 max_cpu = cpu;
769 }
770 return max_cpu;
771 }
772 };
773 void *operator new(size_t n);
774 void operator delete(void *p);
775 // Need virtual destructor
776 virtual ~KMPAffinity() = default;
777 // Determine if affinity is capable
778 virtual void determine_capable(const char *env_var) {}
779 // Bind the current thread to os proc
780 virtual void bind_thread(int proc) {}
781 // Factory functions to allocate/deallocate a mask
782 virtual Mask *allocate_mask() { return nullptr; }
783 virtual void deallocate_mask(Mask *m) {}
784 virtual Mask *allocate_mask_array(int num) { return nullptr; }
785 virtual void deallocate_mask_array(Mask *m) {}
786 virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
787 static void pick_api();
788 static void destroy_api();
789 enum api_type {
790 NATIVE_OS
791#if KMP_USE_HWLOC
792 ,
793 HWLOC
794#endif
795 };
796 virtual api_type get_api_type() const {
797 KMP_ASSERT(0);
798 return NATIVE_OS;
799 }
800
801private:
802 static bool picked_api;
803};
804
805typedef KMPAffinity::Mask kmp_affin_mask_t;
806extern KMPAffinity *__kmp_affinity_dispatch;
807
808class kmp_affinity_raii_t {
809 kmp_affin_mask_t *mask;
810 bool restored;
811
812public:
813 kmp_affinity_raii_t(const kmp_affin_mask_t *new_mask = nullptr)
814 : restored(false) {
815 if (KMP_AFFINITY_CAPABLE()) {
816 KMP_CPU_ALLOC(mask);
817 KMP_ASSERT(mask != NULL);
818 __kmp_get_system_affinity(mask, /*abort_on_error=*/true);
819 if (new_mask)
820 __kmp_set_system_affinity(new_mask, /*abort_on_error=*/true);
821 }
822 }
823 void restore() {
824 if (!restored && KMP_AFFINITY_CAPABLE()) {
825 __kmp_set_system_affinity(mask, /*abort_on_error=*/true);
826 KMP_CPU_FREE(mask);
827 }
828 restored = true;
829 }
830 ~kmp_affinity_raii_t() { restore(); }
831};
832
833// Declare local char buffers with this size for printing debug and info
834// messages, using __kmp_affinity_print_mask().
835#define KMP_AFFIN_MASK_PRINT_LEN 1024
836
837enum affinity_type {
838 affinity_none = 0,
839 affinity_physical,
840 affinity_logical,
841 affinity_compact,
842 affinity_scatter,
843 affinity_explicit,
844 affinity_balanced,
845 affinity_disabled, // not used outsize the env var parser
846 affinity_default
847};
848
849enum affinity_top_method {
850 affinity_top_method_all = 0, // try all (supported) methods, in order
851#if KMP_ARCH_X86 || KMP_ARCH_X86_64
852 affinity_top_method_apicid,
853 affinity_top_method_x2apicid,
854 affinity_top_method_x2apicid_1f,
855#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
856 affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
857#if KMP_GROUP_AFFINITY
858 affinity_top_method_group,
859#endif /* KMP_GROUP_AFFINITY */
860 affinity_top_method_flat,
861#if KMP_USE_HWLOC
862 affinity_top_method_hwloc,
863#endif
864 affinity_top_method_default
865};
866
867#define affinity_respect_mask_default (2)
868
869typedef struct kmp_affinity_flags_t {
870 unsigned dups : 1;
871 unsigned verbose : 1;
872 unsigned warnings : 1;
873 unsigned respect : 2;
874 unsigned reset : 1;
875 unsigned initialized : 1;
876 unsigned core_types_gran : 1;
877 unsigned core_effs_gran : 1;
878 unsigned omp_places : 1;
879 unsigned reserved : 22;
880} kmp_affinity_flags_t;
881KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4);
882
883typedef struct kmp_affinity_ids_t {
884 int ids[KMP_HW_LAST];
885 int operator[](size_t idx) const { return ids[idx]; }
886 int &operator[](size_t idx) { return ids[idx]; }
887 kmp_affinity_ids_t &operator=(const kmp_affinity_ids_t &rhs) {
888 for (int i = 0; i < KMP_HW_LAST; ++i)
889 ids[i] = rhs[i];
890 return *this;
891 }
892} kmp_affinity_ids_t;
893
894typedef struct kmp_affinity_attrs_t {
895 int core_type : 8;
896 int core_eff : 8;
897 unsigned valid : 1;
898 unsigned reserved : 15;
899} kmp_affinity_attrs_t;
900#define KMP_AFFINITY_ATTRS_UNKNOWN \
901 { KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 }
902
903typedef struct kmp_affinity_t {
904 char *proclist;
905 enum affinity_type type;
906 kmp_hw_t gran;
907 int gran_levels;
908 kmp_affinity_attrs_t core_attr_gran;
909 int compact;
910 int offset;
911 kmp_affinity_flags_t flags;
912 unsigned num_masks;
913 kmp_affin_mask_t *masks;
914 kmp_affinity_ids_t *ids;
915 kmp_affinity_attrs_t *attrs;
916 unsigned num_os_id_masks;
917 kmp_affin_mask_t *os_id_masks;
918 const char *env_var;
919} kmp_affinity_t;
920
921#define KMP_AFFINITY_INIT(env) \
922 { \
923 nullptr, affinity_default, KMP_HW_UNKNOWN, -1, KMP_AFFINITY_ATTRS_UNKNOWN, \
924 0, 0, \
925 {TRUE, FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE, \
926 FALSE, FALSE, FALSE}, \
927 0, nullptr, nullptr, nullptr, 0, nullptr, env \
928 }
929
930extern enum affinity_top_method __kmp_affinity_top_method;
931extern kmp_affinity_t __kmp_affinity;
932extern kmp_affinity_t __kmp_hh_affinity;
933extern kmp_affinity_t *__kmp_affinities[2];
934
935extern void __kmp_affinity_bind_thread(int which);
936
937extern kmp_affin_mask_t *__kmp_affin_fullMask;
938extern kmp_affin_mask_t *__kmp_affin_origMask;
939extern char *__kmp_cpuinfo_file;
940
941#endif /* KMP_AFFINITY_SUPPORTED */
942
943// This needs to be kept in sync with the values in omp.h !!!
944typedef enum kmp_proc_bind_t {
945 proc_bind_false = 0,
946 proc_bind_true,
947 proc_bind_primary,
948 proc_bind_close,
949 proc_bind_spread,
950 proc_bind_intel, // use KMP_AFFINITY interface
951 proc_bind_default
952} kmp_proc_bind_t;
953
954typedef struct kmp_nested_proc_bind_t {
955 kmp_proc_bind_t *bind_types;
956 int size;
957 int used;
958} kmp_nested_proc_bind_t;
959
960extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
961extern kmp_proc_bind_t __kmp_teams_proc_bind;
962
963extern int __kmp_display_affinity;
964extern char *__kmp_affinity_format;
965static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
966#if OMPT_SUPPORT
967extern int __kmp_tool;
968extern char *__kmp_tool_libraries;
969#endif // OMPT_SUPPORT
970
971#if KMP_AFFINITY_SUPPORTED
972#define KMP_PLACE_ALL (-1)
973#define KMP_PLACE_UNDEFINED (-2)
974// Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
975#define KMP_AFFINITY_NON_PROC_BIND \
976 ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
977 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
978 (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced))
979#endif /* KMP_AFFINITY_SUPPORTED */
980
981extern int __kmp_affinity_num_places;
982
983typedef enum kmp_cancel_kind_t {
984 cancel_noreq = 0,
985 cancel_parallel = 1,
986 cancel_loop = 2,
987 cancel_sections = 3,
988 cancel_taskgroup = 4
989} kmp_cancel_kind_t;
990
991// KMP_HW_SUBSET support:
992typedef struct kmp_hws_item {
993 int num;
994 int offset;
995} kmp_hws_item_t;
996
997extern kmp_hws_item_t __kmp_hws_socket;
998extern kmp_hws_item_t __kmp_hws_die;
999extern kmp_hws_item_t __kmp_hws_node;
1000extern kmp_hws_item_t __kmp_hws_tile;
1001extern kmp_hws_item_t __kmp_hws_core;
1002extern kmp_hws_item_t __kmp_hws_proc;
1003extern int __kmp_hws_requested;
1004extern int __kmp_hws_abs_flag; // absolute or per-item number requested
1005
1006/* ------------------------------------------------------------------------ */
1007
1008#define KMP_PAD(type, sz) \
1009 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
1010
1011// We need to avoid using -1 as a GTID as +1 is added to the gtid
1012// when storing it in a lock, and the value 0 is reserved.
1013#define KMP_GTID_DNE (-2) /* Does not exist */
1014#define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
1015#define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
1016#define KMP_GTID_UNKNOWN (-5) /* Is not known */
1017#define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
1018
1019/* OpenMP 5.0 Memory Management support */
1020
1021#ifndef __OMP_H
1022// Duplicate type definitions from omp.h
1023typedef uintptr_t omp_uintptr_t;
1024
1025typedef enum {
1026 omp_atk_sync_hint = 1,
1027 omp_atk_alignment = 2,
1028 omp_atk_access = 3,
1029 omp_atk_pool_size = 4,
1030 omp_atk_fallback = 5,
1031 omp_atk_fb_data = 6,
1032 omp_atk_pinned = 7,
1033 omp_atk_partition = 8
1034} omp_alloctrait_key_t;
1035
1036typedef enum {
1037 omp_atv_false = 0,
1038 omp_atv_true = 1,
1039 omp_atv_contended = 3,
1040 omp_atv_uncontended = 4,
1041 omp_atv_serialized = 5,
1042 omp_atv_sequential = omp_atv_serialized, // (deprecated)
1043 omp_atv_private = 6,
1044 omp_atv_all = 7,
1045 omp_atv_thread = 8,
1046 omp_atv_pteam = 9,
1047 omp_atv_cgroup = 10,
1048 omp_atv_default_mem_fb = 11,
1049 omp_atv_null_fb = 12,
1050 omp_atv_abort_fb = 13,
1051 omp_atv_allocator_fb = 14,
1052 omp_atv_environment = 15,
1053 omp_atv_nearest = 16,
1054 omp_atv_blocked = 17,
1055 omp_atv_interleaved = 18
1056} omp_alloctrait_value_t;
1057#define omp_atv_default ((omp_uintptr_t)-1)
1058
1059typedef void *omp_memspace_handle_t;
1060extern omp_memspace_handle_t const omp_default_mem_space;
1061extern omp_memspace_handle_t const omp_large_cap_mem_space;
1062extern omp_memspace_handle_t const omp_const_mem_space;
1063extern omp_memspace_handle_t const omp_high_bw_mem_space;
1064extern omp_memspace_handle_t const omp_low_lat_mem_space;
1065extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
1066extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
1067extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
1068
1069typedef struct {
1070 omp_alloctrait_key_t key;
1071 omp_uintptr_t value;
1072} omp_alloctrait_t;
1073
1074typedef void *omp_allocator_handle_t;
1075extern omp_allocator_handle_t const omp_null_allocator;
1076extern omp_allocator_handle_t const omp_default_mem_alloc;
1077extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
1078extern omp_allocator_handle_t const omp_const_mem_alloc;
1079extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
1080extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
1081extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
1082extern omp_allocator_handle_t const omp_pteam_mem_alloc;
1083extern omp_allocator_handle_t const omp_thread_mem_alloc;
1084extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
1085extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
1086extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
1087extern omp_allocator_handle_t const kmp_max_mem_alloc;
1088extern omp_allocator_handle_t __kmp_def_allocator;
1089
1090// end of duplicate type definitions from omp.h
1091#endif
1092
1093extern int __kmp_memkind_available;
1094
1095typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
1096
1097typedef struct kmp_allocator_t {
1098 omp_memspace_handle_t memspace;
1099 void **memkind; // pointer to memkind
1100 size_t alignment;
1101 omp_alloctrait_value_t fb;
1102 kmp_allocator_t *fb_data;
1103 kmp_uint64 pool_size;
1104 kmp_uint64 pool_used;
1105 bool pinned;
1106} kmp_allocator_t;
1107
1108extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1109 omp_memspace_handle_t,
1110 int ntraits,
1111 omp_alloctrait_t traits[]);
1112extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1113extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1114extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1115// external interfaces, may be used by compiler
1116extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1117extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1118 omp_allocator_handle_t al);
1119extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1120 omp_allocator_handle_t al);
1121extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1122 omp_allocator_handle_t al,
1123 omp_allocator_handle_t free_al);
1124extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1125// internal interfaces, contain real implementation
1126extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1127 omp_allocator_handle_t al);
1128extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1129 omp_allocator_handle_t al);
1130extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1131 omp_allocator_handle_t al,
1132 omp_allocator_handle_t free_al);
1133extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1134
1135extern void __kmp_init_memkind();
1136extern void __kmp_fini_memkind();
1137extern void __kmp_init_target_mem();
1138
1139/* ------------------------------------------------------------------------ */
1140
1141#if ENABLE_LIBOMPTARGET
1142extern void __kmp_init_target_task();
1143#endif
1144
1145/* ------------------------------------------------------------------------ */
1146
1147#define KMP_UINT64_MAX \
1148 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1149
1150#define KMP_MIN_NTH 1
1151
1152#ifndef KMP_MAX_NTH
1153#if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1154#define KMP_MAX_NTH PTHREAD_THREADS_MAX
1155#else
1156#ifdef __ve__
1157// VE's pthread supports only up to 64 threads per a VE process.
1158// Please check p. 14 of following documentation for more details.
1159// https://sxauroratsubasa.sakura.ne.jp/documents/veos/en/VEOS_high_level_design.pdf
1160#define KMP_MAX_NTH 64
1161#else
1162#define KMP_MAX_NTH INT_MAX
1163#endif
1164#endif
1165#endif /* KMP_MAX_NTH */
1166
1167#ifdef PTHREAD_STACK_MIN
1168#define KMP_MIN_STKSIZE ((size_t)PTHREAD_STACK_MIN)
1169#else
1170#define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1171#endif
1172
1173#define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1174
1175#if KMP_ARCH_X86
1176#define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1177#elif KMP_ARCH_X86_64
1178#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1179#define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1180#elif KMP_ARCH_VE
1181// Minimum stack size for pthread for VE is 4MB.
1182// https://www.hpc.nec/documents/veos/en/glibc/Difference_Points_glibc.htm
1183#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1184#else
1185#define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1186#endif
1187
1188#define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1189#define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1190#define KMP_MAX_MALLOC_POOL_INCR \
1191 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1192
1193#define KMP_MIN_STKOFFSET (0)
1194#define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1195#if KMP_OS_DARWIN
1196#define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1197#else
1198#define KMP_DEFAULT_STKOFFSET CACHE_LINE
1199#endif
1200
1201#define KMP_MIN_STKPADDING (0)
1202#define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1203
1204#define KMP_BLOCKTIME_MULTIPLIER \
1205 (1000000) /* number of blocktime units per second */
1206#define KMP_MIN_BLOCKTIME (0)
1207#define KMP_MAX_BLOCKTIME \
1208 (INT_MAX) /* Must be this for "infinite" setting the work */
1209
1210/* __kmp_blocktime is in microseconds */
1211#define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200000))
1212
1213#if KMP_USE_MONITOR
1214#define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1215#define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1216#define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1217
1218/* Calculate new number of monitor wakeups for a specific block time based on
1219 previous monitor_wakeups. Only allow increasing number of wakeups */
1220#define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1221 (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \
1222 : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \
1223 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1224 ? (monitor_wakeups) \
1225 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1226
1227/* Calculate number of intervals for a specific block time based on
1228 monitor_wakeups */
1229#define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1230 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1231 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1232#else
1233#define KMP_BLOCKTIME(team, tid) \
1234 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1235#if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1236// HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1237extern kmp_uint64 __kmp_ticks_per_msec;
1238extern kmp_uint64 __kmp_ticks_per_usec;
1239#if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1240#define KMP_NOW() ((kmp_uint64)_rdtsc())
1241#else
1242#define KMP_NOW() __kmp_hardware_timestamp()
1243#endif
1244#define KMP_BLOCKTIME_INTERVAL(team, tid) \
1245 ((kmp_uint64)KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_usec)
1246#define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1247#else
1248// System time is retrieved sporadically while blocking.
1249extern kmp_uint64 __kmp_now_nsec();
1250#define KMP_NOW() __kmp_now_nsec()
1251#define KMP_BLOCKTIME_INTERVAL(team, tid) \
1252 ((kmp_uint64)KMP_BLOCKTIME(team, tid) * (kmp_uint64)KMP_NSEC_PER_USEC)
1253#define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1254#endif
1255#endif // KMP_USE_MONITOR
1256
1257#define KMP_MIN_STATSCOLS 40
1258#define KMP_MAX_STATSCOLS 4096
1259#define KMP_DEFAULT_STATSCOLS 80
1260
1261#define KMP_MIN_INTERVAL 0
1262#define KMP_MAX_INTERVAL (INT_MAX - 1)
1263#define KMP_DEFAULT_INTERVAL 0
1264
1265#define KMP_MIN_CHUNK 1
1266#define KMP_MAX_CHUNK (INT_MAX - 1)
1267#define KMP_DEFAULT_CHUNK 1
1268
1269#define KMP_MIN_DISP_NUM_BUFF 1
1270#define KMP_DFLT_DISP_NUM_BUFF 7
1271#define KMP_MAX_DISP_NUM_BUFF 4096
1272
1273#define KMP_MAX_ORDERED 8
1274
1275#define KMP_MAX_FIELDS 32
1276
1277#define KMP_MAX_BRANCH_BITS 31
1278
1279#define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1280
1281#define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1282
1283#define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1284
1285/* Minimum number of threads before switch to TLS gtid (experimentally
1286 determined) */
1287/* josh TODO: what about OS X* tuning? */
1288#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1289#define KMP_TLS_GTID_MIN 5
1290#else
1291#define KMP_TLS_GTID_MIN INT_MAX
1292#endif
1293
1294#define KMP_MASTER_TID(tid) (0 == (tid))
1295#define KMP_WORKER_TID(tid) (0 != (tid))
1296
1297#define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1298#define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1299#define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1300
1301#ifndef TRUE
1302#define FALSE 0
1303#define TRUE (!FALSE)
1304#endif
1305
1306/* NOTE: all of the following constants must be even */
1307
1308#if KMP_OS_WINDOWS
1309#define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1310#define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1311#elif KMP_OS_LINUX
1312#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1313#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1314#elif KMP_OS_DARWIN
1315/* TODO: tune for KMP_OS_DARWIN */
1316#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1317#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1318#elif KMP_OS_DRAGONFLY
1319/* TODO: tune for KMP_OS_DRAGONFLY */
1320#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1321#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1322#elif KMP_OS_FREEBSD
1323/* TODO: tune for KMP_OS_FREEBSD */
1324#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1325#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1326#elif KMP_OS_NETBSD
1327/* TODO: tune for KMP_OS_NETBSD */
1328#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1329#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1330#elif KMP_OS_HURD
1331/* TODO: tune for KMP_OS_HURD */
1332#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1333#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1334#elif KMP_OS_OPENBSD
1335/* TODO: tune for KMP_OS_OPENBSD */
1336#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1337#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1338#endif
1339
1340#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1341typedef struct kmp_cpuid {
1342 kmp_uint32 eax;
1343 kmp_uint32 ebx;
1344 kmp_uint32 ecx;
1345 kmp_uint32 edx;
1346} kmp_cpuid_t;
1347
1348typedef struct kmp_cpuinfo_flags_t {
1349 unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1350 unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1351 unsigned hybrid : 1;
1352 unsigned reserved : 29; // Ensure size of 32 bits
1353} kmp_cpuinfo_flags_t;
1354
1355typedef struct kmp_cpuinfo {
1356 int initialized; // If 0, other fields are not initialized.
1357 int signature; // CPUID(1).EAX
1358 int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1359 int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1360 // Model << 4 ) + Model)
1361 int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1362 kmp_cpuinfo_flags_t flags;
1363 int apic_id;
1364 int physical_id;
1365 int logical_id;
1366 kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1367 char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1368} kmp_cpuinfo_t;
1369
1370extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1371
1372#if KMP_OS_UNIX
1373// subleaf is only needed for cache and topology discovery and can be set to
1374// zero in most cases
1375static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1376 __asm__ __volatile__("cpuid"
1377 : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1378 : "a"(leaf), "c"(subleaf));
1379}
1380// Load p into FPU control word
1381static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1382 __asm__ __volatile__("fldcw %0" : : "m"(*p));
1383}
1384// Store FPU control word into p
1385static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1386 __asm__ __volatile__("fstcw %0" : "=m"(*p));
1387}
1388static inline void __kmp_clear_x87_fpu_status_word() {
1389#if KMP_MIC
1390 // 32-bit protected mode x87 FPU state
1391 struct x87_fpu_state {
1392 unsigned cw;
1393 unsigned sw;
1394 unsigned tw;
1395 unsigned fip;
1396 unsigned fips;
1397 unsigned fdp;
1398 unsigned fds;
1399 };
1400 struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1401 __asm__ __volatile__("fstenv %0\n\t" // store FP env
1402 "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1403 "fldenv %0\n\t" // load FP env back
1404 : "+m"(fpu_state), "+m"(fpu_state.sw));
1405#else
1406 __asm__ __volatile__("fnclex");
1407#endif // KMP_MIC
1408}
1409#if __SSE__
1410static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1411static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1412#else
1413static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1414static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1415#endif
1416#else
1417// Windows still has these as external functions in assembly file
1418extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1419extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1420extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1421extern void __kmp_clear_x87_fpu_status_word();
1422static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1423static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1424#endif // KMP_OS_UNIX
1425
1426#define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1427
1428// User-level Monitor/Mwait
1429#if KMP_HAVE_UMWAIT
1430// We always try for UMWAIT first
1431#if KMP_HAVE_WAITPKG_INTRINSICS
1432#if KMP_HAVE_IMMINTRIN_H
1433#include <immintrin.h>
1434#elif KMP_HAVE_INTRIN_H
1435#include <intrin.h>
1436#endif
1437#endif // KMP_HAVE_WAITPKG_INTRINSICS
1438
1439KMP_ATTRIBUTE_TARGET_WAITPKG
1440static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1441#if !KMP_HAVE_WAITPKG_INTRINSICS
1442 uint32_t timeHi = uint32_t(counter >> 32);
1443 uint32_t timeLo = uint32_t(counter & 0xffffffff);
1444 char flag;
1445 __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1446 "setb %0"
1447 // The "=q" restraint means any register accessible as rl
1448 // in 32-bit mode: a, b, c, and d;
1449 // in 64-bit mode: any integer register
1450 : "=q"(flag)
1451 : "a"(timeLo), "d"(timeHi), "c"(hint)
1452 :);
1453 return flag;
1454#else
1455 return _tpause(hint, counter);
1456#endif
1457}
1458KMP_ATTRIBUTE_TARGET_WAITPKG
1459static inline void __kmp_umonitor(void *cacheline) {
1460#if !KMP_HAVE_WAITPKG_INTRINSICS
1461 __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1462 :
1463 : "a"(cacheline)
1464 :);
1465#else
1466 _umonitor(cacheline);
1467#endif
1468}
1469KMP_ATTRIBUTE_TARGET_WAITPKG
1470static inline int __kmp_umwait(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("#umwait\n.byte 0xF2, 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 _umwait(hint, counter);
1486#endif
1487}
1488#elif KMP_HAVE_MWAIT
1489#if KMP_OS_UNIX
1490#include <pmmintrin.h>
1491#else
1492#include <intrin.h>
1493#endif
1494#if KMP_OS_UNIX
1495__attribute__((target("sse3")))
1496#endif
1497static inline void
1498__kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1499 _mm_monitor(cacheline, extensions, hints);
1500}
1501#if KMP_OS_UNIX
1502__attribute__((target("sse3")))
1503#endif
1504static inline void
1505__kmp_mm_mwait(unsigned extensions, unsigned hints) {
1506 _mm_mwait(extensions, hints);
1507}
1508#endif // KMP_HAVE_UMWAIT
1509
1510#if KMP_ARCH_X86
1511extern void __kmp_x86_pause(void);
1512#elif KMP_MIC
1513// Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1514// regression after removal of extra PAUSE from spin loops. Changing
1515// the delay from 100 to 300 showed even better performance than double PAUSE
1516// on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1517static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1518#else
1519static inline void __kmp_x86_pause(void) { _mm_pause(); }
1520#endif
1521#define KMP_CPU_PAUSE() __kmp_x86_pause()
1522#elif KMP_ARCH_PPC64
1523#define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1524#define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1525#define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1526#define KMP_CPU_PAUSE() \
1527 do { \
1528 KMP_PPC64_PRI_LOW(); \
1529 KMP_PPC64_PRI_MED(); \
1530 KMP_PPC64_PRI_LOC_MB(); \
1531 } while (0)
1532#else
1533#define KMP_CPU_PAUSE() /* nothing to do */
1534#endif
1535
1536#define KMP_INIT_YIELD(count) \
1537 { (count) = __kmp_yield_init; }
1538
1539#define KMP_INIT_BACKOFF(time) \
1540 { (time) = __kmp_pause_init; }
1541
1542#define KMP_OVERSUBSCRIBED \
1543 (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1544
1545#define KMP_TRY_YIELD \
1546 ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1547
1548#define KMP_TRY_YIELD_OVERSUB \
1549 ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1550
1551#define KMP_YIELD(cond) \
1552 { \
1553 KMP_CPU_PAUSE(); \
1554 if ((cond) && (KMP_TRY_YIELD)) \
1555 __kmp_yield(); \
1556 }
1557
1558#define KMP_YIELD_OVERSUB() \
1559 { \
1560 KMP_CPU_PAUSE(); \
1561 if ((KMP_TRY_YIELD_OVERSUB)) \
1562 __kmp_yield(); \
1563 }
1564
1565// Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1566// there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1567#define KMP_YIELD_SPIN(count) \
1568 { \
1569 KMP_CPU_PAUSE(); \
1570 if (KMP_TRY_YIELD) { \
1571 (count) -= 2; \
1572 if (!(count)) { \
1573 __kmp_yield(); \
1574 (count) = __kmp_yield_next; \
1575 } \
1576 } \
1577 }
1578
1579// If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1580// (C0.2) state, which improves performance of other SMT threads on the same
1581// core, otherwise, use the fast (C0.1) default state, or whatever the user has
1582// requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1583// available, fall back to the regular CPU pause and yield combination.
1584#if KMP_HAVE_UMWAIT
1585#define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF)
1586#define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1587 { \
1588 if (__kmp_tpause_enabled) { \
1589 if (KMP_OVERSUBSCRIBED) { \
1590 __kmp_tpause(0, (time)); \
1591 } else { \
1592 __kmp_tpause(__kmp_tpause_hint, (time)); \
1593 } \
1594 (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK; \
1595 } else { \
1596 KMP_CPU_PAUSE(); \
1597 if ((KMP_TRY_YIELD_OVERSUB)) { \
1598 __kmp_yield(); \
1599 } else if (__kmp_use_yield == 1) { \
1600 (count) -= 2; \
1601 if (!(count)) { \
1602 __kmp_yield(); \
1603 (count) = __kmp_yield_next; \
1604 } \
1605 } \
1606 } \
1607 }
1608#else
1609#define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1610 { \
1611 KMP_CPU_PAUSE(); \
1612 if ((KMP_TRY_YIELD_OVERSUB)) \
1613 __kmp_yield(); \
1614 else if (__kmp_use_yield == 1) { \
1615 (count) -= 2; \
1616 if (!(count)) { \
1617 __kmp_yield(); \
1618 (count) = __kmp_yield_next; \
1619 } \
1620 } \
1621 }
1622#endif // KMP_HAVE_UMWAIT
1623
1624/* ------------------------------------------------------------------------ */
1625/* Support datatypes for the orphaned construct nesting checks. */
1626/* ------------------------------------------------------------------------ */
1627
1628/* When adding to this enum, add its corresponding string in cons_text_c[]
1629 * array in kmp_error.cpp */
1630enum cons_type {
1631 ct_none,
1632 ct_parallel,
1633 ct_pdo,
1634 ct_pdo_ordered,
1635 ct_psections,
1636 ct_psingle,
1637 ct_critical,
1638 ct_ordered_in_parallel,
1639 ct_ordered_in_pdo,
1640 ct_master,
1641 ct_reduce,
1642 ct_barrier,
1643 ct_masked
1644};
1645
1646#define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1647
1648struct cons_data {
1649 ident_t const *ident;
1650 enum cons_type type;
1651 int prev;
1652 kmp_user_lock_p
1653 name; /* address exclusively for critical section name comparison */
1654};
1655
1656struct cons_header {
1657 int p_top, w_top, s_top;
1658 int stack_size, stack_top;
1659 struct cons_data *stack_data;
1660};
1661
1662struct kmp_region_info {
1663 char *text;
1664 int offset[KMP_MAX_FIELDS];
1665 int length[KMP_MAX_FIELDS];
1666};
1667
1668/* ---------------------------------------------------------------------- */
1669/* ---------------------------------------------------------------------- */
1670
1671#if KMP_OS_WINDOWS
1672typedef HANDLE kmp_thread_t;
1673typedef DWORD kmp_key_t;
1674#endif /* KMP_OS_WINDOWS */
1675
1676#if KMP_OS_UNIX
1677typedef pthread_t kmp_thread_t;
1678typedef pthread_key_t kmp_key_t;
1679#endif
1680
1681extern kmp_key_t __kmp_gtid_threadprivate_key;
1682
1683typedef struct kmp_sys_info {
1684 long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1685 long minflt; /* the number of page faults serviced without any I/O */
1686 long majflt; /* the number of page faults serviced that required I/O */
1687 long nswap; /* the number of times a process was "swapped" out of memory */
1688 long inblock; /* the number of times the file system had to perform input */
1689 long oublock; /* the number of times the file system had to perform output */
1690 long nvcsw; /* the number of times a context switch was voluntarily */
1691 long nivcsw; /* the number of times a context switch was forced */
1692} kmp_sys_info_t;
1693
1694#if USE_ITT_BUILD
1695// We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1696// required type here. Later we will check the type meets requirements.
1697typedef int kmp_itt_mark_t;
1698#define KMP_ITT_DEBUG 0
1699#endif /* USE_ITT_BUILD */
1700
1701typedef kmp_int32 kmp_critical_name[8];
1702
1712typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1713typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1714 ...);
1715
1720/* ---------------------------------------------------------------------------
1721 */
1722/* Threadprivate initialization/finalization function declarations */
1723
1724/* for non-array objects: __kmpc_threadprivate_register() */
1725
1730typedef void *(*kmpc_ctor)(void *);
1731
1736typedef void (*kmpc_dtor)(
1737 void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1738 compiler */
1743typedef void *(*kmpc_cctor)(void *, void *);
1744
1745/* for array objects: __kmpc_threadprivate_register_vec() */
1746/* First arg: "this" pointer */
1747/* Last arg: number of array elements */
1753typedef void *(*kmpc_ctor_vec)(void *, size_t);
1759typedef void (*kmpc_dtor_vec)(void *, size_t);
1765typedef void *(*kmpc_cctor_vec)(void *, void *,
1766 size_t); /* function unused by compiler */
1767
1772/* keeps tracked of threadprivate cache allocations for cleanup later */
1773typedef struct kmp_cached_addr {
1774 void **addr; /* address of allocated cache */
1775 void ***compiler_cache; /* pointer to compiler's cache */
1776 void *data; /* pointer to global data */
1777 struct kmp_cached_addr *next; /* pointer to next cached address */
1778} kmp_cached_addr_t;
1779
1780struct private_data {
1781 struct private_data *next; /* The next descriptor in the list */
1782 void *data; /* The data buffer for this descriptor */
1783 int more; /* The repeat count for this descriptor */
1784 size_t size; /* The data size for this descriptor */
1785};
1786
1787struct private_common {
1788 struct private_common *next;
1789 struct private_common *link;
1790 void *gbl_addr;
1791 void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1792 size_t cmn_size;
1793};
1794
1795struct shared_common {
1796 struct shared_common *next;
1797 struct private_data *pod_init;
1798 void *obj_init;
1799 void *gbl_addr;
1800 union {
1801 kmpc_ctor ctor;
1802 kmpc_ctor_vec ctorv;
1803 } ct;
1804 union {
1805 kmpc_cctor cctor;
1806 kmpc_cctor_vec cctorv;
1807 } cct;
1808 union {
1809 kmpc_dtor dtor;
1810 kmpc_dtor_vec dtorv;
1811 } dt;
1812 size_t vec_len;
1813 int is_vec;
1814 size_t cmn_size;
1815};
1816
1817#define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1818#define KMP_HASH_TABLE_SIZE \
1819 (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1820#define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1821#define KMP_HASH(x) \
1822 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1823
1824struct common_table {
1825 struct private_common *data[KMP_HASH_TABLE_SIZE];
1826};
1827
1828struct shared_table {
1829 struct shared_common *data[KMP_HASH_TABLE_SIZE];
1830};
1831
1832/* ------------------------------------------------------------------------ */
1833
1834#if KMP_USE_HIER_SCHED
1835// Shared barrier data that exists inside a single unit of the scheduling
1836// hierarchy
1837typedef struct kmp_hier_private_bdata_t {
1838 kmp_int32 num_active;
1839 kmp_uint64 index;
1840 kmp_uint64 wait_val[2];
1841} kmp_hier_private_bdata_t;
1842#endif
1843
1844typedef struct kmp_sched_flags {
1845 unsigned ordered : 1;
1846 unsigned nomerge : 1;
1847 unsigned contains_last : 1;
1848#if KMP_USE_HIER_SCHED
1849 unsigned use_hier : 1;
1850 unsigned unused : 28;
1851#else
1852 unsigned unused : 29;
1853#endif
1854} kmp_sched_flags_t;
1855
1856KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1857
1858#if KMP_STATIC_STEAL_ENABLED
1859typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1860 kmp_int32 count;
1861 kmp_int32 ub;
1862 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1863 kmp_int32 lb;
1864 kmp_int32 st;
1865 kmp_int32 tc;
1866 kmp_lock_t *steal_lock; // lock used for chunk stealing
1867 // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1868 // a) parm3 is properly aligned and
1869 // b) all parm1-4 are on the same cache line.
1870 // Because of parm1-4 are used together, performance seems to be better
1871 // if they are on the same cache line (not measured though).
1872
1873 struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1874 kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1875 kmp_int32 parm2; // make no real change at least while padding is off.
1876 kmp_int32 parm3;
1877 kmp_int32 parm4;
1878 };
1879
1880 kmp_uint32 ordered_lower;
1881 kmp_uint32 ordered_upper;
1882#if KMP_OS_WINDOWS
1883 kmp_int32 last_upper;
1884#endif /* KMP_OS_WINDOWS */
1885} dispatch_private_info32_t;
1886
1887typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1888 kmp_int64 count; // current chunk number for static & static-steal scheduling
1889 kmp_int64 ub; /* upper-bound */
1890 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1891 kmp_int64 lb; /* lower-bound */
1892 kmp_int64 st; /* stride */
1893 kmp_int64 tc; /* trip count (number of iterations) */
1894 kmp_lock_t *steal_lock; // lock used for chunk stealing
1895 /* parm[1-4] are used in different ways by different scheduling algorithms */
1896
1897 // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1898 // a) parm3 is properly aligned and
1899 // b) all parm1-4 are in the same cache line.
1900 // Because of parm1-4 are used together, performance seems to be better
1901 // if they are in the same line (not measured though).
1902
1903 struct KMP_ALIGN(32) {
1904 kmp_int64 parm1;
1905 kmp_int64 parm2;
1906 kmp_int64 parm3;
1907 kmp_int64 parm4;
1908 };
1909
1910 kmp_uint64 ordered_lower;
1911 kmp_uint64 ordered_upper;
1912#if KMP_OS_WINDOWS
1913 kmp_int64 last_upper;
1914#endif /* KMP_OS_WINDOWS */
1915} dispatch_private_info64_t;
1916#else /* KMP_STATIC_STEAL_ENABLED */
1917typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1918 kmp_int32 lb;
1919 kmp_int32 ub;
1920 kmp_int32 st;
1921 kmp_int32 tc;
1922
1923 kmp_int32 parm1;
1924 kmp_int32 parm2;
1925 kmp_int32 parm3;
1926 kmp_int32 parm4;
1927
1928 kmp_int32 count;
1929
1930 kmp_uint32 ordered_lower;
1931 kmp_uint32 ordered_upper;
1932#if KMP_OS_WINDOWS
1933 kmp_int32 last_upper;
1934#endif /* KMP_OS_WINDOWS */
1935} dispatch_private_info32_t;
1936
1937typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1938 kmp_int64 lb; /* lower-bound */
1939 kmp_int64 ub; /* upper-bound */
1940 kmp_int64 st; /* stride */
1941 kmp_int64 tc; /* trip count (number of iterations) */
1942
1943 /* parm[1-4] are used in different ways by different scheduling algorithms */
1944 kmp_int64 parm1;
1945 kmp_int64 parm2;
1946 kmp_int64 parm3;
1947 kmp_int64 parm4;
1948
1949 kmp_int64 count; /* current chunk number for static scheduling */
1950
1951 kmp_uint64 ordered_lower;
1952 kmp_uint64 ordered_upper;
1953#if KMP_OS_WINDOWS
1954 kmp_int64 last_upper;
1955#endif /* KMP_OS_WINDOWS */
1956} dispatch_private_info64_t;
1957#endif /* KMP_STATIC_STEAL_ENABLED */
1958
1959typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1960 union private_info {
1961 dispatch_private_info32_t p32;
1962 dispatch_private_info64_t p64;
1963 } u;
1964 enum sched_type schedule; /* scheduling algorithm */
1965 kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1966 std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
1967 kmp_int32 ordered_bumped;
1968 // Stack of buffers for nest of serial regions
1969 struct dispatch_private_info *next;
1970 kmp_int32 type_size; /* the size of types in private_info */
1971#if KMP_USE_HIER_SCHED
1972 kmp_int32 hier_id;
1973 void *parent; /* hierarchical scheduling parent pointer */
1974#endif
1975 enum cons_type pushed_ws;
1976} dispatch_private_info_t;
1977
1978typedef struct dispatch_shared_info32 {
1979 /* chunk index under dynamic, number of idle threads under static-steal;
1980 iteration index otherwise */
1981 volatile kmp_uint32 iteration;
1982 volatile kmp_int32 num_done;
1983 volatile kmp_uint32 ordered_iteration;
1984 // Dummy to retain the structure size after making ordered_iteration scalar
1985 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1986} dispatch_shared_info32_t;
1987
1988typedef struct dispatch_shared_info64 {
1989 /* chunk index under dynamic, number of idle threads under static-steal;
1990 iteration index otherwise */
1991 volatile kmp_uint64 iteration;
1992 volatile kmp_int64 num_done;
1993 volatile kmp_uint64 ordered_iteration;
1994 // Dummy to retain the structure size after making ordered_iteration scalar
1995 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1996} dispatch_shared_info64_t;
1997
1998typedef struct dispatch_shared_info {
1999 union shared_info {
2000 dispatch_shared_info32_t s32;
2001 dispatch_shared_info64_t s64;
2002 } u;
2003 volatile kmp_uint32 buffer_index;
2004 volatile kmp_int32 doacross_buf_idx; // teamwise index
2005 volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
2006 kmp_int32 doacross_num_done; // count finished threads
2007#if KMP_USE_HIER_SCHED
2008 void *hier;
2009#endif
2010#if KMP_USE_HWLOC
2011 // When linking with libhwloc, the ORDERED EPCC test slows down on big
2012 // machines (> 48 cores). Performance analysis showed that a cache thrash
2013 // was occurring and this padding helps alleviate the problem.
2014 char padding[64];
2015#endif
2016} dispatch_shared_info_t;
2017
2018typedef struct kmp_disp {
2019 /* Vector for ORDERED SECTION */
2020 void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
2021 /* Vector for END ORDERED SECTION */
2022 void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
2023
2024 dispatch_shared_info_t *th_dispatch_sh_current;
2025 dispatch_private_info_t *th_dispatch_pr_current;
2026
2027 dispatch_private_info_t *th_disp_buffer;
2028 kmp_uint32 th_disp_index;
2029 kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
2030 volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
2031 kmp_int64 *th_doacross_info; // info on loop bounds
2032#if KMP_USE_INTERNODE_ALIGNMENT
2033 char more_padding[INTERNODE_CACHE_LINE];
2034#endif
2035} kmp_disp_t;
2036
2037/* ------------------------------------------------------------------------ */
2038/* Barrier stuff */
2039
2040/* constants for barrier state update */
2041#define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
2042#define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
2043#define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
2044#define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
2045
2046#define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
2047#define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
2048#define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
2049
2050#if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
2051#error "Barrier sleep bit must be smaller than barrier bump bit"
2052#endif
2053#if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
2054#error "Barrier unused bit must be smaller than barrier bump bit"
2055#endif
2056
2057// Constants for release barrier wait state: currently, hierarchical only
2058#define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
2059#define KMP_BARRIER_OWN_FLAG \
2060 1 // Normal state; worker waiting on own b_go flag in release
2061#define KMP_BARRIER_PARENT_FLAG \
2062 2 // Special state; worker waiting on parent's b_go flag in release
2063#define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
2064 3 // Special state; tells worker to shift from parent to own b_go
2065#define KMP_BARRIER_SWITCHING \
2066 4 // Special state; worker resets appropriate flag on wake-up
2067
2068#define KMP_NOT_SAFE_TO_REAP \
2069 0 // Thread th_reap_state: not safe to reap (tasking)
2070#define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
2071
2072// The flag_type describes the storage used for the flag.
2073enum flag_type {
2074 flag32,
2075 flag64,
2076 atomic_flag64,
2077 flag_oncore,
2078 flag_unset
2079};
2080
2081enum barrier_type {
2082 bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
2083 barriers if enabled) */
2084 bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
2085#if KMP_FAST_REDUCTION_BARRIER
2086 bs_reduction_barrier, /* 2, All barriers that are used in reduction */
2087#endif // KMP_FAST_REDUCTION_BARRIER
2088 bs_last_barrier /* Just a placeholder to mark the end */
2089};
2090
2091// to work with reduction barriers just like with plain barriers
2092#if !KMP_FAST_REDUCTION_BARRIER
2093#define bs_reduction_barrier bs_plain_barrier
2094#endif // KMP_FAST_REDUCTION_BARRIER
2095
2096typedef enum kmp_bar_pat { /* Barrier communication patterns */
2097 bp_linear_bar =
2098 0, /* Single level (degenerate) tree */
2099 bp_tree_bar =
2100 1, /* Balanced tree with branching factor 2^n */
2101 bp_hyper_bar = 2, /* Hypercube-embedded tree with min
2102 branching factor 2^n */
2103 bp_hierarchical_bar = 3, /* Machine hierarchy tree */
2104 bp_dist_bar = 4, /* Distributed barrier */
2105 bp_last_bar /* Placeholder to mark the end */
2106} kmp_bar_pat_e;
2107
2108#define KMP_BARRIER_ICV_PUSH 1
2109
2110/* Record for holding the values of the internal controls stack records */
2111typedef struct kmp_internal_control {
2112 int serial_nesting_level; /* corresponds to the value of the
2113 th_team_serialized field */
2114 kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
2115 thread) */
2116 kmp_int8
2117 bt_set; /* internal control for whether blocktime is explicitly set */
2118 int blocktime; /* internal control for blocktime */
2119#if KMP_USE_MONITOR
2120 int bt_intervals; /* internal control for blocktime intervals */
2121#endif
2122 int nproc; /* internal control for #threads for next parallel region (per
2123 thread) */
2124 int thread_limit; /* internal control for thread-limit-var */
2125 int task_thread_limit; /* internal control for thread-limit-var of a task*/
2126 int max_active_levels; /* internal control for max_active_levels */
2127 kmp_r_sched_t
2128 sched; /* internal control for runtime schedule {sched,chunk} pair */
2129 kmp_proc_bind_t proc_bind; /* internal control for affinity */
2130 kmp_int32 default_device; /* internal control for default device */
2131 struct kmp_internal_control *next;
2132} kmp_internal_control_t;
2133
2134static inline void copy_icvs(kmp_internal_control_t *dst,
2135 kmp_internal_control_t *src) {
2136 *dst = *src;
2137}
2138
2139/* Thread barrier needs volatile barrier fields */
2140typedef struct KMP_ALIGN_CACHE kmp_bstate {
2141 // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2142 // uses of it). It is not explicitly aligned below, because we *don't* want
2143 // it to be padded -- instead, we fit b_go into the same cache line with
2144 // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2145 kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2146 // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2147 // same NGO store
2148 volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2149 KMP_ALIGN_CACHE volatile kmp_uint64
2150 b_arrived; // STATE => task reached synch point.
2151 kmp_uint32 *skip_per_level;
2152 kmp_uint32 my_level;
2153 kmp_int32 parent_tid;
2154 kmp_int32 old_tid;
2155 kmp_uint32 depth;
2156 struct kmp_bstate *parent_bar;
2157 kmp_team_t *team;
2158 kmp_uint64 leaf_state;
2159 kmp_uint32 nproc;
2160 kmp_uint8 base_leaf_kids;
2161 kmp_uint8 leaf_kids;
2162 kmp_uint8 offset;
2163 kmp_uint8 wait_flag;
2164 kmp_uint8 use_oncore_barrier;
2165#if USE_DEBUGGER
2166 // The following field is intended for the debugger solely. Only the worker
2167 // thread itself accesses this field: the worker increases it by 1 when it
2168 // arrives to a barrier.
2169 KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2170#endif /* USE_DEBUGGER */
2171} kmp_bstate_t;
2172
2173union KMP_ALIGN_CACHE kmp_barrier_union {
2174 double b_align; /* use worst case alignment */
2175 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2176 kmp_bstate_t bb;
2177};
2178
2179typedef union kmp_barrier_union kmp_balign_t;
2180
2181/* Team barrier needs only non-volatile arrived counter */
2182union KMP_ALIGN_CACHE kmp_barrier_team_union {
2183 double b_align; /* use worst case alignment */
2184 char b_pad[CACHE_LINE];
2185 struct {
2186 kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2187#if USE_DEBUGGER
2188 // The following two fields are indended for the debugger solely. Only
2189 // primary thread of the team accesses these fields: the first one is
2190 // increased by 1 when the primary thread arrives to a barrier, the second
2191 // one is increased by one when all the threads arrived.
2192 kmp_uint b_master_arrived;
2193 kmp_uint b_team_arrived;
2194#endif
2195 };
2196};
2197
2198typedef union kmp_barrier_team_union kmp_balign_team_t;
2199
2200/* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2201 threads when a condition changes. This is to workaround an NPTL bug where
2202 padding was added to pthread_cond_t which caused the initialization routine
2203 to write outside of the structure if compiled on pre-NPTL threads. */
2204#if KMP_OS_WINDOWS
2205typedef struct kmp_win32_mutex {
2206 /* The Lock */
2207 CRITICAL_SECTION cs;
2208} kmp_win32_mutex_t;
2209
2210typedef struct kmp_win32_cond {
2211 /* Count of the number of waiters. */
2212 int waiters_count_;
2213
2214 /* Serialize access to <waiters_count_> */
2215 kmp_win32_mutex_t waiters_count_lock_;
2216
2217 /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2218 int release_count_;
2219
2220 /* Keeps track of the current "generation" so that we don't allow */
2221 /* one thread to steal all the "releases" from the broadcast. */
2222 int wait_generation_count_;
2223
2224 /* A manual-reset event that's used to block and release waiting threads. */
2225 HANDLE event_;
2226} kmp_win32_cond_t;
2227#endif
2228
2229#if KMP_OS_UNIX
2230
2231union KMP_ALIGN_CACHE kmp_cond_union {
2232 double c_align;
2233 char c_pad[CACHE_LINE];
2234 pthread_cond_t c_cond;
2235};
2236
2237typedef union kmp_cond_union kmp_cond_align_t;
2238
2239union KMP_ALIGN_CACHE kmp_mutex_union {
2240 double m_align;
2241 char m_pad[CACHE_LINE];
2242 pthread_mutex_t m_mutex;
2243};
2244
2245typedef union kmp_mutex_union kmp_mutex_align_t;
2246
2247#endif /* KMP_OS_UNIX */
2248
2249typedef struct kmp_desc_base {
2250 void *ds_stackbase;
2251 size_t ds_stacksize;
2252 int ds_stackgrow;
2253 kmp_thread_t ds_thread;
2254 volatile int ds_tid;
2255 int ds_gtid;
2256#if KMP_OS_WINDOWS
2257 volatile int ds_alive;
2258 DWORD ds_thread_id;
2259/* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2260 However, debugger support (libomp_db) cannot work with handles, because they
2261 uncomparable. For example, debugger requests info about thread with handle h.
2262 h is valid within debugger process, and meaningless within debugee process.
2263 Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2264 within debugee process, but it is a *new* handle which does *not* equal to
2265 any other handle in debugee... The only way to compare handles is convert
2266 them to system-wide ids. GetThreadId() function is available only in
2267 Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2268 on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2269 thread id by call to GetCurrentThreadId() from within the thread and save it
2270 to let libomp_db identify threads. */
2271#endif /* KMP_OS_WINDOWS */
2272} kmp_desc_base_t;
2273
2274typedef union KMP_ALIGN_CACHE kmp_desc {
2275 double ds_align; /* use worst case alignment */
2276 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2277 kmp_desc_base_t ds;
2278} kmp_desc_t;
2279
2280typedef struct kmp_local {
2281 volatile int this_construct; /* count of single's encountered by thread */
2282 void *reduce_data;
2283#if KMP_USE_BGET
2284 void *bget_data;
2285 void *bget_list;
2286#if !USE_CMP_XCHG_FOR_BGET
2287#ifdef USE_QUEUING_LOCK_FOR_BGET
2288 kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2289#else
2290 kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2291// bootstrap lock so we can use it at library
2292// shutdown.
2293#endif /* USE_LOCK_FOR_BGET */
2294#endif /* ! USE_CMP_XCHG_FOR_BGET */
2295#endif /* KMP_USE_BGET */
2296
2297 PACKED_REDUCTION_METHOD_T
2298 packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2299 __kmpc_end_reduce*() */
2300
2301} kmp_local_t;
2302
2303#define KMP_CHECK_UPDATE(a, b) \
2304 if ((a) != (b)) \
2305 (a) = (b)
2306#define KMP_CHECK_UPDATE_SYNC(a, b) \
2307 if ((a) != (b)) \
2308 TCW_SYNC_PTR((a), (b))
2309
2310#define get__blocktime(xteam, xtid) \
2311 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2312#define get__bt_set(xteam, xtid) \
2313 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2314#if KMP_USE_MONITOR
2315#define get__bt_intervals(xteam, xtid) \
2316 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2317#endif
2318
2319#define get__dynamic_2(xteam, xtid) \
2320 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2321#define get__nproc_2(xteam, xtid) \
2322 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2323#define get__sched_2(xteam, xtid) \
2324 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2325
2326#define set__blocktime_team(xteam, xtid, xval) \
2327 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2328 (xval))
2329
2330#if KMP_USE_MONITOR
2331#define set__bt_intervals_team(xteam, xtid, xval) \
2332 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2333 (xval))
2334#endif
2335
2336#define set__bt_set_team(xteam, xtid, xval) \
2337 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2338
2339#define set__dynamic(xthread, xval) \
2340 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2341#define get__dynamic(xthread) \
2342 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2343
2344#define set__nproc(xthread, xval) \
2345 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2346
2347#define set__thread_limit(xthread, xval) \
2348 (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2349
2350#define set__max_active_levels(xthread, xval) \
2351 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2352
2353#define get__max_active_levels(xthread) \
2354 ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2355
2356#define set__sched(xthread, xval) \
2357 (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2358
2359#define set__proc_bind(xthread, xval) \
2360 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2361#define get__proc_bind(xthread) \
2362 ((xthread)->th.th_current_task->td_icvs.proc_bind)
2363
2364// OpenMP tasking data structures
2365
2366typedef enum kmp_tasking_mode {
2367 tskm_immediate_exec = 0,
2368 tskm_extra_barrier = 1,
2369 tskm_task_teams = 2,
2370 tskm_max = 2
2371} kmp_tasking_mode_t;
2372
2373extern kmp_tasking_mode_t
2374 __kmp_tasking_mode; /* determines how/when to execute tasks */
2375extern int __kmp_task_stealing_constraint;
2376extern int __kmp_enable_task_throttling;
2377extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2378// specified, defaults to 0 otherwise
2379// Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2380extern kmp_int32 __kmp_max_task_priority;
2381// Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2382extern kmp_uint64 __kmp_taskloop_min_tasks;
2383
2384/* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2385 taskdata first */
2386#define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2387#define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2388
2389// The tt_found_tasks flag is a signal to all threads in the team that tasks
2390// were spawned and queued since the previous barrier release.
2391#define KMP_TASKING_ENABLED(task_team) \
2392 (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2400typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2401
2402typedef union kmp_cmplrdata {
2403 kmp_int32 priority;
2404 kmp_routine_entry_t
2405 destructors; /* pointer to function to invoke deconstructors of
2406 firstprivate C++ objects */
2407 /* future data */
2408} kmp_cmplrdata_t;
2409
2410/* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2413typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2414 void *shareds;
2415 kmp_routine_entry_t
2416 routine;
2417 kmp_int32 part_id;
2418 kmp_cmplrdata_t
2419 data1; /* Two known optional additions: destructors and priority */
2420 kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2421 /* future data */
2422 /* private vars */
2423} kmp_task_t;
2424
2429typedef struct kmp_taskgroup {
2430 std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2431 std::atomic<kmp_int32>
2432 cancel_request; // request for cancellation of this taskgroup
2433 struct kmp_taskgroup *parent; // parent taskgroup
2434 // Block of data to perform task reduction
2435 void *reduce_data; // reduction related info
2436 kmp_int32 reduce_num_data; // number of data items to reduce
2437 uintptr_t *gomp_data; // gomp reduction data
2438} kmp_taskgroup_t;
2439
2440// forward declarations
2441typedef union kmp_depnode kmp_depnode_t;
2442typedef struct kmp_depnode_list kmp_depnode_list_t;
2443typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2444
2445// macros for checking dep flag as an integer
2446#define KMP_DEP_IN 0x1
2447#define KMP_DEP_OUT 0x2
2448#define KMP_DEP_INOUT 0x3
2449#define KMP_DEP_MTX 0x4
2450#define KMP_DEP_SET 0x8
2451#define KMP_DEP_ALL 0x80
2452// Compiler sends us this info:
2453typedef struct kmp_depend_info {
2454 kmp_intptr_t base_addr;
2455 size_t len;
2456 union {
2457 kmp_uint8 flag; // flag as an unsigned char
2458 struct { // flag as a set of 8 bits
2459#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
2460 /* Same fields as in the #else branch, but in reverse order */
2461 unsigned all : 1;
2462 unsigned unused : 3;
2463 unsigned set : 1;
2464 unsigned mtx : 1;
2465 unsigned out : 1;
2466 unsigned in : 1;
2467#else
2468 unsigned in : 1;
2469 unsigned out : 1;
2470 unsigned mtx : 1;
2471 unsigned set : 1;
2472 unsigned unused : 3;
2473 unsigned all : 1;
2474#endif
2475 } flags;
2476 };
2477} kmp_depend_info_t;
2478
2479// Internal structures to work with task dependencies:
2480struct kmp_depnode_list {
2481 kmp_depnode_t *node;
2482 kmp_depnode_list_t *next;
2483};
2484
2485// Max number of mutexinoutset dependencies per node
2486#define MAX_MTX_DEPS 4
2487
2488typedef struct kmp_base_depnode {
2489 kmp_depnode_list_t *successors; /* used under lock */
2490 kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2491 kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2492 kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2493 kmp_lock_t lock; /* guards shared fields: task, successors */
2494#if KMP_SUPPORT_GRAPH_OUTPUT
2495 kmp_uint32 id;
2496#endif
2497 std::atomic<kmp_int32> npredecessors;
2498 std::atomic<kmp_int32> nrefs;
2499} kmp_base_depnode_t;
2500
2501union KMP_ALIGN_CACHE kmp_depnode {
2502 double dn_align; /* use worst case alignment */
2503 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2504 kmp_base_depnode_t dn;
2505};
2506
2507struct kmp_dephash_entry {
2508 kmp_intptr_t addr;
2509 kmp_depnode_t *last_out;
2510 kmp_depnode_list_t *last_set;
2511 kmp_depnode_list_t *prev_set;
2512 kmp_uint8 last_flag;
2513 kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2514 kmp_dephash_entry_t *next_in_bucket;
2515};
2516
2517typedef struct kmp_dephash {
2518 kmp_dephash_entry_t **buckets;
2519 size_t size;
2520 kmp_depnode_t *last_all;
2521 size_t generation;
2522 kmp_uint32 nelements;
2523 kmp_uint32 nconflicts;
2524} kmp_dephash_t;
2525
2526typedef struct kmp_task_affinity_info {
2527 kmp_intptr_t base_addr;
2528 size_t len;
2529 struct {
2530 bool flag1 : 1;
2531 bool flag2 : 1;
2532 kmp_int32 reserved : 30;
2533 } flags;
2534} kmp_task_affinity_info_t;
2535
2536typedef enum kmp_event_type_t {
2537 KMP_EVENT_UNINITIALIZED = 0,
2538 KMP_EVENT_ALLOW_COMPLETION = 1
2539} kmp_event_type_t;
2540
2541typedef struct {
2542 kmp_event_type_t type;
2543 kmp_tas_lock_t lock;
2544 union {
2545 kmp_task_t *task;
2546 } ed;
2547} kmp_event_t;
2548
2549#if OMPX_TASKGRAPH
2550// Initial number of allocated nodes while recording
2551#define INIT_MAPSIZE 50
2552
2553typedef struct kmp_taskgraph_flags { /*This needs to be exactly 32 bits */
2554 unsigned nowait : 1;
2555 unsigned re_record : 1;
2556 unsigned reserved : 30;
2557} kmp_taskgraph_flags_t;
2558
2560typedef struct kmp_node_info {
2561 kmp_task_t *task; // Pointer to the actual task
2562 kmp_int32 *successors; // Array of the succesors ids
2563 kmp_int32 nsuccessors; // Number of succesors of the node
2564 std::atomic<kmp_int32>
2565 npredecessors_counter; // Number of predessors on the fly
2566 kmp_int32 npredecessors; // Total number of predecessors
2567 kmp_int32 successors_size; // Number of allocated succesors ids
2568 kmp_taskdata_t *parent_task; // Parent implicit task
2569} kmp_node_info_t;
2570
2572typedef enum kmp_tdg_status {
2573 KMP_TDG_NONE = 0,
2574 KMP_TDG_RECORDING = 1,
2575 KMP_TDG_READY = 2
2576} kmp_tdg_status_t;
2577
2579typedef struct kmp_tdg_info {
2580 kmp_int32 tdg_id; // Unique idenfifier of the TDG
2581 kmp_taskgraph_flags_t tdg_flags; // Flags related to a TDG
2582 kmp_int32 map_size; // Number of allocated TDG nodes
2583 kmp_int32 num_roots; // Number of roots tasks int the TDG
2584 kmp_int32 *root_tasks; // Array of tasks identifiers that are roots
2585 kmp_node_info_t *record_map; // Array of TDG nodes
2586 kmp_tdg_status_t tdg_status =
2587 KMP_TDG_NONE; // Status of the TDG (recording, ready...)
2588 std::atomic<kmp_int32> num_tasks; // Number of TDG nodes
2589 kmp_bootstrap_lock_t
2590 graph_lock; // Protect graph attributes when updated via taskloop_recur
2591 // Taskloop reduction related
2592 void *rec_taskred_data; // Data to pass to __kmpc_task_reduction_init or
2593 // __kmpc_taskred_init
2594 kmp_int32 rec_num_taskred;
2595} kmp_tdg_info_t;
2596
2597extern int __kmp_tdg_dot;
2598extern kmp_int32 __kmp_max_tdgs;
2599extern kmp_tdg_info_t **__kmp_global_tdgs;
2600extern kmp_int32 __kmp_curr_tdg_idx;
2601extern kmp_int32 __kmp_successors_size;
2602extern std::atomic<kmp_int32> __kmp_tdg_task_id;
2603extern kmp_int32 __kmp_num_tdg;
2604#endif
2605
2606#ifdef BUILD_TIED_TASK_STACK
2607
2608/* Tied Task stack definitions */
2609typedef struct kmp_stack_block {
2610 kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2611 struct kmp_stack_block *sb_next;
2612 struct kmp_stack_block *sb_prev;
2613} kmp_stack_block_t;
2614
2615typedef struct kmp_task_stack {
2616 kmp_stack_block_t ts_first_block; // first block of stack entries
2617 kmp_taskdata_t **ts_top; // pointer to the top of stack
2618 kmp_int32 ts_entries; // number of entries on the stack
2619} kmp_task_stack_t;
2620
2621#endif // BUILD_TIED_TASK_STACK
2622
2623typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2624#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
2625 /* Same fields as in the #else branch, but in reverse order */
2626#if OMPX_TASKGRAPH
2627 unsigned reserved31 : 6;
2628 unsigned onced : 1;
2629#else
2630 unsigned reserved31 : 7;
2631#endif
2632 unsigned native : 1;
2633 unsigned freed : 1;
2634 unsigned complete : 1;
2635 unsigned executing : 1;
2636 unsigned started : 1;
2637 unsigned team_serial : 1;
2638 unsigned tasking_ser : 1;
2639 unsigned task_serial : 1;
2640 unsigned tasktype : 1;
2641 unsigned reserved : 8;
2642 unsigned hidden_helper : 1;
2643 unsigned detachable : 1;
2644 unsigned priority_specified : 1;
2645 unsigned proxy : 1;
2646 unsigned destructors_thunk : 1;
2647 unsigned merged_if0 : 1;
2648 unsigned final : 1;
2649 unsigned tiedness : 1;
2650#else
2651 /* Compiler flags */ /* Total compiler flags must be 16 bits */
2652 unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2653 unsigned final : 1; /* task is final(1) so execute immediately */
2654 unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2655 code path */
2656 unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2657 invoke destructors from the runtime */
2658 unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2659 context of the RTL) */
2660 unsigned priority_specified : 1; /* set if the compiler provides priority
2661 setting for the task */
2662 unsigned detachable : 1; /* 1 == can detach */
2663 unsigned hidden_helper : 1; /* 1 == hidden helper task */
2664 unsigned reserved : 8; /* reserved for compiler use */
2665
2666 /* Library flags */ /* Total library flags must be 16 bits */
2667 unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2668 unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2669 unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2670 // (1) or may be deferred (0)
2671 unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2672 // (0) [>= 2 threads]
2673 /* If either team_serial or tasking_ser is set, task team may be NULL */
2674 /* Task State Flags: */
2675 unsigned started : 1; /* 1==started, 0==not started */
2676 unsigned executing : 1; /* 1==executing, 0==not executing */
2677 unsigned complete : 1; /* 1==complete, 0==not complete */
2678 unsigned freed : 1; /* 1==freed, 0==allocated */
2679 unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2680#if OMPX_TASKGRAPH
2681 unsigned onced : 1; /* 1==ran once already, 0==never ran, record & replay purposes */
2682 unsigned reserved31 : 6; /* reserved for library use */
2683#else
2684 unsigned reserved31 : 7; /* reserved for library use */
2685#endif
2686#endif
2687} kmp_tasking_flags_t;
2688
2689typedef struct kmp_target_data {
2690 void *async_handle; // libomptarget async handle for task completion query
2691} kmp_target_data_t;
2692
2693struct kmp_taskdata { /* aligned during dynamic allocation */
2694 kmp_int32 td_task_id; /* id, assigned by debugger */
2695 kmp_tasking_flags_t td_flags; /* task flags */
2696 kmp_team_t *td_team; /* team for this task */
2697 kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2698 /* Currently not used except for perhaps IDB */
2699 kmp_taskdata_t *td_parent; /* parent task */
2700 kmp_int32 td_level; /* task nesting level */
2701 std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2702 ident_t *td_ident; /* task identifier */
2703 // Taskwait data.
2704 ident_t *td_taskwait_ident;
2705 kmp_uint32 td_taskwait_counter;
2706 kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2707 KMP_ALIGN_CACHE kmp_internal_control_t
2708 td_icvs; /* Internal control variables for the task */
2709 KMP_ALIGN_CACHE std::atomic<kmp_int32>
2710 td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2711 deallocated */
2712 std::atomic<kmp_int32>
2713 td_incomplete_child_tasks; /* Child tasks not yet complete */
2714 kmp_taskgroup_t
2715 *td_taskgroup; // Each task keeps pointer to its current taskgroup
2716 kmp_dephash_t
2717 *td_dephash; // Dependencies for children tasks are tracked from here
2718 kmp_depnode_t
2719 *td_depnode; // Pointer to graph node if this task has dependencies
2720 kmp_task_team_t *td_task_team;
2721 size_t td_size_alloc; // Size of task structure, including shareds etc.
2722#if defined(KMP_GOMP_COMPAT)
2723 // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2724 kmp_int32 td_size_loop_bounds;
2725#endif
2726 kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2727#if defined(KMP_GOMP_COMPAT)
2728 // GOMP sends in a copy function for copy constructors
2729 void (*td_copy_func)(void *, void *);
2730#endif
2731 kmp_event_t td_allow_completion_event;
2732#if OMPT_SUPPORT
2733 ompt_task_info_t ompt_task_info;
2734#endif
2735#if OMPX_TASKGRAPH
2736 bool is_taskgraph = 0; // whether the task is within a TDG
2737 kmp_tdg_info_t *tdg; // used to associate task with a TDG
2738#endif
2739 kmp_target_data_t td_target_data;
2740}; // struct kmp_taskdata
2741
2742// Make sure padding above worked
2743KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2744
2745// Data for task team but per thread
2746typedef struct kmp_base_thread_data {
2747 kmp_info_p *td_thr; // Pointer back to thread info
2748 // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2749 // queued?
2750 kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2751 kmp_taskdata_t *
2752 *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2753 kmp_int32 td_deque_size; // Size of deck
2754 kmp_uint32 td_deque_head; // Head of deque (will wrap)
2755 kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2756 kmp_int32 td_deque_ntasks; // Number of tasks in deque
2757 // GEH: shouldn't this be volatile since used in while-spin?
2758 kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2759#ifdef BUILD_TIED_TASK_STACK
2760 kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2761// scheduling constraint
2762#endif // BUILD_TIED_TASK_STACK
2763} kmp_base_thread_data_t;
2764
2765#define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2766#define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2767
2768#define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2769#define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2770
2771typedef union KMP_ALIGN_CACHE kmp_thread_data {
2772 kmp_base_thread_data_t td;
2773 double td_align; /* use worst case alignment */
2774 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2775} kmp_thread_data_t;
2776
2777typedef struct kmp_task_pri {
2778 kmp_thread_data_t td;
2779 kmp_int32 priority;
2780 kmp_task_pri *next;
2781} kmp_task_pri_t;
2782
2783// Data for task teams which are used when tasking is enabled for the team
2784typedef struct kmp_base_task_team {
2785 kmp_bootstrap_lock_t
2786 tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2787 /* must be bootstrap lock since used at library shutdown*/
2788
2789 // TODO: check performance vs kmp_tas_lock_t
2790 kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */
2791 kmp_task_pri_t *tt_task_pri_list;
2792
2793 kmp_task_team_t *tt_next; /* For linking the task team free list */
2794 kmp_thread_data_t
2795 *tt_threads_data; /* Array of per-thread structures for task team */
2796 /* Data survives task team deallocation */
2797 kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2798 executing this team? */
2799 /* TRUE means tt_threads_data is set up and initialized */
2800 kmp_int32 tt_nproc; /* #threads in team */
2801 kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2802 kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2803 kmp_int32 tt_untied_task_encountered;
2804 std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued
2805 // There is hidden helper thread encountered in this task team so that we must
2806 // wait when waiting on task team
2807 kmp_int32 tt_hidden_helper_task_encountered;
2808
2809 KMP_ALIGN_CACHE
2810 std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2811
2812 KMP_ALIGN_CACHE
2813 volatile kmp_uint32
2814 tt_active; /* is the team still actively executing tasks */
2815} kmp_base_task_team_t;
2816
2817union KMP_ALIGN_CACHE kmp_task_team {
2818 kmp_base_task_team_t tt;
2819 double tt_align; /* use worst case alignment */
2820 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2821};
2822
2823#if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2824// Free lists keep same-size free memory slots for fast memory allocation
2825// routines
2826typedef struct kmp_free_list {
2827 void *th_free_list_self; // Self-allocated tasks free list
2828 void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2829 // threads
2830 void *th_free_list_other; // Non-self free list (to be returned to owner's
2831 // sync list)
2832} kmp_free_list_t;
2833#endif
2834#if KMP_NESTED_HOT_TEAMS
2835// Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2836// are not put in teams pool, and they don't put threads in threads pool.
2837typedef struct kmp_hot_team_ptr {
2838 kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2839 kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2840} kmp_hot_team_ptr_t;
2841#endif
2842typedef struct kmp_teams_size {
2843 kmp_int32 nteams; // number of teams in a league
2844 kmp_int32 nth; // number of threads in each team of the league
2845} kmp_teams_size_t;
2846
2847// This struct stores a thread that acts as a "root" for a contention
2848// group. Contention groups are rooted at kmp_root threads, but also at
2849// each primary thread of each team created in the teams construct.
2850// This struct therefore also stores a thread_limit associated with
2851// that contention group, and a counter to track the number of threads
2852// active in that contention group. Each thread has a list of these: CG
2853// root threads have an entry in their list in which cg_root refers to
2854// the thread itself, whereas other workers in the CG will have a
2855// single entry where cg_root is same as the entry containing their CG
2856// root. When a thread encounters a teams construct, it will add a new
2857// entry to the front of its list, because it now roots a new CG.
2858typedef struct kmp_cg_root {
2859 kmp_info_p *cg_root; // "root" thread for a contention group
2860 // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2861 // thread_limit clause for teams primary threads
2862 kmp_int32 cg_thread_limit;
2863 kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2864 struct kmp_cg_root *up; // pointer to higher level CG root in list
2865} kmp_cg_root_t;
2866
2867// OpenMP thread data structures
2868
2869typedef struct KMP_ALIGN_CACHE kmp_base_info {
2870 /* Start with the readonly data which is cache aligned and padded. This is
2871 written before the thread starts working by the primary thread. Uber
2872 masters may update themselves later. Usage does not consider serialized
2873 regions. */
2874 kmp_desc_t th_info;
2875 kmp_team_p *th_team; /* team we belong to */
2876 kmp_root_p *th_root; /* pointer to root of task hierarchy */
2877 kmp_info_p *th_next_pool; /* next available thread in the pool */
2878 kmp_disp_t *th_dispatch; /* thread's dispatch data */
2879 int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2880
2881 /* The following are cached from the team info structure */
2882 /* TODO use these in more places as determined to be needed via profiling */
2883 int th_team_nproc; /* number of threads in a team */
2884 kmp_info_p *th_team_master; /* the team's primary thread */
2885 int th_team_serialized; /* team is serialized */
2886 microtask_t th_teams_microtask; /* save entry address for teams construct */
2887 int th_teams_level; /* save initial level of teams construct */
2888/* it is 0 on device but may be any on host */
2889
2890/* The blocktime info is copied from the team struct to the thread struct */
2891/* at the start of a barrier, and the values stored in the team are used */
2892/* at points in the code where the team struct is no longer guaranteed */
2893/* to exist (from the POV of worker threads). */
2894#if KMP_USE_MONITOR
2895 int th_team_bt_intervals;
2896 int th_team_bt_set;
2897#else
2898 kmp_uint64 th_team_bt_intervals;
2899#endif
2900
2901#if KMP_AFFINITY_SUPPORTED
2902 kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2903 kmp_affinity_ids_t th_topology_ids; /* thread's current topology ids */
2904 kmp_affinity_attrs_t th_topology_attrs; /* thread's current topology attrs */
2905#endif
2906 omp_allocator_handle_t th_def_allocator; /* default allocator */
2907 /* The data set by the primary thread at reinit, then R/W by the worker */
2908 KMP_ALIGN_CACHE int
2909 th_set_nproc; /* if > 0, then only use this request for the next fork */
2910#if KMP_NESTED_HOT_TEAMS
2911 kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2912#endif
2913 kmp_proc_bind_t
2914 th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2915 kmp_teams_size_t
2916 th_teams_size; /* number of teams/threads in teams construct */
2917#if KMP_AFFINITY_SUPPORTED
2918 int th_current_place; /* place currently bound to */
2919 int th_new_place; /* place to bind to in par reg */
2920 int th_first_place; /* first place in partition */
2921 int th_last_place; /* last place in partition */
2922#endif
2923 int th_prev_level; /* previous level for affinity format */
2924 int th_prev_num_threads; /* previous num_threads for affinity format */
2925#if USE_ITT_BUILD
2926 kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2927 kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2928 kmp_uint64 th_frame_time; /* frame timestamp */
2929#endif /* USE_ITT_BUILD */
2930 kmp_local_t th_local;
2931 struct private_common *th_pri_head;
2932
2933 /* Now the data only used by the worker (after initial allocation) */
2934 /* TODO the first serial team should actually be stored in the info_t
2935 structure. this will help reduce initial allocation overhead */
2936 KMP_ALIGN_CACHE kmp_team_p
2937 *th_serial_team; /*serialized team held in reserve*/
2938
2939#if OMPT_SUPPORT
2940 ompt_thread_info_t ompt_thread_info;
2941#endif
2942
2943 /* The following are also read by the primary thread during reinit */
2944 struct common_table *th_pri_common;
2945
2946 volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2947 /* while awaiting queuing lock acquire */
2948
2949 volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2950 flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2951
2952 ident_t *th_ident;
2953 unsigned th_x; // Random number generator data
2954 unsigned th_a; // Random number generator data
2955
2956 /* Tasking-related data for the thread */
2957 kmp_task_team_t *th_task_team; // Task team struct
2958 kmp_taskdata_t *th_current_task; // Innermost Task being executed
2959 kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2960 kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2961 // at nested levels
2962 kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2963 kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2964 kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2965 // tasking, thus safe to reap
2966
2967 /* More stuff for keeping track of active/sleeping threads (this part is
2968 written by the worker thread) */
2969 kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2970 int th_active; // ! sleeping; 32 bits for TCR/TCW
2971 std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
2972 // 0 = not used in team; 1 = used in team;
2973 // 2 = transitioning to not used in team; 3 = transitioning to used in team
2974 struct cons_header *th_cons; // used for consistency check
2975#if KMP_USE_HIER_SCHED
2976 // used for hierarchical scheduling
2977 kmp_hier_private_bdata_t *th_hier_bar_data;
2978#endif
2979
2980 /* Add the syncronizing data which is cache aligned and padded. */
2981 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2982
2983 KMP_ALIGN_CACHE volatile kmp_int32
2984 th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2985
2986#if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2987#define NUM_LISTS 4
2988 kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2989// allocation routines
2990#endif
2991
2992#if KMP_OS_WINDOWS
2993 kmp_win32_cond_t th_suspend_cv;
2994 kmp_win32_mutex_t th_suspend_mx;
2995 std::atomic<int> th_suspend_init;
2996#endif
2997#if KMP_OS_UNIX
2998 kmp_cond_align_t th_suspend_cv;
2999 kmp_mutex_align_t th_suspend_mx;
3000 std::atomic<int> th_suspend_init_count;
3001#endif
3002
3003#if USE_ITT_BUILD
3004 kmp_itt_mark_t th_itt_mark_single;
3005// alignment ???
3006#endif /* USE_ITT_BUILD */
3007#if KMP_STATS_ENABLED
3008 kmp_stats_list *th_stats;
3009#endif
3010#if KMP_OS_UNIX
3011 std::atomic<bool> th_blocking;
3012#endif
3013 kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
3014} kmp_base_info_t;
3015
3016typedef union KMP_ALIGN_CACHE kmp_info {
3017 double th_align; /* use worst case alignment */
3018 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
3019 kmp_base_info_t th;
3020} kmp_info_t;
3021
3022// OpenMP thread team data structures
3023
3024typedef struct kmp_base_data {
3025 volatile kmp_uint32 t_value;
3026} kmp_base_data_t;
3027
3028typedef union KMP_ALIGN_CACHE kmp_sleep_team {
3029 double dt_align; /* use worst case alignment */
3030 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3031 kmp_base_data_t dt;
3032} kmp_sleep_team_t;
3033
3034typedef union KMP_ALIGN_CACHE kmp_ordered_team {
3035 double dt_align; /* use worst case alignment */
3036 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3037 kmp_base_data_t dt;
3038} kmp_ordered_team_t;
3039
3040typedef int (*launch_t)(int gtid);
3041
3042/* Minimum number of ARGV entries to malloc if necessary */
3043#define KMP_MIN_MALLOC_ARGV_ENTRIES 100
3044
3045// Set up how many argv pointers will fit in cache lines containing
3046// t_inline_argv. Historically, we have supported at least 96 bytes. Using a
3047// larger value for more space between the primary write/worker read section and
3048// read/write by all section seems to buy more performance on EPCC PARALLEL.
3049#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3050#define KMP_INLINE_ARGV_BYTES \
3051 (4 * CACHE_LINE - \
3052 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
3053 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
3054 CACHE_LINE))
3055#else
3056#define KMP_INLINE_ARGV_BYTES \
3057 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
3058#endif
3059#define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
3060
3061typedef struct KMP_ALIGN_CACHE kmp_base_team {
3062 // Synchronization Data
3063 // ---------------------------------------------------------------------------
3064 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
3065 kmp_balign_team_t t_bar[bs_last_barrier];
3066 std::atomic<int> t_construct; // count of single directive encountered by team
3067 char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
3068
3069 // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
3070 std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
3071 std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
3072
3073 // Primary thread only
3074 // ---------------------------------------------------------------------------
3075 KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
3076 int t_master_this_cons; // "this_construct" single counter of primary thread
3077 // in parent team
3078 ident_t *t_ident; // if volatile, have to change too much other crud to
3079 // volatile too
3080 kmp_team_p *t_parent; // parent team
3081 kmp_team_p *t_next_pool; // next free team in the team pool
3082 kmp_disp_t *t_dispatch; // thread's dispatch data
3083 kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
3084 kmp_proc_bind_t t_proc_bind; // bind type for par region
3085#if USE_ITT_BUILD
3086 kmp_uint64 t_region_time; // region begin timestamp
3087#endif /* USE_ITT_BUILD */
3088
3089 // Primary thread write, workers read
3090 // --------------------------------------------------------------------------
3091 KMP_ALIGN_CACHE void **t_argv;
3092 int t_argc;
3093 int t_nproc; // number of threads in team
3094 microtask_t t_pkfn;
3095 launch_t t_invoke; // procedure to launch the microtask
3096
3097#if OMPT_SUPPORT
3098 ompt_team_info_t ompt_team_info;
3099 ompt_lw_taskteam_t *ompt_serialized_team_info;
3100#endif
3101
3102#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3103 kmp_int8 t_fp_control_saved;
3104 kmp_int8 t_pad2b;
3105 kmp_int16 t_x87_fpu_control_word; // FP control regs
3106 kmp_uint32 t_mxcsr;
3107#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3108
3109 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
3110
3111 KMP_ALIGN_CACHE kmp_info_t **t_threads;
3112 kmp_taskdata_t
3113 *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
3114 int t_level; // nested parallel level
3115
3116 KMP_ALIGN_CACHE int t_max_argc;
3117 int t_max_nproc; // max threads this team can handle (dynamically expandable)
3118 int t_serialized; // levels deep of serialized teams
3119 dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
3120 int t_id; // team's id, assigned by debugger.
3121 int t_active_level; // nested active parallel level
3122 kmp_r_sched_t t_sched; // run-time schedule for the team
3123#if KMP_AFFINITY_SUPPORTED
3124 int t_first_place; // first & last place in parent thread's partition.
3125 int t_last_place; // Restore these values to primary thread after par region.
3126#endif // KMP_AFFINITY_SUPPORTED
3127 int t_display_affinity;
3128 int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
3129 // omp_set_num_threads() call
3130 omp_allocator_handle_t t_def_allocator; /* default allocator */
3131
3132// Read/write by workers as well
3133#if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
3134 // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
3135 // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
3136 // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
3137 // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
3138 char dummy_padding[1024];
3139#endif
3140 // Internal control stack for additional nested teams.
3141 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
3142 // for SERIALIZED teams nested 2 or more levels deep
3143 // typed flag to store request state of cancellation
3144 std::atomic<kmp_int32> t_cancel_request;
3145 int t_master_active; // save on fork, restore on join
3146 void *t_copypriv_data; // team specific pointer to copyprivate data array
3147#if KMP_OS_WINDOWS
3148 std::atomic<kmp_uint32> t_copyin_counter;
3149#endif
3150#if USE_ITT_BUILD
3151 void *t_stack_id; // team specific stack stitching id (for ittnotify)
3152#endif /* USE_ITT_BUILD */
3153 distributedBarrier *b; // Distributed barrier data associated with team
3154} kmp_base_team_t;
3155
3156union KMP_ALIGN_CACHE kmp_team {
3157 kmp_base_team_t t;
3158 double t_align; /* use worst case alignment */
3159 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
3160};
3161
3162typedef union KMP_ALIGN_CACHE kmp_time_global {
3163 double dt_align; /* use worst case alignment */
3164 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3165 kmp_base_data_t dt;
3166} kmp_time_global_t;
3167
3168typedef struct kmp_base_global {
3169 /* cache-aligned */
3170 kmp_time_global_t g_time;
3171
3172 /* non cache-aligned */
3173 volatile int g_abort;
3174 volatile int g_done;
3175
3176 int g_dynamic;
3177 enum dynamic_mode g_dynamic_mode;
3178} kmp_base_global_t;
3179
3180typedef union KMP_ALIGN_CACHE kmp_global {
3181 kmp_base_global_t g;
3182 double g_align; /* use worst case alignment */
3183 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
3184} kmp_global_t;
3185
3186typedef struct kmp_base_root {
3187 // TODO: GEH - combine r_active with r_in_parallel then r_active ==
3188 // (r_in_parallel>= 0)
3189 // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
3190 // the synch overhead or keeping r_active
3191 volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
3192 // keeps a count of active parallel regions per root
3193 std::atomic<int> r_in_parallel;
3194 // GEH: This is misnamed, should be r_active_levels
3195 kmp_team_t *r_root_team;
3196 kmp_team_t *r_hot_team;
3197 kmp_info_t *r_uber_thread;
3198 kmp_lock_t r_begin_lock;
3199 volatile int r_begin;
3200 int r_blocktime; /* blocktime for this root and descendants */
3201#if KMP_AFFINITY_SUPPORTED
3202 int r_affinity_assigned;
3203#endif // KMP_AFFINITY_SUPPORTED
3204} kmp_base_root_t;
3205
3206typedef union KMP_ALIGN_CACHE kmp_root {
3207 kmp_base_root_t r;
3208 double r_align; /* use worst case alignment */
3209 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
3210} kmp_root_t;
3211
3212struct fortran_inx_info {
3213 kmp_int32 data;
3214};
3215
3216// This list type exists to hold old __kmp_threads arrays so that
3217// old references to them may complete while reallocation takes place when
3218// expanding the array. The items in this list are kept alive until library
3219// shutdown.
3220typedef struct kmp_old_threads_list_t {
3221 kmp_info_t **threads;
3222 struct kmp_old_threads_list_t *next;
3223} kmp_old_threads_list_t;
3224
3225/* ------------------------------------------------------------------------ */
3226
3227extern int __kmp_settings;
3228extern int __kmp_duplicate_library_ok;
3229#if USE_ITT_BUILD
3230extern int __kmp_forkjoin_frames;
3231extern int __kmp_forkjoin_frames_mode;
3232#endif
3233extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3234extern int __kmp_determ_red;
3235
3236#ifdef KMP_DEBUG
3237extern int kmp_a_debug;
3238extern int kmp_b_debug;
3239extern int kmp_c_debug;
3240extern int kmp_d_debug;
3241extern int kmp_e_debug;
3242extern int kmp_f_debug;
3243#endif /* KMP_DEBUG */
3244
3245/* For debug information logging using rotating buffer */
3246#define KMP_DEBUG_BUF_LINES_INIT 512
3247#define KMP_DEBUG_BUF_LINES_MIN 1
3248
3249#define KMP_DEBUG_BUF_CHARS_INIT 128
3250#define KMP_DEBUG_BUF_CHARS_MIN 2
3251
3252extern int
3253 __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3254extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3255extern int
3256 __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3257extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3258 entry pointer */
3259
3260extern char *__kmp_debug_buffer; /* Debug buffer itself */
3261extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3262 printed in buffer so far */
3263extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3264 recommended in warnings */
3265/* end rotating debug buffer */
3266
3267#ifdef KMP_DEBUG
3268extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3269
3270#define KMP_PAR_RANGE_ROUTINE_LEN 1024
3271extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3272#define KMP_PAR_RANGE_FILENAME_LEN 1024
3273extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3274extern int __kmp_par_range_lb;
3275extern int __kmp_par_range_ub;
3276#endif
3277
3278/* For printing out dynamic storage map for threads and teams */
3279extern int
3280 __kmp_storage_map; /* True means print storage map for threads and teams */
3281extern int __kmp_storage_map_verbose; /* True means storage map includes
3282 placement info */
3283extern int __kmp_storage_map_verbose_specified;
3284
3285#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3286extern kmp_cpuinfo_t __kmp_cpuinfo;
3287static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3288#elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3289static inline bool __kmp_is_hybrid_cpu() { return true; }
3290#else
3291static inline bool __kmp_is_hybrid_cpu() { return false; }
3292#endif
3293
3294extern volatile int __kmp_init_serial;
3295extern volatile int __kmp_init_gtid;
3296extern volatile int __kmp_init_common;
3297extern volatile int __kmp_need_register_serial;
3298extern volatile int __kmp_init_middle;
3299extern volatile int __kmp_init_parallel;
3300#if KMP_USE_MONITOR
3301extern volatile int __kmp_init_monitor;
3302#endif
3303extern volatile int __kmp_init_user_locks;
3304extern volatile int __kmp_init_hidden_helper_threads;
3305extern int __kmp_init_counter;
3306extern int __kmp_root_counter;
3307extern int __kmp_version;
3308
3309/* list of address of allocated caches for commons */
3310extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3311
3312/* Barrier algorithm types and options */
3313extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3314extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3315extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3316extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3317extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3318extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3319extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3320extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3321extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3322extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3323extern char const *__kmp_barrier_type_name[bs_last_barrier];
3324extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3325
3326/* Global Locks */
3327extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3328extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3329extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3330extern kmp_bootstrap_lock_t
3331 __kmp_exit_lock; /* exit() is not always thread-safe */
3332#if KMP_USE_MONITOR
3333extern kmp_bootstrap_lock_t
3334 __kmp_monitor_lock; /* control monitor thread creation */
3335#endif
3336extern kmp_bootstrap_lock_t
3337 __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3338 __kmp_threads expansion to co-exist */
3339
3340extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
3341extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
3342extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3343
3344extern enum library_type __kmp_library;
3345
3346extern enum sched_type __kmp_sched; /* default runtime scheduling */
3347extern enum sched_type __kmp_static; /* default static scheduling method */
3348extern enum sched_type __kmp_guided; /* default guided scheduling method */
3349extern enum sched_type __kmp_auto; /* default auto scheduling method */
3350extern int __kmp_chunk; /* default runtime chunk size */
3351extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3352
3353extern size_t __kmp_stksize; /* stack size per thread */
3354#if KMP_USE_MONITOR
3355extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3356#endif
3357extern size_t __kmp_stkoffset; /* stack offset per thread */
3358extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3359
3360extern size_t
3361 __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3362extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3363extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3364extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
3365extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3366extern int __kmp_generate_warnings; /* should we issue warnings? */
3367extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3368
3369#ifdef DEBUG_SUSPEND
3370extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3371#endif
3372
3373extern kmp_int32 __kmp_use_yield;
3374extern kmp_int32 __kmp_use_yield_exp_set;
3375extern kmp_uint32 __kmp_yield_init;
3376extern kmp_uint32 __kmp_yield_next;
3377extern kmp_uint64 __kmp_pause_init;
3378
3379/* ------------------------------------------------------------------------- */
3380extern int __kmp_allThreadsSpecified;
3381
3382extern size_t __kmp_align_alloc;
3383/* following data protected by initialization routines */
3384extern int __kmp_xproc; /* number of processors in the system */
3385extern int __kmp_avail_proc; /* number of processors available to the process */
3386extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3387extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3388// maximum total number of concurrently-existing threads on device
3389extern int __kmp_max_nth;
3390// maximum total number of concurrently-existing threads in a contention group
3391extern int __kmp_cg_max_nth;
3392extern int __kmp_task_max_nth; // max threads used in a task
3393extern int __kmp_teams_max_nth; // max threads used in a teams construct
3394extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3395 __kmp_root */
3396extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3397 region a la OMP_NUM_THREADS */
3398extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3399 initialization */
3400extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3401 used (fixed) */
3402extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3403 (__kmpc_threadprivate_cached()) */
3404extern int __kmp_dflt_blocktime; /* number of microseconds to wait before
3405 blocking (env setting) */
3406extern char __kmp_blocktime_units; /* 'm' or 'u' to note units specified */
3407extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */
3408
3409// Convert raw blocktime from ms to us if needed.
3410static inline void __kmp_aux_convert_blocktime(int *bt) {
3411 if (__kmp_blocktime_units == 'm') {
3412 if (*bt > INT_MAX / 1000) {
3413 *bt = INT_MAX / 1000;
3414 KMP_INFORM(MaxValueUsing, "kmp_set_blocktime(ms)", bt);
3415 }
3416 *bt = *bt * 1000;
3417 }
3418}
3419
3420#if KMP_USE_MONITOR
3421extern int
3422 __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3423extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3424 blocking */
3425#endif
3426#ifdef KMP_ADJUST_BLOCKTIME
3427extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3428#endif /* KMP_ADJUST_BLOCKTIME */
3429#ifdef KMP_DFLT_NTH_CORES
3430extern int __kmp_ncores; /* Total number of cores for threads placement */
3431#endif
3432/* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3433extern int __kmp_abort_delay;
3434
3435extern int __kmp_need_register_atfork_specified;
3436extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3437 to install fork handler */
3438extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3439 0 - not set, will be set at runtime
3440 1 - using stack search
3441 2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3442 X*) or TlsGetValue(Windows* OS))
3443 3 - static TLS (__declspec(thread) __kmp_gtid),
3444 Linux* OS .so only. */
3445extern int
3446 __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3447#ifdef KMP_TDATA_GTID
3448extern KMP_THREAD_LOCAL int __kmp_gtid;
3449#endif
3450extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3451extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3452#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3453extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3454extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3455extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3456#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3457
3458// max_active_levels for nested parallelism enabled by default via
3459// OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3460extern int __kmp_dflt_max_active_levels;
3461// Indicates whether value of __kmp_dflt_max_active_levels was already
3462// explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3463extern bool __kmp_dflt_max_active_levels_set;
3464extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3465 concurrent execution per team */
3466#if KMP_NESTED_HOT_TEAMS
3467extern int __kmp_hot_teams_mode;
3468extern int __kmp_hot_teams_max_level;
3469#endif
3470
3471#if KMP_OS_LINUX
3472extern enum clock_function_type __kmp_clock_function;
3473extern int __kmp_clock_function_param;
3474#endif /* KMP_OS_LINUX */
3475
3476#if KMP_MIC_SUPPORTED
3477extern enum mic_type __kmp_mic_type;
3478#endif
3479
3480#ifdef USE_LOAD_BALANCE
3481extern double __kmp_load_balance_interval; // load balance algorithm interval
3482#endif /* USE_LOAD_BALANCE */
3483
3484// OpenMP 3.1 - Nested num threads array
3485typedef struct kmp_nested_nthreads_t {
3486 int *nth;
3487 int size;
3488 int used;
3489} kmp_nested_nthreads_t;
3490
3491extern kmp_nested_nthreads_t __kmp_nested_nth;
3492
3493#if KMP_USE_ADAPTIVE_LOCKS
3494
3495// Parameters for the speculative lock backoff system.
3496struct kmp_adaptive_backoff_params_t {
3497 // Number of soft retries before it counts as a hard retry.
3498 kmp_uint32 max_soft_retries;
3499 // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3500 // the right
3501 kmp_uint32 max_badness;
3502};
3503
3504extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3505
3506#if KMP_DEBUG_ADAPTIVE_LOCKS
3507extern const char *__kmp_speculative_statsfile;
3508#endif
3509
3510#endif // KMP_USE_ADAPTIVE_LOCKS
3511
3512extern int __kmp_display_env; /* TRUE or FALSE */
3513extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3514extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3515extern int __kmp_nteams;
3516extern int __kmp_teams_thread_limit;
3517
3518/* ------------------------------------------------------------------------- */
3519
3520/* the following are protected by the fork/join lock */
3521/* write: lock read: anytime */
3522extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3523/* Holds old arrays of __kmp_threads until library shutdown */
3524extern kmp_old_threads_list_t *__kmp_old_threads_list;
3525/* read/write: lock */
3526extern volatile kmp_team_t *__kmp_team_pool;
3527extern volatile kmp_info_t *__kmp_thread_pool;
3528extern kmp_info_t *__kmp_thread_pool_insert_pt;
3529
3530// total num threads reachable from some root thread including all root threads
3531extern volatile int __kmp_nth;
3532/* total number of threads reachable from some root thread including all root
3533 threads, and those in the thread pool */
3534extern volatile int __kmp_all_nth;
3535extern std::atomic<int> __kmp_thread_pool_active_nth;
3536
3537extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3538/* end data protected by fork/join lock */
3539/* ------------------------------------------------------------------------- */
3540
3541#define __kmp_get_gtid() __kmp_get_global_thread_id()
3542#define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3543#define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3544#define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3545#define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3546
3547// AT: Which way is correct?
3548// AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3549// AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3550#define __kmp_get_team_num_threads(gtid) \
3551 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3552
3553static inline bool KMP_UBER_GTID(int gtid) {
3554 KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3555 KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3556 return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3557 __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3558}
3559
3560static inline int __kmp_tid_from_gtid(int gtid) {
3561 KMP_DEBUG_ASSERT(gtid >= 0);
3562 return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3563}
3564
3565static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3566 KMP_DEBUG_ASSERT(tid >= 0 && team);
3567 return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3568}
3569
3570static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3571 KMP_DEBUG_ASSERT(thr);
3572 return thr->th.th_info.ds.ds_gtid;
3573}
3574
3575static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3576 KMP_DEBUG_ASSERT(gtid >= 0);
3577 return __kmp_threads[gtid];
3578}
3579
3580static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3581 KMP_DEBUG_ASSERT(gtid >= 0);
3582 return __kmp_threads[gtid]->th.th_team;
3583}
3584
3585static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3586 if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3587 KMP_FATAL(ThreadIdentInvalid);
3588}
3589
3590#if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3591extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3592extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3593extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3594extern int __kmp_mwait_hints; // Hints to pass in to mwait
3595#endif
3596
3597#if KMP_HAVE_UMWAIT
3598extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3599extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3600extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3601extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3602#endif
3603
3604/* ------------------------------------------------------------------------- */
3605
3606extern kmp_global_t __kmp_global; /* global status */
3607
3608extern kmp_info_t __kmp_monitor;
3609// For Debugging Support Library
3610extern std::atomic<kmp_int32> __kmp_team_counter;
3611// For Debugging Support Library
3612extern std::atomic<kmp_int32> __kmp_task_counter;
3613
3614#if USE_DEBUGGER
3615#define _KMP_GEN_ID(counter) \
3616 (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3617#else
3618#define _KMP_GEN_ID(counter) (~0)
3619#endif /* USE_DEBUGGER */
3620
3621#define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3622#define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3623
3624/* ------------------------------------------------------------------------ */
3625
3626extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3627 size_t size, char const *format, ...);
3628
3629extern void __kmp_serial_initialize(void);
3630extern void __kmp_middle_initialize(void);
3631extern void __kmp_parallel_initialize(void);
3632
3633extern void __kmp_internal_begin(void);
3634extern void __kmp_internal_end_library(int gtid);
3635extern void __kmp_internal_end_thread(int gtid);
3636extern void __kmp_internal_end_atexit(void);
3637extern void __kmp_internal_end_dtor(void);
3638extern void __kmp_internal_end_dest(void *);
3639
3640extern int __kmp_register_root(int initial_thread);
3641extern void __kmp_unregister_root(int gtid);
3642extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3643
3644extern int __kmp_ignore_mppbeg(void);
3645extern int __kmp_ignore_mppend(void);
3646
3647extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3648extern void __kmp_exit_single(int gtid);
3649
3650extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3651extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3652
3653#ifdef USE_LOAD_BALANCE
3654extern int __kmp_get_load_balance(int);
3655#endif
3656
3657extern int __kmp_get_global_thread_id(void);
3658extern int __kmp_get_global_thread_id_reg(void);
3659extern void __kmp_exit_thread(int exit_status);
3660extern void __kmp_abort(char const *format, ...);
3661extern void __kmp_abort_thread(void);
3662KMP_NORETURN extern void __kmp_abort_process(void);
3663extern void __kmp_warn(char const *format, ...);
3664
3665extern void __kmp_set_num_threads(int new_nth, int gtid);
3666
3667extern bool __kmp_detect_shm();
3668extern bool __kmp_detect_tmp();
3669
3670// Returns current thread (pointer to kmp_info_t). Current thread *must* be
3671// registered.
3672static inline kmp_info_t *__kmp_entry_thread() {
3673 int gtid = __kmp_entry_gtid();
3674
3675 return __kmp_threads[gtid];
3676}
3677
3678extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3679extern int __kmp_get_max_active_levels(int gtid);
3680extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3681extern int __kmp_get_team_size(int gtid, int level);
3682extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3683extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3684
3685extern unsigned short __kmp_get_random(kmp_info_t *thread);
3686extern void __kmp_init_random(kmp_info_t *thread);
3687
3688extern kmp_r_sched_t __kmp_get_schedule_global(void);
3689extern void __kmp_adjust_num_threads(int new_nproc);
3690extern void __kmp_check_stksize(size_t *val);
3691
3692extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3693extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3694extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3695#define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3696#define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3697#define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3698
3699#if USE_FAST_MEMORY
3700extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3701 size_t size KMP_SRC_LOC_DECL);
3702extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3703extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3704extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3705#define __kmp_fast_allocate(this_thr, size) \
3706 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3707#define __kmp_fast_free(this_thr, ptr) \
3708 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3709#endif
3710
3711extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3712extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3713 size_t elsize KMP_SRC_LOC_DECL);
3714extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3715 size_t size KMP_SRC_LOC_DECL);
3716extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3717#define __kmp_thread_malloc(th, size) \
3718 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3719#define __kmp_thread_calloc(th, nelem, elsize) \
3720 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3721#define __kmp_thread_realloc(th, ptr, size) \
3722 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3723#define __kmp_thread_free(th, ptr) \
3724 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3725
3726extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3727
3728extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3729 kmp_proc_bind_t proc_bind);
3730extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3731 int num_threads);
3732extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3733 int num_teams_ub, int num_threads);
3734
3735extern void __kmp_yield();
3736
3737extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3738 enum sched_type schedule, kmp_int32 lb,
3739 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3740extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3741 enum sched_type schedule, kmp_uint32 lb,
3742 kmp_uint32 ub, kmp_int32 st,
3743 kmp_int32 chunk);
3744extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3745 enum sched_type schedule, kmp_int64 lb,
3746 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3747extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3748 enum sched_type schedule, kmp_uint64 lb,
3749 kmp_uint64 ub, kmp_int64 st,
3750 kmp_int64 chunk);
3751
3752extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3753 kmp_int32 *p_last, kmp_int32 *p_lb,
3754 kmp_int32 *p_ub, kmp_int32 *p_st);
3755extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3756 kmp_int32 *p_last, kmp_uint32 *p_lb,
3757 kmp_uint32 *p_ub, kmp_int32 *p_st);
3758extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3759 kmp_int32 *p_last, kmp_int64 *p_lb,
3760 kmp_int64 *p_ub, kmp_int64 *p_st);
3761extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3762 kmp_int32 *p_last, kmp_uint64 *p_lb,
3763 kmp_uint64 *p_ub, kmp_int64 *p_st);
3764
3765extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3766extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3767extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3768extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3769
3770#ifdef KMP_GOMP_COMPAT
3771
3772extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3773 enum sched_type schedule, kmp_int32 lb,
3774 kmp_int32 ub, kmp_int32 st,
3775 kmp_int32 chunk, int push_ws);
3776extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3777 enum sched_type schedule, kmp_uint32 lb,
3778 kmp_uint32 ub, kmp_int32 st,
3779 kmp_int32 chunk, int push_ws);
3780extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3781 enum sched_type schedule, kmp_int64 lb,
3782 kmp_int64 ub, kmp_int64 st,
3783 kmp_int64 chunk, int push_ws);
3784extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3785 enum sched_type schedule, kmp_uint64 lb,
3786 kmp_uint64 ub, kmp_int64 st,
3787 kmp_int64 chunk, int push_ws);
3788extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3789extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3790extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3791extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3792
3793#endif /* KMP_GOMP_COMPAT */
3794
3795extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3796extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3797extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3798extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3799extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3800extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3801 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3802 void *obj);
3803extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3804 kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3805
3806extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3807 int final_spin
3808#if USE_ITT_BUILD
3809 ,
3810 void *itt_sync_obj
3811#endif
3812);
3813extern void __kmp_release_64(kmp_flag_64<> *flag);
3814
3815extern void __kmp_infinite_loop(void);
3816
3817extern void __kmp_cleanup(void);
3818
3819#if KMP_HANDLE_SIGNALS
3820extern int __kmp_handle_signals;
3821extern void __kmp_install_signals(int parallel_init);
3822extern void __kmp_remove_signals(void);
3823#endif
3824
3825extern void __kmp_clear_system_time(void);
3826extern void __kmp_read_system_time(double *delta);
3827
3828extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3829
3830extern void __kmp_expand_host_name(char *buffer, size_t size);
3831extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3832
3833#if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM))
3834extern void
3835__kmp_initialize_system_tick(void); /* Initialize timer tick value */
3836#endif
3837
3838extern void
3839__kmp_runtime_initialize(void); /* machine specific initialization */
3840extern void __kmp_runtime_destroy(void);
3841
3842#if KMP_AFFINITY_SUPPORTED
3843extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3844 kmp_affin_mask_t *mask);
3845extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3846 kmp_affin_mask_t *mask);
3847extern void __kmp_affinity_initialize(kmp_affinity_t &affinity);
3848extern void __kmp_affinity_uninitialize(void);
3849extern void __kmp_affinity_set_init_mask(
3850 int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3851void __kmp_affinity_bind_init_mask(int gtid);
3852extern void __kmp_affinity_bind_place(int gtid);
3853extern void __kmp_affinity_determine_capable(const char *env_var);
3854extern int __kmp_aux_set_affinity(void **mask);
3855extern int __kmp_aux_get_affinity(void **mask);
3856extern int __kmp_aux_get_affinity_max_proc();
3857extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3858extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3859extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3860extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3861#if KMP_OS_LINUX || KMP_OS_FREEBSD
3862extern int kmp_set_thread_affinity_mask_initial(void);
3863#endif
3864static inline void __kmp_assign_root_init_mask() {
3865 int gtid = __kmp_entry_gtid();
3866 kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3867 if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3868 __kmp_affinity_set_init_mask(gtid, /*isa_root=*/TRUE);
3869 __kmp_affinity_bind_init_mask(gtid);
3870 r->r.r_affinity_assigned = TRUE;
3871 }
3872}
3873static inline void __kmp_reset_root_init_mask(int gtid) {
3874 if (!KMP_AFFINITY_CAPABLE())
3875 return;
3876 kmp_info_t *th = __kmp_threads[gtid];
3877 kmp_root_t *r = th->th.th_root;
3878 if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3879 __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3880 KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3881 r->r.r_affinity_assigned = FALSE;
3882 }
3883}
3884#else /* KMP_AFFINITY_SUPPORTED */
3885#define __kmp_assign_root_init_mask() /* Nothing */
3886static inline void __kmp_reset_root_init_mask(int gtid) {}
3887#endif /* KMP_AFFINITY_SUPPORTED */
3888// No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3889// format string is for affinity, so platforms that do not support
3890// affinity can still use the other fields, e.g., %n for num_threads
3891extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3892 kmp_str_buf_t *buffer);
3893extern void __kmp_aux_display_affinity(int gtid, const char *format);
3894
3895extern void __kmp_cleanup_hierarchy();
3896extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3897
3898#if KMP_USE_FUTEX
3899
3900extern int __kmp_futex_determine_capable(void);
3901
3902#endif // KMP_USE_FUTEX
3903
3904extern void __kmp_gtid_set_specific(int gtid);
3905extern int __kmp_gtid_get_specific(void);
3906
3907extern double __kmp_read_cpu_time(void);
3908
3909extern int __kmp_read_system_info(struct kmp_sys_info *info);
3910
3911#if KMP_USE_MONITOR
3912extern void __kmp_create_monitor(kmp_info_t *th);
3913#endif
3914
3915extern void *__kmp_launch_thread(kmp_info_t *thr);
3916
3917extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3918
3919#if KMP_OS_WINDOWS
3920extern int __kmp_still_running(kmp_info_t *th);
3921extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3922extern void __kmp_free_handle(kmp_thread_t tHandle);
3923#endif
3924
3925#if KMP_USE_MONITOR
3926extern void __kmp_reap_monitor(kmp_info_t *th);
3927#endif
3928extern void __kmp_reap_worker(kmp_info_t *th);
3929extern void __kmp_terminate_thread(int gtid);
3930
3931extern int __kmp_try_suspend_mx(kmp_info_t *th);
3932extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3933extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3934
3935extern void __kmp_elapsed(double *);
3936extern void __kmp_elapsed_tick(double *);
3937
3938extern void __kmp_enable(int old_state);
3939extern void __kmp_disable(int *old_state);
3940
3941extern void __kmp_thread_sleep(int millis);
3942
3943extern void __kmp_common_initialize(void);
3944extern void __kmp_common_destroy(void);
3945extern void __kmp_common_destroy_gtid(int gtid);
3946
3947#if KMP_OS_UNIX
3948extern void __kmp_register_atfork(void);
3949#endif
3950extern void __kmp_suspend_initialize(void);
3951extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3952extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3953
3954extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3955 int tid);
3956extern kmp_team_t *
3957__kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3958#if OMPT_SUPPORT
3959 ompt_data_t ompt_parallel_data,
3960#endif
3961 kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3962 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3963extern void __kmp_free_thread(kmp_info_t *);
3964extern void __kmp_free_team(kmp_root_t *,
3965 kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3966extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3967
3968/* ------------------------------------------------------------------------ */
3969
3970extern void __kmp_initialize_bget(kmp_info_t *th);
3971extern void __kmp_finalize_bget(kmp_info_t *th);
3972
3973KMP_EXPORT void *kmpc_malloc(size_t size);
3974KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3975KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3976KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3977KMP_EXPORT void kmpc_free(void *ptr);
3978
3979/* declarations for internal use */
3980
3981extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3982 size_t reduce_size, void *reduce_data,
3983 void (*reduce)(void *, void *));
3984extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3985extern int __kmp_barrier_gomp_cancel(int gtid);
3986
3991enum fork_context_e {
3992 fork_context_gnu,
3994 fork_context_intel,
3995 fork_context_last
3996};
3997extern int __kmp_fork_call(ident_t *loc, int gtid,
3998 enum fork_context_e fork_context, kmp_int32 argc,
3999 microtask_t microtask, launch_t invoker,
4000 kmp_va_list ap);
4001
4002extern void __kmp_join_call(ident_t *loc, int gtid
4003#if OMPT_SUPPORT
4004 ,
4005 enum fork_context_e fork_context
4006#endif
4007 ,
4008 int exit_teams = 0);
4009
4010extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
4011extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
4012extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
4013extern int __kmp_invoke_task_func(int gtid);
4014extern void __kmp_run_before_invoked_task(int gtid, int tid,
4015 kmp_info_t *this_thr,
4016 kmp_team_t *team);
4017extern void __kmp_run_after_invoked_task(int gtid, int tid,
4018 kmp_info_t *this_thr,
4019 kmp_team_t *team);
4020
4021// should never have been exported
4022KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
4023extern int __kmp_invoke_teams_master(int gtid);
4024extern void __kmp_teams_master(int gtid);
4025extern int __kmp_aux_get_team_num();
4026extern int __kmp_aux_get_num_teams();
4027extern void __kmp_save_internal_controls(kmp_info_t *thread);
4028extern void __kmp_user_set_library(enum library_type arg);
4029extern void __kmp_aux_set_library(enum library_type arg);
4030extern void __kmp_aux_set_stacksize(size_t arg);
4031extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
4032extern void __kmp_aux_set_defaults(char const *str, size_t len);
4033
4034/* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
4035void kmpc_set_blocktime(int arg);
4036void ompc_set_nested(int flag);
4037void ompc_set_dynamic(int flag);
4038void ompc_set_num_threads(int arg);
4039
4040extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
4041 kmp_team_t *team, int tid);
4042extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
4043extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4044 kmp_tasking_flags_t *flags,
4045 size_t sizeof_kmp_task_t,
4046 size_t sizeof_shareds,
4047 kmp_routine_entry_t task_entry);
4048extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
4049 kmp_team_t *team, int tid,
4050 int set_curr_task);
4051extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
4052extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
4053
4054extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
4055 int gtid,
4056 kmp_task_t *task);
4057extern void __kmp_fulfill_event(kmp_event_t *event);
4058
4059extern void __kmp_free_task_team(kmp_info_t *thread,
4060 kmp_task_team_t *task_team);
4061extern void __kmp_reap_task_teams(void);
4062extern void __kmp_wait_to_unref_task_teams(void);
4063extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
4064 int always);
4065extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
4066extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
4067#if USE_ITT_BUILD
4068 ,
4069 void *itt_sync_obj
4070#endif /* USE_ITT_BUILD */
4071 ,
4072 int wait = 1);
4073extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
4074 int gtid);
4075
4076extern int __kmp_is_address_mapped(void *addr);
4077extern kmp_uint64 __kmp_hardware_timestamp(void);
4078
4079#if KMP_OS_UNIX
4080extern int __kmp_read_from_file(char const *path, char const *format, ...);
4081#endif
4082
4083/* ------------------------------------------------------------------------ */
4084//
4085// Assembly routines that have no compiler intrinsic replacement
4086//
4087
4088extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
4089 void *argv[]
4090#if OMPT_SUPPORT
4091 ,
4092 void **exit_frame_ptr
4093#endif
4094);
4095
4096/* ------------------------------------------------------------------------ */
4097
4098KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
4099KMP_EXPORT void __kmpc_end(ident_t *);
4100
4101KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
4102 kmpc_ctor_vec ctor,
4103 kmpc_cctor_vec cctor,
4104 kmpc_dtor_vec dtor,
4105 size_t vector_length);
4106KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
4107 kmpc_ctor ctor, kmpc_cctor cctor,
4108 kmpc_dtor dtor);
4109KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
4110 void *data, size_t size);
4111
4112KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
4113KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
4114KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
4115KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
4116
4117KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
4118KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
4119 kmpc_micro microtask, ...);
4120KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs,
4121 kmpc_micro microtask, kmp_int32 cond,
4122 void *args);
4123
4124KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
4125KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
4126
4127KMP_EXPORT void __kmpc_flush(ident_t *);
4128KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
4129KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
4130KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
4131KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
4132 kmp_int32 filter);
4133KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
4134KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
4135KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
4136KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
4137 kmp_critical_name *);
4138KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
4139 kmp_critical_name *);
4140KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
4141 kmp_critical_name *, uint32_t hint);
4142
4143KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
4144KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
4145
4146KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
4147 kmp_int32 global_tid);
4148
4149KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
4150KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
4151
4152KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid);
4153KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid,
4154 kmp_int32 numberOfSections);
4155KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid);
4156
4157KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
4158 kmp_int32 schedtype, kmp_int32 *plastiter,
4159 kmp_int *plower, kmp_int *pupper,
4160 kmp_int *pstride, kmp_int incr,
4161 kmp_int chunk);
4162
4163KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
4164
4165KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
4166 size_t cpy_size, void *cpy_data,
4167 void (*cpy_func)(void *, void *),
4168 kmp_int32 didit);
4169
4170KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid,
4171 void *cpy_data);
4172
4173extern void KMPC_SET_NUM_THREADS(int arg);
4174extern void KMPC_SET_DYNAMIC(int flag);
4175extern void KMPC_SET_NESTED(int flag);
4176
4177/* OMP 3.0 tasking interface routines */
4178KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
4179 kmp_task_t *new_task);
4180KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4181 kmp_int32 flags,
4182 size_t sizeof_kmp_task_t,
4183 size_t sizeof_shareds,
4184 kmp_routine_entry_t task_entry);
4185KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
4186 ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
4187 size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
4188KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
4189 kmp_task_t *task);
4190KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
4191 kmp_task_t *task);
4192KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
4193 kmp_task_t *new_task);
4194KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
4195KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
4196 int end_part);
4197
4198#if TASK_UNUSED
4199void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
4200void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
4201 kmp_task_t *task);
4202#endif // TASK_UNUSED
4203
4204/* ------------------------------------------------------------------------ */
4205
4206KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
4207KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
4208
4209KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
4210 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
4211 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
4212 kmp_depend_info_t *noalias_dep_list);
4213KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
4214 kmp_int32 ndeps,
4215 kmp_depend_info_t *dep_list,
4216 kmp_int32 ndeps_noalias,
4217 kmp_depend_info_t *noalias_dep_list);
4218/* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause.
4219 * Placeholder for taskwait with nowait clause.*/
4220KMP_EXPORT void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid,
4221 kmp_int32 ndeps,
4222 kmp_depend_info_t *dep_list,
4223 kmp_int32 ndeps_noalias,
4224 kmp_depend_info_t *noalias_dep_list,
4225 kmp_int32 has_no_wait);
4226
4227extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
4228 bool serialize_immediate);
4229
4230KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
4231 kmp_int32 cncl_kind);
4232KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
4233 kmp_int32 cncl_kind);
4234KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
4235KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
4236
4237KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
4238KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
4239KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
4240 kmp_int32 if_val, kmp_uint64 *lb,
4241 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
4242 kmp_int32 sched, kmp_uint64 grainsize,
4243 void *task_dup);
4244KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
4245 kmp_task_t *task, kmp_int32 if_val,
4246 kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4247 kmp_int32 nogroup, kmp_int32 sched,
4248 kmp_uint64 grainsize, kmp_int32 modifier,
4249 void *task_dup);
4250KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
4251KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
4252KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
4253KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
4254 int is_ws, int num,
4255 void *data);
4256KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
4257 int num, void *data);
4258KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
4259 int is_ws);
4260KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
4261 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4262 kmp_task_affinity_info_t *affin_list);
4263KMP_EXPORT void __kmp_set_num_teams(int num_teams);
4264KMP_EXPORT int __kmp_get_max_teams(void);
4265KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
4266KMP_EXPORT int __kmp_get_teams_thread_limit(void);
4267
4268/* Interface target task integration */
4269KMP_EXPORT void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid);
4270KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid);
4271
4272/* Lock interface routines (fast versions with gtid passed in) */
4273KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
4274 void **user_lock);
4275KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
4276 void **user_lock);
4277KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
4278 void **user_lock);
4279KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
4280 void **user_lock);
4281KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4282KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
4283 void **user_lock);
4284KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
4285 void **user_lock);
4286KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
4287 void **user_lock);
4288KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4289KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
4290 void **user_lock);
4291
4292KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4293 void **user_lock, uintptr_t hint);
4294KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4295 void **user_lock,
4296 uintptr_t hint);
4297
4298#if OMPX_TASKGRAPH
4299// Taskgraph's Record & Replay mechanism
4300// __kmp_tdg_is_recording: check whether a given TDG is recording
4301// status: the tdg's current status
4302static inline bool __kmp_tdg_is_recording(kmp_tdg_status_t status) {
4303 return status == KMP_TDG_RECORDING;
4304}
4305
4306KMP_EXPORT kmp_int32 __kmpc_start_record_task(ident_t *loc, kmp_int32 gtid,
4307 kmp_int32 input_flags,
4308 kmp_int32 tdg_id);
4309KMP_EXPORT void __kmpc_end_record_task(ident_t *loc, kmp_int32 gtid,
4310 kmp_int32 input_flags, kmp_int32 tdg_id);
4311#endif
4312/* Interface to fast scalable reduce methods routines */
4313
4314KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
4315 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4316 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4317 kmp_critical_name *lck);
4318KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4319 kmp_critical_name *lck);
4320KMP_EXPORT kmp_int32 __kmpc_reduce(
4321 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4322 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4323 kmp_critical_name *lck);
4324KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4325 kmp_critical_name *lck);
4326
4327/* Internal fast reduction routines */
4328
4329extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4330 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4331 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4332 kmp_critical_name *lck);
4333
4334// this function is for testing set/get/determine reduce method
4335KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4336
4337KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4338KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4339
4340// C++ port
4341// missing 'extern "C"' declarations
4342
4343KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4344KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4345KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4346 kmp_int32 num_threads);
4347
4348KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4349 int proc_bind);
4350KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4351 kmp_int32 num_teams,
4352 kmp_int32 num_threads);
4353KMP_EXPORT void __kmpc_set_thread_limit(ident_t *loc, kmp_int32 global_tid,
4354 kmp_int32 thread_limit);
4355/* Function for OpenMP 5.1 num_teams clause */
4356KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4357 kmp_int32 num_teams_lb,
4358 kmp_int32 num_teams_ub,
4359 kmp_int32 num_threads);
4360KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4361 kmpc_micro microtask, ...);
4362struct kmp_dim { // loop bounds info casted to kmp_int64
4363 kmp_int64 lo; // lower
4364 kmp_int64 up; // upper
4365 kmp_int64 st; // stride
4366};
4367KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4368 kmp_int32 num_dims,
4369 const struct kmp_dim *dims);
4370KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4371 const kmp_int64 *vec);
4372KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4373 const kmp_int64 *vec);
4374KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4375
4376KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4377 void *data, size_t size,
4378 void ***cache);
4379
4380// The routines below are not exported.
4381// Consider making them 'static' in corresponding source files.
4382void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4383 void *data_addr, size_t pc_size);
4384struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4385 void *data_addr,
4386 size_t pc_size);
4387void __kmp_threadprivate_resize_cache(int newCapacity);
4388void __kmp_cleanup_threadprivate_caches();
4389
4390// ompc_, kmpc_ entries moved from omp.h.
4391#if KMP_OS_WINDOWS
4392#define KMPC_CONVENTION __cdecl
4393#else
4394#define KMPC_CONVENTION
4395#endif
4396
4397#ifndef __OMP_H
4398typedef enum omp_sched_t {
4399 omp_sched_static = 1,
4400 omp_sched_dynamic = 2,
4401 omp_sched_guided = 3,
4402 omp_sched_auto = 4
4403} omp_sched_t;
4404typedef void *kmp_affinity_mask_t;
4405#endif
4406
4407KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4408KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4409KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4410KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4411KMP_EXPORT int KMPC_CONVENTION
4412kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4413KMP_EXPORT int KMPC_CONVENTION
4414kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4415KMP_EXPORT int KMPC_CONVENTION
4416kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4417
4418KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4419KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4420KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4421KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4422KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4423void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4424size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4425void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4426size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4427 char const *format);
4428
4429enum kmp_target_offload_kind {
4430 tgt_disabled = 0,
4431 tgt_default = 1,
4432 tgt_mandatory = 2
4433};
4434typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4435// Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4436extern kmp_target_offload_kind_t __kmp_target_offload;
4437extern int __kmpc_get_target_offload();
4438
4439// Constants used in libomptarget
4440#define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4441#define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4442
4443// OMP Pause Resource
4444
4445// The following enum is used both to set the status in __kmp_pause_status, and
4446// as the internal equivalent of the externally-visible omp_pause_resource_t.
4447typedef enum kmp_pause_status_t {
4448 kmp_not_paused = 0, // status is not paused, or, requesting resume
4449 kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4450 kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4451} kmp_pause_status_t;
4452
4453// This stores the pause state of the runtime
4454extern kmp_pause_status_t __kmp_pause_status;
4455extern int __kmpc_pause_resource(kmp_pause_status_t level);
4456extern int __kmp_pause_resource(kmp_pause_status_t level);
4457// Soft resume sets __kmp_pause_status, and wakes up all threads.
4458extern void __kmp_resume_if_soft_paused();
4459// Hard resume simply resets the status to not paused. Library will appear to
4460// be uninitialized after hard pause. Let OMP constructs trigger required
4461// initializations.
4462static inline void __kmp_resume_if_hard_paused() {
4463 if (__kmp_pause_status == kmp_hard_paused) {
4464 __kmp_pause_status = kmp_not_paused;
4465 }
4466}
4467
4468extern void __kmp_omp_display_env(int verbose);
4469
4470// 1: it is initializing hidden helper team
4471extern volatile int __kmp_init_hidden_helper;
4472// 1: the hidden helper team is done
4473extern volatile int __kmp_hidden_helper_team_done;
4474// 1: enable hidden helper task
4475extern kmp_int32 __kmp_enable_hidden_helper;
4476// Main thread of hidden helper team
4477extern kmp_info_t *__kmp_hidden_helper_main_thread;
4478// Descriptors for the hidden helper threads
4479extern kmp_info_t **__kmp_hidden_helper_threads;
4480// Number of hidden helper threads
4481extern kmp_int32 __kmp_hidden_helper_threads_num;
4482// Number of hidden helper tasks that have not been executed yet
4483extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4484
4485extern void __kmp_hidden_helper_initialize();
4486extern void __kmp_hidden_helper_threads_initz_routine();
4487extern void __kmp_do_initialize_hidden_helper_threads();
4488extern void __kmp_hidden_helper_threads_initz_wait();
4489extern void __kmp_hidden_helper_initz_release();
4490extern void __kmp_hidden_helper_threads_deinitz_wait();
4491extern void __kmp_hidden_helper_threads_deinitz_release();
4492extern void __kmp_hidden_helper_main_thread_wait();
4493extern void __kmp_hidden_helper_worker_thread_wait();
4494extern void __kmp_hidden_helper_worker_thread_signal();
4495extern void __kmp_hidden_helper_main_thread_release();
4496
4497// Check whether a given thread is a hidden helper thread
4498#define KMP_HIDDEN_HELPER_THREAD(gtid) \
4499 ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4500
4501#define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4502 ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4503
4504#define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid) \
4505 ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4506
4507#define KMP_HIDDEN_HELPER_TEAM(team) \
4508 (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4509
4510// Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4511// main thread, is skipped.
4512#define KMP_GTID_TO_SHADOW_GTID(gtid) \
4513 ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4514
4515// Return the adjusted gtid value by subtracting from gtid the number
4516// of hidden helper threads. This adjusted value is the gtid the thread would
4517// have received if there were no hidden helper threads.
4518static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4519 int adjusted_gtid = gtid;
4520 if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4521 gtid - __kmp_hidden_helper_threads_num >= 0) {
4522 adjusted_gtid -= __kmp_hidden_helper_threads_num;
4523 }
4524 return adjusted_gtid;
4525}
4526
4527// Support for error directive
4528typedef enum kmp_severity_t {
4529 severity_warning = 1,
4530 severity_fatal = 2
4531} kmp_severity_t;
4532extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4533
4534// Support for scope directive
4535KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4536KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4537
4538#ifdef __cplusplus
4539}
4540#endif
4541
4542template <bool C, bool S>
4543extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4544template <bool C, bool S>
4545extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4546template <bool C, bool S>
4547extern void __kmp_atomic_suspend_64(int th_gtid,
4548 kmp_atomic_flag_64<C, S> *flag);
4549extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4550#if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4551template <bool C, bool S>
4552extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4553template <bool C, bool S>
4554extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4555template <bool C, bool S>
4556extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4557extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4558#endif
4559template <bool C, bool S>
4560extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4561template <bool C, bool S>
4562extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4563template <bool C, bool S>
4564extern void __kmp_atomic_resume_64(int target_gtid,
4565 kmp_atomic_flag_64<C, S> *flag);
4566extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4567
4568template <bool C, bool S>
4569int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4570 kmp_flag_32<C, S> *flag, int final_spin,
4571 int *thread_finished,
4572#if USE_ITT_BUILD
4573 void *itt_sync_obj,
4574#endif /* USE_ITT_BUILD */
4575 kmp_int32 is_constrained);
4576template <bool C, bool S>
4577int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4578 kmp_flag_64<C, S> *flag, int final_spin,
4579 int *thread_finished,
4580#if USE_ITT_BUILD
4581 void *itt_sync_obj,
4582#endif /* USE_ITT_BUILD */
4583 kmp_int32 is_constrained);
4584template <bool C, bool S>
4585int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4586 kmp_atomic_flag_64<C, S> *flag,
4587 int final_spin, int *thread_finished,
4588#if USE_ITT_BUILD
4589 void *itt_sync_obj,
4590#endif /* USE_ITT_BUILD */
4591 kmp_int32 is_constrained);
4592int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4593 kmp_flag_oncore *flag, int final_spin,
4594 int *thread_finished,
4595#if USE_ITT_BUILD
4596 void *itt_sync_obj,
4597#endif /* USE_ITT_BUILD */
4598 kmp_int32 is_constrained);
4599
4600extern int __kmp_nesting_mode;
4601extern int __kmp_nesting_mode_nlevels;
4602extern int *__kmp_nesting_nth_level;
4603extern void __kmp_init_nesting_mode();
4604extern void __kmp_set_nesting_mode_threads();
4605
4613 FILE *f;
4614
4615 void close() {
4616 if (f && f != stdout && f != stderr) {
4617 fclose(f);
4618 f = nullptr;
4619 }
4620 }
4621
4622public:
4623 kmp_safe_raii_file_t() : f(nullptr) {}
4624 kmp_safe_raii_file_t(const char *filename, const char *mode,
4625 const char *env_var = nullptr)
4626 : f(nullptr) {
4627 open(filename, mode, env_var);
4628 }
4629 ~kmp_safe_raii_file_t() { close(); }
4630
4634 void open(const char *filename, const char *mode,
4635 const char *env_var = nullptr) {
4636 KMP_ASSERT(!f);
4637 f = fopen(filename, mode);
4638 if (!f) {
4639 int code = errno;
4640 if (env_var) {
4641 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4642 KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4643 } else {
4644 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4645 __kmp_msg_null);
4646 }
4647 }
4648 }
4651 int try_open(const char *filename, const char *mode) {
4652 KMP_ASSERT(!f);
4653 f = fopen(filename, mode);
4654 if (!f)
4655 return errno;
4656 return 0;
4657 }
4660 void set_stdout() {
4661 KMP_ASSERT(!f);
4662 f = stdout;
4663 }
4666 void set_stderr() {
4667 KMP_ASSERT(!f);
4668 f = stderr;
4669 }
4670 operator bool() { return bool(f); }
4671 operator FILE *() { return f; }
4672};
4673
4674template <typename SourceType, typename TargetType,
4675 bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4676 bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4677 bool isSourceSigned = std::is_signed<SourceType>::value,
4678 bool isTargetSigned = std::is_signed<TargetType>::value>
4679struct kmp_convert {};
4680
4681// Both types are signed; Source smaller
4682template <typename SourceType, typename TargetType>
4683struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4684 static TargetType to(SourceType src) { return (TargetType)src; }
4685};
4686// Source equal
4687template <typename SourceType, typename TargetType>
4688struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4689 static TargetType to(SourceType src) { return src; }
4690};
4691// Source bigger
4692template <typename SourceType, typename TargetType>
4693struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4694 static TargetType to(SourceType src) {
4695 KMP_ASSERT(src <= static_cast<SourceType>(
4696 (std::numeric_limits<TargetType>::max)()));
4697 KMP_ASSERT(src >= static_cast<SourceType>(
4698 (std::numeric_limits<TargetType>::min)()));
4699 return (TargetType)src;
4700 }
4701};
4702
4703// Source signed, Target unsigned
4704// Source smaller
4705template <typename SourceType, typename TargetType>
4706struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4707 static TargetType to(SourceType src) {
4708 KMP_ASSERT(src >= 0);
4709 return (TargetType)src;
4710 }
4711};
4712// Source equal
4713template <typename SourceType, typename TargetType>
4714struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4715 static TargetType to(SourceType src) {
4716 KMP_ASSERT(src >= 0);
4717 return (TargetType)src;
4718 }
4719};
4720// Source bigger
4721template <typename SourceType, typename TargetType>
4722struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4723 static TargetType to(SourceType src) {
4724 KMP_ASSERT(src >= 0);
4725 KMP_ASSERT(src <= static_cast<SourceType>(
4726 (std::numeric_limits<TargetType>::max)()));
4727 return (TargetType)src;
4728 }
4729};
4730
4731// Source unsigned, Target signed
4732// Source smaller
4733template <typename SourceType, typename TargetType>
4734struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4735 static TargetType to(SourceType src) { return (TargetType)src; }
4736};
4737// Source equal
4738template <typename SourceType, typename TargetType>
4739struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4740 static TargetType to(SourceType src) {
4741 KMP_ASSERT(src <= static_cast<SourceType>(
4742 (std::numeric_limits<TargetType>::max)()));
4743 return (TargetType)src;
4744 }
4745};
4746// Source bigger
4747template <typename SourceType, typename TargetType>
4748struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4749 static TargetType to(SourceType src) {
4750 KMP_ASSERT(src <= static_cast<SourceType>(
4751 (std::numeric_limits<TargetType>::max)()));
4752 return (TargetType)src;
4753 }
4754};
4755
4756// Source unsigned, Target unsigned
4757// Source smaller
4758template <typename SourceType, typename TargetType>
4759struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4760 static TargetType to(SourceType src) { return (TargetType)src; }
4761};
4762// Source equal
4763template <typename SourceType, typename TargetType>
4764struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4765 static TargetType to(SourceType src) { return src; }
4766};
4767// Source bigger
4768template <typename SourceType, typename TargetType>
4769struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4770 static TargetType to(SourceType src) {
4771 KMP_ASSERT(src <= static_cast<SourceType>(
4772 (std::numeric_limits<TargetType>::max)()));
4773 return (TargetType)src;
4774 }
4775};
4776
4777template <typename T1, typename T2>
4778static inline void __kmp_type_convert(T1 src, T2 *dest) {
4779 *dest = kmp_convert<T1, T2>::to(src);
4780}
4781
4782#endif /* KMP_H */
void set_stdout()
Definition kmp.h:4660
void set_stderr()
Definition kmp.h:4666
int try_open(const char *filename, const char *mode)
Definition kmp.h:4651
void open(const char *filename, const char *mode, const char *env_var=nullptr)
Definition kmp.h:4634
struct ident ident_t
@ KMP_IDENT_KMPC
Definition kmp.h:197
@ KMP_IDENT_IMB
Definition kmp.h:195
@ KMP_IDENT_WORK_LOOP
Definition kmp.h:215
@ KMP_IDENT_BARRIER_IMPL
Definition kmp.h:206
@ KMP_IDENT_WORK_SECTIONS
Definition kmp.h:217
@ KMP_IDENT_AUTOPAR
Definition kmp.h:200
@ KMP_IDENT_ATOMIC_HINT_MASK
Definition kmp.h:224
@ KMP_IDENT_WORK_DISTRIBUTE
Definition kmp.h:219
@ KMP_IDENT_BARRIER_EXPL
Definition kmp.h:204
@ KMP_IDENT_ATOMIC_REDUCE
Definition kmp.h:202
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:1712
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:1736
void *(* kmpc_cctor)(void *, void *)
Definition kmp.h:1743
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:1765
void *(* kmpc_ctor)(void *)
Definition kmp.h:1730
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:1753
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:1759
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:358
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:409
@ kmp_sch_runtime_simd
Definition kmp.h:380
@ kmp_nm_ord_auto
Definition kmp.h:428
@ kmp_sch_auto
Definition kmp.h:365
@ kmp_nm_auto
Definition kmp.h:411
@ kmp_distribute_static_chunked
Definition kmp.h:396
@ kmp_sch_static
Definition kmp.h:361
@ kmp_sch_guided_simd
Definition kmp.h:379
@ kmp_sch_modifier_monotonic
Definition kmp.h:446
@ kmp_sch_default
Definition kmp.h:466
@ kmp_sch_modifier_nonmonotonic
Definition kmp.h:448
@ kmp_nm_ord_static
Definition kmp.h:424
@ kmp_distribute_static
Definition kmp.h:397
@ kmp_sch_guided_chunked
Definition kmp.h:363
@ kmp_nm_static
Definition kmp.h:407
@ kmp_sch_lower
Definition kmp.h:359
@ kmp_nm_upper
Definition kmp.h:430
@ kmp_ord_lower
Definition kmp.h:385
@ kmp_ord_static
Definition kmp.h:387
@ kmp_sch_upper
Definition kmp.h:383
@ kmp_ord_upper
Definition kmp.h:393
@ kmp_nm_lower
Definition kmp.h:403
@ kmp_ord_auto
Definition kmp.h:391
Definition kmp.h:235
kmp_int32 reserved_1
Definition kmp.h:236
char const * psource
Definition kmp.h:245
kmp_int32 reserved_2
Definition kmp.h:239
kmp_int32 reserved_3
Definition kmp.h:244
kmp_int32 flags
Definition kmp.h:237