17#include "kmp_config.h"
27#ifndef KMP_STATIC_STEAL_ENABLED
28#define KMP_STATIC_STEAL_ENABLED 1
31#define TASK_CURRENT_NOT_QUEUED 0
32#define TASK_CURRENT_QUEUED 1
34#ifdef BUILD_TIED_TASK_STACK
35#define TASK_STACK_EMPTY 0
36#define TASK_STACK_BLOCK_BITS 5
38#define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
40#define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
43#define TASK_NOT_PUSHED 1
44#define TASK_SUCCESSFULLY_PUSHED 0
47#define TASK_EXPLICIT 1
48#define TASK_IMPLICIT 0
51#define TASK_DETACHABLE 1
52#define TASK_UNDETACHABLE 0
54#define KMP_CANCEL_THREADS
55#define KMP_THREAD_ATTR
59#if defined(__ANDROID__)
60#undef KMP_CANCEL_THREADS
83#include "kmp_safe_c_api.h"
91#undef KMP_USE_HIER_SCHED
92#define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
95#if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
97#ifndef HWLOC_OBJ_NUMANODE
98#define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
100#ifndef HWLOC_OBJ_PACKAGE
101#define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
105#if KMP_ARCH_X86 || KMP_ARCH_X86_64
106#include <xmmintrin.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))
115#include "kmp_debug.h"
117#include "kmp_version.h"
118#include "kmp_barrier.h"
120#include "kmp_debugger.h"
124#define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
126#include "kmp_wrapper_malloc.h"
129#if !defined NSIG && defined _NSIG
135#pragma weak clock_gettime
139#include "ompt-internal.h"
143#include "ompd-specific.h"
147#define UNLIKELY(x) (x)
156#ifndef USE_FAST_MEMORY
157#define USE_FAST_MEMORY 3
160#ifndef KMP_NESTED_HOT_TEAMS
161#define KMP_NESTED_HOT_TEAMS 0
162#define USE_NESTED_HOT_ARG(x)
164#if KMP_NESTED_HOT_TEAMS
165#define USE_NESTED_HOT_ARG(x) , x
167#define USE_NESTED_HOT_ARG(x)
172#ifndef USE_CMP_XCHG_FOR_BGET
173#define USE_CMP_XCHG_FOR_BGET 1
181#define KMP_NSEC_PER_SEC 1000000000L
182#define KMP_USEC_PER_SEC 1000000L
206 KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
207 KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
208 KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
210 KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
211 KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
224 KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
225 KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
226 KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
227 KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
228 KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
249 kmp_int32 get_openmp_version() {
250 return (((
flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
258typedef union kmp_team kmp_team_t;
259typedef struct kmp_taskdata kmp_taskdata_t;
260typedef union kmp_task_team kmp_task_team_t;
261typedef union kmp_team kmp_team_p;
262typedef union kmp_info kmp_info_p;
263typedef union kmp_root kmp_root_p;
265template <
bool C = false,
bool S = true>
class kmp_flag_32;
266template <
bool C = false,
bool S = true>
class kmp_flag_64;
267template <
bool C = false,
bool S = true>
class kmp_atomic_flag_64;
268class kmp_flag_oncore;
278#define KMP_PACK_64(HIGH_32, LOW_32) \
279 ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
284 while (*(_x) == ' ' || *(_x) == '\t') \
287#define SKIP_DIGITS(_x) \
289 while (*(_x) >= '0' && *(_x) <= '9') \
292#define SKIP_TOKEN(_x) \
294 while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
295 (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
298#define SKIP_TO(_x, _c) \
300 while (*(_x) != '\0' && *(_x) != (_c)) \
306#define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
307#define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
312enum kmp_state_timer {
322#ifdef USE_LOAD_BALANCE
323 dynamic_load_balance,
326 dynamic_thread_limit,
332#ifndef KMP_SCHED_TYPE_DEFINED
333#define KMP_SCHED_TYPE_DEFINED
334typedef enum kmp_sched {
337 kmp_sched_static = 1,
338 kmp_sched_dynamic = 2,
339 kmp_sched_guided = 3,
341 kmp_sched_upper_std = 5,
342 kmp_sched_lower_ext = 100,
343 kmp_sched_trapezoidal = 101,
344#if KMP_STATIC_STEAL_ENABLED
345 kmp_sched_static_steal = 102,
348 kmp_sched_default = kmp_sched_static,
349 kmp_sched_monotonic = 0x80000000
359 kmp_sch_static_chunked = 33,
361 kmp_sch_dynamic_chunked = 35,
363 kmp_sch_runtime = 37,
365 kmp_sch_trapezoidal = 39,
368 kmp_sch_static_greedy = 40,
369 kmp_sch_static_balanced = 41,
371 kmp_sch_guided_iterative_chunked = 42,
372 kmp_sch_guided_analytical_chunked = 43,
374 kmp_sch_static_steal = 44,
377 kmp_sch_static_balanced_chunked = 45,
385 kmp_ord_static_chunked = 65,
387 kmp_ord_dynamic_chunked = 67,
388 kmp_ord_guided_chunked = 68,
389 kmp_ord_runtime = 69,
391 kmp_ord_trapezoidal = 71,
404 kmp_nm_static_chunked =
407 kmp_nm_dynamic_chunked = 163,
409 kmp_nm_runtime = 165,
411 kmp_nm_trapezoidal = 167,
414 kmp_nm_static_greedy = 168,
415 kmp_nm_static_balanced = 169,
417 kmp_nm_guided_iterative_chunked = 170,
418 kmp_nm_guided_analytical_chunked = 171,
419 kmp_nm_static_steal =
422 kmp_nm_ord_static_chunked = 193,
424 kmp_nm_ord_dynamic_chunked = 195,
425 kmp_nm_ord_guided_chunked = 196,
426 kmp_nm_ord_runtime = 197,
428 kmp_nm_ord_trapezoidal = 199,
450#define SCHEDULE_WITHOUT_MODIFIERS(s) \
453#define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
454#define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
455#define SCHEDULE_HAS_NO_MODIFIERS(s) \
456 (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
457#define SCHEDULE_GET_MODIFIERS(s) \
458 ((enum sched_type)( \
459 (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
460#define SCHEDULE_SET_MODIFIERS(s, m) \
461 (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
462#define SCHEDULE_NONMONOTONIC 0
463#define SCHEDULE_MONOTONIC 1
470__kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
472 if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
473 *kind = (kmp_sched_t)((
int)*kind | (int)kmp_sched_monotonic);
479__kmp_sched_apply_mods_intkind(kmp_sched_t kind,
481 if ((
int)kind & (
int)kmp_sched_monotonic) {
482 *internal_kind = (
enum sched_type)((
int)*internal_kind |
488static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
489 return (kmp_sched_t)((int)kind & ~((
int)kmp_sched_monotonic));
493typedef union kmp_r_sched {
512enum clock_function_type {
513 clock_function_gettimeofday,
514 clock_function_clock_gettime
519enum mic_type { non_mic, mic1, mic2, mic3, dummy };
524#undef KMP_FAST_REDUCTION_BARRIER
525#define KMP_FAST_REDUCTION_BARRIER 1
527#undef KMP_FAST_REDUCTION_CORE_DUO
528#if KMP_ARCH_X86 || KMP_ARCH_X86_64
529#define KMP_FAST_REDUCTION_CORE_DUO 1
532enum _reduction_method {
533 reduction_method_not_defined = 0,
534 critical_reduce_block = (1 << 8),
535 atomic_reduce_block = (2 << 8),
536 tree_reduce_block = (3 << 8),
537 empty_reduce_block = (4 << 8)
552#if KMP_FAST_REDUCTION_BARRIER
553#define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
554 ((reduction_method) | (barrier_type))
556#define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
557 ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
559#define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
560 ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
562#define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
565#define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
566 (packed_reduction_method)
568#define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
571#define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
572 ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
573 (which_reduction_block))
575#if KMP_FAST_REDUCTION_BARRIER
576#define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
577 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
579#define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
580 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
583typedef int PACKED_REDUCTION_METHOD_T;
591#pragma warning(disable : 271 310)
621typedef enum kmp_hw_core_type_t {
622 KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
623#if KMP_ARCH_X86 || KMP_ARCH_X86_64
624 KMP_HW_CORE_TYPE_ATOM = 0x20,
625 KMP_HW_CORE_TYPE_CORE = 0x40,
626 KMP_HW_MAX_NUM_CORE_TYPES = 3,
628 KMP_HW_MAX_NUM_CORE_TYPES = 1,
632#define KMP_HW_MAX_NUM_CORE_EFFS 8
634#define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \
635 KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
636#define KMP_ASSERT_VALID_HW_TYPE(type) \
637 KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
639#define KMP_FOREACH_HW_TYPE(type) \
640 for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \
641 type = (kmp_hw_t)((int)type + 1))
643const char *__kmp_hw_get_keyword(kmp_hw_t type,
bool plural =
false);
644const char *__kmp_hw_get_catalog_string(kmp_hw_t type,
bool plural =
false);
645const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
648#if KMP_AFFINITY_SUPPORTED
652#if _MSC_VER < 1600 && KMP_MSVC_COMPAT
653typedef struct GROUP_AFFINITY {
659#if KMP_GROUP_AFFINITY
660extern int __kmp_num_proc_groups;
662static const int __kmp_num_proc_groups = 1;
664typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
665extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
667typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
668extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
670typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
671extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
673typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE,
const GROUP_AFFINITY *,
675extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
679extern hwloc_topology_t __kmp_hwloc_topology;
680extern int __kmp_hwloc_error;
683extern size_t __kmp_affin_mask_size;
684#define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
685#define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
686#define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
687#define KMP_CPU_SET_ITERATE(i, mask) \
688 for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
689#define KMP_CPU_SET(i, mask) (mask)->set(i)
690#define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
691#define KMP_CPU_CLR(i, mask) (mask)->clear(i)
692#define KMP_CPU_ZERO(mask) (mask)->zero()
693#define KMP_CPU_COPY(dest, src) (dest)->copy(src)
694#define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
695#define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
696#define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
697#define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
698#define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
699#define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
700#define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
701#define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
702#define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
703#define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
704#define KMP_CPU_ALLOC_ARRAY(arr, n) \
705 (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
706#define KMP_CPU_FREE_ARRAY(arr, n) \
707 __kmp_affinity_dispatch->deallocate_mask_array(arr)
708#define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
709#define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
710#define __kmp_get_system_affinity(mask, abort_bool) \
711 (mask)->get_system_affinity(abort_bool)
712#define __kmp_set_system_affinity(mask, abort_bool) \
713 (mask)->set_system_affinity(abort_bool)
714#define __kmp_get_proc_group(mask) (mask)->get_proc_group()
720 void *
operator new(
size_t n);
721 void operator delete(
void *p);
722 void *
operator new[](
size_t n);
723 void operator delete[](
void *p);
726 virtual void set(
int i) {}
728 virtual bool is_set(
int i)
const {
return false; }
730 virtual void clear(
int i) {}
732 virtual void zero() {}
734 virtual void copy(
const Mask *src) {}
736 virtual void bitwise_and(
const Mask *rhs) {}
738 virtual void bitwise_or(
const Mask *rhs) {}
740 virtual void bitwise_not() {}
743 virtual int begin()
const {
return 0; }
744 virtual int end()
const {
return 0; }
745 virtual int next(
int previous)
const {
return 0; }
747 virtual int set_process_affinity(
bool abort_on_error)
const {
return -1; }
750 virtual int set_system_affinity(
bool abort_on_error)
const {
return -1; }
752 virtual int get_system_affinity(
bool abort_on_error) {
return -1; }
755 virtual int get_proc_group()
const {
return -1; }
756 int get_max_cpu()
const {
759 KMP_CPU_SET_ITERATE(cpu,
this) {
766 void *
operator new(
size_t n);
767 void operator delete(
void *p);
769 virtual ~KMPAffinity() =
default;
771 virtual void determine_capable(
const char *env_var) {}
773 virtual void bind_thread(
int proc) {}
775 virtual Mask *allocate_mask() {
return nullptr; }
776 virtual void deallocate_mask(Mask *m) {}
777 virtual Mask *allocate_mask_array(
int num) {
return nullptr; }
778 virtual void deallocate_mask_array(Mask *m) {}
779 virtual Mask *index_mask_array(Mask *m,
int index) {
return nullptr; }
780 static void pick_api();
781 static void destroy_api();
789 virtual api_type get_api_type()
const {
795 static bool picked_api;
798typedef KMPAffinity::Mask kmp_affin_mask_t;
799extern KMPAffinity *__kmp_affinity_dispatch;
803#define KMP_AFFIN_MASK_PRINT_LEN 1024
817enum affinity_top_method {
818 affinity_top_method_all = 0,
819#if KMP_ARCH_X86 || KMP_ARCH_X86_64
820 affinity_top_method_apicid,
821 affinity_top_method_x2apicid,
822 affinity_top_method_x2apicid_1f,
824 affinity_top_method_cpuinfo,
825#if KMP_GROUP_AFFINITY
826 affinity_top_method_group,
828 affinity_top_method_flat,
830 affinity_top_method_hwloc,
832 affinity_top_method_default
835#define affinity_respect_mask_default (2)
837typedef struct kmp_affinity_flags_t {
839 unsigned verbose : 1;
840 unsigned warnings : 1;
841 unsigned respect : 2;
843 unsigned initialized : 1;
844 unsigned reserved : 25;
845} kmp_affinity_flags_t;
846KMP_BUILD_ASSERT(
sizeof(kmp_affinity_flags_t) == 4);
848typedef struct kmp_affinity_ids_t {
849 int ids[KMP_HW_LAST];
850 int operator[](
size_t idx)
const {
return ids[idx]; }
851 int &operator[](
size_t idx) {
return ids[idx]; }
852 kmp_affinity_ids_t &operator=(
const kmp_affinity_ids_t &rhs) {
853 for (
int i = 0; i < KMP_HW_LAST; ++i)
859typedef struct kmp_affinity_attrs_t {
863 unsigned reserved : 15;
864} kmp_affinity_attrs_t;
865#define KMP_AFFINITY_ATTRS_UNKNOWN \
866 { KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 }
868typedef struct kmp_affinity_t {
870 enum affinity_type type;
875 kmp_affinity_flags_t flags;
877 kmp_affin_mask_t *masks;
878 kmp_affinity_ids_t *ids;
879 kmp_affinity_attrs_t *attrs;
880 unsigned num_os_id_masks;
881 kmp_affin_mask_t *os_id_masks;
885#define KMP_AFFINITY_INIT(env) \
887 nullptr, affinity_default, KMP_HW_UNKNOWN, -1, 0, 0, \
888 {TRUE, FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE}, 0, \
889 nullptr, nullptr, nullptr, 0, nullptr, env \
892extern enum affinity_top_method __kmp_affinity_top_method;
893extern kmp_affinity_t __kmp_affinity;
894extern kmp_affinity_t __kmp_hh_affinity;
895extern kmp_affinity_t *__kmp_affinities[2];
897extern void __kmp_affinity_bind_thread(
int which);
899extern kmp_affin_mask_t *__kmp_affin_fullMask;
900extern kmp_affin_mask_t *__kmp_affin_origMask;
901extern char *__kmp_cpuinfo_file;
906typedef enum kmp_proc_bind_t {
916typedef struct kmp_nested_proc_bind_t {
917 kmp_proc_bind_t *bind_types;
920} kmp_nested_proc_bind_t;
922extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
923extern kmp_proc_bind_t __kmp_teams_proc_bind;
925extern int __kmp_display_affinity;
926extern char *__kmp_affinity_format;
927static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
929extern int __kmp_tool;
930extern char *__kmp_tool_libraries;
933#if KMP_AFFINITY_SUPPORTED
934#define KMP_PLACE_ALL (-1)
935#define KMP_PLACE_UNDEFINED (-2)
937#define KMP_AFFINITY_NON_PROC_BIND \
938 ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
939 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
940 (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced))
943extern int __kmp_affinity_num_places;
945typedef enum kmp_cancel_kind_t {
954typedef struct kmp_hws_item {
959extern kmp_hws_item_t __kmp_hws_socket;
960extern kmp_hws_item_t __kmp_hws_die;
961extern kmp_hws_item_t __kmp_hws_node;
962extern kmp_hws_item_t __kmp_hws_tile;
963extern kmp_hws_item_t __kmp_hws_core;
964extern kmp_hws_item_t __kmp_hws_proc;
965extern int __kmp_hws_requested;
966extern int __kmp_hws_abs_flag;
970#define KMP_PAD(type, sz) \
971 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
975#define KMP_GTID_DNE (-2)
976#define KMP_GTID_SHUTDOWN (-3)
977#define KMP_GTID_MONITOR (-4)
978#define KMP_GTID_UNKNOWN (-5)
979#define KMP_GTID_MIN (-6)
985typedef uintptr_t omp_uintptr_t;
988 omp_atk_sync_hint = 1,
989 omp_atk_alignment = 2,
991 omp_atk_pool_size = 4,
992 omp_atk_fallback = 5,
995 omp_atk_partition = 8
996} omp_alloctrait_key_t;
1001 omp_atv_contended = 3,
1002 omp_atv_uncontended = 4,
1003 omp_atv_serialized = 5,
1004 omp_atv_sequential = omp_atv_serialized,
1005 omp_atv_private = 6,
1009 omp_atv_cgroup = 10,
1010 omp_atv_default_mem_fb = 11,
1011 omp_atv_null_fb = 12,
1012 omp_atv_abort_fb = 13,
1013 omp_atv_allocator_fb = 14,
1014 omp_atv_environment = 15,
1015 omp_atv_nearest = 16,
1016 omp_atv_blocked = 17,
1017 omp_atv_interleaved = 18
1018} omp_alloctrait_value_t;
1019#define omp_atv_default ((omp_uintptr_t)-1)
1021typedef void *omp_memspace_handle_t;
1022extern omp_memspace_handle_t
const omp_default_mem_space;
1023extern omp_memspace_handle_t
const omp_large_cap_mem_space;
1024extern omp_memspace_handle_t
const omp_const_mem_space;
1025extern omp_memspace_handle_t
const omp_high_bw_mem_space;
1026extern omp_memspace_handle_t
const omp_low_lat_mem_space;
1027extern omp_memspace_handle_t
const llvm_omp_target_host_mem_space;
1028extern omp_memspace_handle_t
const llvm_omp_target_shared_mem_space;
1029extern omp_memspace_handle_t
const llvm_omp_target_device_mem_space;
1032 omp_alloctrait_key_t key;
1033 omp_uintptr_t value;
1036typedef void *omp_allocator_handle_t;
1037extern omp_allocator_handle_t
const omp_null_allocator;
1038extern omp_allocator_handle_t
const omp_default_mem_alloc;
1039extern omp_allocator_handle_t
const omp_large_cap_mem_alloc;
1040extern omp_allocator_handle_t
const omp_const_mem_alloc;
1041extern omp_allocator_handle_t
const omp_high_bw_mem_alloc;
1042extern omp_allocator_handle_t
const omp_low_lat_mem_alloc;
1043extern omp_allocator_handle_t
const omp_cgroup_mem_alloc;
1044extern omp_allocator_handle_t
const omp_pteam_mem_alloc;
1045extern omp_allocator_handle_t
const omp_thread_mem_alloc;
1046extern omp_allocator_handle_t
const llvm_omp_target_host_mem_alloc;
1047extern omp_allocator_handle_t
const llvm_omp_target_shared_mem_alloc;
1048extern omp_allocator_handle_t
const llvm_omp_target_device_mem_alloc;
1049extern omp_allocator_handle_t
const kmp_max_mem_alloc;
1050extern omp_allocator_handle_t __kmp_def_allocator;
1055extern int __kmp_memkind_available;
1057typedef omp_memspace_handle_t kmp_memspace_t;
1059typedef struct kmp_allocator_t {
1060 omp_memspace_handle_t memspace;
1063 omp_alloctrait_value_t fb;
1064 kmp_allocator_t *fb_data;
1065 kmp_uint64 pool_size;
1066 kmp_uint64 pool_used;
1070extern omp_allocator_handle_t __kmpc_init_allocator(
int gtid,
1071 omp_memspace_handle_t,
1073 omp_alloctrait_t traits[]);
1074extern void __kmpc_destroy_allocator(
int gtid, omp_allocator_handle_t al);
1075extern void __kmpc_set_default_allocator(
int gtid, omp_allocator_handle_t al);
1076extern omp_allocator_handle_t __kmpc_get_default_allocator(
int gtid);
1078extern void *__kmpc_alloc(
int gtid,
size_t sz, omp_allocator_handle_t al);
1079extern void *__kmpc_aligned_alloc(
int gtid,
size_t align,
size_t sz,
1080 omp_allocator_handle_t al);
1081extern void *__kmpc_calloc(
int gtid,
size_t nmemb,
size_t sz,
1082 omp_allocator_handle_t al);
1083extern void *__kmpc_realloc(
int gtid,
void *ptr,
size_t sz,
1084 omp_allocator_handle_t al,
1085 omp_allocator_handle_t free_al);
1086extern void __kmpc_free(
int gtid,
void *ptr, omp_allocator_handle_t al);
1088extern void *__kmp_alloc(
int gtid,
size_t align,
size_t sz,
1089 omp_allocator_handle_t al);
1090extern void *__kmp_calloc(
int gtid,
size_t align,
size_t nmemb,
size_t sz,
1091 omp_allocator_handle_t al);
1092extern void *__kmp_realloc(
int gtid,
void *ptr,
size_t sz,
1093 omp_allocator_handle_t al,
1094 omp_allocator_handle_t free_al);
1095extern void ___kmpc_free(
int gtid,
void *ptr, omp_allocator_handle_t al);
1097extern void __kmp_init_memkind();
1098extern void __kmp_fini_memkind();
1099extern void __kmp_init_target_mem();
1103#define KMP_UINT64_MAX \
1104 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1106#define KMP_MIN_NTH 1
1109#if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1110#define KMP_MAX_NTH PTHREAD_THREADS_MAX
1112#define KMP_MAX_NTH INT_MAX
1116#ifdef PTHREAD_STACK_MIN
1117#define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
1119#define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1122#define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1125#define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1126#elif KMP_ARCH_X86_64
1127#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1128#define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1130#define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1133#define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1134#define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1135#define KMP_MAX_MALLOC_POOL_INCR \
1136 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1138#define KMP_MIN_STKOFFSET (0)
1139#define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1141#define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1143#define KMP_DEFAULT_STKOFFSET CACHE_LINE
1146#define KMP_MIN_STKPADDING (0)
1147#define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1149#define KMP_BLOCKTIME_MULTIPLIER \
1151#define KMP_MIN_BLOCKTIME (0)
1152#define KMP_MAX_BLOCKTIME \
1156#define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200))
1159#define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1160#define KMP_MIN_MONITOR_WAKEUPS (1)
1161#define KMP_MAX_MONITOR_WAKEUPS (1000)
1165#define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1166 (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \
1167 : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \
1168 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1169 ? (monitor_wakeups) \
1170 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1174#define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1175 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1176 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1178#define KMP_BLOCKTIME(team, tid) \
1179 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1180#if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1182extern kmp_uint64 __kmp_ticks_per_msec;
1183#if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1184#define KMP_NOW() ((kmp_uint64)_rdtsc())
1186#define KMP_NOW() __kmp_hardware_timestamp()
1188#define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1189#define KMP_BLOCKTIME_INTERVAL(team, tid) \
1190 (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1191#define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1194extern kmp_uint64 __kmp_now_nsec();
1195#define KMP_NOW() __kmp_now_nsec()
1196#define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1197#define KMP_BLOCKTIME_INTERVAL(team, tid) \
1198 (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1199#define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1203#define KMP_MIN_STATSCOLS 40
1204#define KMP_MAX_STATSCOLS 4096
1205#define KMP_DEFAULT_STATSCOLS 80
1207#define KMP_MIN_INTERVAL 0
1208#define KMP_MAX_INTERVAL (INT_MAX - 1)
1209#define KMP_DEFAULT_INTERVAL 0
1211#define KMP_MIN_CHUNK 1
1212#define KMP_MAX_CHUNK (INT_MAX - 1)
1213#define KMP_DEFAULT_CHUNK 1
1215#define KMP_MIN_DISP_NUM_BUFF 1
1216#define KMP_DFLT_DISP_NUM_BUFF 7
1217#define KMP_MAX_DISP_NUM_BUFF 4096
1219#define KMP_MAX_ORDERED 8
1221#define KMP_MAX_FIELDS 32
1223#define KMP_MAX_BRANCH_BITS 31
1225#define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1227#define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1229#define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1234#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1235#define KMP_TLS_GTID_MIN 5
1237#define KMP_TLS_GTID_MIN INT_MAX
1240#define KMP_MASTER_TID(tid) (0 == (tid))
1241#define KMP_WORKER_TID(tid) (0 != (tid))
1243#define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1244#define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1245#define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1249#define TRUE (!FALSE)
1255#define KMP_INIT_WAIT 64U
1256#define KMP_NEXT_WAIT 32U
1258#define KMP_INIT_WAIT 1024U
1259#define KMP_NEXT_WAIT 512U
1262#define KMP_INIT_WAIT 1024U
1263#define KMP_NEXT_WAIT 512U
1264#elif KMP_OS_DRAGONFLY
1266#define KMP_INIT_WAIT 1024U
1267#define KMP_NEXT_WAIT 512U
1270#define KMP_INIT_WAIT 1024U
1271#define KMP_NEXT_WAIT 512U
1274#define KMP_INIT_WAIT 1024U
1275#define KMP_NEXT_WAIT 512U
1278#define KMP_INIT_WAIT 1024U
1279#define KMP_NEXT_WAIT 512U
1282#define KMP_INIT_WAIT 1024U
1283#define KMP_NEXT_WAIT 512U
1286#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1287typedef struct kmp_cpuid {
1294typedef struct kmp_cpuinfo_flags_t {
1297 unsigned hybrid : 1;
1298 unsigned reserved : 29;
1299} kmp_cpuinfo_flags_t;
1301typedef struct kmp_cpuinfo {
1308 kmp_cpuinfo_flags_t flags;
1312 kmp_uint64 frequency;
1313 char name[3 *
sizeof(kmp_cpuid_t)];
1316extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1321static inline void __kmp_x86_cpuid(
int leaf,
int subleaf,
struct kmp_cpuid *p) {
1322 __asm__ __volatile__(
"cpuid"
1323 :
"=a"(p->eax),
"=b"(p->ebx),
"=c"(p->ecx),
"=d"(p->edx)
1324 :
"a"(leaf),
"c"(subleaf));
1327static inline void __kmp_load_x87_fpu_control_word(
const kmp_int16 *p) {
1328 __asm__ __volatile__(
"fldcw %0" : :
"m"(*p));
1331static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1332 __asm__ __volatile__(
"fstcw %0" :
"=m"(*p));
1334static inline void __kmp_clear_x87_fpu_status_word() {
1337 struct x87_fpu_state {
1346 struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1347 __asm__ __volatile__(
"fstenv %0\n\t"
1348 "andw $0x7f00, %1\n\t"
1350 :
"+m"(fpu_state),
"+m"(fpu_state.sw));
1352 __asm__ __volatile__(
"fnclex");
1356static inline void __kmp_load_mxcsr(
const kmp_uint32 *p) { _mm_setcsr(*p); }
1357static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1359static inline void __kmp_load_mxcsr(
const kmp_uint32 *p) {}
1360static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1364extern void __kmp_x86_cpuid(
int mode,
int mode2,
struct kmp_cpuid *p);
1365extern void __kmp_load_x87_fpu_control_word(
const kmp_int16 *p);
1366extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1367extern void __kmp_clear_x87_fpu_status_word();
1368static inline void __kmp_load_mxcsr(
const kmp_uint32 *p) { _mm_setcsr(*p); }
1369static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1372#define KMP_X86_MXCSR_MASK 0xffffffc0
1377#if KMP_HAVE_WAITPKG_INTRINSICS
1378#if KMP_HAVE_IMMINTRIN_H
1379#include <immintrin.h>
1380#elif KMP_HAVE_INTRIN_H
1385KMP_ATTRIBUTE_TARGET_WAITPKG
1386static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1387#if !KMP_HAVE_WAITPKG_INTRINSICS
1388 uint32_t timeHi = uint32_t(counter >> 32);
1389 uint32_t timeLo = uint32_t(counter & 0xffffffff);
1391 __asm__
volatile(
"#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1397 :
"a"(timeLo),
"d"(timeHi),
"c"(hint)
1401 return _tpause(hint, counter);
1404KMP_ATTRIBUTE_TARGET_WAITPKG
1405static inline void __kmp_umonitor(
void *cacheline) {
1406#if !KMP_HAVE_WAITPKG_INTRINSICS
1407 __asm__
volatile(
"# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1412 _umonitor(cacheline);
1415KMP_ATTRIBUTE_TARGET_WAITPKG
1416static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1417#if !KMP_HAVE_WAITPKG_INTRINSICS
1418 uint32_t timeHi = uint32_t(counter >> 32);
1419 uint32_t timeLo = uint32_t(counter & 0xffffffff);
1421 __asm__
volatile(
"#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1427 :
"a"(timeLo),
"d"(timeHi),
"c"(hint)
1431 return _umwait(hint, counter);
1436#include <pmmintrin.h>
1441__attribute__((target(
"sse3")))
1444__kmp_mm_monitor(
void *cacheline,
unsigned extensions,
unsigned hints) {
1445 _mm_monitor(cacheline, extensions, hints);
1448__attribute__((target(
"sse3")))
1451__kmp_mm_mwait(
unsigned extensions,
unsigned hints) {
1452 _mm_mwait(extensions, hints);
1457extern void __kmp_x86_pause(
void);
1463static inline void __kmp_x86_pause(
void) { _mm_delay_32(300); }
1465static inline void __kmp_x86_pause(
void) { _mm_pause(); }
1467#define KMP_CPU_PAUSE() __kmp_x86_pause()
1469#define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1470#define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1471#define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1472#define KMP_CPU_PAUSE() \
1474 KMP_PPC64_PRI_LOW(); \
1475 KMP_PPC64_PRI_MED(); \
1476 KMP_PPC64_PRI_LOC_MB(); \
1479#define KMP_CPU_PAUSE()
1482#define KMP_INIT_YIELD(count) \
1483 { (count) = __kmp_yield_init; }
1485#define KMP_INIT_BACKOFF(time) \
1486 { (time) = __kmp_pause_init; }
1488#define KMP_OVERSUBSCRIBED \
1489 (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1491#define KMP_TRY_YIELD \
1492 ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1494#define KMP_TRY_YIELD_OVERSUB \
1495 ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1497#define KMP_YIELD(cond) \
1500 if ((cond) && (KMP_TRY_YIELD)) \
1504#define KMP_YIELD_OVERSUB() \
1507 if ((KMP_TRY_YIELD_OVERSUB)) \
1513#define KMP_YIELD_SPIN(count) \
1516 if (KMP_TRY_YIELD) { \
1520 (count) = __kmp_yield_next; \
1531#define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF)
1532#define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1534 if (__kmp_tpause_enabled) { \
1535 if (KMP_OVERSUBSCRIBED) { \
1536 __kmp_tpause(0, (time)); \
1538 __kmp_tpause(__kmp_tpause_hint, (time)); \
1540 (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK; \
1543 if ((KMP_TRY_YIELD_OVERSUB)) { \
1545 } else if (__kmp_use_yield == 1) { \
1549 (count) = __kmp_yield_next; \
1555#define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1558 if ((KMP_TRY_YIELD_OVERSUB)) \
1560 else if (__kmp_use_yield == 1) { \
1564 (count) = __kmp_yield_next; \
1584 ct_ordered_in_parallel,
1592#define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1596 enum cons_type type;
1603 int p_top, w_top, s_top;
1604 int stack_size, stack_top;
1605 struct cons_data *stack_data;
1608struct kmp_region_info {
1610 int offset[KMP_MAX_FIELDS];
1611 int length[KMP_MAX_FIELDS];
1618typedef HANDLE kmp_thread_t;
1619typedef DWORD kmp_key_t;
1623typedef pthread_t kmp_thread_t;
1624typedef pthread_key_t kmp_key_t;
1627extern kmp_key_t __kmp_gtid_threadprivate_key;
1629typedef struct kmp_sys_info {
1643typedef int kmp_itt_mark_t;
1644#define KMP_ITT_DEBUG 0
1647typedef kmp_int32 kmp_critical_name[8];
1658typedef void (*
kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1659typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1676typedef void *(*kmpc_ctor)(
void *);
1689typedef void *(*kmpc_cctor)(
void *,
void *);
1699typedef void *(*kmpc_ctor_vec)(
void *, size_t);
1711typedef void *(*kmpc_cctor_vec)(
void *,
void *,
1719typedef struct kmp_cached_addr {
1721 void ***compiler_cache;
1723 struct kmp_cached_addr *next;
1726struct private_data {
1727 struct private_data *next;
1733struct private_common {
1734 struct private_common *next;
1735 struct private_common *link;
1741struct shared_common {
1742 struct shared_common *next;
1743 struct private_data *pod_init;
1763#define KMP_HASH_TABLE_LOG2 9
1764#define KMP_HASH_TABLE_SIZE \
1765 (1 << KMP_HASH_TABLE_LOG2)
1766#define KMP_HASH_SHIFT 3
1767#define KMP_HASH(x) \
1768 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1770struct common_table {
1771 struct private_common *data[KMP_HASH_TABLE_SIZE];
1774struct shared_table {
1775 struct shared_common *data[KMP_HASH_TABLE_SIZE];
1780#if KMP_USE_HIER_SCHED
1783typedef struct kmp_hier_private_bdata_t {
1784 kmp_int32 num_active;
1786 kmp_uint64 wait_val[2];
1787} kmp_hier_private_bdata_t;
1790typedef struct kmp_sched_flags {
1791 unsigned ordered : 1;
1792 unsigned nomerge : 1;
1793 unsigned contains_last : 1;
1794#if KMP_USE_HIER_SCHED
1795 unsigned use_hier : 1;
1796 unsigned unused : 28;
1798 unsigned unused : 29;
1802KMP_BUILD_ASSERT(
sizeof(kmp_sched_flags_t) == 4);
1804#if KMP_STATIC_STEAL_ENABLED
1805typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1812 kmp_lock_t *steal_lock;
1819 struct KMP_ALIGN(32) {
1826 kmp_uint32 ordered_lower;
1827 kmp_uint32 ordered_upper;
1829 kmp_int32 last_upper;
1831} dispatch_private_info32_t;
1833typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1840 kmp_lock_t *steal_lock;
1849 struct KMP_ALIGN(32) {
1856 kmp_uint64 ordered_lower;
1857 kmp_uint64 ordered_upper;
1859 kmp_int64 last_upper;
1861} dispatch_private_info64_t;
1863typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1876 kmp_uint32 ordered_lower;
1877 kmp_uint32 ordered_upper;
1879 kmp_int32 last_upper;
1881} dispatch_private_info32_t;
1883typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1897 kmp_uint64 ordered_lower;
1898 kmp_uint64 ordered_upper;
1900 kmp_int64 last_upper;
1902} dispatch_private_info64_t;
1905typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1906 union private_info {
1907 dispatch_private_info32_t p32;
1908 dispatch_private_info64_t p64;
1911 kmp_sched_flags_t flags;
1912 std::atomic<kmp_uint32> steal_flag;
1913 kmp_int32 ordered_bumped;
1915 struct dispatch_private_info *next;
1916 kmp_int32 type_size;
1917#if KMP_USE_HIER_SCHED
1921 enum cons_type pushed_ws;
1922} dispatch_private_info_t;
1924typedef struct dispatch_shared_info32 {
1927 volatile kmp_uint32 iteration;
1928 volatile kmp_int32 num_done;
1929 volatile kmp_uint32 ordered_iteration;
1931 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1932} dispatch_shared_info32_t;
1934typedef struct dispatch_shared_info64 {
1937 volatile kmp_uint64 iteration;
1938 volatile kmp_int64 num_done;
1939 volatile kmp_uint64 ordered_iteration;
1941 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1942} dispatch_shared_info64_t;
1944typedef struct dispatch_shared_info {
1946 dispatch_shared_info32_t s32;
1947 dispatch_shared_info64_t s64;
1949 volatile kmp_uint32 buffer_index;
1950 volatile kmp_int32 doacross_buf_idx;
1951 volatile kmp_uint32 *doacross_flags;
1952 kmp_int32 doacross_num_done;
1953#if KMP_USE_HIER_SCHED
1962} dispatch_shared_info_t;
1964typedef struct kmp_disp {
1966 void (*th_deo_fcn)(
int *gtid,
int *cid,
ident_t *);
1968 void (*th_dxo_fcn)(
int *gtid,
int *cid,
ident_t *);
1970 dispatch_shared_info_t *th_dispatch_sh_current;
1971 dispatch_private_info_t *th_dispatch_pr_current;
1973 dispatch_private_info_t *th_disp_buffer;
1974 kmp_uint32 th_disp_index;
1975 kmp_int32 th_doacross_buf_idx;
1976 volatile kmp_uint32 *th_doacross_flags;
1977 kmp_int64 *th_doacross_info;
1978#if KMP_USE_INTERNODE_ALIGNMENT
1979 char more_padding[INTERNODE_CACHE_LINE];
1987#define KMP_INIT_BARRIER_STATE 0
1988#define KMP_BARRIER_SLEEP_BIT 0
1989#define KMP_BARRIER_UNUSED_BIT 1
1990#define KMP_BARRIER_BUMP_BIT 2
1992#define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1993#define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1994#define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1996#if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1997#error "Barrier sleep bit must be smaller than barrier bump bit"
1999#if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
2000#error "Barrier unused bit must be smaller than barrier bump bit"
2004#define KMP_BARRIER_NOT_WAITING 0
2005#define KMP_BARRIER_OWN_FLAG \
2007#define KMP_BARRIER_PARENT_FLAG \
2009#define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
2011#define KMP_BARRIER_SWITCHING \
2014#define KMP_NOT_SAFE_TO_REAP \
2016#define KMP_SAFE_TO_REAP 1
2028 bs_plain_barrier = 0,
2030 bs_forkjoin_barrier,
2031#if KMP_FAST_REDUCTION_BARRIER
2032 bs_reduction_barrier,
2038#if !KMP_FAST_REDUCTION_BARRIER
2039#define bs_reduction_barrier bs_plain_barrier
2042typedef enum kmp_bar_pat {
2049 bp_hierarchical_bar = 3,
2054#define KMP_BARRIER_ICV_PUSH 1
2057typedef struct kmp_internal_control {
2058 int serial_nesting_level;
2071 int max_active_levels;
2074 kmp_proc_bind_t proc_bind;
2075 kmp_int32 default_device;
2076 struct kmp_internal_control *next;
2077} kmp_internal_control_t;
2079static inline void copy_icvs(kmp_internal_control_t *dst,
2080 kmp_internal_control_t *src) {
2085typedef struct KMP_ALIGN_CACHE kmp_bstate {
2090 kmp_internal_control_t th_fixed_icvs;
2093 volatile kmp_uint64 b_go;
2094 KMP_ALIGN_CACHE
volatile kmp_uint64
2096 kmp_uint32 *skip_per_level;
2097 kmp_uint32 my_level;
2098 kmp_int32 parent_tid;
2101 struct kmp_bstate *parent_bar;
2103 kmp_uint64 leaf_state;
2105 kmp_uint8 base_leaf_kids;
2106 kmp_uint8 leaf_kids;
2108 kmp_uint8 wait_flag;
2109 kmp_uint8 use_oncore_barrier;
2114 KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2118union KMP_ALIGN_CACHE kmp_barrier_union {
2120 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2124typedef union kmp_barrier_union kmp_balign_t;
2127union KMP_ALIGN_CACHE kmp_barrier_team_union {
2129 char b_pad[CACHE_LINE];
2131 kmp_uint64 b_arrived;
2137 kmp_uint b_master_arrived;
2138 kmp_uint b_team_arrived;
2143typedef union kmp_barrier_team_union kmp_balign_team_t;
2150typedef struct kmp_win32_mutex {
2152 CRITICAL_SECTION cs;
2155typedef struct kmp_win32_cond {
2160 kmp_win32_mutex_t waiters_count_lock_;
2167 int wait_generation_count_;
2176union KMP_ALIGN_CACHE kmp_cond_union {
2178 char c_pad[CACHE_LINE];
2179 pthread_cond_t c_cond;
2182typedef union kmp_cond_union kmp_cond_align_t;
2184union KMP_ALIGN_CACHE kmp_mutex_union {
2186 char m_pad[CACHE_LINE];
2187 pthread_mutex_t m_mutex;
2190typedef union kmp_mutex_union kmp_mutex_align_t;
2194typedef struct kmp_desc_base {
2196 size_t ds_stacksize;
2198 kmp_thread_t ds_thread;
2199 volatile int ds_tid;
2202 volatile int ds_alive;
2219typedef union KMP_ALIGN_CACHE kmp_desc {
2221 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2225typedef struct kmp_local {
2226 volatile int this_construct;
2231#if !USE_CMP_XCHG_FOR_BGET
2232#ifdef USE_QUEUING_LOCK_FOR_BGET
2233 kmp_lock_t bget_lock;
2235 kmp_bootstrap_lock_t bget_lock;
2242 PACKED_REDUCTION_METHOD_T
2243 packed_reduction_method;
2248#define KMP_CHECK_UPDATE(a, b) \
2251#define KMP_CHECK_UPDATE_SYNC(a, b) \
2253 TCW_SYNC_PTR((a), (b))
2255#define get__blocktime(xteam, xtid) \
2256 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2257#define get__bt_set(xteam, xtid) \
2258 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2260#define get__bt_intervals(xteam, xtid) \
2261 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2264#define get__dynamic_2(xteam, xtid) \
2265 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2266#define get__nproc_2(xteam, xtid) \
2267 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2268#define get__sched_2(xteam, xtid) \
2269 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2271#define set__blocktime_team(xteam, xtid, xval) \
2272 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2276#define set__bt_intervals_team(xteam, xtid, xval) \
2277 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2281#define set__bt_set_team(xteam, xtid, xval) \
2282 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2284#define set__dynamic(xthread, xval) \
2285 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2286#define get__dynamic(xthread) \
2287 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2289#define set__nproc(xthread, xval) \
2290 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2292#define set__thread_limit(xthread, xval) \
2293 (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2295#define set__max_active_levels(xthread, xval) \
2296 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2298#define get__max_active_levels(xthread) \
2299 ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2301#define set__sched(xthread, xval) \
2302 (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2304#define set__proc_bind(xthread, xval) \
2305 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2306#define get__proc_bind(xthread) \
2307 ((xthread)->th.th_current_task->td_icvs.proc_bind)
2311typedef enum kmp_tasking_mode {
2312 tskm_immediate_exec = 0,
2313 tskm_extra_barrier = 1,
2314 tskm_task_teams = 2,
2316} kmp_tasking_mode_t;
2318extern kmp_tasking_mode_t
2320extern int __kmp_task_stealing_constraint;
2321extern int __kmp_enable_task_throttling;
2322extern kmp_int32 __kmp_default_device;
2325extern kmp_int32 __kmp_max_task_priority;
2327extern kmp_uint64 __kmp_taskloop_min_tasks;
2331#define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2332#define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2336#define KMP_TASKING_ENABLED(task_team) \
2337 (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2345typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32,
void *);
2347typedef union kmp_cmplrdata {
2358typedef struct kmp_task {
2365 kmp_cmplrdata_t data2;
2374typedef struct kmp_taskgroup {
2375 std::atomic<kmp_int32> count;
2376 std::atomic<kmp_int32>
2378 struct kmp_taskgroup *parent;
2381 kmp_int32 reduce_num_data;
2382 uintptr_t *gomp_data;
2386typedef union kmp_depnode kmp_depnode_t;
2387typedef struct kmp_depnode_list kmp_depnode_list_t;
2388typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2391#define KMP_DEP_IN 0x1
2392#define KMP_DEP_OUT 0x2
2393#define KMP_DEP_INOUT 0x3
2394#define KMP_DEP_MTX 0x4
2395#define KMP_DEP_SET 0x8
2396#define KMP_DEP_ALL 0x80
2398typedef struct kmp_depend_info {
2399 kmp_intptr_t base_addr;
2408 unsigned unused : 3;
2415struct kmp_depnode_list {
2416 kmp_depnode_t *node;
2417 kmp_depnode_list_t *next;
2421#define MAX_MTX_DEPS 4
2423typedef struct kmp_base_depnode {
2424 kmp_depnode_list_t *successors;
2426 kmp_lock_t *mtx_locks[MAX_MTX_DEPS];
2427 kmp_int32 mtx_num_locks;
2429#if KMP_SUPPORT_GRAPH_OUTPUT
2432 std::atomic<kmp_int32> npredecessors;
2433 std::atomic<kmp_int32> nrefs;
2434} kmp_base_depnode_t;
2436union KMP_ALIGN_CACHE kmp_depnode {
2438 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2439 kmp_base_depnode_t dn;
2442struct kmp_dephash_entry {
2444 kmp_depnode_t *last_out;
2445 kmp_depnode_list_t *last_set;
2446 kmp_depnode_list_t *prev_set;
2447 kmp_uint8 last_flag;
2448 kmp_lock_t *mtx_lock;
2449 kmp_dephash_entry_t *next_in_bucket;
2452typedef struct kmp_dephash {
2453 kmp_dephash_entry_t **buckets;
2455 kmp_depnode_t *last_all;
2457 kmp_uint32 nelements;
2458 kmp_uint32 nconflicts;
2461typedef struct kmp_task_affinity_info {
2462 kmp_intptr_t base_addr;
2467 kmp_int32 reserved : 30;
2469} kmp_task_affinity_info_t;
2471typedef enum kmp_event_type_t {
2472 KMP_EVENT_UNINITIALIZED = 0,
2473 KMP_EVENT_ALLOW_COMPLETION = 1
2477 kmp_event_type_t type;
2478 kmp_tas_lock_t lock;
2484#ifdef BUILD_TIED_TASK_STACK
2487typedef struct kmp_stack_block {
2488 kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2489 struct kmp_stack_block *sb_next;
2490 struct kmp_stack_block *sb_prev;
2493typedef struct kmp_task_stack {
2494 kmp_stack_block_t ts_first_block;
2495 kmp_taskdata_t **ts_top;
2496 kmp_int32 ts_entries;
2501typedef struct kmp_tasking_flags {
2503 unsigned tiedness : 1;
2505 unsigned merged_if0 : 1;
2507 unsigned destructors_thunk : 1;
2511 unsigned priority_specified : 1;
2513 unsigned detachable : 1;
2514 unsigned hidden_helper : 1;
2515 unsigned reserved : 8;
2518 unsigned tasktype : 1;
2519 unsigned task_serial : 1;
2520 unsigned tasking_ser : 1;
2522 unsigned team_serial : 1;
2526 unsigned started : 1;
2527 unsigned executing : 1;
2528 unsigned complete : 1;
2530 unsigned native : 1;
2531 unsigned reserved31 : 7;
2533} kmp_tasking_flags_t;
2535typedef struct kmp_target_data {
2539struct kmp_taskdata {
2540 kmp_int32 td_task_id;
2541 kmp_tasking_flags_t td_flags;
2542 kmp_team_t *td_team;
2543 kmp_info_p *td_alloc_thread;
2545 kmp_taskdata_t *td_parent;
2547 std::atomic<kmp_int32> td_untied_count;
2551 kmp_uint32 td_taskwait_counter;
2552 kmp_int32 td_taskwait_thread;
2553 KMP_ALIGN_CACHE kmp_internal_control_t
2555 KMP_ALIGN_CACHE std::atomic<kmp_int32>
2556 td_allocated_child_tasks;
2558 std::atomic<kmp_int32>
2559 td_incomplete_child_tasks;
2566 kmp_task_team_t *td_task_team;
2567 size_t td_size_alloc;
2568#if defined(KMP_GOMP_COMPAT)
2570 kmp_int32 td_size_loop_bounds;
2572 kmp_taskdata_t *td_last_tied;
2573#if defined(KMP_GOMP_COMPAT)
2575 void (*td_copy_func)(
void *,
void *);
2577 kmp_event_t td_allow_completion_event;
2579 ompt_task_info_t ompt_task_info;
2581 kmp_target_data_t td_target_data;
2585KMP_BUILD_ASSERT(
sizeof(kmp_taskdata_t) %
sizeof(
void *) == 0);
2588typedef struct kmp_base_thread_data {
2592 kmp_bootstrap_lock_t td_deque_lock;
2595 kmp_int32 td_deque_size;
2596 kmp_uint32 td_deque_head;
2597 kmp_uint32 td_deque_tail;
2598 kmp_int32 td_deque_ntasks;
2600 kmp_int32 td_deque_last_stolen;
2601#ifdef BUILD_TIED_TASK_STACK
2602 kmp_task_stack_t td_susp_tied_tasks;
2605} kmp_base_thread_data_t;
2607#define TASK_DEQUE_BITS 8
2608#define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2610#define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2611#define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2613typedef union KMP_ALIGN_CACHE kmp_thread_data {
2614 kmp_base_thread_data_t td;
2616 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2619typedef struct kmp_task_pri {
2620 kmp_thread_data_t td;
2626typedef struct kmp_base_task_team {
2627 kmp_bootstrap_lock_t
2632 kmp_bootstrap_lock_t tt_task_pri_lock;
2633 kmp_task_pri_t *tt_task_pri_list;
2635 kmp_task_team_t *tt_next;
2639 kmp_int32 tt_found_tasks;
2643 kmp_int32 tt_max_threads;
2644 kmp_int32 tt_found_proxy_tasks;
2645 kmp_int32 tt_untied_task_encountered;
2646 std::atomic<kmp_int32> tt_num_task_pri;
2649 kmp_int32 tt_hidden_helper_task_encountered;
2652 std::atomic<kmp_int32> tt_unfinished_threads;
2657} kmp_base_task_team_t;
2659union KMP_ALIGN_CACHE kmp_task_team {
2660 kmp_base_task_team_t tt;
2662 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2665#if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2668typedef struct kmp_free_list {
2669 void *th_free_list_self;
2670 void *th_free_list_sync;
2672 void *th_free_list_other;
2676#if KMP_NESTED_HOT_TEAMS
2679typedef struct kmp_hot_team_ptr {
2680 kmp_team_p *hot_team;
2681 kmp_int32 hot_team_nth;
2682} kmp_hot_team_ptr_t;
2684typedef struct kmp_teams_size {
2700typedef struct kmp_cg_root {
2701 kmp_info_p *cg_root;
2704 kmp_int32 cg_thread_limit;
2705 kmp_int32 cg_nthreads;
2706 struct kmp_cg_root *up;
2711typedef struct KMP_ALIGN_CACHE kmp_base_info {
2717 kmp_team_p *th_team;
2718 kmp_root_p *th_root;
2719 kmp_info_p *th_next_pool;
2720 kmp_disp_t *th_dispatch;
2726 kmp_info_p *th_team_master;
2727 int th_team_serialized;
2728 microtask_t th_teams_microtask;
2737 int th_team_bt_intervals;
2740 kmp_uint64 th_team_bt_intervals;
2743#if KMP_AFFINITY_SUPPORTED
2744 kmp_affin_mask_t *th_affin_mask;
2745 kmp_affinity_ids_t th_topology_ids;
2746 kmp_affinity_attrs_t th_topology_attrs;
2748 omp_allocator_handle_t th_def_allocator;
2752#if KMP_NESTED_HOT_TEAMS
2753 kmp_hot_team_ptr_t *th_hot_teams;
2759#if KMP_AFFINITY_SUPPORTED
2760 int th_current_place;
2766 int th_prev_num_threads;
2768 kmp_uint64 th_bar_arrive_time;
2769 kmp_uint64 th_bar_min_time;
2770 kmp_uint64 th_frame_time;
2772 kmp_local_t th_local;
2773 struct private_common *th_pri_head;
2778 KMP_ALIGN_CACHE kmp_team_p
2782 ompt_thread_info_t ompt_thread_info;
2786 struct common_table *th_pri_common;
2788 volatile kmp_uint32 th_spin_here;
2791 volatile void *th_sleep_loc;
2792 flag_type th_sleep_loc_type;
2799 kmp_task_team_t *th_task_team;
2800 kmp_taskdata_t *th_current_task;
2801 kmp_uint8 th_task_state;
2802 kmp_uint8 *th_task_state_memo_stack;
2804 kmp_uint32 th_task_state_top;
2805 kmp_uint32 th_task_state_stack_sz;
2806 kmp_uint32 th_reap_state;
2811 kmp_uint8 th_active_in_pool;
2813 std::atomic<kmp_uint32> th_used_in_team;
2816 struct cons_header *th_cons;
2817#if KMP_USE_HIER_SCHED
2819 kmp_hier_private_bdata_t *th_hier_bar_data;
2823 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2825 KMP_ALIGN_CACHE
volatile kmp_int32
2828#if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2830 kmp_free_list_t th_free_lists[NUM_LISTS];
2835 kmp_win32_cond_t th_suspend_cv;
2836 kmp_win32_mutex_t th_suspend_mx;
2837 std::atomic<int> th_suspend_init;
2840 kmp_cond_align_t th_suspend_cv;
2841 kmp_mutex_align_t th_suspend_mx;
2842 std::atomic<int> th_suspend_init_count;
2846 kmp_itt_mark_t th_itt_mark_single;
2849#if KMP_STATS_ENABLED
2850 kmp_stats_list *th_stats;
2853 std::atomic<bool> th_blocking;
2855 kmp_cg_root_t *th_cg_roots;
2858typedef union KMP_ALIGN_CACHE kmp_info {
2860 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2866typedef struct kmp_base_data {
2867 volatile kmp_uint32 t_value;
2870typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2872 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2876typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2878 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2880} kmp_ordered_team_t;
2882typedef int (*launch_t)(
int gtid);
2885#define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2891#if KMP_ARCH_X86 || KMP_ARCH_X86_64
2892#define KMP_INLINE_ARGV_BYTES \
2894 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2895 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2898#define KMP_INLINE_ARGV_BYTES \
2899 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2901#define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2903typedef struct KMP_ALIGN_CACHE kmp_base_team {
2906 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2907 kmp_balign_team_t t_bar[bs_last_barrier];
2908 std::atomic<int> t_construct;
2909 char pad[
sizeof(kmp_lock_t)];
2912 std::atomic<void *> t_tg_reduce_data[2];
2913 std::atomic<int> t_tg_fini_counter[2];
2917 KMP_ALIGN_CACHE
int t_master_tid;
2918 int t_master_this_cons;
2922 kmp_team_p *t_parent;
2923 kmp_team_p *t_next_pool;
2924 kmp_disp_t *t_dispatch;
2925 kmp_task_team_t *t_task_team[2];
2926 kmp_proc_bind_t t_proc_bind;
2928 kmp_uint64 t_region_time;
2933 KMP_ALIGN_CACHE
void **t_argv;
2940 ompt_team_info_t ompt_team_info;
2941 ompt_lw_taskteam_t *ompt_serialized_team_info;
2944#if KMP_ARCH_X86 || KMP_ARCH_X86_64
2945 kmp_int8 t_fp_control_saved;
2947 kmp_int16 t_x87_fpu_control_word;
2951 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2953 KMP_ALIGN_CACHE kmp_info_t **t_threads;
2955 *t_implicit_task_taskdata;
2958 KMP_ALIGN_CACHE
int t_max_argc;
2961 dispatch_shared_info_t *t_disp_buffer;
2964 kmp_r_sched_t t_sched;
2965#if KMP_AFFINITY_SUPPORTED
2969 int t_display_affinity;
2972 omp_allocator_handle_t t_def_allocator;
2975#if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2980 char dummy_padding[1024];
2983 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2986 std::atomic<kmp_int32> t_cancel_request;
2987 int t_master_active;
2988 void *t_copypriv_data;
2990 std::atomic<kmp_uint32> t_copyin_counter;
2995 distributedBarrier *b;
2998union KMP_ALIGN_CACHE kmp_team {
3001 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
3004typedef union KMP_ALIGN_CACHE kmp_time_global {
3006 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3010typedef struct kmp_base_global {
3012 kmp_time_global_t g_time;
3015 volatile int g_abort;
3016 volatile int g_done;
3019 enum dynamic_mode g_dynamic_mode;
3022typedef union KMP_ALIGN_CACHE kmp_global {
3023 kmp_base_global_t g;
3025 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
3028typedef struct kmp_base_root {
3033 volatile int r_active;
3035 std::atomic<int> r_in_parallel;
3037 kmp_team_t *r_root_team;
3038 kmp_team_t *r_hot_team;
3039 kmp_info_t *r_uber_thread;
3040 kmp_lock_t r_begin_lock;
3041 volatile int r_begin;
3043#if KMP_AFFINITY_SUPPORTED
3044 int r_affinity_assigned;
3048typedef union KMP_ALIGN_CACHE kmp_root {
3051 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
3054struct fortran_inx_info {
3062typedef struct kmp_old_threads_list_t {
3063 kmp_info_t **threads;
3064 struct kmp_old_threads_list_t *next;
3065} kmp_old_threads_list_t;
3069extern int __kmp_settings;
3070extern int __kmp_duplicate_library_ok;
3072extern int __kmp_forkjoin_frames;
3073extern int __kmp_forkjoin_frames_mode;
3075extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3076extern int __kmp_determ_red;
3079extern int kmp_a_debug;
3080extern int kmp_b_debug;
3081extern int kmp_c_debug;
3082extern int kmp_d_debug;
3083extern int kmp_e_debug;
3084extern int kmp_f_debug;
3088#define KMP_DEBUG_BUF_LINES_INIT 512
3089#define KMP_DEBUG_BUF_LINES_MIN 1
3091#define KMP_DEBUG_BUF_CHARS_INIT 128
3092#define KMP_DEBUG_BUF_CHARS_MIN 2
3096extern int __kmp_debug_buf_lines;
3098 __kmp_debug_buf_chars;
3099extern int __kmp_debug_buf_atomic;
3102extern char *__kmp_debug_buffer;
3103extern std::atomic<int> __kmp_debug_count;
3105extern int __kmp_debug_buf_warn_chars;
3110extern int __kmp_par_range;
3112#define KMP_PAR_RANGE_ROUTINE_LEN 1024
3113extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3114#define KMP_PAR_RANGE_FILENAME_LEN 1024
3115extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3116extern int __kmp_par_range_lb;
3117extern int __kmp_par_range_ub;
3123extern int __kmp_storage_map_verbose;
3125extern int __kmp_storage_map_verbose_specified;
3127#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3128extern kmp_cpuinfo_t __kmp_cpuinfo;
3129static inline bool __kmp_is_hybrid_cpu() {
return __kmp_cpuinfo.flags.hybrid; }
3130#elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3131static inline bool __kmp_is_hybrid_cpu() {
return true; }
3133static inline bool __kmp_is_hybrid_cpu() {
return false; }
3136extern volatile int __kmp_init_serial;
3137extern volatile int __kmp_init_gtid;
3138extern volatile int __kmp_init_common;
3139extern volatile int __kmp_need_register_serial;
3140extern volatile int __kmp_init_middle;
3141extern volatile int __kmp_init_parallel;
3143extern volatile int __kmp_init_monitor;
3145extern volatile int __kmp_init_user_locks;
3146extern volatile int __kmp_init_hidden_helper_threads;
3147extern int __kmp_init_counter;
3148extern int __kmp_root_counter;
3149extern int __kmp_version;
3152extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3155extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3156extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3157extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3158extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3159extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3160extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3161extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3162extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3163extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3164extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3165extern char const *__kmp_barrier_type_name[bs_last_barrier];
3166extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3169extern kmp_bootstrap_lock_t __kmp_initz_lock;
3170extern kmp_bootstrap_lock_t __kmp_forkjoin_lock;
3171extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3172extern kmp_bootstrap_lock_t
3175extern kmp_bootstrap_lock_t
3178extern kmp_bootstrap_lock_t
3179 __kmp_tp_cached_lock;
3182extern kmp_lock_t __kmp_global_lock;
3183extern kmp_queuing_lock_t __kmp_dispatch_lock;
3184extern kmp_lock_t __kmp_debug_lock;
3186extern enum library_type __kmp_library;
3192extern int __kmp_chunk;
3193extern int __kmp_force_monotonic;
3195extern size_t __kmp_stksize;
3197extern size_t __kmp_monitor_stksize;
3199extern size_t __kmp_stkoffset;
3200extern int __kmp_stkpadding;
3203 __kmp_malloc_pool_incr;
3204extern int __kmp_env_stksize;
3205extern int __kmp_env_blocktime;
3206extern int __kmp_env_checks;
3207extern int __kmp_env_consistency_check;
3208extern int __kmp_generate_warnings;
3209extern int __kmp_reserve_warn;
3212extern int __kmp_suspend_count;
3215extern kmp_int32 __kmp_use_yield;
3216extern kmp_int32 __kmp_use_yield_exp_set;
3217extern kmp_uint32 __kmp_yield_init;
3218extern kmp_uint32 __kmp_yield_next;
3219extern kmp_uint64 __kmp_pause_init;
3222extern int __kmp_allThreadsSpecified;
3224extern size_t __kmp_align_alloc;
3226extern int __kmp_xproc;
3227extern int __kmp_avail_proc;
3228extern size_t __kmp_sys_min_stksize;
3229extern int __kmp_sys_max_nth;
3231extern int __kmp_max_nth;
3233extern int __kmp_cg_max_nth;
3234extern int __kmp_teams_max_nth;
3235extern int __kmp_threads_capacity;
3237extern int __kmp_dflt_team_nth;
3239extern int __kmp_dflt_team_nth_ub;
3241extern int __kmp_tp_capacity;
3243extern int __kmp_tp_cached;
3245extern int __kmp_dflt_blocktime;
3247extern bool __kmp_wpolicy_passive;
3250 __kmp_monitor_wakeups;
3251extern int __kmp_bt_intervals;
3254#ifdef KMP_ADJUST_BLOCKTIME
3255extern int __kmp_zero_bt;
3257#ifdef KMP_DFLT_NTH_CORES
3258extern int __kmp_ncores;
3261extern int __kmp_abort_delay;
3263extern int __kmp_need_register_atfork_specified;
3264extern int __kmp_need_register_atfork;
3266extern int __kmp_gtid_mode;
3274 __kmp_adjust_gtid_mode;
3275#ifdef KMP_TDATA_GTID
3276extern KMP_THREAD_LOCAL
int __kmp_gtid;
3278extern int __kmp_tls_gtid_min;
3279extern int __kmp_foreign_tp;
3280#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3281extern int __kmp_inherit_fp_control;
3282extern kmp_int16 __kmp_init_x87_fpu_control_word;
3283extern kmp_uint32 __kmp_init_mxcsr;
3288extern int __kmp_dflt_max_active_levels;
3291extern bool __kmp_dflt_max_active_levels_set;
3292extern int __kmp_dispatch_num_buffers;
3294#if KMP_NESTED_HOT_TEAMS
3295extern int __kmp_hot_teams_mode;
3296extern int __kmp_hot_teams_max_level;
3300extern enum clock_function_type __kmp_clock_function;
3301extern int __kmp_clock_function_param;
3304#if KMP_MIC_SUPPORTED
3305extern enum mic_type __kmp_mic_type;
3308#ifdef USE_LOAD_BALANCE
3309extern double __kmp_load_balance_interval;
3313typedef struct kmp_nested_nthreads_t {
3317} kmp_nested_nthreads_t;
3319extern kmp_nested_nthreads_t __kmp_nested_nth;
3321#if KMP_USE_ADAPTIVE_LOCKS
3324struct kmp_adaptive_backoff_params_t {
3326 kmp_uint32 max_soft_retries;
3329 kmp_uint32 max_badness;
3332extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3334#if KMP_DEBUG_ADAPTIVE_LOCKS
3335extern const char *__kmp_speculative_statsfile;
3340extern int __kmp_display_env;
3341extern int __kmp_display_env_verbose;
3342extern int __kmp_omp_cancellation;
3343extern int __kmp_nteams;
3344extern int __kmp_teams_thread_limit;
3350extern kmp_info_t **__kmp_threads;
3352extern kmp_old_threads_list_t *__kmp_old_threads_list;
3354extern volatile kmp_team_t *__kmp_team_pool;
3355extern volatile kmp_info_t *__kmp_thread_pool;
3356extern kmp_info_t *__kmp_thread_pool_insert_pt;
3359extern volatile int __kmp_nth;
3362extern volatile int __kmp_all_nth;
3363extern std::atomic<int> __kmp_thread_pool_active_nth;
3365extern kmp_root_t **__kmp_root;
3369#define __kmp_get_gtid() __kmp_get_global_thread_id()
3370#define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3371#define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3372#define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3373#define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3378#define __kmp_get_team_num_threads(gtid) \
3379 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3381static inline bool KMP_UBER_GTID(
int gtid) {
3382 KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3383 KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3384 return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3385 __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3388static inline int __kmp_tid_from_gtid(
int gtid) {
3389 KMP_DEBUG_ASSERT(gtid >= 0);
3390 return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3393static inline int __kmp_gtid_from_tid(
int tid,
const kmp_team_t *team) {
3394 KMP_DEBUG_ASSERT(tid >= 0 && team);
3395 return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3398static inline int __kmp_gtid_from_thread(
const kmp_info_t *thr) {
3399 KMP_DEBUG_ASSERT(thr);
3400 return thr->th.th_info.ds.ds_gtid;
3403static inline kmp_info_t *__kmp_thread_from_gtid(
int gtid) {
3404 KMP_DEBUG_ASSERT(gtid >= 0);
3405 return __kmp_threads[gtid];
3408static inline kmp_team_t *__kmp_team_from_gtid(
int gtid) {
3409 KMP_DEBUG_ASSERT(gtid >= 0);
3410 return __kmp_threads[gtid]->th.th_team;
3413static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3414 if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3415 KMP_FATAL(ThreadIdentInvalid);
3418#if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3419extern int __kmp_user_level_mwait;
3420extern int __kmp_umwait_enabled;
3421extern int __kmp_mwait_enabled;
3422extern int __kmp_mwait_hints;
3426extern int __kmp_waitpkg_enabled;
3427extern int __kmp_tpause_state;
3428extern int __kmp_tpause_hint;
3429extern int __kmp_tpause_enabled;
3434extern kmp_global_t __kmp_global;
3436extern kmp_info_t __kmp_monitor;
3438extern std::atomic<kmp_int32> __kmp_team_counter;
3440extern std::atomic<kmp_int32> __kmp_task_counter;
3443#define _KMP_GEN_ID(counter) \
3444 (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3446#define _KMP_GEN_ID(counter) (~0)
3449#define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3450#define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3454extern void __kmp_print_storage_map_gtid(
int gtid,
void *p1,
void *p2,
3455 size_t size,
char const *format, ...);
3457extern void __kmp_serial_initialize(
void);
3458extern void __kmp_middle_initialize(
void);
3459extern void __kmp_parallel_initialize(
void);
3461extern void __kmp_internal_begin(
void);
3462extern void __kmp_internal_end_library(
int gtid);
3463extern void __kmp_internal_end_thread(
int gtid);
3464extern void __kmp_internal_end_atexit(
void);
3465extern void __kmp_internal_end_dtor(
void);
3466extern void __kmp_internal_end_dest(
void *);
3468extern int __kmp_register_root(
int initial_thread);
3469extern void __kmp_unregister_root(
int gtid);
3470extern void __kmp_unregister_library(
void);
3472extern int __kmp_ignore_mppbeg(
void);
3473extern int __kmp_ignore_mppend(
void);
3475extern int __kmp_enter_single(
int gtid,
ident_t *id_ref,
int push_ws);
3476extern void __kmp_exit_single(
int gtid);
3478extern void __kmp_parallel_deo(
int *gtid_ref,
int *cid_ref,
ident_t *loc_ref);
3479extern void __kmp_parallel_dxo(
int *gtid_ref,
int *cid_ref,
ident_t *loc_ref);
3481#ifdef USE_LOAD_BALANCE
3482extern int __kmp_get_load_balance(
int);
3485extern int __kmp_get_global_thread_id(
void);
3486extern int __kmp_get_global_thread_id_reg(
void);
3487extern void __kmp_exit_thread(
int exit_status);
3488extern void __kmp_abort(
char const *format, ...);
3489extern void __kmp_abort_thread(
void);
3490KMP_NORETURN
extern void __kmp_abort_process(
void);
3491extern void __kmp_warn(
char const *format, ...);
3493extern void __kmp_set_num_threads(
int new_nth,
int gtid);
3497static inline kmp_info_t *__kmp_entry_thread() {
3498 int gtid = __kmp_entry_gtid();
3500 return __kmp_threads[gtid];
3503extern void __kmp_set_max_active_levels(
int gtid,
int new_max_active_levels);
3504extern int __kmp_get_max_active_levels(
int gtid);
3505extern int __kmp_get_ancestor_thread_num(
int gtid,
int level);
3506extern int __kmp_get_team_size(
int gtid,
int level);
3507extern void __kmp_set_schedule(
int gtid, kmp_sched_t new_sched,
int chunk);
3508extern void __kmp_get_schedule(
int gtid, kmp_sched_t *sched,
int *chunk);
3510extern unsigned short __kmp_get_random(kmp_info_t *thread);
3511extern void __kmp_init_random(kmp_info_t *thread);
3513extern kmp_r_sched_t __kmp_get_schedule_global(
void);
3514extern void __kmp_adjust_num_threads(
int new_nproc);
3515extern void __kmp_check_stksize(
size_t *val);
3517extern void *___kmp_allocate(
size_t size KMP_SRC_LOC_DECL);
3518extern void *___kmp_page_allocate(
size_t size KMP_SRC_LOC_DECL);
3519extern void ___kmp_free(
void *ptr KMP_SRC_LOC_DECL);
3520#define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3521#define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3522#define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3525extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3526 size_t size KMP_SRC_LOC_DECL);
3527extern void ___kmp_fast_free(kmp_info_t *this_thr,
void *ptr KMP_SRC_LOC_DECL);
3528extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3529extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3530#define __kmp_fast_allocate(this_thr, size) \
3531 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3532#define __kmp_fast_free(this_thr, ptr) \
3533 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3536extern void *___kmp_thread_malloc(kmp_info_t *th,
size_t size KMP_SRC_LOC_DECL);
3537extern void *___kmp_thread_calloc(kmp_info_t *th,
size_t nelem,
3538 size_t elsize KMP_SRC_LOC_DECL);
3539extern void *___kmp_thread_realloc(kmp_info_t *th,
void *ptr,
3540 size_t size KMP_SRC_LOC_DECL);
3541extern void ___kmp_thread_free(kmp_info_t *th,
void *ptr KMP_SRC_LOC_DECL);
3542#define __kmp_thread_malloc(th, size) \
3543 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3544#define __kmp_thread_calloc(th, nelem, elsize) \
3545 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3546#define __kmp_thread_realloc(th, ptr, size) \
3547 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3548#define __kmp_thread_free(th, ptr) \
3549 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3551extern void __kmp_push_num_threads(
ident_t *loc,
int gtid,
int num_threads);
3553extern void __kmp_push_proc_bind(
ident_t *loc,
int gtid,
3554 kmp_proc_bind_t proc_bind);
3555extern void __kmp_push_num_teams(
ident_t *loc,
int gtid,
int num_teams,
3557extern void __kmp_push_num_teams_51(
ident_t *loc,
int gtid,
int num_teams_lb,
3558 int num_teams_ub,
int num_threads);
3560extern void __kmp_yield();
3564 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3567 kmp_uint32 ub, kmp_int32 st,
3571 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3574 kmp_uint64 ub, kmp_int64 st,
3578 kmp_int32 *p_last, kmp_int32 *p_lb,
3579 kmp_int32 *p_ub, kmp_int32 *p_st);
3581 kmp_int32 *p_last, kmp_uint32 *p_lb,
3582 kmp_uint32 *p_ub, kmp_int32 *p_st);
3584 kmp_int32 *p_last, kmp_int64 *p_lb,
3585 kmp_int64 *p_ub, kmp_int64 *p_st);
3587 kmp_int32 *p_last, kmp_uint64 *p_lb,
3588 kmp_uint64 *p_ub, kmp_int64 *p_st);
3595#ifdef KMP_GOMP_COMPAT
3597extern void __kmp_aux_dispatch_init_4(
ident_t *loc, kmp_int32 gtid,
3599 kmp_int32 ub, kmp_int32 st,
3600 kmp_int32 chunk,
int push_ws);
3601extern void __kmp_aux_dispatch_init_4u(
ident_t *loc, kmp_int32 gtid,
3603 kmp_uint32 ub, kmp_int32 st,
3604 kmp_int32 chunk,
int push_ws);
3605extern void __kmp_aux_dispatch_init_8(
ident_t *loc, kmp_int32 gtid,
3607 kmp_int64 ub, kmp_int64 st,
3608 kmp_int64 chunk,
int push_ws);
3609extern void __kmp_aux_dispatch_init_8u(
ident_t *loc, kmp_int32 gtid,
3611 kmp_uint64 ub, kmp_int64 st,
3612 kmp_int64 chunk,
int push_ws);
3613extern void __kmp_aux_dispatch_fini_chunk_4(
ident_t *loc, kmp_int32 gtid);
3614extern void __kmp_aux_dispatch_fini_chunk_8(
ident_t *loc, kmp_int32 gtid);
3615extern void __kmp_aux_dispatch_fini_chunk_4u(
ident_t *loc, kmp_int32 gtid);
3616extern void __kmp_aux_dispatch_fini_chunk_8u(
ident_t *loc, kmp_int32 gtid);
3620extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3621extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3622extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3623extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3624extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3625extern kmp_uint32 __kmp_wait_4(kmp_uint32
volatile *spinner, kmp_uint32 checker,
3626 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3628extern void __kmp_wait_4_ptr(
void *spinner, kmp_uint32 checker,
3629 kmp_uint32 (*pred)(
void *, kmp_uint32),
void *obj);
3631extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3638extern void __kmp_release_64(kmp_flag_64<> *flag);
3640extern void __kmp_infinite_loop(
void);
3642extern void __kmp_cleanup(
void);
3644#if KMP_HANDLE_SIGNALS
3645extern int __kmp_handle_signals;
3646extern void __kmp_install_signals(
int parallel_init);
3647extern void __kmp_remove_signals(
void);
3650extern void __kmp_clear_system_time(
void);
3651extern void __kmp_read_system_time(
double *delta);
3653extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3655extern void __kmp_expand_host_name(
char *buffer,
size_t size);
3656extern void __kmp_expand_file_name(
char *result,
size_t rlen,
char *pattern);
3658#if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM))
3660__kmp_initialize_system_tick(
void);
3664__kmp_runtime_initialize(
void);
3665extern void __kmp_runtime_destroy(
void);
3667#if KMP_AFFINITY_SUPPORTED
3668extern char *__kmp_affinity_print_mask(
char *buf,
int buf_len,
3669 kmp_affin_mask_t *mask);
3670extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3671 kmp_affin_mask_t *mask);
3672extern void __kmp_affinity_initialize(kmp_affinity_t &affinity);
3673extern void __kmp_affinity_uninitialize(
void);
3674extern void __kmp_affinity_set_init_mask(
3675 int gtid,
int isa_root);
3676extern void __kmp_affinity_set_place(
int gtid);
3677extern void __kmp_affinity_determine_capable(
const char *env_var);
3678extern int __kmp_aux_set_affinity(
void **mask);
3679extern int __kmp_aux_get_affinity(
void **mask);
3680extern int __kmp_aux_get_affinity_max_proc();
3681extern int __kmp_aux_set_affinity_mask_proc(
int proc,
void **mask);
3682extern int __kmp_aux_unset_affinity_mask_proc(
int proc,
void **mask);
3683extern int __kmp_aux_get_affinity_mask_proc(
int proc,
void **mask);
3684extern void __kmp_balanced_affinity(kmp_info_t *th,
int team_size);
3685#if KMP_OS_LINUX || KMP_OS_FREEBSD
3686extern int kmp_set_thread_affinity_mask_initial(
void);
3688static inline void __kmp_assign_root_init_mask() {
3689 int gtid = __kmp_entry_gtid();
3690 kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3691 if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3692 __kmp_affinity_set_init_mask(gtid, TRUE);
3693 r->r.r_affinity_assigned = TRUE;
3696static inline void __kmp_reset_root_init_mask(
int gtid) {
3697 if (!KMP_AFFINITY_CAPABLE())
3699 kmp_info_t *th = __kmp_threads[gtid];
3700 kmp_root_t *r = th->th.th_root;
3701 if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3702 __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3703 KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3704 r->r.r_affinity_assigned = FALSE;
3708#define __kmp_assign_root_init_mask()
3709static inline void __kmp_reset_root_init_mask(
int gtid) {}
3714extern size_t __kmp_aux_capture_affinity(
int gtid,
const char *format,
3715 kmp_str_buf_t *buffer);
3716extern void __kmp_aux_display_affinity(
int gtid,
const char *format);
3718extern void __kmp_cleanup_hierarchy();
3719extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3723extern int __kmp_futex_determine_capable(
void);
3727extern void __kmp_gtid_set_specific(
int gtid);
3728extern int __kmp_gtid_get_specific(
void);
3730extern double __kmp_read_cpu_time(
void);
3732extern int __kmp_read_system_info(
struct kmp_sys_info *info);
3735extern void __kmp_create_monitor(kmp_info_t *th);
3738extern void *__kmp_launch_thread(kmp_info_t *thr);
3740extern void __kmp_create_worker(
int gtid, kmp_info_t *th,
size_t stack_size);
3743extern int __kmp_still_running(kmp_info_t *th);
3744extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3745extern void __kmp_free_handle(kmp_thread_t tHandle);
3749extern void __kmp_reap_monitor(kmp_info_t *th);
3751extern void __kmp_reap_worker(kmp_info_t *th);
3752extern void __kmp_terminate_thread(
int gtid);
3754extern int __kmp_try_suspend_mx(kmp_info_t *th);
3755extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3756extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3758extern void __kmp_elapsed(
double *);
3759extern void __kmp_elapsed_tick(
double *);
3761extern void __kmp_enable(
int old_state);
3762extern void __kmp_disable(
int *old_state);
3764extern void __kmp_thread_sleep(
int millis);
3766extern void __kmp_common_initialize(
void);
3767extern void __kmp_common_destroy(
void);
3768extern void __kmp_common_destroy_gtid(
int gtid);
3771extern void __kmp_register_atfork(
void);
3773extern void __kmp_suspend_initialize(
void);
3774extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3775extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3777extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3780__kmp_allocate_team(kmp_root_t *root,
int new_nproc,
int max_nproc,
3782 ompt_data_t ompt_parallel_data,
3784 kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3785 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3786extern void __kmp_free_thread(kmp_info_t *);
3787extern void __kmp_free_team(kmp_root_t *,
3788 kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3789extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3793extern void __kmp_initialize_bget(kmp_info_t *th);
3794extern void __kmp_finalize_bget(kmp_info_t *th);
3796KMP_EXPORT
void *kmpc_malloc(
size_t size);
3797KMP_EXPORT
void *kmpc_aligned_malloc(
size_t size,
size_t alignment);
3798KMP_EXPORT
void *kmpc_calloc(
size_t nelem,
size_t elsize);
3799KMP_EXPORT
void *kmpc_realloc(
void *ptr,
size_t size);
3800KMP_EXPORT
void kmpc_free(
void *ptr);
3804extern int __kmp_barrier(
enum barrier_type bt,
int gtid,
int is_split,
3805 size_t reduce_size,
void *reduce_data,
3806 void (*reduce)(
void *,
void *));
3807extern void __kmp_end_split_barrier(
enum barrier_type bt,
int gtid);
3808extern int __kmp_barrier_gomp_cancel(
int gtid);
3814enum fork_context_e {
3820extern int __kmp_fork_call(
ident_t *loc,
int gtid,
3821 enum fork_context_e fork_context, kmp_int32 argc,
3822 microtask_t microtask, launch_t invoker,
3825extern void __kmp_join_call(
ident_t *loc,
int gtid
3828 enum fork_context_e fork_context
3831 int exit_teams = 0);
3833extern void __kmp_serialized_parallel(
ident_t *
id, kmp_int32 gtid);
3834extern void __kmp_internal_fork(
ident_t *
id,
int gtid, kmp_team_t *team);
3835extern void __kmp_internal_join(
ident_t *
id,
int gtid, kmp_team_t *team);
3836extern int __kmp_invoke_task_func(
int gtid);
3837extern void __kmp_run_before_invoked_task(
int gtid,
int tid,
3838 kmp_info_t *this_thr,
3840extern void __kmp_run_after_invoked_task(
int gtid,
int tid,
3841 kmp_info_t *this_thr,
3845KMP_EXPORT
int __kmpc_invoke_task_func(
int gtid);
3846extern int __kmp_invoke_teams_master(
int gtid);
3847extern void __kmp_teams_master(
int gtid);
3848extern int __kmp_aux_get_team_num();
3849extern int __kmp_aux_get_num_teams();
3850extern void __kmp_save_internal_controls(kmp_info_t *thread);
3851extern void __kmp_user_set_library(
enum library_type arg);
3852extern void __kmp_aux_set_library(
enum library_type arg);
3853extern void __kmp_aux_set_stacksize(
size_t arg);
3854extern void __kmp_aux_set_blocktime(
int arg, kmp_info_t *thread,
int tid);
3855extern void __kmp_aux_set_defaults(
char const *str,
size_t len);
3858void kmpc_set_blocktime(
int arg);
3859void ompc_set_nested(
int flag);
3860void ompc_set_dynamic(
int flag);
3861void ompc_set_num_threads(
int arg);
3863extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3864 kmp_team_t *team,
int tid);
3865extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3866extern kmp_task_t *__kmp_task_alloc(
ident_t *loc_ref, kmp_int32 gtid,
3867 kmp_tasking_flags_t *flags,
3868 size_t sizeof_kmp_task_t,
3869 size_t sizeof_shareds,
3870 kmp_routine_entry_t task_entry);
3871extern void __kmp_init_implicit_task(
ident_t *loc_ref, kmp_info_t *this_thr,
3872 kmp_team_t *team,
int tid,
3874extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3875extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3877extern kmp_event_t *__kmpc_task_allow_completion_event(
ident_t *loc_ref,
3880extern void __kmp_fulfill_event(kmp_event_t *event);
3882extern void __kmp_free_task_team(kmp_info_t *thread,
3883 kmp_task_team_t *task_team);
3884extern void __kmp_reap_task_teams(
void);
3885extern void __kmp_wait_to_unref_task_teams(
void);
3886extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3888extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3889extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3896extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3899extern int __kmp_is_address_mapped(
void *addr);
3900extern kmp_uint64 __kmp_hardware_timestamp(
void);
3903extern int __kmp_read_from_file(
char const *path,
char const *format, ...);
3911extern int __kmp_invoke_microtask(microtask_t pkfn,
int gtid,
int npr,
int argc,
3915 void **exit_frame_ptr
3928 size_t vector_length);
3932KMP_EXPORT
void *__kmpc_threadprivate(
ident_t *, kmp_int32 global_tid,
3933 void *data,
size_t size);
3960 kmp_critical_name *);
3962 kmp_critical_name *);
3964 kmp_critical_name *, uint32_t hint);
3970 kmp_int32 global_tid);
3977 kmp_int32 numberOfSections);
3980KMP_EXPORT
void KMPC_FOR_STATIC_INIT(
ident_t *loc, kmp_int32 global_tid,
3981 kmp_int32 schedtype, kmp_int32 *plastiter,
3982 kmp_int *plower, kmp_int *pupper,
3983 kmp_int *pstride, kmp_int incr,
3989 size_t cpy_size,
void *cpy_data,
3990 void (*cpy_func)(
void *,
void *),
3996extern void KMPC_SET_NUM_THREADS(
int arg);
3997extern void KMPC_SET_DYNAMIC(
int flag);
3998extern void KMPC_SET_NESTED(
int flag);
4001KMP_EXPORT kmp_int32 __kmpc_omp_task(
ident_t *loc_ref, kmp_int32 gtid,
4002 kmp_task_t *new_task);
4003KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(
ident_t *loc_ref, kmp_int32 gtid,
4005 size_t sizeof_kmp_task_t,
4006 size_t sizeof_shareds,
4007 kmp_routine_entry_t task_entry);
4008KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
4009 ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags,
size_t sizeof_kmp_task_t,
4010 size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
4011KMP_EXPORT
void __kmpc_omp_task_begin_if0(
ident_t *loc_ref, kmp_int32 gtid,
4013KMP_EXPORT
void __kmpc_omp_task_complete_if0(
ident_t *loc_ref, kmp_int32 gtid,
4015KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(
ident_t *loc_ref, kmp_int32 gtid,
4016 kmp_task_t *new_task);
4017KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(
ident_t *loc_ref, kmp_int32 gtid);
4018KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(
ident_t *loc_ref, kmp_int32 gtid,
4022void __kmpc_omp_task_begin(
ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
4023void __kmpc_omp_task_complete(
ident_t *loc_ref, kmp_int32 gtid,
4029KMP_EXPORT
void __kmpc_taskgroup(
ident_t *loc,
int gtid);
4030KMP_EXPORT
void __kmpc_end_taskgroup(
ident_t *loc,
int gtid);
4033 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
4034 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
4035 kmp_depend_info_t *noalias_dep_list);
4038 kmp_depend_info_t *dep_list,
4039 kmp_int32 ndeps_noalias,
4040 kmp_depend_info_t *noalias_dep_list);
4043KMP_EXPORT
void __kmpc_omp_taskwait_deps_51(
ident_t *loc_ref, kmp_int32 gtid,
4045 kmp_depend_info_t *dep_list,
4046 kmp_int32 ndeps_noalias,
4047 kmp_depend_info_t *noalias_dep_list,
4048 kmp_int32 has_no_wait);
4050extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
4051 bool serialize_immediate);
4053KMP_EXPORT kmp_int32 __kmpc_cancel(
ident_t *loc_ref, kmp_int32 gtid,
4054 kmp_int32 cncl_kind);
4055KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(
ident_t *loc_ref, kmp_int32 gtid,
4056 kmp_int32 cncl_kind);
4057KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(
ident_t *loc_ref, kmp_int32 gtid);
4058KMP_EXPORT
int __kmp_get_cancellation_status(
int cancel_kind);
4063 kmp_int32 if_val, kmp_uint64 *lb,
4064 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
4065 kmp_int32 sched, kmp_uint64 grainsize,
4068 kmp_task_t *task, kmp_int32 if_val,
4069 kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4070 kmp_int32 nogroup, kmp_int32 sched,
4071 kmp_uint64 grainsize, kmp_int32 modifier,
4080 int num,
void *data);
4084 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4085 kmp_task_affinity_info_t *affin_list);
4086KMP_EXPORT
void __kmp_set_num_teams(
int num_teams);
4087KMP_EXPORT
int __kmp_get_max_teams(
void);
4088KMP_EXPORT
void __kmp_set_teams_thread_limit(
int limit);
4089KMP_EXPORT
int __kmp_get_teams_thread_limit(
void);
4096KMP_EXPORT
void __kmpc_init_lock(
ident_t *loc, kmp_int32 gtid,
4098KMP_EXPORT
void __kmpc_init_nest_lock(
ident_t *loc, kmp_int32 gtid,
4100KMP_EXPORT
void __kmpc_destroy_lock(
ident_t *loc, kmp_int32 gtid,
4102KMP_EXPORT
void __kmpc_destroy_nest_lock(
ident_t *loc, kmp_int32 gtid,
4104KMP_EXPORT
void __kmpc_set_lock(
ident_t *loc, kmp_int32 gtid,
void **user_lock);
4105KMP_EXPORT
void __kmpc_set_nest_lock(
ident_t *loc, kmp_int32 gtid,
4107KMP_EXPORT
void __kmpc_unset_lock(
ident_t *loc, kmp_int32 gtid,
4109KMP_EXPORT
void __kmpc_unset_nest_lock(
ident_t *loc, kmp_int32 gtid,
4111KMP_EXPORT
int __kmpc_test_lock(
ident_t *loc, kmp_int32 gtid,
void **user_lock);
4112KMP_EXPORT
int __kmpc_test_nest_lock(
ident_t *loc, kmp_int32 gtid,
4115KMP_EXPORT
void __kmpc_init_lock_with_hint(
ident_t *loc, kmp_int32 gtid,
4116 void **user_lock, uintptr_t hint);
4117KMP_EXPORT
void __kmpc_init_nest_lock_with_hint(
ident_t *loc, kmp_int32 gtid,
4124 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
size_t reduce_size,
4125 void *reduce_data,
void (*reduce_func)(
void *lhs_data,
void *rhs_data),
4126 kmp_critical_name *lck);
4128 kmp_critical_name *lck);
4130 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
size_t reduce_size,
4131 void *reduce_data,
void (*reduce_func)(
void *lhs_data,
void *rhs_data),
4132 kmp_critical_name *lck);
4134 kmp_critical_name *lck);
4138extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4139 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
size_t reduce_size,
4140 void *reduce_data,
void (*reduce_func)(
void *lhs_data,
void *rhs_data),
4141 kmp_critical_name *lck);
4144KMP_EXPORT kmp_int32 __kmp_get_reduce_method(
void);
4146KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4147KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4153KMP_EXPORT
void __kmpc_pop_num_threads(
ident_t *loc, kmp_int32 global_tid);
4155 kmp_int32 num_threads);
4157KMP_EXPORT
void __kmpc_push_proc_bind(
ident_t *loc, kmp_int32 global_tid,
4160 kmp_int32 num_teams,
4161 kmp_int32 num_threads);
4164 kmp_int32 num_teams_lb,
4165 kmp_int32 num_teams_ub,
4166 kmp_int32 num_threads);
4176 const struct kmp_dim *dims);
4177KMP_EXPORT
void __kmpc_doacross_wait(
ident_t *loc, kmp_int32 gtid,
4178 const kmp_int64 *vec);
4179KMP_EXPORT
void __kmpc_doacross_post(
ident_t *loc, kmp_int32 gtid,
4180 const kmp_int64 *vec);
4181KMP_EXPORT
void __kmpc_doacross_fini(
ident_t *loc, kmp_int32 gtid);
4184 void *data,
size_t size,
4188extern int _You_must_link_with_exactly_one_OpenMP_library;
4189extern int _You_must_link_with_Intel_OpenMP_library;
4190#if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
4191extern int _You_must_link_with_Microsoft_OpenMP_library;
4196void kmp_threadprivate_insert_private_data(
int gtid,
void *pc_addr,
4197 void *data_addr,
size_t pc_size);
4198struct private_common *kmp_threadprivate_insert(
int gtid,
void *pc_addr,
4201void __kmp_threadprivate_resize_cache(
int newCapacity);
4202void __kmp_cleanup_threadprivate_caches();
4206#define KMPC_CONVENTION __cdecl
4208#define KMPC_CONVENTION
4212typedef enum omp_sched_t {
4213 omp_sched_static = 1,
4214 omp_sched_dynamic = 2,
4215 omp_sched_guided = 3,
4218typedef void *kmp_affinity_mask_t;
4221KMP_EXPORT
void KMPC_CONVENTION ompc_set_max_active_levels(
int);
4222KMP_EXPORT
void KMPC_CONVENTION ompc_set_schedule(omp_sched_t,
int);
4223KMP_EXPORT
int KMPC_CONVENTION ompc_get_ancestor_thread_num(
int);
4224KMP_EXPORT
int KMPC_CONVENTION ompc_get_team_size(
int);
4225KMP_EXPORT
int KMPC_CONVENTION
4226kmpc_set_affinity_mask_proc(
int, kmp_affinity_mask_t *);
4227KMP_EXPORT
int KMPC_CONVENTION
4228kmpc_unset_affinity_mask_proc(
int, kmp_affinity_mask_t *);
4229KMP_EXPORT
int KMPC_CONVENTION
4230kmpc_get_affinity_mask_proc(
int, kmp_affinity_mask_t *);
4232KMP_EXPORT
void KMPC_CONVENTION kmpc_set_stacksize(
int);
4233KMP_EXPORT
void KMPC_CONVENTION kmpc_set_stacksize_s(
size_t);
4234KMP_EXPORT
void KMPC_CONVENTION kmpc_set_library(
int);
4235KMP_EXPORT
void KMPC_CONVENTION kmpc_set_defaults(
char const *);
4236KMP_EXPORT
void KMPC_CONVENTION kmpc_set_disp_num_buffers(
int);
4237void KMP_EXPAND_NAME(ompc_set_affinity_format)(
char const *format);
4238size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(
char *buffer,
size_t size);
4239void KMP_EXPAND_NAME(ompc_display_affinity)(
char const *format);
4240size_t KMP_EXPAND_NAME(ompc_capture_affinity)(
char *buffer,
size_t buf_size,
4241 char const *format);
4243enum kmp_target_offload_kind {
4248typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4250extern kmp_target_offload_kind_t __kmp_target_offload;
4251extern int __kmpc_get_target_offload();
4254#define KMP_DEVICE_DEFAULT -1
4255#define KMP_DEVICE_ALL -11
4261typedef enum kmp_pause_status_t {
4263 kmp_soft_paused = 1,
4265} kmp_pause_status_t;
4268extern kmp_pause_status_t __kmp_pause_status;
4269extern int __kmpc_pause_resource(kmp_pause_status_t level);
4270extern int __kmp_pause_resource(kmp_pause_status_t level);
4272extern void __kmp_resume_if_soft_paused();
4276static inline void __kmp_resume_if_hard_paused() {
4277 if (__kmp_pause_status == kmp_hard_paused) {
4278 __kmp_pause_status = kmp_not_paused;
4282extern void __kmp_omp_display_env(
int verbose);
4285extern volatile int __kmp_init_hidden_helper;
4287extern volatile int __kmp_hidden_helper_team_done;
4289extern kmp_int32 __kmp_enable_hidden_helper;
4291extern kmp_info_t *__kmp_hidden_helper_main_thread;
4293extern kmp_info_t **__kmp_hidden_helper_threads;
4295extern kmp_int32 __kmp_hidden_helper_threads_num;
4297extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4299extern void __kmp_hidden_helper_initialize();
4300extern void __kmp_hidden_helper_threads_initz_routine();
4301extern void __kmp_do_initialize_hidden_helper_threads();
4302extern void __kmp_hidden_helper_threads_initz_wait();
4303extern void __kmp_hidden_helper_initz_release();
4304extern void __kmp_hidden_helper_threads_deinitz_wait();
4305extern void __kmp_hidden_helper_threads_deinitz_release();
4306extern void __kmp_hidden_helper_main_thread_wait();
4307extern void __kmp_hidden_helper_worker_thread_wait();
4308extern void __kmp_hidden_helper_worker_thread_signal();
4309extern void __kmp_hidden_helper_main_thread_release();
4312#define KMP_HIDDEN_HELPER_THREAD(gtid) \
4313 ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4315#define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4316 ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4318#define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid) \
4319 ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4321#define KMP_HIDDEN_HELPER_TEAM(team) \
4322 (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4326#define KMP_GTID_TO_SHADOW_GTID(gtid) \
4327 ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4332static inline int __kmp_adjust_gtid_for_hidden_helpers(
int gtid) {
4333 int adjusted_gtid = gtid;
4334 if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4335 gtid - __kmp_hidden_helper_threads_num >= 0) {
4336 adjusted_gtid -= __kmp_hidden_helper_threads_num;
4338 return adjusted_gtid;
4342typedef enum kmp_severity_t {
4343 severity_warning = 1,
4346extern void __kmpc_error(
ident_t *loc,
int severity,
const char *message);
4349KMP_EXPORT
void __kmpc_scope(
ident_t *loc, kmp_int32 gtid,
void *reserved);
4350KMP_EXPORT
void __kmpc_end_scope(
ident_t *loc, kmp_int32 gtid,
void *reserved);
4356template <
bool C,
bool S>
4357extern void __kmp_suspend_32(
int th_gtid, kmp_flag_32<C, S> *flag);
4358template <
bool C,
bool S>
4359extern void __kmp_suspend_64(
int th_gtid, kmp_flag_64<C, S> *flag);
4360template <
bool C,
bool S>
4361extern void __kmp_atomic_suspend_64(
int th_gtid,
4362 kmp_atomic_flag_64<C, S> *flag);
4363extern void __kmp_suspend_oncore(
int th_gtid, kmp_flag_oncore *flag);
4364#if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4365template <
bool C,
bool S>
4366extern void __kmp_mwait_32(
int th_gtid, kmp_flag_32<C, S> *flag);
4367template <
bool C,
bool S>
4368extern void __kmp_mwait_64(
int th_gtid, kmp_flag_64<C, S> *flag);
4369template <
bool C,
bool S>
4370extern void __kmp_atomic_mwait_64(
int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4371extern void __kmp_mwait_oncore(
int th_gtid, kmp_flag_oncore *flag);
4373template <
bool C,
bool S>
4374extern void __kmp_resume_32(
int target_gtid, kmp_flag_32<C, S> *flag);
4375template <
bool C,
bool S>
4376extern void __kmp_resume_64(
int target_gtid, kmp_flag_64<C, S> *flag);
4377template <
bool C,
bool S>
4378extern void __kmp_atomic_resume_64(
int target_gtid,
4379 kmp_atomic_flag_64<C, S> *flag);
4380extern void __kmp_resume_oncore(
int target_gtid, kmp_flag_oncore *flag);
4382template <
bool C,
bool S>
4383int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4384 kmp_flag_32<C, S> *flag,
int final_spin,
4385 int *thread_finished,
4389 kmp_int32 is_constrained);
4390template <
bool C,
bool S>
4391int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4392 kmp_flag_64<C, S> *flag,
int final_spin,
4393 int *thread_finished,
4397 kmp_int32 is_constrained);
4398template <
bool C,
bool S>
4399int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4400 kmp_atomic_flag_64<C, S> *flag,
4401 int final_spin,
int *thread_finished,
4405 kmp_int32 is_constrained);
4406int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4407 kmp_flag_oncore *flag,
int final_spin,
4408 int *thread_finished,
4412 kmp_int32 is_constrained);
4414extern int __kmp_nesting_mode;
4415extern int __kmp_nesting_mode_nlevels;
4416extern int *__kmp_nesting_nth_level;
4417extern void __kmp_init_nesting_mode();
4418extern void __kmp_set_nesting_mode_threads();
4430 if (f && f != stdout && f != stderr) {
4439 const char *env_var =
nullptr)
4441 open(filename, mode, env_var);
4448 void open(
const char *filename,
const char *mode,
4449 const char *env_var =
nullptr) {
4451 f = fopen(filename, mode);
4455 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4456 KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4458 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4467 f = fopen(filename, mode);
4484 operator bool() {
return bool(f); }
4485 operator FILE *() {
return f; }
4488template <
typename SourceType,
typename TargetType,
4489 bool isSourceSmaller = (
sizeof(SourceType) <
sizeof(TargetType)),
4490 bool isSourceEqual = (
sizeof(SourceType) ==
sizeof(TargetType)),
4491 bool isSourceSigned = std::is_signed<SourceType>::value,
4492 bool isTargetSigned = std::is_signed<TargetType>::value>
4493struct kmp_convert {};
4496template <
typename SourceType,
typename TargetType>
4497struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4498 static TargetType to(SourceType src) {
return (TargetType)src; }
4501template <
typename SourceType,
typename TargetType>
4502struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4503 static TargetType to(SourceType src) {
return src; }
4506template <
typename SourceType,
typename TargetType>
4507struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4508 static TargetType to(SourceType src) {
4509 KMP_ASSERT(src <=
static_cast<SourceType
>(
4510 (std::numeric_limits<TargetType>::max)()));
4511 KMP_ASSERT(src >=
static_cast<SourceType
>(
4512 (std::numeric_limits<TargetType>::min)()));
4513 return (TargetType)src;
4519template <
typename SourceType,
typename TargetType>
4520struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4521 static TargetType to(SourceType src) {
4522 KMP_ASSERT(src >= 0);
4523 return (TargetType)src;
4527template <
typename SourceType,
typename TargetType>
4528struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4529 static TargetType to(SourceType src) {
4530 KMP_ASSERT(src >= 0);
4531 return (TargetType)src;
4535template <
typename SourceType,
typename TargetType>
4536struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4537 static TargetType to(SourceType src) {
4538 KMP_ASSERT(src >= 0);
4539 KMP_ASSERT(src <=
static_cast<SourceType
>(
4540 (std::numeric_limits<TargetType>::max)()));
4541 return (TargetType)src;
4547template <
typename SourceType,
typename TargetType>
4548struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4549 static TargetType to(SourceType src) {
return (TargetType)src; }
4552template <
typename SourceType,
typename TargetType>
4553struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4554 static TargetType to(SourceType src) {
4555 KMP_ASSERT(src <=
static_cast<SourceType
>(
4556 (std::numeric_limits<TargetType>::max)()));
4557 return (TargetType)src;
4561template <
typename SourceType,
typename TargetType>
4562struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4563 static TargetType to(SourceType src) {
4564 KMP_ASSERT(src <=
static_cast<SourceType
>(
4565 (std::numeric_limits<TargetType>::max)()));
4566 return (TargetType)src;
4572template <
typename SourceType,
typename TargetType>
4573struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4574 static TargetType to(SourceType src) {
return (TargetType)src; }
4577template <
typename SourceType,
typename TargetType>
4578struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4579 static TargetType to(SourceType src) {
return src; }
4582template <
typename SourceType,
typename TargetType>
4583struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4584 static TargetType to(SourceType src) {
4585 KMP_ASSERT(src <=
static_cast<SourceType
>(
4586 (std::numeric_limits<TargetType>::max)()));
4587 return (TargetType)src;
4591template <
typename T1,
typename T2>
4592static inline void __kmp_type_convert(T1 src, T2 *dest) {
4593 *dest = kmp_convert<T1, T2>::to(src);
int try_open(const char *filename, const char *mode)
void open(const char *filename, const char *mode, const char *env_var=nullptr)
@ KMP_IDENT_WORK_SECTIONS
@ KMP_IDENT_ATOMIC_HINT_MASK
@ KMP_IDENT_WORK_DISTRIBUTE
@ KMP_IDENT_ATOMIC_REDUCE
KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *)
KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro microtask,...)
KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs, kmpc_micro microtask, kmp_int32 cond, void *args)
KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs, kmpc_micro microtask,...)
KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid)
void(* kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams_lb, kmp_int32 num_teams_ub, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags)
KMP_EXPORT void __kmpc_end(ident_t *)
KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
KMP_EXPORT kmp_int32 __kmpc_reduce(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void(*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck)
KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_flush(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void(*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck)
KMP_EXPORT void * __kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d)
void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int sched, kmp_uint64 grainsize, void *task_dup)
KMP_EXPORT void * __kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
KMP_EXPORT void * __kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid)
KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask)
KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, int is_ws)
KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins, kmp_task_affinity_info_t *affin_list)
void __kmpc_taskloop_5(ident_t *loc, int gtid, kmp_task_t *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int sched, kmp_uint64 grainsize, int modifier, void *task_dup)
KMP_EXPORT void * __kmpc_task_reduction_init(int gtid, int num_data, void *data)
KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask)
KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
KMP_EXPORT void * __kmpc_taskred_init(int gtid, int num_data, void *data)
KMP_EXPORT void ** __kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid)
void(* kmpc_dtor)(void *)
void *(* kmpc_cctor)(void *, void *)
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)
void *(* kmpc_ctor)(void *)
KMP_EXPORT void * __kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid, void *cpy_data)
void *(* kmpc_ctor_vec)(void *, size_t)
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)
KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data, kmpc_ctor_vec ctor, kmpc_cctor_vec cctor, kmpc_dtor_vec dtor, size_t vector_length)
KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc)
KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *)
KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid)
int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_int32 *p_lb, kmp_int32 *p_ub, kmp_int32 *p_st)
KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid)
void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid, kmp_critical_name *, uint32_t hint)
KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid, kmp_int32 numberOfSections)
void __kmpc_doacross_init(ident_t *loc, int gtid, int num_dims, const struct kmp_dim *dims)
int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_uint32 *p_lb, kmp_uint32 *p_ub, kmp_int32 *p_st)
KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid)
int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_uint64 *p_lb, kmp_uint64 *p_ub, kmp_int64 *p_st)
void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid)
int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_int64 *p_lb, kmp_int64 *p_ub, kmp_int64 *p_st)
void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid, kmp_int32 filter)
void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_int32 lb, kmp_int32 ub, kmp_int32 st, kmp_int32 chunk)
void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_uint32 lb, kmp_uint32 ub, kmp_int32 st, kmp_int32 chunk)
void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_uint64 lb, kmp_uint64 ub, kmp_int64 st, kmp_int64 chunk)
void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid)
void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_int64 lb, kmp_int64 ub, kmp_int64 st, kmp_int64 chunk)
KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
@ kmp_distribute_static_chunked
@ kmp_sch_modifier_monotonic
@ kmp_sch_modifier_nonmonotonic