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 // entries when the stack is empty
36 #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
38 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
40 #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
41 #endif // BUILD_TIED_TASK_STACK
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
72 #include <sys/types.h>
81 #include "kmp_safe_c_api.h"
87 #if KMP_USE_HIER_SCHED
89 #undef KMP_USE_HIER_SCHED
90 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
93 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
95 #ifndef HWLOC_OBJ_NUMANODE
96 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
98 #ifndef HWLOC_OBJ_PACKAGE
99 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
101 #if HWLOC_API_VERSION >= 0x00020000
103 typedef int kmp_hwloc_depth_t;
105 typedef unsigned int kmp_hwloc_depth_t;
109 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
110 #include <xmmintrin.h>
113 #include "kmp_debug.h"
114 #include "kmp_lock.h"
115 #include "kmp_version.h"
117 #include "kmp_debugger.h"
119 #include "kmp_i18n.h"
121 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
123 #include "kmp_wrapper_malloc.h"
126 #if !defined NSIG && defined _NSIG
132 #pragma weak clock_gettime
136 #include "ompt-internal.h"
145 #ifndef USE_FAST_MEMORY
146 #define USE_FAST_MEMORY 3
149 #ifndef KMP_NESTED_HOT_TEAMS
150 #define KMP_NESTED_HOT_TEAMS 0
151 #define USE_NESTED_HOT_ARG(x)
153 #if KMP_NESTED_HOT_TEAMS
154 #define USE_NESTED_HOT_ARG(x) , x
156 #define USE_NESTED_HOT_ARG(x)
161 #ifndef USE_CMP_XCHG_FOR_BGET
162 #define USE_CMP_XCHG_FOR_BGET 1
170 #define KMP_NSEC_PER_SEC 1000000000L
171 #define KMP_USEC_PER_SEC 1000000L
195 KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
196 KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
197 KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
199 KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
200 KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
213 KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
214 KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
215 KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
216 KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
242 typedef union kmp_team kmp_team_t;
243 typedef struct kmp_taskdata kmp_taskdata_t;
244 typedef union kmp_task_team kmp_task_team_t;
245 typedef union kmp_team kmp_team_p;
246 typedef union kmp_info kmp_info_p;
247 typedef union kmp_root kmp_root_p;
257 #define KMP_PACK_64(HIGH_32, LOW_32) \
258 ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
261 #define SKIP_WS(_x) \
263 while (*(_x) == ' ' || *(_x) == '\t') \
266 #define SKIP_DIGITS(_x) \
268 while (*(_x) >= '0' && *(_x) <= '9') \
271 #define SKIP_TOKEN(_x) \
273 while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
274 (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
277 #define SKIP_TO(_x, _c) \
279 while (*(_x) != '\0' && *(_x) != (_c)) \
285 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
286 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
291 enum kmp_state_timer {
301 #ifdef USE_LOAD_BALANCE
302 dynamic_load_balance,
305 dynamic_thread_limit,
311 #ifndef KMP_SCHED_TYPE_DEFINED
312 #define KMP_SCHED_TYPE_DEFINED
313 typedef enum kmp_sched {
316 kmp_sched_static = 1,
317 kmp_sched_dynamic = 2,
318 kmp_sched_guided = 3,
320 kmp_sched_upper_std = 5,
321 kmp_sched_lower_ext = 100,
322 kmp_sched_trapezoidal = 101,
323 #if KMP_STATIC_STEAL_ENABLED
324 kmp_sched_static_steal = 102,
327 kmp_sched_default = kmp_sched_static,
328 kmp_sched_monotonic = 0x80000000
338 kmp_sch_static_chunked = 33,
340 kmp_sch_dynamic_chunked = 35,
342 kmp_sch_runtime = 37,
344 kmp_sch_trapezoidal = 39,
347 kmp_sch_static_greedy = 40,
348 kmp_sch_static_balanced = 41,
350 kmp_sch_guided_iterative_chunked = 42,
351 kmp_sch_guided_analytical_chunked = 43,
353 kmp_sch_static_steal = 44,
356 kmp_sch_static_balanced_chunked = 45,
364 kmp_ord_static_chunked = 65,
366 kmp_ord_dynamic_chunked = 67,
367 kmp_ord_guided_chunked = 68,
368 kmp_ord_runtime = 69,
370 kmp_ord_trapezoidal = 71,
383 kmp_nm_static_chunked =
386 kmp_nm_dynamic_chunked = 163,
388 kmp_nm_runtime = 165,
390 kmp_nm_trapezoidal = 167,
393 kmp_nm_static_greedy = 168,
394 kmp_nm_static_balanced = 169,
396 kmp_nm_guided_iterative_chunked = 170,
397 kmp_nm_guided_analytical_chunked = 171,
398 kmp_nm_static_steal =
401 kmp_nm_ord_static_chunked = 193,
403 kmp_nm_ord_dynamic_chunked = 195,
404 kmp_nm_ord_guided_chunked = 196,
405 kmp_nm_ord_runtime = 197,
407 kmp_nm_ord_trapezoidal = 199,
429 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
432 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
433 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
434 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
435 (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
436 #define SCHEDULE_GET_MODIFIERS(s) \
437 ((enum sched_type)( \
438 (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
439 #define SCHEDULE_SET_MODIFIERS(s, m) \
440 (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
441 #define SCHEDULE_NONMONOTONIC 0
442 #define SCHEDULE_MONOTONIC 1
449 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
451 if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
452 *kind = (kmp_sched_t)((
int)*kind | (int)kmp_sched_monotonic);
458 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
460 if ((
int)kind & (
int)kmp_sched_monotonic) {
461 *internal_kind = (
enum sched_type)((
int)*internal_kind |
467 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
468 return (kmp_sched_t)((int)kind & ~((
int)kmp_sched_monotonic));
472 typedef union kmp_r_sched {
491 enum clock_function_type {
492 clock_function_gettimeofday,
493 clock_function_clock_gettime
497 #if KMP_MIC_SUPPORTED
498 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
503 #undef KMP_FAST_REDUCTION_BARRIER
504 #define KMP_FAST_REDUCTION_BARRIER 1
506 #undef KMP_FAST_REDUCTION_CORE_DUO
507 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
508 #define KMP_FAST_REDUCTION_CORE_DUO 1
511 enum _reduction_method {
512 reduction_method_not_defined = 0,
513 critical_reduce_block = (1 << 8),
514 atomic_reduce_block = (2 << 8),
515 tree_reduce_block = (3 << 8),
516 empty_reduce_block = (4 << 8)
531 #if KMP_FAST_REDUCTION_BARRIER
532 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
533 ((reduction_method) | (barrier_type))
535 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
536 ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
538 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
539 ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
541 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
544 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
545 (packed_reduction_method)
547 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
550 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
551 ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
552 (which_reduction_block))
554 #if KMP_FAST_REDUCTION_BARRIER
555 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
556 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
558 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
559 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
562 typedef int PACKED_REDUCTION_METHOD_T;
569 #pragma warning(push)
570 #pragma warning(disable : 271 310)
584 #if KMP_AFFINITY_SUPPORTED
588 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
589 typedef struct GROUP_AFFINITY {
595 #if KMP_GROUP_AFFINITY
596 extern int __kmp_num_proc_groups;
598 static const int __kmp_num_proc_groups = 1;
600 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
601 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
603 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
604 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
606 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
607 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
609 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE,
const GROUP_AFFINITY *,
611 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
615 extern hwloc_topology_t __kmp_hwloc_topology;
616 extern int __kmp_hwloc_error;
617 extern int __kmp_numa_detected;
618 extern int __kmp_tile_depth;
621 extern size_t __kmp_affin_mask_size;
622 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
623 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
624 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
625 #define KMP_CPU_SET_ITERATE(i, mask) \
626 for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
627 #define KMP_CPU_SET(i, mask) (mask)->set(i)
628 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
629 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
630 #define KMP_CPU_ZERO(mask) (mask)->zero()
631 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
632 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
633 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
634 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
635 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
636 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
637 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
638 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
639 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
640 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
641 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
642 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
643 (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
644 #define KMP_CPU_FREE_ARRAY(arr, n) \
645 __kmp_affinity_dispatch->deallocate_mask_array(arr)
646 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
647 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
648 #define __kmp_get_system_affinity(mask, abort_bool) \
649 (mask)->get_system_affinity(abort_bool)
650 #define __kmp_set_system_affinity(mask, abort_bool) \
651 (mask)->set_system_affinity(abort_bool)
652 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
658 void *
operator new(
size_t n);
659 void operator delete(
void *p);
660 void *
operator new[](
size_t n);
661 void operator delete[](
void *p);
664 virtual void set(
int i) {}
666 virtual bool is_set(
int i)
const {
return false; }
668 virtual void clear(
int i) {}
670 virtual void zero() {}
672 virtual void copy(
const Mask *src) {}
674 virtual void bitwise_and(
const Mask *rhs) {}
676 virtual void bitwise_or(
const Mask *rhs) {}
678 virtual void bitwise_not() {}
681 virtual int begin()
const {
return 0; }
682 virtual int end()
const {
return 0; }
683 virtual int next(
int previous)
const {
return 0; }
685 virtual int set_system_affinity(
bool abort_on_error)
const {
return -1; }
687 virtual int get_system_affinity(
bool abort_on_error) {
return -1; }
690 virtual int get_proc_group()
const {
return -1; }
692 void *
operator new(
size_t n);
693 void operator delete(
void *p);
695 virtual ~KMPAffinity() =
default;
697 virtual void determine_capable(
const char *env_var) {}
699 virtual void bind_thread(
int proc) {}
701 virtual Mask *allocate_mask() {
return nullptr; }
702 virtual void deallocate_mask(Mask *m) {}
703 virtual Mask *allocate_mask_array(
int num) {
return nullptr; }
704 virtual void deallocate_mask_array(Mask *m) {}
705 virtual Mask *index_mask_array(Mask *m,
int index) {
return nullptr; }
706 static void pick_api();
707 static void destroy_api();
715 virtual api_type get_api_type()
const {
721 static bool picked_api;
724 typedef KMPAffinity::Mask kmp_affin_mask_t;
725 extern KMPAffinity *__kmp_affinity_dispatch;
729 #define KMP_AFFIN_MASK_PRINT_LEN 1024
744 affinity_gran_fine = 0,
745 affinity_gran_thread,
749 affinity_gran_package,
751 #if KMP_GROUP_AFFINITY
756 affinity_gran_default
759 enum affinity_top_method {
760 affinity_top_method_all = 0,
761 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
762 affinity_top_method_apicid,
763 affinity_top_method_x2apicid,
765 affinity_top_method_cpuinfo,
766 #if KMP_GROUP_AFFINITY
767 affinity_top_method_group,
769 affinity_top_method_flat,
771 affinity_top_method_hwloc,
773 affinity_top_method_default
776 #define affinity_respect_mask_default (-1)
778 extern enum affinity_type __kmp_affinity_type;
779 extern enum affinity_gran __kmp_affinity_gran;
780 extern int __kmp_affinity_gran_levels;
781 extern int __kmp_affinity_dups;
782 extern enum affinity_top_method __kmp_affinity_top_method;
783 extern int __kmp_affinity_compact;
784 extern int __kmp_affinity_offset;
785 extern int __kmp_affinity_verbose;
786 extern int __kmp_affinity_warnings;
787 extern int __kmp_affinity_respect_mask;
788 extern char *__kmp_affinity_proclist;
789 extern kmp_affin_mask_t *__kmp_affinity_masks;
790 extern unsigned __kmp_affinity_num_masks;
791 extern void __kmp_affinity_bind_thread(
int which);
793 extern kmp_affin_mask_t *__kmp_affin_fullMask;
794 extern char *__kmp_cpuinfo_file;
799 typedef enum kmp_proc_bind_t {
809 typedef struct kmp_nested_proc_bind_t {
810 kmp_proc_bind_t *bind_types;
813 } kmp_nested_proc_bind_t;
815 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
817 extern int __kmp_display_affinity;
818 extern char *__kmp_affinity_format;
819 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
821 #if KMP_AFFINITY_SUPPORTED
822 #define KMP_PLACE_ALL (-1)
823 #define KMP_PLACE_UNDEFINED (-2)
825 #define KMP_AFFINITY_NON_PROC_BIND \
826 ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
827 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
828 (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
831 extern int __kmp_affinity_num_places;
833 typedef enum kmp_cancel_kind_t {
842 typedef struct kmp_hws_item {
847 extern kmp_hws_item_t __kmp_hws_socket;
848 extern kmp_hws_item_t __kmp_hws_node;
849 extern kmp_hws_item_t __kmp_hws_tile;
850 extern kmp_hws_item_t __kmp_hws_core;
851 extern kmp_hws_item_t __kmp_hws_proc;
852 extern int __kmp_hws_requested;
853 extern int __kmp_hws_abs_flag;
857 #define KMP_PAD(type, sz) \
858 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
862 #define KMP_GTID_DNE (-2)
863 #define KMP_GTID_SHUTDOWN (-3)
864 #define KMP_GTID_MONITOR (-4)
865 #define KMP_GTID_UNKNOWN (-5)
866 #define KMP_GTID_MIN (-6)
872 typedef uintptr_t omp_uintptr_t;
875 omp_atk_threadmodel = 1,
876 omp_atk_alignment = 2,
878 omp_atk_pool_size = 4,
879 omp_atk_fallback = 5,
882 omp_atk_partition = 8
883 } omp_alloctrait_key_t;
889 omp_atv_contended = 3,
890 omp_atv_uncontended = 4,
891 omp_atv_sequential = 5,
897 omp_atv_default_mem_fb = 11,
898 omp_atv_null_fb = 12,
899 omp_atv_abort_fb = 13,
900 omp_atv_allocator_fb = 14,
901 omp_atv_environment = 15,
902 omp_atv_nearest = 16,
903 omp_atv_blocked = 17,
904 omp_atv_interleaved = 18
905 } omp_alloctrait_value_t;
907 typedef void *omp_memspace_handle_t;
908 extern omp_memspace_handle_t
const omp_default_mem_space;
909 extern omp_memspace_handle_t
const omp_large_cap_mem_space;
910 extern omp_memspace_handle_t
const omp_const_mem_space;
911 extern omp_memspace_handle_t
const omp_high_bw_mem_space;
912 extern omp_memspace_handle_t
const omp_low_lat_mem_space;
915 omp_alloctrait_key_t key;
919 typedef void *omp_allocator_handle_t;
920 extern omp_allocator_handle_t
const omp_null_allocator;
921 extern omp_allocator_handle_t
const omp_default_mem_alloc;
922 extern omp_allocator_handle_t
const omp_large_cap_mem_alloc;
923 extern omp_allocator_handle_t
const omp_const_mem_alloc;
924 extern omp_allocator_handle_t
const omp_high_bw_mem_alloc;
925 extern omp_allocator_handle_t
const omp_low_lat_mem_alloc;
926 extern omp_allocator_handle_t
const omp_cgroup_mem_alloc;
927 extern omp_allocator_handle_t
const omp_pteam_mem_alloc;
928 extern omp_allocator_handle_t
const omp_thread_mem_alloc;
929 extern omp_allocator_handle_t
const kmp_max_mem_alloc;
930 extern omp_allocator_handle_t __kmp_def_allocator;
935 extern int __kmp_memkind_available;
937 typedef omp_memspace_handle_t kmp_memspace_t;
939 typedef struct kmp_allocator_t {
940 omp_memspace_handle_t memspace;
943 omp_alloctrait_value_t fb;
944 kmp_allocator_t *fb_data;
945 kmp_uint64 pool_size;
946 kmp_uint64 pool_used;
949 extern omp_allocator_handle_t __kmpc_init_allocator(
int gtid,
950 omp_memspace_handle_t,
952 omp_alloctrait_t traits[]);
953 extern void __kmpc_destroy_allocator(
int gtid, omp_allocator_handle_t al);
954 extern void __kmpc_set_default_allocator(
int gtid, omp_allocator_handle_t al);
955 extern omp_allocator_handle_t __kmpc_get_default_allocator(
int gtid);
956 extern void *__kmpc_alloc(
int gtid,
size_t sz, omp_allocator_handle_t al);
957 extern void __kmpc_free(
int gtid,
void *ptr, omp_allocator_handle_t al);
959 extern void __kmp_init_memkind();
960 extern void __kmp_fini_memkind();
964 #define KMP_UINT64_MAX \
965 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
967 #define KMP_MIN_NTH 1
970 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
971 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
973 #define KMP_MAX_NTH INT_MAX
977 #ifdef PTHREAD_STACK_MIN
978 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
980 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
983 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
986 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
987 #elif KMP_ARCH_X86_64
988 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
989 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
991 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
994 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
995 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
996 #define KMP_MAX_MALLOC_POOL_INCR \
997 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
999 #define KMP_MIN_STKOFFSET (0)
1000 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1002 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1004 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1007 #define KMP_MIN_STKPADDING (0)
1008 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1010 #define KMP_BLOCKTIME_MULTIPLIER \
1012 #define KMP_MIN_BLOCKTIME (0)
1013 #define KMP_MAX_BLOCKTIME \
1015 #define KMP_DEFAULT_BLOCKTIME (200)
1018 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1019 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1020 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1024 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1025 (((blocktime) == KMP_MAX_BLOCKTIME) \
1026 ? (monitor_wakeups) \
1027 : ((blocktime) == KMP_MIN_BLOCKTIME) \
1028 ? KMP_MAX_MONITOR_WAKEUPS \
1029 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1030 ? (monitor_wakeups) \
1031 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1035 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1036 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1037 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1039 #define KMP_BLOCKTIME(team, tid) \
1040 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1041 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1043 extern kmp_uint64 __kmp_ticks_per_msec;
1044 #if KMP_COMPILER_ICC
1045 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1047 #define KMP_NOW() __kmp_hardware_timestamp()
1049 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1050 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1051 (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1052 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1055 extern kmp_uint64 __kmp_now_nsec();
1056 #define KMP_NOW() __kmp_now_nsec()
1057 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1058 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1059 (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1060 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1062 #endif // KMP_USE_MONITOR
1064 #define KMP_MIN_STATSCOLS 40
1065 #define KMP_MAX_STATSCOLS 4096
1066 #define KMP_DEFAULT_STATSCOLS 80
1068 #define KMP_MIN_INTERVAL 0
1069 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1070 #define KMP_DEFAULT_INTERVAL 0
1072 #define KMP_MIN_CHUNK 1
1073 #define KMP_MAX_CHUNK (INT_MAX - 1)
1074 #define KMP_DEFAULT_CHUNK 1
1076 #define KMP_DFLT_DISP_NUM_BUFF 7
1077 #define KMP_MAX_ORDERED 8
1079 #define KMP_MAX_FIELDS 32
1081 #define KMP_MAX_BRANCH_BITS 31
1083 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1085 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1087 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1092 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1093 #define KMP_TLS_GTID_MIN 5
1095 #define KMP_TLS_GTID_MIN INT_MAX
1098 #define KMP_MASTER_TID(tid) ((tid) == 0)
1099 #define KMP_WORKER_TID(tid) ((tid) != 0)
1101 #define KMP_MASTER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) == 0)
1102 #define KMP_WORKER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) != 0)
1103 #define KMP_INITIAL_GTID(gtid) ((gtid) == 0)
1107 #define TRUE (!FALSE)
1113 #define KMP_INIT_WAIT 64U
1114 #define KMP_NEXT_WAIT 32U
1116 #define KMP_INIT_WAIT 16U
1117 #define KMP_NEXT_WAIT 8U
1119 #define KMP_INIT_WAIT 1024U
1120 #define KMP_NEXT_WAIT 512U
1123 #define KMP_INIT_WAIT 1024U
1124 #define KMP_NEXT_WAIT 512U
1125 #elif KMP_OS_DRAGONFLY
1127 #define KMP_INIT_WAIT 1024U
1128 #define KMP_NEXT_WAIT 512U
1129 #elif KMP_OS_FREEBSD
1131 #define KMP_INIT_WAIT 1024U
1132 #define KMP_NEXT_WAIT 512U
1135 #define KMP_INIT_WAIT 1024U
1136 #define KMP_NEXT_WAIT 512U
1139 #define KMP_INIT_WAIT 1024U
1140 #define KMP_NEXT_WAIT 512U
1141 #elif KMP_OS_OPENBSD
1143 #define KMP_INIT_WAIT 1024U
1144 #define KMP_NEXT_WAIT 512U
1147 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1148 typedef struct kmp_cpuid {
1155 typedef struct kmp_cpuinfo {
1164 int cpu_stackoffset;
1168 kmp_uint64 frequency;
1169 char name[3 *
sizeof(kmp_cpuid_t)];
1172 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1177 static inline void __kmp_x86_cpuid(
int leaf,
int subleaf,
struct kmp_cpuid *p) {
1178 __asm__ __volatile__(
"cpuid"
1179 :
"=a"(p->eax),
"=b"(p->ebx),
"=c"(p->ecx),
"=d"(p->edx)
1180 :
"a"(leaf),
"c"(subleaf));
1183 static inline void __kmp_load_x87_fpu_control_word(
const kmp_int16 *p) {
1184 __asm__ __volatile__(
"fldcw %0" : :
"m"(*p));
1187 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1188 __asm__ __volatile__(
"fstcw %0" :
"=m"(*p));
1190 static inline void __kmp_clear_x87_fpu_status_word() {
1193 struct x87_fpu_state {
1202 struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1203 __asm__ __volatile__(
"fstenv %0\n\t"
1204 "andw $0x7f00, %1\n\t"
1206 :
"+m"(fpu_state),
"+m"(fpu_state.sw));
1208 __asm__ __volatile__(
"fnclex");
1212 static inline void __kmp_load_mxcsr(
const kmp_uint32 *p) { _mm_setcsr(*p); }
1213 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1215 static inline void __kmp_load_mxcsr(
const kmp_uint32 *p) {}
1216 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1220 extern void __kmp_x86_cpuid(
int mode,
int mode2,
struct kmp_cpuid *p);
1221 extern void __kmp_load_x87_fpu_control_word(
const kmp_int16 *p);
1222 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1223 extern void __kmp_clear_x87_fpu_status_word();
1224 static inline void __kmp_load_mxcsr(
const kmp_uint32 *p) { _mm_setcsr(*p); }
1225 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1226 #endif // KMP_OS_UNIX
1228 #define KMP_X86_MXCSR_MASK 0xffffffc0
1231 extern void __kmp_x86_pause(
void);
1237 static inline void __kmp_x86_pause(
void) { _mm_delay_32(300); }
1239 static inline void __kmp_x86_pause(
void) { _mm_pause(); }
1241 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1242 #elif KMP_ARCH_PPC64
1243 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1244 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1245 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1246 #define KMP_CPU_PAUSE() \
1248 KMP_PPC64_PRI_LOW(); \
1249 KMP_PPC64_PRI_MED(); \
1250 KMP_PPC64_PRI_LOC_MB(); \
1253 #define KMP_CPU_PAUSE()
1256 #define KMP_INIT_YIELD(count) \
1257 { (count) = __kmp_yield_init; }
1259 #define KMP_OVERSUBSCRIBED \
1260 (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1262 #define KMP_TRY_YIELD \
1263 ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1265 #define KMP_TRY_YIELD_OVERSUB \
1266 ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1268 #define KMP_YIELD(cond) \
1271 if ((cond) && (KMP_TRY_YIELD)) \
1275 #define KMP_YIELD_OVERSUB() \
1278 if ((KMP_TRY_YIELD_OVERSUB)) \
1284 #define KMP_YIELD_SPIN(count) \
1287 if (KMP_TRY_YIELD) { \
1291 (count) = __kmp_yield_next; \
1296 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count) \
1299 if ((KMP_TRY_YIELD_OVERSUB)) \
1301 else if (__kmp_use_yield == 1) { \
1305 (count) = __kmp_yield_next; \
1322 ct_ordered_in_parallel,
1329 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1333 enum cons_type type;
1339 struct cons_header {
1340 int p_top, w_top, s_top;
1341 int stack_size, stack_top;
1342 struct cons_data *stack_data;
1345 struct kmp_region_info {
1347 int offset[KMP_MAX_FIELDS];
1348 int length[KMP_MAX_FIELDS];
1355 typedef HANDLE kmp_thread_t;
1356 typedef DWORD kmp_key_t;
1360 typedef pthread_t kmp_thread_t;
1361 typedef pthread_key_t kmp_key_t;
1364 extern kmp_key_t __kmp_gtid_threadprivate_key;
1366 typedef struct kmp_sys_info {
1380 typedef int kmp_itt_mark_t;
1381 #define KMP_ITT_DEBUG 0
1384 typedef kmp_int32 kmp_critical_name[8];
1395 typedef void (*
kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1396 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1413 typedef void *(*kmpc_ctor)(
void *);
1426 typedef void *(*kmpc_cctor)(
void *,
void *);
1436 typedef void *(*kmpc_ctor_vec)(
void *, size_t);
1448 typedef void *(*kmpc_cctor_vec)(
void *,
void *,
1456 typedef struct kmp_cached_addr {
1458 void ***compiler_cache;
1460 struct kmp_cached_addr *next;
1461 } kmp_cached_addr_t;
1463 struct private_data {
1464 struct private_data *next;
1470 struct private_common {
1471 struct private_common *next;
1472 struct private_common *link;
1478 struct shared_common {
1479 struct shared_common *next;
1480 struct private_data *pod_init;
1500 #define KMP_HASH_TABLE_LOG2 9
1501 #define KMP_HASH_TABLE_SIZE \
1502 (1 << KMP_HASH_TABLE_LOG2)
1503 #define KMP_HASH_SHIFT 3
1504 #define KMP_HASH(x) \
1505 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1507 struct common_table {
1508 struct private_common *data[KMP_HASH_TABLE_SIZE];
1511 struct shared_table {
1512 struct shared_common *data[KMP_HASH_TABLE_SIZE];
1517 #if KMP_USE_HIER_SCHED
1520 typedef struct kmp_hier_private_bdata_t {
1521 kmp_int32 num_active;
1523 kmp_uint64 wait_val[2];
1524 } kmp_hier_private_bdata_t;
1527 typedef struct kmp_sched_flags {
1528 unsigned ordered : 1;
1529 unsigned nomerge : 1;
1530 unsigned contains_last : 1;
1531 #if KMP_USE_HIER_SCHED
1532 unsigned use_hier : 1;
1533 unsigned unused : 28;
1535 unsigned unused : 29;
1537 } kmp_sched_flags_t;
1539 KMP_BUILD_ASSERT(
sizeof(kmp_sched_flags_t) == 4);
1541 #if KMP_STATIC_STEAL_ENABLED
1542 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1549 kmp_int32 static_steal_counter;
1551 kmp_lock_t *th_steal_lock;
1558 struct KMP_ALIGN(32) {
1565 kmp_uint32 ordered_lower;
1566 kmp_uint32 ordered_upper;
1571 kmp_int32 last_upper;
1573 } dispatch_private_info32_t;
1575 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1582 kmp_int64 static_steal_counter;
1584 kmp_lock_t *th_steal_lock;
1593 struct KMP_ALIGN(32) {
1600 kmp_uint64 ordered_lower;
1601 kmp_uint64 ordered_upper;
1606 kmp_int64 last_upper;
1608 } dispatch_private_info64_t;
1610 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1623 kmp_uint32 ordered_lower;
1624 kmp_uint32 ordered_upper;
1626 kmp_int32 last_upper;
1628 } dispatch_private_info32_t;
1630 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1644 kmp_uint64 ordered_lower;
1645 kmp_uint64 ordered_upper;
1647 kmp_int64 last_upper;
1649 } dispatch_private_info64_t;
1652 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1653 union private_info {
1654 dispatch_private_info32_t p32;
1655 dispatch_private_info64_t p64;
1658 kmp_sched_flags_t flags;
1659 kmp_int32 ordered_bumped;
1661 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 3];
1663 struct dispatch_private_info *next;
1664 kmp_int32 type_size;
1665 #if KMP_USE_HIER_SCHED
1669 enum cons_type pushed_ws;
1670 } dispatch_private_info_t;
1672 typedef struct dispatch_shared_info32 {
1675 volatile kmp_uint32 iteration;
1676 volatile kmp_uint32 num_done;
1677 volatile kmp_uint32 ordered_iteration;
1679 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1680 } dispatch_shared_info32_t;
1682 typedef struct dispatch_shared_info64 {
1685 volatile kmp_uint64 iteration;
1686 volatile kmp_uint64 num_done;
1687 volatile kmp_uint64 ordered_iteration;
1689 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1690 } dispatch_shared_info64_t;
1692 typedef struct dispatch_shared_info {
1694 dispatch_shared_info32_t s32;
1695 dispatch_shared_info64_t s64;
1697 volatile kmp_uint32 buffer_index;
1698 volatile kmp_int32 doacross_buf_idx;
1699 volatile kmp_uint32 *doacross_flags;
1700 kmp_int32 doacross_num_done;
1701 #if KMP_USE_HIER_SCHED
1710 } dispatch_shared_info_t;
1712 typedef struct kmp_disp {
1714 void (*th_deo_fcn)(
int *gtid,
int *cid,
ident_t *);
1716 void (*th_dxo_fcn)(
int *gtid,
int *cid,
ident_t *);
1718 dispatch_shared_info_t *th_dispatch_sh_current;
1719 dispatch_private_info_t *th_dispatch_pr_current;
1721 dispatch_private_info_t *th_disp_buffer;
1722 kmp_int32 th_disp_index;
1723 kmp_int32 th_doacross_buf_idx;
1724 volatile kmp_uint32 *th_doacross_flags;
1725 kmp_int64 *th_doacross_info;
1726 #if KMP_USE_INTERNODE_ALIGNMENT
1727 char more_padding[INTERNODE_CACHE_LINE];
1735 #define KMP_INIT_BARRIER_STATE 0
1736 #define KMP_BARRIER_SLEEP_BIT 0
1737 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1738 #define KMP_BARRIER_BUMP_BIT 2
1740 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1741 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1742 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1744 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1745 #error "Barrier sleep bit must be smaller than barrier bump bit"
1747 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1748 #error "Barrier unused bit must be smaller than barrier bump bit"
1752 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1753 #define KMP_BARRIER_OWN_FLAG \
1754 1 // Normal state; worker waiting on own b_go flag in release
1755 #define KMP_BARRIER_PARENT_FLAG \
1756 2 // Special state; worker waiting on parent's b_go flag in release
1757 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1758 3 // Special state; tells worker to shift from parent to own b_go
1759 #define KMP_BARRIER_SWITCHING \
1760 4 // Special state; worker resets appropriate flag on wake-up
1762 #define KMP_NOT_SAFE_TO_REAP \
1763 0 // Thread th_reap_state: not safe to reap (tasking)
1764 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1767 bs_plain_barrier = 0,
1769 bs_forkjoin_barrier,
1770 #if KMP_FAST_REDUCTION_BARRIER
1771 bs_reduction_barrier,
1772 #endif // KMP_FAST_REDUCTION_BARRIER
1777 #if !KMP_FAST_REDUCTION_BARRIER
1778 #define bs_reduction_barrier bs_plain_barrier
1779 #endif // KMP_FAST_REDUCTION_BARRIER
1781 typedef enum kmp_bar_pat {
1789 bp_hierarchical_bar = 3,
1793 #define KMP_BARRIER_ICV_PUSH 1
1796 typedef struct kmp_internal_control {
1797 int serial_nesting_level;
1810 int max_active_levels;
1813 kmp_proc_bind_t proc_bind;
1814 kmp_int32 default_device;
1815 struct kmp_internal_control *next;
1816 } kmp_internal_control_t;
1818 static inline void copy_icvs(kmp_internal_control_t *dst,
1819 kmp_internal_control_t *src) {
1824 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1829 kmp_internal_control_t th_fixed_icvs;
1832 volatile kmp_uint64 b_go;
1833 KMP_ALIGN_CACHE
volatile kmp_uint64
1835 kmp_uint32 *skip_per_level;
1836 kmp_uint32 my_level;
1837 kmp_int32 parent_tid;
1840 struct kmp_bstate *parent_bar;
1842 kmp_uint64 leaf_state;
1844 kmp_uint8 base_leaf_kids;
1845 kmp_uint8 leaf_kids;
1847 kmp_uint8 wait_flag;
1848 kmp_uint8 use_oncore_barrier;
1853 KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1857 union KMP_ALIGN_CACHE kmp_barrier_union {
1859 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1863 typedef union kmp_barrier_union kmp_balign_t;
1866 union KMP_ALIGN_CACHE kmp_barrier_team_union {
1868 char b_pad[CACHE_LINE];
1870 kmp_uint64 b_arrived;
1876 kmp_uint b_master_arrived;
1877 kmp_uint b_team_arrived;
1882 typedef union kmp_barrier_team_union kmp_balign_team_t;
1889 typedef struct kmp_win32_mutex {
1891 CRITICAL_SECTION cs;
1892 } kmp_win32_mutex_t;
1894 typedef struct kmp_win32_cond {
1899 kmp_win32_mutex_t waiters_count_lock_;
1906 int wait_generation_count_;
1915 union KMP_ALIGN_CACHE kmp_cond_union {
1917 char c_pad[CACHE_LINE];
1918 pthread_cond_t c_cond;
1921 typedef union kmp_cond_union kmp_cond_align_t;
1923 union KMP_ALIGN_CACHE kmp_mutex_union {
1925 char m_pad[CACHE_LINE];
1926 pthread_mutex_t m_mutex;
1929 typedef union kmp_mutex_union kmp_mutex_align_t;
1933 typedef struct kmp_desc_base {
1935 size_t ds_stacksize;
1937 kmp_thread_t ds_thread;
1938 volatile int ds_tid;
1941 volatile int ds_alive;
1958 typedef union KMP_ALIGN_CACHE kmp_desc {
1960 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
1964 typedef struct kmp_local {
1965 volatile int this_construct;
1970 #if !USE_CMP_XCHG_FOR_BGET
1971 #ifdef USE_QUEUING_LOCK_FOR_BGET
1972 kmp_lock_t bget_lock;
1974 kmp_bootstrap_lock_t bget_lock;
1981 PACKED_REDUCTION_METHOD_T
1982 packed_reduction_method;
1987 #define KMP_CHECK_UPDATE(a, b) \
1990 #define KMP_CHECK_UPDATE_SYNC(a, b) \
1992 TCW_SYNC_PTR((a), (b))
1994 #define get__blocktime(xteam, xtid) \
1995 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
1996 #define get__bt_set(xteam, xtid) \
1997 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
1999 #define get__bt_intervals(xteam, xtid) \
2000 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2003 #define get__dynamic_2(xteam, xtid) \
2004 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2005 #define get__nproc_2(xteam, xtid) \
2006 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2007 #define get__sched_2(xteam, xtid) \
2008 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2010 #define set__blocktime_team(xteam, xtid, xval) \
2011 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2015 #define set__bt_intervals_team(xteam, xtid, xval) \
2016 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2020 #define set__bt_set_team(xteam, xtid, xval) \
2021 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2023 #define set__dynamic(xthread, xval) \
2024 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2025 #define get__dynamic(xthread) \
2026 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2028 #define set__nproc(xthread, xval) \
2029 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2031 #define set__thread_limit(xthread, xval) \
2032 (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2034 #define set__max_active_levels(xthread, xval) \
2035 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2037 #define get__max_active_levels(xthread) \
2038 ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2040 #define set__sched(xthread, xval) \
2041 (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2043 #define set__proc_bind(xthread, xval) \
2044 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2045 #define get__proc_bind(xthread) \
2046 ((xthread)->th.th_current_task->td_icvs.proc_bind)
2050 typedef enum kmp_tasking_mode {
2051 tskm_immediate_exec = 0,
2052 tskm_extra_barrier = 1,
2053 tskm_task_teams = 2,
2055 } kmp_tasking_mode_t;
2057 extern kmp_tasking_mode_t
2059 extern int __kmp_task_stealing_constraint;
2060 extern int __kmp_enable_task_throttling;
2061 extern kmp_int32 __kmp_default_device;
2064 extern kmp_int32 __kmp_max_task_priority;
2066 extern kmp_uint64 __kmp_taskloop_min_tasks;
2070 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2071 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2075 #define KMP_TASKING_ENABLED(task_team) \
2076 (TCR_SYNC_4((task_team)->tt.tt_found_tasks) == TRUE)
2084 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32,
void *);
2086 typedef union kmp_cmplrdata {
2097 typedef struct kmp_task {
2104 kmp_cmplrdata_t data2;
2113 typedef struct kmp_taskgroup {
2114 std::atomic<kmp_int32> count;
2115 std::atomic<kmp_int32>
2117 struct kmp_taskgroup *parent;
2120 kmp_int32 reduce_num_data;
2124 typedef union kmp_depnode kmp_depnode_t;
2125 typedef struct kmp_depnode_list kmp_depnode_list_t;
2126 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2129 typedef struct kmp_depend_info {
2130 kmp_intptr_t base_addr;
2137 } kmp_depend_info_t;
2140 struct kmp_depnode_list {
2141 kmp_depnode_t *node;
2142 kmp_depnode_list_t *next;
2146 #define MAX_MTX_DEPS 4
2148 typedef struct kmp_base_depnode {
2149 kmp_depnode_list_t *successors;
2151 kmp_lock_t *mtx_locks[MAX_MTX_DEPS];
2152 kmp_int32 mtx_num_locks;
2154 #if KMP_SUPPORT_GRAPH_OUTPUT
2157 std::atomic<kmp_int32> npredecessors;
2158 std::atomic<kmp_int32> nrefs;
2159 } kmp_base_depnode_t;
2161 union KMP_ALIGN_CACHE kmp_depnode {
2163 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2164 kmp_base_depnode_t dn;
2167 struct kmp_dephash_entry {
2169 kmp_depnode_t *last_out;
2170 kmp_depnode_list_t *last_ins;
2171 kmp_depnode_list_t *last_mtxs;
2172 kmp_int32 last_flag;
2173 kmp_lock_t *mtx_lock;
2174 kmp_dephash_entry_t *next_in_bucket;
2177 typedef struct kmp_dephash {
2178 kmp_dephash_entry_t **buckets;
2181 kmp_uint32 nelements;
2182 kmp_uint32 nconflicts;
2185 typedef struct kmp_task_affinity_info {
2186 kmp_intptr_t base_addr;
2191 kmp_int32 reserved : 30;
2193 } kmp_task_affinity_info_t;
2195 typedef enum kmp_event_type_t {
2196 KMP_EVENT_UNINITIALIZED = 0,
2197 KMP_EVENT_ALLOW_COMPLETION = 1
2201 kmp_event_type_t type;
2202 kmp_tas_lock_t lock;
2208 #ifdef BUILD_TIED_TASK_STACK
2211 typedef struct kmp_stack_block {
2212 kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2213 struct kmp_stack_block *sb_next;
2214 struct kmp_stack_block *sb_prev;
2215 } kmp_stack_block_t;
2217 typedef struct kmp_task_stack {
2218 kmp_stack_block_t ts_first_block;
2219 kmp_taskdata_t **ts_top;
2220 kmp_int32 ts_entries;
2223 #endif // BUILD_TIED_TASK_STACK
2225 typedef struct kmp_tasking_flags {
2227 unsigned tiedness : 1;
2229 unsigned merged_if0 : 1;
2231 unsigned destructors_thunk : 1;
2235 unsigned priority_specified : 1;
2237 unsigned detachable : 1;
2238 unsigned reserved : 9;
2241 unsigned tasktype : 1;
2242 unsigned task_serial : 1;
2243 unsigned tasking_ser : 1;
2245 unsigned team_serial : 1;
2249 unsigned started : 1;
2250 unsigned executing : 1;
2251 unsigned complete : 1;
2253 unsigned native : 1;
2254 unsigned reserved31 : 7;
2256 } kmp_tasking_flags_t;
2258 struct kmp_taskdata {
2259 kmp_int32 td_task_id;
2260 kmp_tasking_flags_t td_flags;
2261 kmp_team_t *td_team;
2262 kmp_info_p *td_alloc_thread;
2264 kmp_taskdata_t *td_parent;
2266 std::atomic<kmp_int32> td_untied_count;
2270 kmp_uint32 td_taskwait_counter;
2271 kmp_int32 td_taskwait_thread;
2272 KMP_ALIGN_CACHE kmp_internal_control_t
2274 KMP_ALIGN_CACHE std::atomic<kmp_int32>
2275 td_allocated_child_tasks;
2277 std::atomic<kmp_int32>
2278 td_incomplete_child_tasks;
2285 kmp_task_team_t *td_task_team;
2286 kmp_int32 td_size_alloc;
2287 #if defined(KMP_GOMP_COMPAT)
2289 kmp_int32 td_size_loop_bounds;
2291 kmp_taskdata_t *td_last_tied;
2292 #if defined(KMP_GOMP_COMPAT)
2294 void (*td_copy_func)(
void *,
void *);
2296 kmp_event_t td_allow_completion_event;
2298 ompt_task_info_t ompt_task_info;
2303 KMP_BUILD_ASSERT(
sizeof(kmp_taskdata_t) %
sizeof(
void *) == 0);
2306 typedef struct kmp_base_thread_data {
2310 kmp_bootstrap_lock_t td_deque_lock;
2313 kmp_int32 td_deque_size;
2314 kmp_uint32 td_deque_head;
2315 kmp_uint32 td_deque_tail;
2316 kmp_int32 td_deque_ntasks;
2318 kmp_int32 td_deque_last_stolen;
2319 #ifdef BUILD_TIED_TASK_STACK
2320 kmp_task_stack_t td_susp_tied_tasks;
2322 #endif // BUILD_TIED_TASK_STACK
2323 } kmp_base_thread_data_t;
2325 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2326 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2328 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2329 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2331 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2332 kmp_base_thread_data_t td;
2334 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2335 } kmp_thread_data_t;
2338 typedef struct kmp_base_task_team {
2339 kmp_bootstrap_lock_t
2342 kmp_task_team_t *tt_next;
2346 kmp_int32 tt_found_tasks;
2350 kmp_int32 tt_max_threads;
2351 kmp_int32 tt_found_proxy_tasks;
2352 kmp_int32 tt_untied_task_encountered;
2355 std::atomic<kmp_int32> tt_unfinished_threads;
2360 } kmp_base_task_team_t;
2362 union KMP_ALIGN_CACHE kmp_task_team {
2363 kmp_base_task_team_t tt;
2365 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2368 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2371 typedef struct kmp_free_list {
2372 void *th_free_list_self;
2373 void *th_free_list_sync;
2375 void *th_free_list_other;
2379 #if KMP_NESTED_HOT_TEAMS
2382 typedef struct kmp_hot_team_ptr {
2383 kmp_team_p *hot_team;
2384 kmp_int32 hot_team_nth;
2385 } kmp_hot_team_ptr_t;
2387 typedef struct kmp_teams_size {
2403 typedef struct kmp_cg_root {
2404 kmp_info_p *cg_root;
2407 kmp_int32 cg_thread_limit;
2408 kmp_int32 cg_nthreads;
2409 struct kmp_cg_root *up;
2414 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2419 kmp_team_p *th_team;
2420 kmp_root_p *th_root;
2421 kmp_info_p *th_next_pool;
2422 kmp_disp_t *th_dispatch;
2428 kmp_info_p *th_team_master;
2429 int th_team_serialized;
2430 microtask_t th_teams_microtask;
2439 int th_team_bt_intervals;
2442 kmp_uint64 th_team_bt_intervals;
2445 #if KMP_AFFINITY_SUPPORTED
2446 kmp_affin_mask_t *th_affin_mask;
2448 omp_allocator_handle_t th_def_allocator;
2452 #if KMP_NESTED_HOT_TEAMS
2453 kmp_hot_team_ptr_t *th_hot_teams;
2459 #if KMP_AFFINITY_SUPPORTED
2460 int th_current_place;
2466 int th_prev_num_threads;
2468 kmp_uint64 th_bar_arrive_time;
2469 kmp_uint64 th_bar_min_time;
2470 kmp_uint64 th_frame_time;
2472 kmp_local_t th_local;
2473 struct private_common *th_pri_head;
2478 KMP_ALIGN_CACHE kmp_team_p
2482 ompt_thread_info_t ompt_thread_info;
2486 struct common_table *th_pri_common;
2488 volatile kmp_uint32 th_spin_here;
2491 volatile void *th_sleep_loc;
2498 kmp_task_team_t *th_task_team;
2499 kmp_taskdata_t *th_current_task;
2500 kmp_uint8 th_task_state;
2501 kmp_uint8 *th_task_state_memo_stack;
2503 kmp_uint32 th_task_state_top;
2504 kmp_uint32 th_task_state_stack_sz;
2505 kmp_uint32 th_reap_state;
2510 kmp_uint8 th_active_in_pool;
2512 struct cons_header *th_cons;
2513 #if KMP_USE_HIER_SCHED
2515 kmp_hier_private_bdata_t *th_hier_bar_data;
2519 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2521 KMP_ALIGN_CACHE
volatile kmp_int32
2524 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2526 kmp_free_list_t th_free_lists[NUM_LISTS];
2531 kmp_win32_cond_t th_suspend_cv;
2532 kmp_win32_mutex_t th_suspend_mx;
2533 std::atomic<int> th_suspend_init;
2536 kmp_cond_align_t th_suspend_cv;
2537 kmp_mutex_align_t th_suspend_mx;
2538 std::atomic<int> th_suspend_init_count;
2542 kmp_itt_mark_t th_itt_mark_single;
2545 #if KMP_STATS_ENABLED
2546 kmp_stats_list *th_stats;
2549 std::atomic<bool> th_blocking;
2551 kmp_cg_root_t *th_cg_roots;
2554 typedef union KMP_ALIGN_CACHE kmp_info {
2556 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2562 typedef struct kmp_base_data {
volatile kmp_uint32 t_value; } kmp_base_data_t;
2564 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2566 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2570 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2572 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2574 } kmp_ordered_team_t;
2576 typedef int (*launch_t)(
int gtid);
2579 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2585 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2586 #define KMP_INLINE_ARGV_BYTES \
2588 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2589 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2592 #define KMP_INLINE_ARGV_BYTES \
2593 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2595 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2597 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2600 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2601 kmp_balign_team_t t_bar[bs_last_barrier];
2602 std::atomic<int> t_construct;
2603 char pad[
sizeof(kmp_lock_t)];
2606 std::atomic<void *> t_tg_reduce_data[2];
2607 std::atomic<int> t_tg_fini_counter[2];
2611 KMP_ALIGN_CACHE
int t_master_tid;
2612 int t_master_this_cons;
2616 kmp_team_p *t_parent;
2617 kmp_team_p *t_next_pool;
2618 kmp_disp_t *t_dispatch;
2619 kmp_task_team_t *t_task_team[2];
2620 kmp_proc_bind_t t_proc_bind;
2622 kmp_uint64 t_region_time;
2627 KMP_ALIGN_CACHE
void **t_argv;
2634 ompt_team_info_t ompt_team_info;
2635 ompt_lw_taskteam_t *ompt_serialized_team_info;
2638 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2639 kmp_int8 t_fp_control_saved;
2641 kmp_int16 t_x87_fpu_control_word;
2645 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2647 KMP_ALIGN_CACHE kmp_info_t **t_threads;
2649 *t_implicit_task_taskdata;
2652 KMP_ALIGN_CACHE
int t_max_argc;
2655 dispatch_shared_info_t *t_disp_buffer;
2658 kmp_r_sched_t t_sched;
2659 #if KMP_AFFINITY_SUPPORTED
2662 #endif // KMP_AFFINITY_SUPPORTED
2663 int t_display_affinity;
2666 omp_allocator_handle_t t_def_allocator;
2669 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2674 char dummy_padding[1024];
2677 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2680 std::atomic<kmp_int32> t_cancel_request;
2681 int t_master_active;
2682 void *t_copypriv_data;
2684 std::atomic<kmp_uint32> t_copyin_counter;
2691 union KMP_ALIGN_CACHE kmp_team {
2694 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2697 typedef union KMP_ALIGN_CACHE kmp_time_global {
2699 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2701 } kmp_time_global_t;
2703 typedef struct kmp_base_global {
2705 kmp_time_global_t g_time;
2708 volatile int g_abort;
2709 volatile int g_done;
2712 enum dynamic_mode g_dynamic_mode;
2713 } kmp_base_global_t;
2715 typedef union KMP_ALIGN_CACHE kmp_global {
2716 kmp_base_global_t g;
2718 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2721 typedef struct kmp_base_root {
2726 volatile int r_active;
2728 std::atomic<int> r_in_parallel;
2730 kmp_team_t *r_root_team;
2731 kmp_team_t *r_hot_team;
2732 kmp_info_t *r_uber_thread;
2733 kmp_lock_t r_begin_lock;
2734 volatile int r_begin;
2738 typedef union KMP_ALIGN_CACHE kmp_root {
2741 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2744 struct fortran_inx_info {
2750 extern int __kmp_settings;
2751 extern int __kmp_duplicate_library_ok;
2753 extern int __kmp_forkjoin_frames;
2754 extern int __kmp_forkjoin_frames_mode;
2756 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2757 extern int __kmp_determ_red;
2760 extern int kmp_a_debug;
2761 extern int kmp_b_debug;
2762 extern int kmp_c_debug;
2763 extern int kmp_d_debug;
2764 extern int kmp_e_debug;
2765 extern int kmp_f_debug;
2769 #define KMP_DEBUG_BUF_LINES_INIT 512
2770 #define KMP_DEBUG_BUF_LINES_MIN 1
2772 #define KMP_DEBUG_BUF_CHARS_INIT 128
2773 #define KMP_DEBUG_BUF_CHARS_MIN 2
2777 extern int __kmp_debug_buf_lines;
2779 __kmp_debug_buf_chars;
2780 extern int __kmp_debug_buf_atomic;
2783 extern char *__kmp_debug_buffer;
2784 extern std::atomic<int> __kmp_debug_count;
2786 extern int __kmp_debug_buf_warn_chars;
2791 extern int __kmp_par_range;
2793 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2794 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2795 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2796 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2797 extern int __kmp_par_range_lb;
2798 extern int __kmp_par_range_ub;
2804 extern int __kmp_storage_map_verbose;
2806 extern int __kmp_storage_map_verbose_specified;
2808 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2809 extern kmp_cpuinfo_t __kmp_cpuinfo;
2812 extern volatile int __kmp_init_serial;
2813 extern volatile int __kmp_init_gtid;
2814 extern volatile int __kmp_init_common;
2815 extern volatile int __kmp_init_middle;
2816 extern volatile int __kmp_init_parallel;
2818 extern volatile int __kmp_init_monitor;
2820 extern volatile int __kmp_init_user_locks;
2821 extern int __kmp_init_counter;
2822 extern int __kmp_root_counter;
2823 extern int __kmp_version;
2826 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2829 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
2830 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
2831 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
2832 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
2833 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
2834 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
2835 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
2836 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
2837 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
2838 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
2839 extern char const *__kmp_barrier_type_name[bs_last_barrier];
2840 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
2843 extern kmp_bootstrap_lock_t __kmp_initz_lock;
2844 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock;
2845 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
2846 extern kmp_bootstrap_lock_t
2849 extern kmp_bootstrap_lock_t
2852 extern kmp_bootstrap_lock_t
2853 __kmp_tp_cached_lock;
2856 extern kmp_lock_t __kmp_global_lock;
2857 extern kmp_queuing_lock_t __kmp_dispatch_lock;
2858 extern kmp_lock_t __kmp_debug_lock;
2860 extern enum library_type __kmp_library;
2866 extern int __kmp_chunk;
2868 extern size_t __kmp_stksize;
2870 extern size_t __kmp_monitor_stksize;
2872 extern size_t __kmp_stkoffset;
2873 extern int __kmp_stkpadding;
2876 __kmp_malloc_pool_incr;
2877 extern int __kmp_env_stksize;
2878 extern int __kmp_env_blocktime;
2879 extern int __kmp_env_checks;
2880 extern int __kmp_env_consistency_check;
2881 extern int __kmp_generate_warnings;
2882 extern int __kmp_reserve_warn;
2884 #ifdef DEBUG_SUSPEND
2885 extern int __kmp_suspend_count;
2888 extern kmp_int32 __kmp_use_yield;
2889 extern kmp_int32 __kmp_use_yield_exp_set;
2890 extern kmp_uint32 __kmp_yield_init;
2891 extern kmp_uint32 __kmp_yield_next;
2894 extern int __kmp_allThreadsSpecified;
2896 extern size_t __kmp_align_alloc;
2898 extern int __kmp_xproc;
2899 extern int __kmp_avail_proc;
2900 extern size_t __kmp_sys_min_stksize;
2901 extern int __kmp_sys_max_nth;
2903 extern int __kmp_max_nth;
2905 extern int __kmp_cg_max_nth;
2906 extern int __kmp_teams_max_nth;
2907 extern int __kmp_threads_capacity;
2909 extern int __kmp_dflt_team_nth;
2911 extern int __kmp_dflt_team_nth_ub;
2913 extern int __kmp_tp_capacity;
2915 extern int __kmp_tp_cached;
2917 extern int __kmp_dflt_blocktime;
2921 __kmp_monitor_wakeups;
2922 extern int __kmp_bt_intervals;
2925 #ifdef KMP_ADJUST_BLOCKTIME
2926 extern int __kmp_zero_bt;
2928 #ifdef KMP_DFLT_NTH_CORES
2929 extern int __kmp_ncores;
2932 extern int __kmp_abort_delay;
2934 extern int __kmp_need_register_atfork_specified;
2936 __kmp_need_register_atfork;
2938 extern int __kmp_gtid_mode;
2946 __kmp_adjust_gtid_mode;
2947 #ifdef KMP_TDATA_GTID
2948 extern KMP_THREAD_LOCAL
int __kmp_gtid;
2950 extern int __kmp_tls_gtid_min;
2951 extern int __kmp_foreign_tp;
2952 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2953 extern int __kmp_inherit_fp_control;
2954 extern kmp_int16 __kmp_init_x87_fpu_control_word;
2955 extern kmp_uint32 __kmp_init_mxcsr;
2960 extern int __kmp_dflt_max_active_levels;
2963 extern bool __kmp_dflt_max_active_levels_set;
2964 extern int __kmp_dispatch_num_buffers;
2966 #if KMP_NESTED_HOT_TEAMS
2967 extern int __kmp_hot_teams_mode;
2968 extern int __kmp_hot_teams_max_level;
2972 extern enum clock_function_type __kmp_clock_function;
2973 extern int __kmp_clock_function_param;
2976 #if KMP_MIC_SUPPORTED
2977 extern enum mic_type __kmp_mic_type;
2980 #ifdef USE_LOAD_BALANCE
2981 extern double __kmp_load_balance_interval;
2985 typedef struct kmp_nested_nthreads_t {
2989 } kmp_nested_nthreads_t;
2991 extern kmp_nested_nthreads_t __kmp_nested_nth;
2993 #if KMP_USE_ADAPTIVE_LOCKS
2996 struct kmp_adaptive_backoff_params_t {
2998 kmp_uint32 max_soft_retries;
3001 kmp_uint32 max_badness;
3004 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3006 #if KMP_DEBUG_ADAPTIVE_LOCKS
3007 extern const char *__kmp_speculative_statsfile;
3010 #endif // KMP_USE_ADAPTIVE_LOCKS
3012 extern int __kmp_display_env;
3013 extern int __kmp_display_env_verbose;
3014 extern int __kmp_omp_cancellation;
3020 extern kmp_info_t **__kmp_threads;
3022 extern volatile kmp_team_t *__kmp_team_pool;
3023 extern volatile kmp_info_t *__kmp_thread_pool;
3024 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3027 extern volatile int __kmp_nth;
3030 extern volatile int __kmp_all_nth;
3031 extern std::atomic<int> __kmp_thread_pool_active_nth;
3033 extern kmp_root_t **__kmp_root;
3037 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3038 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3039 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3040 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3041 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3046 #define __kmp_get_team_num_threads(gtid) \
3047 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3049 static inline bool KMP_UBER_GTID(
int gtid) {
3050 KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3051 KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3052 return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3053 __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3056 static inline int __kmp_tid_from_gtid(
int gtid) {
3057 KMP_DEBUG_ASSERT(gtid >= 0);
3058 return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3061 static inline int __kmp_gtid_from_tid(
int tid,
const kmp_team_t *team) {
3062 KMP_DEBUG_ASSERT(tid >= 0 && team);
3063 return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3066 static inline int __kmp_gtid_from_thread(
const kmp_info_t *thr) {
3067 KMP_DEBUG_ASSERT(thr);
3068 return thr->th.th_info.ds.ds_gtid;
3071 static inline kmp_info_t *__kmp_thread_from_gtid(
int gtid) {
3072 KMP_DEBUG_ASSERT(gtid >= 0);
3073 return __kmp_threads[gtid];
3076 static inline kmp_team_t *__kmp_team_from_gtid(
int gtid) {
3077 KMP_DEBUG_ASSERT(gtid >= 0);
3078 return __kmp_threads[gtid]->th.th_team;
3083 extern kmp_global_t __kmp_global;
3085 extern kmp_info_t __kmp_monitor;
3087 extern std::atomic<kmp_int32> __kmp_team_counter;
3089 extern std::atomic<kmp_int32> __kmp_task_counter;
3092 #define _KMP_GEN_ID(counter) \
3093 (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3095 #define _KMP_GEN_ID(counter) (~0)
3098 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3099 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3103 extern void __kmp_print_storage_map_gtid(
int gtid,
void *p1,
void *p2,
3104 size_t size,
char const *format, ...);
3106 extern void __kmp_serial_initialize(
void);
3107 extern void __kmp_middle_initialize(
void);
3108 extern void __kmp_parallel_initialize(
void);
3110 extern void __kmp_internal_begin(
void);
3111 extern void __kmp_internal_end_library(
int gtid);
3112 extern void __kmp_internal_end_thread(
int gtid);
3113 extern void __kmp_internal_end_atexit(
void);
3114 extern void __kmp_internal_end_dtor(
void);
3115 extern void __kmp_internal_end_dest(
void *);
3117 extern int __kmp_register_root(
int initial_thread);
3118 extern void __kmp_unregister_root(
int gtid);
3120 extern int __kmp_ignore_mppbeg(
void);
3121 extern int __kmp_ignore_mppend(
void);
3123 extern int __kmp_enter_single(
int gtid,
ident_t *id_ref,
int push_ws);
3124 extern void __kmp_exit_single(
int gtid);
3126 extern void __kmp_parallel_deo(
int *gtid_ref,
int *cid_ref,
ident_t *loc_ref);
3127 extern void __kmp_parallel_dxo(
int *gtid_ref,
int *cid_ref,
ident_t *loc_ref);
3129 #ifdef USE_LOAD_BALANCE
3130 extern int __kmp_get_load_balance(
int);
3133 extern int __kmp_get_global_thread_id(
void);
3134 extern int __kmp_get_global_thread_id_reg(
void);
3135 extern void __kmp_exit_thread(
int exit_status);
3136 extern void __kmp_abort(
char const *format, ...);
3137 extern void __kmp_abort_thread(
void);
3138 KMP_NORETURN
extern void __kmp_abort_process(
void);
3139 extern void __kmp_warn(
char const *format, ...);
3141 extern void __kmp_set_num_threads(
int new_nth,
int gtid);
3145 static inline kmp_info_t *__kmp_entry_thread() {
3146 int gtid = __kmp_entry_gtid();
3148 return __kmp_threads[gtid];
3151 extern void __kmp_set_max_active_levels(
int gtid,
int new_max_active_levels);
3152 extern int __kmp_get_max_active_levels(
int gtid);
3153 extern int __kmp_get_ancestor_thread_num(
int gtid,
int level);
3154 extern int __kmp_get_team_size(
int gtid,
int level);
3155 extern void __kmp_set_schedule(
int gtid, kmp_sched_t new_sched,
int chunk);
3156 extern void __kmp_get_schedule(
int gtid, kmp_sched_t *sched,
int *chunk);
3158 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3159 extern void __kmp_init_random(kmp_info_t *thread);
3161 extern kmp_r_sched_t __kmp_get_schedule_global(
void);
3162 extern void __kmp_adjust_num_threads(
int new_nproc);
3163 extern void __kmp_check_stksize(
size_t *val);
3165 extern void *___kmp_allocate(
size_t size KMP_SRC_LOC_DECL);
3166 extern void *___kmp_page_allocate(
size_t size KMP_SRC_LOC_DECL);
3167 extern void ___kmp_free(
void *ptr KMP_SRC_LOC_DECL);
3168 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3169 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3170 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3173 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3174 size_t size KMP_SRC_LOC_DECL);
3175 extern void ___kmp_fast_free(kmp_info_t *this_thr,
void *ptr KMP_SRC_LOC_DECL);
3176 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3177 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3178 #define __kmp_fast_allocate(this_thr, size) \
3179 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3180 #define __kmp_fast_free(this_thr, ptr) \
3181 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3184 extern void *___kmp_thread_malloc(kmp_info_t *th,
size_t size KMP_SRC_LOC_DECL);
3185 extern void *___kmp_thread_calloc(kmp_info_t *th,
size_t nelem,
3186 size_t elsize KMP_SRC_LOC_DECL);
3187 extern void *___kmp_thread_realloc(kmp_info_t *th,
void *ptr,
3188 size_t size KMP_SRC_LOC_DECL);
3189 extern void ___kmp_thread_free(kmp_info_t *th,
void *ptr KMP_SRC_LOC_DECL);
3190 #define __kmp_thread_malloc(th, size) \
3191 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3192 #define __kmp_thread_calloc(th, nelem, elsize) \
3193 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3194 #define __kmp_thread_realloc(th, ptr, size) \
3195 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3196 #define __kmp_thread_free(th, ptr) \
3197 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3199 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3200 #define KMP_INTERNAL_FREE(p) free(p)
3201 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3202 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3204 extern void __kmp_push_num_threads(
ident_t *loc,
int gtid,
int num_threads);
3206 extern void __kmp_push_proc_bind(
ident_t *loc,
int gtid,
3207 kmp_proc_bind_t proc_bind);
3208 extern void __kmp_push_num_teams(
ident_t *loc,
int gtid,
int num_teams,
3211 extern void __kmp_yield();
3215 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3218 kmp_uint32 ub, kmp_int32 st,
3222 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3225 kmp_uint64 ub, kmp_int64 st,
3229 kmp_int32 *p_last, kmp_int32 *p_lb,
3230 kmp_int32 *p_ub, kmp_int32 *p_st);
3232 kmp_int32 *p_last, kmp_uint32 *p_lb,
3233 kmp_uint32 *p_ub, kmp_int32 *p_st);
3235 kmp_int32 *p_last, kmp_int64 *p_lb,
3236 kmp_int64 *p_ub, kmp_int64 *p_st);
3238 kmp_int32 *p_last, kmp_uint64 *p_lb,
3239 kmp_uint64 *p_ub, kmp_int64 *p_st);
3246 #ifdef KMP_GOMP_COMPAT
3248 extern void __kmp_aux_dispatch_init_4(
ident_t *loc, kmp_int32 gtid,
3250 kmp_int32 ub, kmp_int32 st,
3251 kmp_int32 chunk,
int push_ws);
3252 extern void __kmp_aux_dispatch_init_4u(
ident_t *loc, kmp_int32 gtid,
3254 kmp_uint32 ub, kmp_int32 st,
3255 kmp_int32 chunk,
int push_ws);
3256 extern void __kmp_aux_dispatch_init_8(
ident_t *loc, kmp_int32 gtid,
3258 kmp_int64 ub, kmp_int64 st,
3259 kmp_int64 chunk,
int push_ws);
3260 extern void __kmp_aux_dispatch_init_8u(
ident_t *loc, kmp_int32 gtid,
3262 kmp_uint64 ub, kmp_int64 st,
3263 kmp_int64 chunk,
int push_ws);
3264 extern void __kmp_aux_dispatch_fini_chunk_4(
ident_t *loc, kmp_int32 gtid);
3265 extern void __kmp_aux_dispatch_fini_chunk_8(
ident_t *loc, kmp_int32 gtid);
3266 extern void __kmp_aux_dispatch_fini_chunk_4u(
ident_t *loc, kmp_int32 gtid);
3267 extern void __kmp_aux_dispatch_fini_chunk_8u(
ident_t *loc, kmp_int32 gtid);
3271 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3272 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3273 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3274 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3275 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3276 extern kmp_uint32 __kmp_wait_4(kmp_uint32
volatile *spinner, kmp_uint32 checker,
3277 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3279 extern void __kmp_wait_4_ptr(
void *spinner, kmp_uint32 checker,
3280 kmp_uint32 (*pred)(
void *, kmp_uint32),
void *obj);
3284 class kmp_flag_oncore;
3285 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64 *flag,
3292 extern void __kmp_release_64(kmp_flag_64 *flag);
3294 extern void __kmp_infinite_loop(
void);
3296 extern void __kmp_cleanup(
void);
3298 #if KMP_HANDLE_SIGNALS
3299 extern int __kmp_handle_signals;
3300 extern void __kmp_install_signals(
int parallel_init);
3301 extern void __kmp_remove_signals(
void);
3304 extern void __kmp_clear_system_time(
void);
3305 extern void __kmp_read_system_time(
double *delta);
3307 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3309 extern void __kmp_expand_host_name(
char *buffer,
size_t size);
3310 extern void __kmp_expand_file_name(
char *result,
size_t rlen,
char *pattern);
3312 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3314 __kmp_initialize_system_tick(
void);
3318 __kmp_runtime_initialize(
void);
3319 extern void __kmp_runtime_destroy(
void);
3321 #if KMP_AFFINITY_SUPPORTED
3322 extern char *__kmp_affinity_print_mask(
char *buf,
int buf_len,
3323 kmp_affin_mask_t *mask);
3324 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3325 kmp_affin_mask_t *mask);
3326 extern void __kmp_affinity_initialize(
void);
3327 extern void __kmp_affinity_uninitialize(
void);
3328 extern void __kmp_affinity_set_init_mask(
3329 int gtid,
int isa_root);
3330 extern void __kmp_affinity_set_place(
int gtid);
3331 extern void __kmp_affinity_determine_capable(
const char *env_var);
3332 extern int __kmp_aux_set_affinity(
void **mask);
3333 extern int __kmp_aux_get_affinity(
void **mask);
3334 extern int __kmp_aux_get_affinity_max_proc();
3335 extern int __kmp_aux_set_affinity_mask_proc(
int proc,
void **mask);
3336 extern int __kmp_aux_unset_affinity_mask_proc(
int proc,
void **mask);
3337 extern int __kmp_aux_get_affinity_mask_proc(
int proc,
void **mask);
3338 extern void __kmp_balanced_affinity(kmp_info_t *th,
int team_size);
3339 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3340 extern int kmp_set_thread_affinity_mask_initial(
void);
3346 extern size_t __kmp_aux_capture_affinity(
int gtid,
const char *format,
3347 kmp_str_buf_t *buffer);
3348 extern void __kmp_aux_display_affinity(
int gtid,
const char *format);
3350 extern void __kmp_cleanup_hierarchy();
3351 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3355 extern int __kmp_futex_determine_capable(
void);
3357 #endif // KMP_USE_FUTEX
3359 extern void __kmp_gtid_set_specific(
int gtid);
3360 extern int __kmp_gtid_get_specific(
void);
3362 extern double __kmp_read_cpu_time(
void);
3364 extern int __kmp_read_system_info(
struct kmp_sys_info *info);
3367 extern void __kmp_create_monitor(kmp_info_t *th);
3370 extern void *__kmp_launch_thread(kmp_info_t *thr);
3372 extern void __kmp_create_worker(
int gtid, kmp_info_t *th,
size_t stack_size);
3375 extern int __kmp_still_running(kmp_info_t *th);
3376 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3377 extern void __kmp_free_handle(kmp_thread_t tHandle);
3381 extern void __kmp_reap_monitor(kmp_info_t *th);
3383 extern void __kmp_reap_worker(kmp_info_t *th);
3384 extern void __kmp_terminate_thread(
int gtid);
3386 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3387 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3388 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3390 extern void __kmp_suspend_32(
int th_gtid, kmp_flag_32 *flag);
3391 extern void __kmp_suspend_64(
int th_gtid, kmp_flag_64 *flag);
3392 extern void __kmp_suspend_oncore(
int th_gtid, kmp_flag_oncore *flag);
3393 extern void __kmp_resume_32(
int target_gtid, kmp_flag_32 *flag);
3394 extern void __kmp_resume_64(
int target_gtid, kmp_flag_64 *flag);
3395 extern void __kmp_resume_oncore(
int target_gtid, kmp_flag_oncore *flag);
3397 extern void __kmp_elapsed(
double *);
3398 extern void __kmp_elapsed_tick(
double *);
3400 extern void __kmp_enable(
int old_state);
3401 extern void __kmp_disable(
int *old_state);
3403 extern void __kmp_thread_sleep(
int millis);
3405 extern void __kmp_common_initialize(
void);
3406 extern void __kmp_common_destroy(
void);
3407 extern void __kmp_common_destroy_gtid(
int gtid);
3410 extern void __kmp_register_atfork(
void);
3412 extern void __kmp_suspend_initialize(
void);
3413 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3414 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3416 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3419 __kmp_allocate_team(kmp_root_t *root,
int new_nproc,
int max_nproc,
3421 ompt_data_t ompt_parallel_data,
3423 kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3424 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3425 extern void __kmp_free_thread(kmp_info_t *);
3426 extern void __kmp_free_team(kmp_root_t *,
3427 kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3428 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3432 extern void __kmp_initialize_bget(kmp_info_t *th);
3433 extern void __kmp_finalize_bget(kmp_info_t *th);
3435 KMP_EXPORT
void *kmpc_malloc(
size_t size);
3436 KMP_EXPORT
void *kmpc_aligned_malloc(
size_t size,
size_t alignment);
3437 KMP_EXPORT
void *kmpc_calloc(
size_t nelem,
size_t elsize);
3438 KMP_EXPORT
void *kmpc_realloc(
void *ptr,
size_t size);
3439 KMP_EXPORT
void kmpc_free(
void *ptr);
3443 extern int __kmp_barrier(
enum barrier_type bt,
int gtid,
int is_split,
3444 size_t reduce_size,
void *reduce_data,
3445 void (*reduce)(
void *,
void *));
3446 extern void __kmp_end_split_barrier(
enum barrier_type bt,
int gtid);
3447 extern int __kmp_barrier_gomp_cancel(
int gtid);
3453 enum fork_context_e {
3459 extern int __kmp_fork_call(
ident_t *loc,
int gtid,
3460 enum fork_context_e fork_context, kmp_int32 argc,
3461 microtask_t microtask, launch_t invoker,
3463 #
if (KMP_ARCH_ARM || KMP_ARCH_X86_64 || KMP_ARCH_AARCH64) && KMP_OS_LINUX
3470 extern void __kmp_join_call(
ident_t *loc,
int gtid
3473 enum fork_context_e fork_context
3476 int exit_teams = 0);
3478 extern void __kmp_serialized_parallel(
ident_t *
id, kmp_int32 gtid);
3479 extern void __kmp_internal_fork(
ident_t *
id,
int gtid, kmp_team_t *team);
3480 extern void __kmp_internal_join(
ident_t *
id,
int gtid, kmp_team_t *team);
3481 extern int __kmp_invoke_task_func(
int gtid);
3482 extern void __kmp_run_before_invoked_task(
int gtid,
int tid,
3483 kmp_info_t *this_thr,
3485 extern void __kmp_run_after_invoked_task(
int gtid,
int tid,
3486 kmp_info_t *this_thr,
3490 KMP_EXPORT
int __kmpc_invoke_task_func(
int gtid);
3491 extern int __kmp_invoke_teams_master(
int gtid);
3492 extern void __kmp_teams_master(
int gtid);
3493 extern int __kmp_aux_get_team_num();
3494 extern int __kmp_aux_get_num_teams();
3495 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3496 extern void __kmp_user_set_library(
enum library_type arg);
3497 extern void __kmp_aux_set_library(
enum library_type arg);
3498 extern void __kmp_aux_set_stacksize(
size_t arg);
3499 extern void __kmp_aux_set_blocktime(
int arg, kmp_info_t *thread,
int tid);
3500 extern void __kmp_aux_set_defaults(
char const *str,
int len);
3503 void kmpc_set_blocktime(
int arg);
3504 void ompc_set_nested(
int flag);
3505 void ompc_set_dynamic(
int flag);
3506 void ompc_set_num_threads(
int arg);
3508 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3509 kmp_team_t *team,
int tid);
3510 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3511 extern kmp_task_t *__kmp_task_alloc(
ident_t *loc_ref, kmp_int32 gtid,
3512 kmp_tasking_flags_t *flags,
3513 size_t sizeof_kmp_task_t,
3514 size_t sizeof_shareds,
3515 kmp_routine_entry_t task_entry);
3516 extern void __kmp_init_implicit_task(
ident_t *loc_ref, kmp_info_t *this_thr,
3517 kmp_team_t *team,
int tid,
3519 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3520 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3522 extern kmp_event_t *__kmpc_task_allow_completion_event(
ident_t *loc_ref,
3525 extern void __kmp_fulfill_event(kmp_event_t *event);
3527 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
3528 kmp_flag_32 *flag,
int final_spin,
3529 int *thread_finished,
3533 kmp_int32 is_constrained);
3534 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
3535 kmp_flag_64 *flag,
int final_spin,
3536 int *thread_finished,
3540 kmp_int32 is_constrained);
3541 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
3542 kmp_flag_oncore *flag,
int final_spin,
3543 int *thread_finished,
3547 kmp_int32 is_constrained);
3549 extern void __kmp_free_task_team(kmp_info_t *thread,
3550 kmp_task_team_t *task_team);
3551 extern void __kmp_reap_task_teams(
void);
3552 extern void __kmp_wait_to_unref_task_teams(
void);
3553 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3555 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3556 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3563 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3566 extern int __kmp_is_address_mapped(
void *addr);
3567 extern kmp_uint64 __kmp_hardware_timestamp(
void);
3570 extern int __kmp_read_from_file(
char const *path,
char const *format, ...);
3578 extern int __kmp_invoke_microtask(microtask_t pkfn,
int gtid,
int npr,
int argc,
3582 void **exit_frame_ptr
3595 size_t vector_length);
3599 KMP_EXPORT
void *__kmpc_threadprivate(
ident_t *, kmp_int32 global_tid,
3600 void *data,
size_t size);
3621 kmp_critical_name *);
3623 kmp_critical_name *);
3625 kmp_critical_name *, uint32_t hint);
3631 kmp_int32 global_tid);
3636 KMP_EXPORT
void KMPC_FOR_STATIC_INIT(
ident_t *loc, kmp_int32 global_tid,
3637 kmp_int32 schedtype, kmp_int32 *plastiter,
3638 kmp_int *plower, kmp_int *pupper,
3639 kmp_int *pstride, kmp_int incr,
3645 size_t cpy_size,
void *cpy_data,
3646 void (*cpy_func)(
void *,
void *),
3649 extern void KMPC_SET_NUM_THREADS(
int arg);
3650 extern void KMPC_SET_DYNAMIC(
int flag);
3651 extern void KMPC_SET_NESTED(
int flag);
3654 KMP_EXPORT kmp_int32 __kmpc_omp_task(
ident_t *loc_ref, kmp_int32 gtid,
3655 kmp_task_t *new_task);
3656 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(
ident_t *loc_ref, kmp_int32 gtid,
3658 size_t sizeof_kmp_task_t,
3659 size_t sizeof_shareds,
3660 kmp_routine_entry_t task_entry);
3661 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
ident_t *loc_ref, kmp_int32 gtid,
3663 size_t sizeof_kmp_task_t,
3664 size_t sizeof_shareds,
3665 kmp_routine_entry_t task_entry,
3666 kmp_int64 device_id);
3667 KMP_EXPORT
void __kmpc_omp_task_begin_if0(
ident_t *loc_ref, kmp_int32 gtid,
3669 KMP_EXPORT
void __kmpc_omp_task_complete_if0(
ident_t *loc_ref, kmp_int32 gtid,
3671 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(
ident_t *loc_ref, kmp_int32 gtid,
3672 kmp_task_t *new_task);
3673 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(
ident_t *loc_ref, kmp_int32 gtid);
3675 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(
ident_t *loc_ref, kmp_int32 gtid,
3679 void __kmpc_omp_task_begin(
ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3680 void __kmpc_omp_task_complete(
ident_t *loc_ref, kmp_int32 gtid,
3682 #endif // TASK_UNUSED
3686 KMP_EXPORT
void __kmpc_taskgroup(
ident_t *loc,
int gtid);
3687 KMP_EXPORT
void __kmpc_end_taskgroup(
ident_t *loc,
int gtid);
3690 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3691 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3692 kmp_depend_info_t *noalias_dep_list);
3695 kmp_depend_info_t *dep_list,
3696 kmp_int32 ndeps_noalias,
3697 kmp_depend_info_t *noalias_dep_list);
3698 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3699 bool serialize_immediate);
3701 KMP_EXPORT kmp_int32 __kmpc_cancel(
ident_t *loc_ref, kmp_int32 gtid,
3702 kmp_int32 cncl_kind);
3703 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(
ident_t *loc_ref, kmp_int32 gtid,
3704 kmp_int32 cncl_kind);
3705 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(
ident_t *loc_ref, kmp_int32 gtid);
3706 KMP_EXPORT
int __kmp_get_cancellation_status(
int cancel_kind);
3711 kmp_int32 if_val, kmp_uint64 *lb,
3712 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3713 kmp_int32 sched, kmp_uint64 grainsize,
3722 int num,
void *data);
3726 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3727 kmp_task_affinity_info_t *affin_list);
3730 KMP_EXPORT
void __kmpc_init_lock(
ident_t *loc, kmp_int32 gtid,
3732 KMP_EXPORT
void __kmpc_init_nest_lock(
ident_t *loc, kmp_int32 gtid,
3734 KMP_EXPORT
void __kmpc_destroy_lock(
ident_t *loc, kmp_int32 gtid,
3736 KMP_EXPORT
void __kmpc_destroy_nest_lock(
ident_t *loc, kmp_int32 gtid,
3738 KMP_EXPORT
void __kmpc_set_lock(
ident_t *loc, kmp_int32 gtid,
void **user_lock);
3739 KMP_EXPORT
void __kmpc_set_nest_lock(
ident_t *loc, kmp_int32 gtid,
3741 KMP_EXPORT
void __kmpc_unset_lock(
ident_t *loc, kmp_int32 gtid,
3743 KMP_EXPORT
void __kmpc_unset_nest_lock(
ident_t *loc, kmp_int32 gtid,
3745 KMP_EXPORT
int __kmpc_test_lock(
ident_t *loc, kmp_int32 gtid,
void **user_lock);
3746 KMP_EXPORT
int __kmpc_test_nest_lock(
ident_t *loc, kmp_int32 gtid,
3749 KMP_EXPORT
void __kmpc_init_lock_with_hint(
ident_t *loc, kmp_int32 gtid,
3750 void **user_lock, uintptr_t hint);
3751 KMP_EXPORT
void __kmpc_init_nest_lock_with_hint(
ident_t *loc, kmp_int32 gtid,
3758 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
size_t reduce_size,
3759 void *reduce_data,
void (*reduce_func)(
void *lhs_data,
void *rhs_data),
3760 kmp_critical_name *lck);
3762 kmp_critical_name *lck);
3764 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
size_t reduce_size,
3765 void *reduce_data,
void (*reduce_func)(
void *lhs_data,
void *rhs_data),
3766 kmp_critical_name *lck);
3768 kmp_critical_name *lck);
3772 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3773 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
size_t reduce_size,
3774 void *reduce_data,
void (*reduce_func)(
void *lhs_data,
void *rhs_data),
3775 kmp_critical_name *lck);
3778 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(
void);
3780 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3781 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3787 KMP_EXPORT
void __kmpc_pop_num_threads(
ident_t *loc, kmp_int32 global_tid);
3789 kmp_int32 num_threads);
3791 KMP_EXPORT
void __kmpc_push_proc_bind(
ident_t *loc, kmp_int32 global_tid,
3794 kmp_int32 num_teams,
3795 kmp_int32 num_threads);
3805 const struct kmp_dim *dims);
3806 KMP_EXPORT
void __kmpc_doacross_wait(
ident_t *loc, kmp_int32 gtid,
3807 const kmp_int64 *vec);
3808 KMP_EXPORT
void __kmpc_doacross_post(
ident_t *loc, kmp_int32 gtid,
3809 const kmp_int64 *vec);
3810 KMP_EXPORT
void __kmpc_doacross_fini(
ident_t *loc, kmp_int32 gtid);
3813 void *data,
size_t size,
3817 extern int _You_must_link_with_exactly_one_OpenMP_library;
3818 extern int _You_must_link_with_Intel_OpenMP_library;
3819 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3820 extern int _You_must_link_with_Microsoft_OpenMP_library;
3825 void kmp_threadprivate_insert_private_data(
int gtid,
void *pc_addr,
3826 void *data_addr,
size_t pc_size);
3827 struct private_common *kmp_threadprivate_insert(
int gtid,
void *pc_addr,
3830 void __kmp_threadprivate_resize_cache(
int newCapacity);
3831 void __kmp_cleanup_threadprivate_caches();
3835 #define KMPC_CONVENTION __cdecl
3837 #define KMPC_CONVENTION
3841 typedef enum omp_sched_t {
3842 omp_sched_static = 1,
3843 omp_sched_dynamic = 2,
3844 omp_sched_guided = 3,
3847 typedef void *kmp_affinity_mask_t;
3850 KMP_EXPORT
void KMPC_CONVENTION ompc_set_max_active_levels(
int);
3851 KMP_EXPORT
void KMPC_CONVENTION ompc_set_schedule(omp_sched_t,
int);
3852 KMP_EXPORT
int KMPC_CONVENTION ompc_get_ancestor_thread_num(
int);
3853 KMP_EXPORT
int KMPC_CONVENTION ompc_get_team_size(
int);
3854 KMP_EXPORT
int KMPC_CONVENTION
3855 kmpc_set_affinity_mask_proc(
int, kmp_affinity_mask_t *);
3856 KMP_EXPORT
int KMPC_CONVENTION
3857 kmpc_unset_affinity_mask_proc(
int, kmp_affinity_mask_t *);
3858 KMP_EXPORT
int KMPC_CONVENTION
3859 kmpc_get_affinity_mask_proc(
int, kmp_affinity_mask_t *);
3861 KMP_EXPORT
void KMPC_CONVENTION kmpc_set_stacksize(
int);
3862 KMP_EXPORT
void KMPC_CONVENTION kmpc_set_stacksize_s(
size_t);
3863 KMP_EXPORT
void KMPC_CONVENTION kmpc_set_library(
int);
3864 KMP_EXPORT
void KMPC_CONVENTION kmpc_set_defaults(
char const *);
3865 KMP_EXPORT
void KMPC_CONVENTION kmpc_set_disp_num_buffers(
int);
3867 enum kmp_target_offload_kind {
3872 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
3874 extern kmp_target_offload_kind_t __kmp_target_offload;
3875 extern int __kmpc_get_target_offload();
3878 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
3879 #define KMP_HOST_DEVICE -10 // This is what it is in libomptarget, go figure.
3880 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
3886 typedef enum kmp_pause_status_t {
3888 kmp_soft_paused = 1,
3890 } kmp_pause_status_t;
3893 extern kmp_pause_status_t __kmp_pause_status;
3894 extern int __kmpc_pause_resource(kmp_pause_status_t level);
3895 extern int __kmp_pause_resource(kmp_pause_status_t level);
3897 extern void __kmp_resume_if_soft_paused();
3901 static inline void __kmp_resume_if_hard_paused() {
3902 if (__kmp_pause_status == kmp_hard_paused) {
3903 __kmp_pause_status = kmp_not_paused;
3907 extern void __kmp_omp_display_env(
int verbose);