18 #include "kmp_config.h"
28 #ifndef KMP_STATIC_STEAL_ENABLED
29 #define KMP_STATIC_STEAL_ENABLED 1
32 #define TASK_CURRENT_NOT_QUEUED 0
33 #define TASK_CURRENT_QUEUED 1
35 #ifdef BUILD_TIED_TASK_STACK
36 #define TASK_STACK_EMPTY 0 // entries when the stack is empty
37 #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
39 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
41 #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
42 #endif // BUILD_TIED_TASK_STACK
44 #define TASK_NOT_PUSHED 1
45 #define TASK_SUCCESSFULLY_PUSHED 0
48 #define TASK_EXPLICIT 1
49 #define TASK_IMPLICIT 0
53 #define KMP_CANCEL_THREADS
54 #define KMP_THREAD_ATTR
58 #if defined(__ANDROID__)
59 #undef KMP_CANCEL_THREADS
71 #include <sys/types.h>
80 #include "kmp_safe_c_api.h"
86 #if KMP_USE_HIER_SCHED
88 #undef KMP_USE_HIER_SCHED
89 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
92 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
94 #ifndef HWLOC_OBJ_NUMANODE
95 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
97 #ifndef HWLOC_OBJ_PACKAGE
98 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
102 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
103 #include <xmmintrin.h>
106 #include "kmp_debug.h"
107 #include "kmp_lock.h"
108 #include "kmp_version.h"
110 #include "kmp_debugger.h"
112 #include "kmp_i18n.h"
114 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
116 #include "kmp_wrapper_malloc.h"
119 #if !defined NSIG && defined _NSIG
125 #pragma weak clock_gettime
129 #include "ompt-internal.h"
135 #ifndef USE_FAST_MEMORY
136 #define USE_FAST_MEMORY 3
139 #ifndef KMP_NESTED_HOT_TEAMS
140 #define KMP_NESTED_HOT_TEAMS 0
141 #define USE_NESTED_HOT_ARG(x)
143 #if KMP_NESTED_HOT_TEAMS
145 #define USE_NESTED_HOT_ARG(x) , x
148 #undef KMP_NESTED_HOT_TEAMS
149 #define KMP_NESTED_HOT_TEAMS 0
150 #define USE_NESTED_HOT_ARG(x)
153 #define USE_NESTED_HOT_ARG(x)
158 #ifndef USE_CMP_XCHG_FOR_BGET
159 #define USE_CMP_XCHG_FOR_BGET 1
167 #define KMP_NSEC_PER_SEC 1000000000L
168 #define KMP_USEC_PER_SEC 1000000L
181 #define KMP_IDENT_IMB 0x01
183 #define KMP_IDENT_KMPC 0x02
186 #define KMP_IDENT_AUTOPAR 0x08
188 #define KMP_IDENT_ATOMIC_REDUCE 0x10
190 #define KMP_IDENT_BARRIER_EXPL 0x20
192 #define KMP_IDENT_BARRIER_IMPL 0x0040
193 #define KMP_IDENT_BARRIER_IMPL_MASK 0x01C0
194 #define KMP_IDENT_BARRIER_IMPL_FOR 0x0040
195 #define KMP_IDENT_BARRIER_IMPL_SECTIONS 0x00C0
197 #define KMP_IDENT_BARRIER_IMPL_SINGLE 0x0140
198 #define KMP_IDENT_BARRIER_IMPL_WORKSHARE 0x01C0
200 #define KMP_IDENT_WORK_LOOP 0x200 // static loop
201 #define KMP_IDENT_WORK_SECTIONS 0x400 // sections
202 #define KMP_IDENT_WORK_DISTRIBUTE 0x800 // distribute
227 typedef union kmp_team kmp_team_t;
228 typedef struct kmp_taskdata kmp_taskdata_t;
229 typedef union kmp_task_team kmp_task_team_t;
230 typedef union kmp_team kmp_team_p;
231 typedef union kmp_info kmp_info_p;
232 typedef union kmp_root kmp_root_p;
242 #define KMP_PACK_64(HIGH_32, LOW_32) \
243 ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
246 #define SKIP_WS(_x) \
248 while (*(_x) == ' ' || *(_x) == '\t') \
251 #define SKIP_DIGITS(_x) \
253 while (*(_x) >= '0' && *(_x) <= '9') \
256 #define SKIP_TOKEN(_x) \
258 while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
259 (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
262 #define SKIP_TO(_x, _c) \
264 while (*(_x) != '\0' && *(_x) != (_c)) \
270 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
271 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
276 enum kmp_state_timer {
286 #ifdef USE_LOAD_BALANCE
287 dynamic_load_balance,
290 dynamic_thread_limit,
296 #ifndef KMP_SCHED_TYPE_DEFINED
297 #define KMP_SCHED_TYPE_DEFINED
298 typedef enum kmp_sched {
301 kmp_sched_static = 1,
302 kmp_sched_dynamic = 2,
303 kmp_sched_guided = 3,
305 kmp_sched_upper_std = 5,
306 kmp_sched_lower_ext = 100,
307 kmp_sched_trapezoidal = 101,
308 #if KMP_STATIC_STEAL_ENABLED
309 kmp_sched_static_steal = 102,
312 kmp_sched_default = kmp_sched_static
322 kmp_sch_static_chunked = 33,
324 kmp_sch_dynamic_chunked = 35,
326 kmp_sch_runtime = 37,
328 kmp_sch_trapezoidal = 39,
331 kmp_sch_static_greedy = 40,
332 kmp_sch_static_balanced = 41,
334 kmp_sch_guided_iterative_chunked = 42,
335 kmp_sch_guided_analytical_chunked = 43,
337 kmp_sch_static_steal = 44,
341 kmp_sch_static_balanced_chunked = 45,
342 kmp_sch_guided_simd = 46,
343 kmp_sch_runtime_simd = 47,
350 kmp_ord_static_chunked = 65,
352 kmp_ord_dynamic_chunked = 67,
353 kmp_ord_guided_chunked = 68,
354 kmp_ord_runtime = 69,
356 kmp_ord_trapezoidal = 71,
371 kmp_nm_static_chunked =
374 kmp_nm_dynamic_chunked = 163,
376 kmp_nm_runtime = 165,
378 kmp_nm_trapezoidal = 167,
381 kmp_nm_static_greedy = 168,
382 kmp_nm_static_balanced = 169,
384 kmp_nm_guided_iterative_chunked = 170,
385 kmp_nm_guided_analytical_chunked = 171,
386 kmp_nm_static_steal =
389 kmp_nm_ord_static_chunked = 193,
391 kmp_nm_ord_dynamic_chunked = 195,
392 kmp_nm_ord_guided_chunked = 196,
393 kmp_nm_ord_runtime = 197,
395 kmp_nm_ord_trapezoidal = 199,
413 kmp_sch_modifier_monotonic =
415 kmp_sch_modifier_nonmonotonic =
418 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
420 (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))
421 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
422 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
423 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
424 (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
429 #define SCHEDULE_WITHOUT_MODIFIERS(s) (s)
430 #define SCHEDULE_HAS_MONOTONIC(s) false
431 #define SCHEDULE_HAS_NONMONOTONIC(s) false
432 #define SCHEDULE_HAS_NO_MODIFIERS(s) true
439 typedef union kmp_r_sched {
458 enum clock_function_type {
459 clock_function_gettimeofday,
460 clock_function_clock_gettime
464 #if KMP_MIC_SUPPORTED
465 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
470 #undef KMP_FAST_REDUCTION_BARRIER
471 #define KMP_FAST_REDUCTION_BARRIER 1
473 #undef KMP_FAST_REDUCTION_CORE_DUO
474 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
475 #define KMP_FAST_REDUCTION_CORE_DUO 1
478 enum _reduction_method {
479 reduction_method_not_defined = 0,
480 critical_reduce_block = (1 << 8),
481 atomic_reduce_block = (2 << 8),
482 tree_reduce_block = (3 << 8),
483 empty_reduce_block = (4 << 8)
498 #if KMP_FAST_REDUCTION_BARRIER
499 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
500 ((reduction_method) | (barrier_type))
502 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
503 ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
505 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
506 ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
508 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
511 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
512 (packed_reduction_method)
514 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
517 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
518 ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
519 (which_reduction_block))
521 #if KMP_FAST_REDUCTION_BARRIER
522 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
523 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
525 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
526 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
529 typedef int PACKED_REDUCTION_METHOD_T;
535 #pragma warning(push)
536 #pragma warning(disable : 271 310)
547 #if KMP_AFFINITY_SUPPORTED
552 typedef struct GROUP_AFFINITY {
558 #if KMP_GROUP_AFFINITY
559 extern int __kmp_num_proc_groups;
561 static const int __kmp_num_proc_groups = 1;
563 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
564 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
566 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
567 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
569 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
570 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
572 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE,
const GROUP_AFFINITY *,
574 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
578 extern hwloc_topology_t __kmp_hwloc_topology;
579 extern int __kmp_hwloc_error;
580 extern int __kmp_numa_detected;
581 extern int __kmp_tile_depth;
584 extern size_t __kmp_affin_mask_size;
585 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
586 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
587 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
588 #define KMP_CPU_SET_ITERATE(i, mask) \
589 for (i = (mask)->begin(); i != (mask)->end(); i = (mask)->next(i))
590 #define KMP_CPU_SET(i, mask) (mask)->set(i)
591 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
592 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
593 #define KMP_CPU_ZERO(mask) (mask)->zero()
594 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
595 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
596 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
597 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
598 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
599 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
600 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
601 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
602 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
603 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
604 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
605 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
606 (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
607 #define KMP_CPU_FREE_ARRAY(arr, n) \
608 __kmp_affinity_dispatch->deallocate_mask_array(arr)
609 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
610 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
611 #define __kmp_get_system_affinity(mask, abort_bool) \
612 (mask)->get_system_affinity(abort_bool)
613 #define __kmp_set_system_affinity(mask, abort_bool) \
614 (mask)->set_system_affinity(abort_bool)
615 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
621 void *
operator new(
size_t n);
622 void operator delete(
void *p);
623 void *
operator new[](
size_t n);
624 void operator delete[](
void *p);
627 virtual void set(
int i) {}
629 virtual bool is_set(
int i)
const {
return false; }
631 virtual void clear(
int i) {}
633 virtual void zero() {}
635 virtual void copy(
const Mask *src) {}
637 virtual void bitwise_and(
const Mask *rhs) {}
639 virtual void bitwise_or(
const Mask *rhs) {}
641 virtual void bitwise_not() {}
644 virtual int begin()
const {
return 0; }
645 virtual int end()
const {
return 0; }
646 virtual int next(
int previous)
const {
return 0; }
648 virtual int set_system_affinity(
bool abort_on_error)
const {
return -1; }
650 virtual int get_system_affinity(
bool abort_on_error) {
return -1; }
653 virtual int get_proc_group()
const {
return -1; }
655 void *
operator new(
size_t n);
656 void operator delete(
void *p);
658 virtual ~KMPAffinity() =
default;
660 virtual void determine_capable(
const char *env_var) {}
662 virtual void bind_thread(
int proc) {}
664 virtual Mask *allocate_mask() {
return nullptr; }
665 virtual void deallocate_mask(Mask *m) {}
666 virtual Mask *allocate_mask_array(
int num) {
return nullptr; }
667 virtual void deallocate_mask_array(Mask *m) {}
668 virtual Mask *index_mask_array(Mask *m,
int index) {
return nullptr; }
669 static void pick_api();
670 static void destroy_api();
678 virtual api_type get_api_type()
const {
684 static bool picked_api;
687 typedef KMPAffinity::Mask kmp_affin_mask_t;
688 extern KMPAffinity *__kmp_affinity_dispatch;
692 #define KMP_AFFIN_MASK_PRINT_LEN 1024
707 affinity_gran_fine = 0,
708 affinity_gran_thread,
712 affinity_gran_package,
714 #if KMP_GROUP_AFFINITY
719 affinity_gran_default
722 enum affinity_top_method {
723 affinity_top_method_all = 0,
724 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
725 affinity_top_method_apicid,
726 affinity_top_method_x2apicid,
728 affinity_top_method_cpuinfo,
729 #if KMP_GROUP_AFFINITY
730 affinity_top_method_group,
732 affinity_top_method_flat,
734 affinity_top_method_hwloc,
736 affinity_top_method_default
739 #define affinity_respect_mask_default (-1)
741 extern enum affinity_type __kmp_affinity_type;
742 extern enum affinity_gran __kmp_affinity_gran;
743 extern int __kmp_affinity_gran_levels;
744 extern int __kmp_affinity_dups;
745 extern enum affinity_top_method __kmp_affinity_top_method;
746 extern int __kmp_affinity_compact;
747 extern int __kmp_affinity_offset;
748 extern int __kmp_affinity_verbose;
749 extern int __kmp_affinity_warnings;
750 extern int __kmp_affinity_respect_mask;
751 extern char *__kmp_affinity_proclist;
752 extern kmp_affin_mask_t *__kmp_affinity_masks;
753 extern unsigned __kmp_affinity_num_masks;
754 extern void __kmp_affinity_bind_thread(
int which);
756 extern kmp_affin_mask_t *__kmp_affin_fullMask;
757 extern char *__kmp_cpuinfo_file;
764 typedef enum kmp_proc_bind_t {
774 typedef struct kmp_nested_proc_bind_t {
775 kmp_proc_bind_t *bind_types;
778 } kmp_nested_proc_bind_t;
780 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
784 #if KMP_AFFINITY_SUPPORTED
785 #define KMP_PLACE_ALL (-1)
786 #define KMP_PLACE_UNDEFINED (-2)
788 #define KMP_AFFINITY_NON_PROC_BIND \
789 ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
790 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
791 (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
794 extern int __kmp_affinity_num_places;
797 typedef enum kmp_cancel_kind_t {
804 #endif // OMP_40_ENABLED
807 typedef struct kmp_hws_item {
812 extern kmp_hws_item_t __kmp_hws_socket;
813 extern kmp_hws_item_t __kmp_hws_node;
814 extern kmp_hws_item_t __kmp_hws_tile;
815 extern kmp_hws_item_t __kmp_hws_core;
816 extern kmp_hws_item_t __kmp_hws_proc;
817 extern int __kmp_hws_requested;
818 extern int __kmp_hws_abs_flag;
822 #define KMP_PAD(type, sz) \
823 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
827 #define KMP_GTID_DNE (-2)
828 #define KMP_GTID_SHUTDOWN (-3)
829 #define KMP_GTID_MONITOR (-4)
830 #define KMP_GTID_UNKNOWN (-5)
831 #define KMP_GTID_MIN (-6)
833 #define __kmp_get_gtid() __kmp_get_global_thread_id()
834 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
836 #define __kmp_tid_from_gtid(gtid) \
837 (KMP_DEBUG_ASSERT((gtid) >= 0), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid)
839 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
840 #define __kmp_gtid_from_tid(tid, team) \
841 (KMP_DEBUG_ASSERT((tid) >= 0 && (team) != NULL), \
842 team->t.t_threads[(tid)]->th.th_info.ds.ds_gtid)
844 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
845 #define __kmp_team_from_gtid(gtid) \
846 (KMP_DEBUG_ASSERT((gtid) >= 0), __kmp_threads[(gtid)]->th.th_team)
848 #define __kmp_thread_from_gtid(gtid) \
849 (KMP_DEBUG_ASSERT((gtid) >= 0), __kmp_threads[(gtid)])
850 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
854 #define __kmp_gtid_from_thread(thr) \
855 (KMP_DEBUG_ASSERT((thr) != NULL), (thr)->th.th_info.ds.ds_gtid)
860 #define __kmp_get_team_num_threads(gtid) \
861 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
865 #define KMP_UINT64_MAX \
866 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
868 #define KMP_MIN_NTH 1
871 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
872 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
874 #define KMP_MAX_NTH INT_MAX
878 #ifdef PTHREAD_STACK_MIN
879 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
881 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
884 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
887 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
888 #elif KMP_ARCH_X86_64
889 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
890 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
892 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
895 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
896 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
897 #define KMP_MAX_MALLOC_POOL_INCR \
898 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
900 #define KMP_MIN_STKOFFSET (0)
901 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
903 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
905 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
908 #define KMP_MIN_STKPADDING (0)
909 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
911 #define KMP_BLOCKTIME_MULTIPLIER \
913 #define KMP_MIN_BLOCKTIME (0)
914 #define KMP_MAX_BLOCKTIME \
916 #define KMP_DEFAULT_BLOCKTIME (200)
919 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
920 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
921 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
925 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
926 (((blocktime) == KMP_MAX_BLOCKTIME) \
927 ? (monitor_wakeups) \
928 : ((blocktime) == KMP_MIN_BLOCKTIME) \
929 ? KMP_MAX_MONITOR_WAKEUPS \
930 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
931 ? (monitor_wakeups) \
932 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
936 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
937 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
938 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
940 #define KMP_BLOCKTIME(team, tid) \
941 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
942 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
944 extern kmp_uint64 __kmp_ticks_per_msec;
946 #define KMP_NOW() ((kmp_uint64)_rdtsc())
948 #define KMP_NOW() __kmp_hardware_timestamp()
950 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
951 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
952 (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
953 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
956 extern kmp_uint64 __kmp_now_nsec();
957 #define KMP_NOW() __kmp_now_nsec()
958 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
959 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
960 (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
961 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
963 #define KMP_YIELD_NOW() \
964 (KMP_NOW_MSEC() / KMP_MAX(__kmp_dflt_blocktime, 1) % \
965 (__kmp_yield_on_count + __kmp_yield_off_count) < \
966 (kmp_uint32)__kmp_yield_on_count)
967 #endif // KMP_USE_MONITOR
969 #define KMP_MIN_STATSCOLS 40
970 #define KMP_MAX_STATSCOLS 4096
971 #define KMP_DEFAULT_STATSCOLS 80
973 #define KMP_MIN_INTERVAL 0
974 #define KMP_MAX_INTERVAL (INT_MAX - 1)
975 #define KMP_DEFAULT_INTERVAL 0
977 #define KMP_MIN_CHUNK 1
978 #define KMP_MAX_CHUNK (INT_MAX - 1)
979 #define KMP_DEFAULT_CHUNK 1
981 #define KMP_MIN_INIT_WAIT 1
982 #define KMP_MAX_INIT_WAIT (INT_MAX / 2)
983 #define KMP_DEFAULT_INIT_WAIT 2048U
985 #define KMP_MIN_NEXT_WAIT 1
986 #define KMP_MAX_NEXT_WAIT (INT_MAX / 2)
987 #define KMP_DEFAULT_NEXT_WAIT 1024U
989 #define KMP_DFLT_DISP_NUM_BUFF 7
990 #define KMP_MAX_ORDERED 8
992 #define KMP_MAX_FIELDS 32
994 #define KMP_MAX_BRANCH_BITS 31
996 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
998 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1000 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1005 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1006 #define KMP_TLS_GTID_MIN 5
1008 #define KMP_TLS_GTID_MIN INT_MAX
1011 #define KMP_MASTER_TID(tid) ((tid) == 0)
1012 #define KMP_WORKER_TID(tid) ((tid) != 0)
1014 #define KMP_MASTER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) == 0)
1015 #define KMP_WORKER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) != 0)
1016 #define KMP_UBER_GTID(gtid) \
1017 (KMP_DEBUG_ASSERT((gtid) >= KMP_GTID_MIN), \
1018 KMP_DEBUG_ASSERT((gtid) < __kmp_threads_capacity), \
1019 (gtid) >= 0 && __kmp_root[(gtid)] && __kmp_threads[(gtid)] && \
1020 (__kmp_threads[(gtid)] == __kmp_root[(gtid)]->r.r_uber_thread))
1021 #define KMP_INITIAL_GTID(gtid) ((gtid) == 0)
1025 #define TRUE (!FALSE)
1031 #define KMP_INIT_WAIT 64U
1032 #define KMP_NEXT_WAIT 32U
1034 #define KMP_INIT_WAIT 16U
1035 #define KMP_NEXT_WAIT 8U
1037 #define KMP_INIT_WAIT 1024U
1038 #define KMP_NEXT_WAIT 512U
1041 #define KMP_INIT_WAIT 1024U
1042 #define KMP_NEXT_WAIT 512U
1043 #elif KMP_OS_FREEBSD
1045 #define KMP_INIT_WAIT 1024U
1046 #define KMP_NEXT_WAIT 512U
1049 #define KMP_INIT_WAIT 1024U
1050 #define KMP_NEXT_WAIT 512U
1053 #define KMP_INIT_WAIT 1024U
1054 #define KMP_NEXT_WAIT 512U
1055 #elif KMP_OS_KFREEBSD
1057 #define KMP_INIT_WAIT 1024U
1058 #define KMP_NEXT_WAIT 512U
1061 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1062 typedef struct kmp_cpuid {
1068 extern void __kmp_x86_cpuid(
int mode,
int mode2,
struct kmp_cpuid *p);
1070 extern void __kmp_x86_pause(
void);
1076 static void __kmp_x86_pause(
void) { _mm_delay_32(300); }
1078 static void __kmp_x86_pause(
void) { _mm_pause(); }
1080 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1081 #elif KMP_ARCH_PPC64
1082 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1083 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1084 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1085 #define KMP_CPU_PAUSE() \
1087 KMP_PPC64_PRI_LOW(); \
1088 KMP_PPC64_PRI_MED(); \
1089 KMP_PPC64_PRI_LOC_MB(); \
1092 #define KMP_CPU_PAUSE()
1095 #define KMP_INIT_YIELD(count) \
1096 { (count) = __kmp_yield_init; }
1098 #define KMP_YIELD(cond) \
1101 __kmp_yield((cond)); \
1107 #define KMP_YIELD_WHEN(cond, count) \
1112 __kmp_yield(cond); \
1113 (count) = __kmp_yield_next; \
1116 #define KMP_YIELD_SPIN(count) \
1122 (count) = __kmp_yield_next; \
1146 ct_ordered_in_parallel,
1148 ct_ordered_in_taskq,
1155 #define IS_CONS_TYPE_TASKQ(ct) \
1156 (((int)(ct)) >= ((int)ct_taskq) && ((int)(ct)) <= ((int)ct_task_ordered))
1157 #define IS_CONS_TYPE_ORDERED(ct) \
1158 ((ct) == ct_pdo_ordered || (ct) == ct_task_ordered)
1162 enum cons_type type;
1168 struct cons_header {
1169 int p_top, w_top, s_top;
1170 int stack_size, stack_top;
1171 struct cons_data *stack_data;
1174 struct kmp_region_info {
1176 int offset[KMP_MAX_FIELDS];
1177 int length[KMP_MAX_FIELDS];
1184 typedef HANDLE kmp_thread_t;
1185 typedef DWORD kmp_key_t;
1189 typedef pthread_t kmp_thread_t;
1190 typedef pthread_key_t kmp_key_t;
1193 extern kmp_key_t __kmp_gtid_threadprivate_key;
1195 typedef struct kmp_sys_info {
1206 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1207 typedef struct kmp_cpuinfo {
1216 int cpu_stackoffset;
1220 kmp_uint64 frequency;
1221 char name[3 *
sizeof(kmp_cpuid_t)];
1228 typedef int kmp_itt_mark_t;
1229 #define KMP_ITT_DEBUG 0
1234 #define HIGH_WATER_MARK(nslots) (((nslots)*3) / 4)
1236 #define __KMP_TASKQ_THUNKS_PER_TH 1
1241 #define TQF_IS_ORDERED 0x0001 // __kmpc_taskq interface, taskq ordered
1243 #define TQF_IS_LASTPRIVATE 0x0002
1244 #define TQF_IS_NOWAIT 0x0004 // __kmpc_taskq interface, end taskq nowait
1246 #define TQF_HEURISTICS 0x0008
1249 #define TQF_INTERFACE_RESERVED1 0x0010
1251 #define TQF_INTERFACE_RESERVED2 0x0020
1253 #define TQF_INTERFACE_RESERVED3 0x0040
1255 #define TQF_INTERFACE_RESERVED4 0x0080
1257 #define TQF_INTERFACE_FLAGS 0x00ff // all the __kmpc_taskq interface flags
1259 #define TQF_IS_LAST_TASK 0x0100
1261 #define TQF_TASKQ_TASK 0x0200
1264 #define TQF_RELEASE_WORKERS 0x0400
1266 #define TQF_ALL_TASKS_QUEUED 0x0800
1268 #define TQF_PARALLEL_CONTEXT 0x1000
1270 #define TQF_DEALLOCATED 0x2000
1272 #define TQF_INTERNAL_FLAGS 0x3f00 // all the internal use only flags
1274 typedef struct KMP_ALIGN_CACHE kmpc_aligned_int32_t {
1276 } kmpc_aligned_int32_t;
1278 typedef struct KMP_ALIGN_CACHE kmpc_aligned_queue_slot_t {
1279 struct kmpc_thunk_t *qs_thunk;
1280 } kmpc_aligned_queue_slot_t;
1282 typedef struct kmpc_task_queue_t {
1285 kmp_lock_t tq_link_lck;
1288 struct kmpc_task_queue_t *tq_parent;
1290 struct kmpc_task_queue_t *tq_next_free;
1293 volatile struct kmpc_task_queue_t *tq_first_child;
1295 struct kmpc_task_queue_t *tq_next_child;
1297 struct kmpc_task_queue_t *tq_prev_child;
1299 volatile kmp_int32 tq_ref_count;
1305 struct kmpc_aligned_shared_vars_t *tq_shareds;
1309 kmp_uint32 tq_tasknum_queuing;
1311 volatile kmp_uint32 tq_tasknum_serving;
1314 kmp_lock_t tq_free_thunks_lck;
1316 struct kmpc_thunk_t *tq_free_thunks;
1318 struct kmpc_thunk_t *tq_thunk_space;
1321 kmp_lock_t tq_queue_lck;
1324 kmpc_aligned_queue_slot_t *tq_queue;
1325 volatile struct kmpc_thunk_t *tq_taskq_slot;
1327 kmp_int32 tq_nslots;
1331 volatile kmp_int32 tq_nfull;
1333 volatile kmp_int32 tq_flags;
1338 struct kmpc_aligned_int32_t *tq_th_thunks;
1343 } kmpc_task_queue_t;
1345 typedef void (*kmpc_task_t)(kmp_int32 global_tid,
struct kmpc_thunk_t *thunk);
1348 typedef struct kmpc_shared_vars_t {
1349 kmpc_task_queue_t *sv_queue;
1350 } kmpc_shared_vars_t;
1352 typedef struct KMP_ALIGN_CACHE kmpc_aligned_shared_vars_t {
1353 volatile struct kmpc_shared_vars_t *ai_data;
1354 } kmpc_aligned_shared_vars_t;
1357 typedef struct kmpc_thunk_t {
1359 kmpc_shared_vars_t *th_shareds;
1360 struct kmpc_thunk_t *th_next_free;
1363 kmpc_task_t th_task;
1364 struct kmpc_thunk_t *th_encl_thunk;
1369 kmp_int32 th_status;
1370 kmp_uint32 th_tasknum;
1375 typedef struct KMP_ALIGN_CACHE kmp_taskq {
1376 int tq_curr_thunk_capacity;
1378 kmpc_task_queue_t *tq_root;
1379 kmp_int32 tq_global_flags;
1381 kmp_lock_t tq_freelist_lck;
1382 kmpc_task_queue_t *tq_freelist;
1384 kmpc_thunk_t **tq_curr_thunk;
1389 typedef kmp_int32 kmp_critical_name[8];
1400 typedef void (*
kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1401 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1418 typedef void *(*kmpc_ctor)(
void *);
1431 typedef void *(*kmpc_cctor)(
void *,
void *);
1441 typedef void *(*kmpc_ctor_vec)(
void *, size_t);
1453 typedef void *(*kmpc_cctor_vec)(
void *,
void *,
1461 typedef struct kmp_cached_addr {
1463 void ***compiler_cache;
1465 struct kmp_cached_addr *next;
1466 } kmp_cached_addr_t;
1468 struct private_data {
1469 struct private_data *next;
1475 struct private_common {
1476 struct private_common *next;
1477 struct private_common *link;
1483 struct shared_common {
1484 struct shared_common *next;
1485 struct private_data *pod_init;
1505 #define KMP_HASH_TABLE_LOG2 9
1506 #define KMP_HASH_TABLE_SIZE \
1507 (1 << KMP_HASH_TABLE_LOG2)
1508 #define KMP_HASH_SHIFT 3
1509 #define KMP_HASH(x) \
1510 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1512 struct common_table {
1513 struct private_common *data[KMP_HASH_TABLE_SIZE];
1516 struct shared_table {
1517 struct shared_common *data[KMP_HASH_TABLE_SIZE];
1522 #if KMP_USE_HIER_SCHED
1525 typedef struct kmp_hier_private_bdata_t {
1526 kmp_int32 num_active;
1528 kmp_uint64 wait_val[2];
1529 } kmp_hier_private_bdata_t;
1532 typedef struct kmp_sched_flags {
1533 unsigned ordered : 1;
1534 unsigned nomerge : 1;
1535 unsigned contains_last : 1;
1536 #if KMP_USE_HIER_SCHED
1537 unsigned use_hier : 1;
1538 unsigned unused : 28;
1540 unsigned unused : 29;
1542 } kmp_sched_flags_t;
1544 KMP_BUILD_ASSERT(
sizeof(kmp_sched_flags_t) == 4);
1546 #if KMP_STATIC_STEAL_ENABLED
1547 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1554 kmp_int32 static_steal_counter;
1563 struct KMP_ALIGN(32) {
1570 kmp_uint32 ordered_lower;
1571 kmp_uint32 ordered_upper;
1576 kmp_int32 last_upper;
1578 } dispatch_private_info32_t;
1580 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1587 kmp_int64 static_steal_counter;
1598 struct KMP_ALIGN(32) {
1605 kmp_uint64 ordered_lower;
1606 kmp_uint64 ordered_upper;
1611 kmp_int64 last_upper;
1613 } dispatch_private_info64_t;
1615 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1628 kmp_uint32 ordered_lower;
1629 kmp_uint32 ordered_upper;
1631 kmp_int32 last_upper;
1633 } dispatch_private_info32_t;
1635 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1649 kmp_uint64 ordered_lower;
1650 kmp_uint64 ordered_upper;
1652 kmp_int64 last_upper;
1654 } dispatch_private_info64_t;
1657 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1658 union private_info {
1659 dispatch_private_info32_t p32;
1660 dispatch_private_info64_t p64;
1663 kmp_sched_flags_t flags;
1664 kmp_int32 ordered_bumped;
1666 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 3];
1668 struct dispatch_private_info *next;
1669 kmp_int32 type_size;
1670 #if KMP_USE_HIER_SCHED
1674 enum cons_type pushed_ws;
1675 } dispatch_private_info_t;
1677 typedef struct dispatch_shared_info32 {
1680 volatile kmp_uint32 iteration;
1681 volatile kmp_uint32 num_done;
1682 volatile kmp_uint32 ordered_iteration;
1684 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1685 } dispatch_shared_info32_t;
1687 typedef struct dispatch_shared_info64 {
1690 volatile kmp_uint64 iteration;
1691 volatile kmp_uint64 num_done;
1692 volatile kmp_uint64 ordered_iteration;
1694 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1695 } dispatch_shared_info64_t;
1697 typedef struct dispatch_shared_info {
1699 dispatch_shared_info32_t s32;
1700 dispatch_shared_info64_t s64;
1702 volatile kmp_uint32 buffer_index;
1704 volatile kmp_int32 doacross_buf_idx;
1705 volatile kmp_uint32 *doacross_flags;
1706 kmp_int32 doacross_num_done;
1708 #if KMP_USE_HIER_SCHED
1717 } dispatch_shared_info_t;
1719 typedef struct kmp_disp {
1721 void (*th_deo_fcn)(
int *gtid,
int *cid,
ident_t *);
1723 void (*th_dxo_fcn)(
int *gtid,
int *cid,
ident_t *);
1725 dispatch_shared_info_t *th_dispatch_sh_current;
1726 dispatch_private_info_t *th_dispatch_pr_current;
1728 dispatch_private_info_t *th_disp_buffer;
1729 kmp_int32 th_disp_index;
1731 kmp_int32 th_doacross_buf_idx;
1732 volatile kmp_uint32 *th_doacross_flags;
1735 kmp_int64 *th_doacross_info;
1736 kmp_lock_t *th_steal_lock;
1739 #if KMP_STATIC_STEAL_ENABLED
1740 kmp_lock_t *th_steal_lock;
1741 void *dummy_padding[1];
1743 void *dummy_padding[2];
1746 #if KMP_USE_INTERNODE_ALIGNMENT
1747 char more_padding[INTERNODE_CACHE_LINE];
1755 #define KMP_INIT_BARRIER_STATE 0
1756 #define KMP_BARRIER_SLEEP_BIT 0
1757 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1758 #define KMP_BARRIER_BUMP_BIT 2
1760 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1761 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1762 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1764 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1765 #error "Barrier sleep bit must be smaller than barrier bump bit"
1767 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1768 #error "Barrier unused bit must be smaller than barrier bump bit"
1772 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1773 #define KMP_BARRIER_OWN_FLAG \
1774 1 // Normal state; worker waiting on own b_go flag in release
1775 #define KMP_BARRIER_PARENT_FLAG \
1776 2 // Special state; worker waiting on parent's b_go flag in release
1777 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1778 3 // Special state; tells worker to shift from parent to own b_go
1779 #define KMP_BARRIER_SWITCHING \
1780 4 // Special state; worker resets appropriate flag on wake-up
1782 #define KMP_NOT_SAFE_TO_REAP \
1783 0 // Thread th_reap_state: not safe to reap (tasking)
1784 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1787 bs_plain_barrier = 0,
1789 bs_forkjoin_barrier,
1790 #if KMP_FAST_REDUCTION_BARRIER
1791 bs_reduction_barrier,
1792 #endif // KMP_FAST_REDUCTION_BARRIER
1797 #if !KMP_FAST_REDUCTION_BARRIER
1798 #define bs_reduction_barrier bs_plain_barrier
1799 #endif // KMP_FAST_REDUCTION_BARRIER
1801 typedef enum kmp_bar_pat {
1809 bp_hierarchical_bar = 3,
1813 #define KMP_BARRIER_ICV_PUSH 1
1816 typedef struct kmp_internal_control {
1817 int serial_nesting_level;
1830 int max_active_levels;
1834 kmp_proc_bind_t proc_bind;
1835 kmp_int32 default_device;
1836 #endif // OMP_40_ENABLED
1837 struct kmp_internal_control *next;
1838 } kmp_internal_control_t;
1840 static inline void copy_icvs(kmp_internal_control_t *dst,
1841 kmp_internal_control_t *src) {
1846 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1851 kmp_internal_control_t th_fixed_icvs;
1854 volatile kmp_uint64 b_go;
1855 KMP_ALIGN_CACHE
volatile kmp_uint64
1857 kmp_uint32 *skip_per_level;
1858 kmp_uint32 my_level;
1859 kmp_int32 parent_tid;
1862 struct kmp_bstate *parent_bar;
1864 kmp_uint64 leaf_state;
1866 kmp_uint8 base_leaf_kids;
1867 kmp_uint8 leaf_kids;
1869 kmp_uint8 wait_flag;
1870 kmp_uint8 use_oncore_barrier;
1875 KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1879 union KMP_ALIGN_CACHE kmp_barrier_union {
1881 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1885 typedef union kmp_barrier_union kmp_balign_t;
1888 union KMP_ALIGN_CACHE kmp_barrier_team_union {
1890 char b_pad[CACHE_LINE];
1892 kmp_uint64 b_arrived;
1898 kmp_uint b_master_arrived;
1899 kmp_uint b_team_arrived;
1904 typedef union kmp_barrier_team_union kmp_balign_team_t;
1911 typedef struct kmp_win32_mutex {
1913 CRITICAL_SECTION cs;
1914 } kmp_win32_mutex_t;
1916 typedef struct kmp_win32_cond {
1921 kmp_win32_mutex_t waiters_count_lock_;
1928 int wait_generation_count_;
1937 union KMP_ALIGN_CACHE kmp_cond_union {
1939 char c_pad[CACHE_LINE];
1940 pthread_cond_t c_cond;
1943 typedef union kmp_cond_union kmp_cond_align_t;
1945 union KMP_ALIGN_CACHE kmp_mutex_union {
1947 char m_pad[CACHE_LINE];
1948 pthread_mutex_t m_mutex;
1951 typedef union kmp_mutex_union kmp_mutex_align_t;
1955 typedef struct kmp_desc_base {
1957 size_t ds_stacksize;
1959 kmp_thread_t ds_thread;
1960 volatile int ds_tid;
1963 volatile int ds_alive;
1980 typedef union KMP_ALIGN_CACHE kmp_desc {
1982 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
1986 typedef struct kmp_local {
1987 volatile int this_construct;
1992 #if !USE_CMP_XCHG_FOR_BGET
1993 #ifdef USE_QUEUING_LOCK_FOR_BGET
1994 kmp_lock_t bget_lock;
1996 kmp_bootstrap_lock_t bget_lock;
2003 PACKED_REDUCTION_METHOD_T
2004 packed_reduction_method;
2009 #define KMP_CHECK_UPDATE(a, b) \
2012 #define KMP_CHECK_UPDATE_SYNC(a, b) \
2014 TCW_SYNC_PTR((a), (b))
2016 #define get__blocktime(xteam, xtid) \
2017 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2018 #define get__bt_set(xteam, xtid) \
2019 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2021 #define get__bt_intervals(xteam, xtid) \
2022 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2025 #define get__nested_2(xteam, xtid) \
2026 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nested)
2027 #define get__dynamic_2(xteam, xtid) \
2028 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2029 #define get__nproc_2(xteam, xtid) \
2030 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2031 #define get__sched_2(xteam, xtid) \
2032 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2034 #define set__blocktime_team(xteam, xtid, xval) \
2035 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2039 #define set__bt_intervals_team(xteam, xtid, xval) \
2040 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2044 #define set__bt_set_team(xteam, xtid, xval) \
2045 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2047 #define set__nested(xthread, xval) \
2048 (((xthread)->th.th_current_task->td_icvs.nested) = (xval))
2049 #define get__nested(xthread) \
2050 (((xthread)->th.th_current_task->td_icvs.nested) ? (FTN_TRUE) : (FTN_FALSE))
2052 #define set__dynamic(xthread, xval) \
2053 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2054 #define get__dynamic(xthread) \
2055 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2057 #define set__nproc(xthread, xval) \
2058 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2060 #define set__max_active_levels(xthread, xval) \
2061 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2063 #define set__sched(xthread, xval) \
2064 (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2068 #define set__proc_bind(xthread, xval) \
2069 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2070 #define get__proc_bind(xthread) \
2071 ((xthread)->th.th_current_task->td_icvs.proc_bind)
2077 typedef enum kmp_tasking_mode {
2078 tskm_immediate_exec = 0,
2079 tskm_extra_barrier = 1,
2080 tskm_task_teams = 2,
2082 } kmp_tasking_mode_t;
2084 extern kmp_tasking_mode_t
2086 extern int __kmp_task_stealing_constraint;
2088 extern kmp_int32 __kmp_default_device;
2093 extern kmp_int32 __kmp_max_task_priority;
2095 extern kmp_uint64 __kmp_taskloop_min_tasks;
2100 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2101 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2105 #define KMP_TASKING_ENABLED(task_team) \
2106 (TCR_SYNC_4((task_team)->tt.tt_found_tasks) == TRUE)
2114 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32,
void *);
2116 #if OMP_40_ENABLED || OMP_45_ENABLED
2117 typedef union kmp_cmplrdata {
2120 #endif // OMP_45_ENABLED
2125 #endif // OMP_40_ENABLED
2133 typedef struct kmp_task {
2138 #if OMP_40_ENABLED || OMP_45_ENABLED
2141 kmp_cmplrdata_t data2;
2152 typedef struct kmp_taskgroup {
2153 std::atomic<kmp_int32> count;
2154 std::atomic<kmp_int32>
2156 struct kmp_taskgroup *parent;
2161 kmp_int32 reduce_num_data;
2166 typedef union kmp_depnode kmp_depnode_t;
2167 typedef struct kmp_depnode_list kmp_depnode_list_t;
2168 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2170 typedef struct kmp_depend_info {
2171 kmp_intptr_t base_addr;
2177 } kmp_depend_info_t;
2179 struct kmp_depnode_list {
2180 kmp_depnode_t *node;
2181 kmp_depnode_list_t *next;
2184 typedef struct kmp_base_depnode {
2185 kmp_depnode_list_t *successors;
2190 #if KMP_SUPPORT_GRAPH_OUTPUT
2194 std::atomic<kmp_int32> npredecessors;
2195 std::atomic<kmp_int32> nrefs;
2196 } kmp_base_depnode_t;
2198 union KMP_ALIGN_CACHE kmp_depnode {
2200 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2201 kmp_base_depnode_t dn;
2204 struct kmp_dephash_entry {
2206 kmp_depnode_t *last_out;
2207 kmp_depnode_list_t *last_ins;
2208 kmp_dephash_entry_t *next_in_bucket;
2211 typedef struct kmp_dephash {
2212 kmp_dephash_entry_t **buckets;
2215 kmp_uint32 nelements;
2216 kmp_uint32 nconflicts;
2222 #ifdef BUILD_TIED_TASK_STACK
2225 typedef struct kmp_stack_block {
2226 kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2227 struct kmp_stack_block *sb_next;
2228 struct kmp_stack_block *sb_prev;
2229 } kmp_stack_block_t;
2231 typedef struct kmp_task_stack {
2232 kmp_stack_block_t ts_first_block;
2233 kmp_taskdata_t **ts_top;
2234 kmp_int32 ts_entries;
2237 #endif // BUILD_TIED_TASK_STACK
2239 typedef struct kmp_tasking_flags {
2241 unsigned tiedness : 1;
2243 unsigned merged_if0 : 1;
2246 unsigned destructors_thunk : 1;
2251 unsigned priority_specified : 1;
2253 unsigned reserved : 10;
2255 unsigned reserved : 12;
2257 #else // OMP_40_ENABLED
2258 unsigned reserved : 13;
2259 #endif // OMP_40_ENABLED
2262 unsigned tasktype : 1;
2263 unsigned task_serial : 1;
2264 unsigned tasking_ser : 1;
2266 unsigned team_serial : 1;
2270 unsigned started : 1;
2271 unsigned executing : 1;
2272 unsigned complete : 1;
2274 unsigned native : 1;
2275 unsigned reserved31 : 7;
2277 } kmp_tasking_flags_t;
2279 struct kmp_taskdata {
2280 kmp_int32 td_task_id;
2281 kmp_tasking_flags_t td_flags;
2282 kmp_team_t *td_team;
2283 kmp_info_p *td_alloc_thread;
2285 kmp_taskdata_t *td_parent;
2287 std::atomic<kmp_int32> td_untied_count;
2291 kmp_uint32 td_taskwait_counter;
2292 kmp_int32 td_taskwait_thread;
2293 KMP_ALIGN_CACHE kmp_internal_control_t
2295 KMP_ALIGN_CACHE std::atomic<kmp_int32>
2296 td_allocated_child_tasks;
2298 std::atomic<kmp_int32>
2299 td_incomplete_child_tasks;
2307 #endif // OMP_40_ENABLED
2309 kmp_task_team_t *td_task_team;
2310 kmp_int32 td_size_alloc;
2311 #if defined(KMP_GOMP_COMPAT)
2313 kmp_int32 td_size_loop_bounds;
2315 #endif // OMP_45_ENABLED
2316 kmp_taskdata_t *td_last_tied;
2317 #if defined(KMP_GOMP_COMPAT) && OMP_45_ENABLED
2319 void (*td_copy_func)(
void *,
void *);
2322 ompt_task_info_t ompt_task_info;
2327 KMP_BUILD_ASSERT(
sizeof(kmp_taskdata_t) %
sizeof(
void *) == 0);
2330 typedef struct kmp_base_thread_data {
2334 kmp_bootstrap_lock_t td_deque_lock;
2337 kmp_int32 td_deque_size;
2338 kmp_uint32 td_deque_head;
2339 kmp_uint32 td_deque_tail;
2340 kmp_int32 td_deque_ntasks;
2342 kmp_int32 td_deque_last_stolen;
2343 #ifdef BUILD_TIED_TASK_STACK
2344 kmp_task_stack_t td_susp_tied_tasks;
2346 #endif // BUILD_TIED_TASK_STACK
2347 } kmp_base_thread_data_t;
2349 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2350 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2352 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2353 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2355 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2356 kmp_base_thread_data_t td;
2358 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2359 } kmp_thread_data_t;
2362 typedef struct kmp_base_task_team {
2363 kmp_bootstrap_lock_t
2366 kmp_task_team_t *tt_next;
2370 kmp_int32 tt_found_tasks;
2378 tt_found_proxy_tasks;
2380 kmp_int32 tt_untied_task_encountered;
2383 std::atomic<kmp_int32> tt_unfinished_threads;
2388 } kmp_base_task_team_t;
2390 union KMP_ALIGN_CACHE kmp_task_team {
2391 kmp_base_task_team_t tt;
2393 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2396 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2399 typedef struct kmp_free_list {
2400 void *th_free_list_self;
2401 void *th_free_list_sync;
2403 void *th_free_list_other;
2407 #if KMP_NESTED_HOT_TEAMS
2410 typedef struct kmp_hot_team_ptr {
2411 kmp_team_p *hot_team;
2412 kmp_int32 hot_team_nth;
2413 } kmp_hot_team_ptr_t;
2416 typedef struct kmp_teams_size {
2424 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2429 kmp_team_p *th_team;
2430 kmp_root_p *th_root;
2431 kmp_info_p *th_next_pool;
2432 kmp_disp_t *th_dispatch;
2438 kmp_info_p *th_team_master;
2439 int th_team_serialized;
2441 microtask_t th_teams_microtask;
2451 int th_team_bt_intervals;
2454 kmp_uint64 th_team_bt_intervals;
2457 #if KMP_AFFINITY_SUPPORTED
2458 kmp_affin_mask_t *th_affin_mask;
2464 #if KMP_NESTED_HOT_TEAMS
2465 kmp_hot_team_ptr_t *th_hot_teams;
2472 #if KMP_AFFINITY_SUPPORTED
2473 int th_current_place;
2480 kmp_uint64 th_bar_arrive_time;
2481 kmp_uint64 th_bar_min_time;
2482 kmp_uint64 th_frame_time;
2484 kmp_local_t th_local;
2485 struct private_common *th_pri_head;
2490 KMP_ALIGN_CACHE kmp_team_p
2494 ompt_thread_info_t ompt_thread_info;
2498 struct common_table *th_pri_common;
2500 volatile kmp_uint32 th_spin_here;
2503 volatile void *th_sleep_loc;
2510 kmp_task_team_t *th_task_team;
2511 kmp_taskdata_t *th_current_task;
2512 kmp_uint8 th_task_state;
2513 kmp_uint8 *th_task_state_memo_stack;
2515 kmp_uint32 th_task_state_top;
2516 kmp_uint32 th_task_state_stack_sz;
2517 kmp_uint32 th_reap_state;
2522 kmp_uint8 th_active_in_pool;
2524 struct cons_header *th_cons;
2525 #if KMP_USE_HIER_SCHED
2527 kmp_hier_private_bdata_t *th_hier_bar_data;
2531 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2533 KMP_ALIGN_CACHE
volatile kmp_int32
2536 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2538 kmp_free_list_t th_free_lists[NUM_LISTS];
2543 kmp_win32_cond_t th_suspend_cv;
2544 kmp_win32_mutex_t th_suspend_mx;
2545 int th_suspend_init;
2548 kmp_cond_align_t th_suspend_cv;
2549 kmp_mutex_align_t th_suspend_mx;
2550 int th_suspend_init_count;
2554 kmp_itt_mark_t th_itt_mark_single;
2557 #if KMP_STATS_ENABLED
2558 kmp_stats_list *th_stats;
2561 std::atomic<bool> th_blocking;
2565 typedef union KMP_ALIGN_CACHE kmp_info {
2567 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2573 typedef struct kmp_base_data {
volatile kmp_uint32 t_value; } kmp_base_data_t;
2575 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2577 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2581 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2583 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2585 } kmp_ordered_team_t;
2587 typedef int (*launch_t)(
int gtid);
2590 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2596 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2597 #define KMP_INLINE_ARGV_BYTES \
2599 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2600 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2603 #define KMP_INLINE_ARGV_BYTES \
2604 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2606 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2608 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2611 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2612 kmp_balign_team_t t_bar[bs_last_barrier];
2613 std::atomic<int> t_construct;
2614 char pad[
sizeof(kmp_lock_t)];
2618 KMP_ALIGN_CACHE
int t_master_tid;
2619 int t_master_this_cons;
2623 kmp_team_p *t_parent;
2624 kmp_team_p *t_next_pool;
2625 kmp_disp_t *t_dispatch;
2626 kmp_task_team_t *t_task_team[2];
2628 kmp_proc_bind_t t_proc_bind;
2629 #endif // OMP_40_ENABLED
2631 kmp_uint64 t_region_time;
2636 KMP_ALIGN_CACHE
void **t_argv;
2643 ompt_team_info_t ompt_team_info;
2644 ompt_lw_taskteam_t *ompt_serialized_team_info;
2647 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2648 kmp_int8 t_fp_control_saved;
2650 kmp_int16 t_x87_fpu_control_word;
2654 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2656 KMP_ALIGN_CACHE kmp_info_t **t_threads;
2658 *t_implicit_task_taskdata;
2661 KMP_ALIGN_CACHE
int t_max_argc;
2664 dispatch_shared_info_t *t_disp_buffer;
2667 kmp_r_sched_t t_sched;
2668 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2671 #endif // OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2676 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2681 char dummy_padding[1024];
2684 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2688 std::atomic<kmp_int32> t_cancel_request;
2690 int t_master_active;
2691 kmp_taskq_t t_taskq;
2692 void *t_copypriv_data;
2694 std::atomic<kmp_uint32> t_copyin_counter;
2701 union KMP_ALIGN_CACHE kmp_team {
2704 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2707 typedef union KMP_ALIGN_CACHE kmp_time_global {
2709 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2711 } kmp_time_global_t;
2713 typedef struct kmp_base_global {
2715 kmp_time_global_t g_time;
2718 volatile int g_abort;
2719 volatile int g_done;
2722 enum dynamic_mode g_dynamic_mode;
2723 } kmp_base_global_t;
2725 typedef union KMP_ALIGN_CACHE kmp_global {
2726 kmp_base_global_t g;
2728 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2731 typedef struct kmp_base_root {
2736 volatile int r_active;
2738 volatile int r_nested;
2740 std::atomic<int> r_in_parallel;
2742 kmp_team_t *r_root_team;
2743 kmp_team_t *r_hot_team;
2744 kmp_info_t *r_uber_thread;
2745 kmp_lock_t r_begin_lock;
2746 volatile int r_begin;
2751 typedef union KMP_ALIGN_CACHE kmp_root {
2754 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2757 struct fortran_inx_info {
2763 extern int __kmp_settings;
2764 extern int __kmp_duplicate_library_ok;
2766 extern int __kmp_forkjoin_frames;
2767 extern int __kmp_forkjoin_frames_mode;
2769 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2770 extern int __kmp_determ_red;
2773 extern int kmp_a_debug;
2774 extern int kmp_b_debug;
2775 extern int kmp_c_debug;
2776 extern int kmp_d_debug;
2777 extern int kmp_e_debug;
2778 extern int kmp_f_debug;
2782 #define KMP_DEBUG_BUF_LINES_INIT 512
2783 #define KMP_DEBUG_BUF_LINES_MIN 1
2785 #define KMP_DEBUG_BUF_CHARS_INIT 128
2786 #define KMP_DEBUG_BUF_CHARS_MIN 2
2790 extern int __kmp_debug_buf_lines;
2792 __kmp_debug_buf_chars;
2793 extern int __kmp_debug_buf_atomic;
2796 extern char *__kmp_debug_buffer;
2797 extern std::atomic<int> __kmp_debug_count;
2799 extern int __kmp_debug_buf_warn_chars;
2804 extern int __kmp_par_range;
2806 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2807 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2808 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2809 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2810 extern int __kmp_par_range_lb;
2811 extern int __kmp_par_range_ub;
2817 extern int __kmp_storage_map_verbose;
2819 extern int __kmp_storage_map_verbose_specified;
2821 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2822 extern kmp_cpuinfo_t __kmp_cpuinfo;
2825 extern volatile int __kmp_init_serial;
2826 extern volatile int __kmp_init_gtid;
2827 extern volatile int __kmp_init_common;
2828 extern volatile int __kmp_init_middle;
2829 extern volatile int __kmp_init_parallel;
2831 extern volatile int __kmp_init_monitor;
2833 extern volatile int __kmp_init_user_locks;
2834 extern int __kmp_init_counter;
2835 extern int __kmp_root_counter;
2836 extern int __kmp_version;
2839 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2842 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
2843 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
2844 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
2845 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
2846 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
2847 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
2848 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
2849 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
2850 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
2851 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
2852 extern char const *__kmp_barrier_type_name[bs_last_barrier];
2853 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
2856 extern kmp_bootstrap_lock_t __kmp_initz_lock;
2857 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock;
2858 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
2859 extern kmp_bootstrap_lock_t
2862 extern kmp_bootstrap_lock_t
2865 extern kmp_bootstrap_lock_t
2866 __kmp_tp_cached_lock;
2869 extern kmp_lock_t __kmp_global_lock;
2870 extern kmp_queuing_lock_t __kmp_dispatch_lock;
2871 extern kmp_lock_t __kmp_debug_lock;
2874 extern unsigned int __kmp_init_wait;
2875 extern unsigned int __kmp_next_wait;
2877 extern enum library_type __kmp_library;
2883 extern int __kmp_chunk;
2885 extern size_t __kmp_stksize;
2887 extern size_t __kmp_monitor_stksize;
2889 extern size_t __kmp_stkoffset;
2890 extern int __kmp_stkpadding;
2893 __kmp_malloc_pool_incr;
2894 extern int __kmp_env_stksize;
2895 extern int __kmp_env_blocktime;
2896 extern int __kmp_env_checks;
2897 extern int __kmp_env_consistency_check;
2898 extern int __kmp_generate_warnings;
2899 extern int __kmp_reserve_warn;
2901 #ifdef DEBUG_SUSPEND
2902 extern int __kmp_suspend_count;
2905 extern kmp_uint32 __kmp_yield_init;
2906 extern kmp_uint32 __kmp_yield_next;
2909 extern kmp_uint32 __kmp_yielding_on;
2911 extern kmp_uint32 __kmp_yield_cycle;
2912 extern kmp_int32 __kmp_yield_on_count;
2913 extern kmp_int32 __kmp_yield_off_count;
2916 extern int __kmp_allThreadsSpecified;
2918 extern size_t __kmp_align_alloc;
2920 extern int __kmp_xproc;
2921 extern int __kmp_avail_proc;
2922 extern size_t __kmp_sys_min_stksize;
2923 extern int __kmp_sys_max_nth;
2925 extern int __kmp_max_nth;
2927 extern int __kmp_cg_max_nth;
2928 extern int __kmp_teams_max_nth;
2929 extern int __kmp_threads_capacity;
2931 extern int __kmp_dflt_team_nth;
2933 extern int __kmp_dflt_team_nth_ub;
2935 extern int __kmp_tp_capacity;
2937 extern int __kmp_tp_cached;
2939 extern int __kmp_dflt_nested;
2941 extern int __kmp_dflt_blocktime;
2945 __kmp_monitor_wakeups;
2946 extern int __kmp_bt_intervals;
2949 #ifdef KMP_ADJUST_BLOCKTIME
2950 extern int __kmp_zero_bt;
2952 #ifdef KMP_DFLT_NTH_CORES
2953 extern int __kmp_ncores;
2956 extern int __kmp_abort_delay;
2958 extern int __kmp_need_register_atfork_specified;
2960 __kmp_need_register_atfork;
2962 extern int __kmp_gtid_mode;
2970 __kmp_adjust_gtid_mode;
2971 #ifdef KMP_TDATA_GTID
2972 extern KMP_THREAD_LOCAL
int __kmp_gtid;
2974 extern int __kmp_tls_gtid_min;
2975 extern int __kmp_foreign_tp;
2976 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2977 extern int __kmp_inherit_fp_control;
2978 extern kmp_int16 __kmp_init_x87_fpu_control_word;
2979 extern kmp_uint32 __kmp_init_mxcsr;
2982 extern int __kmp_dflt_max_active_levels;
2985 extern int __kmp_dispatch_num_buffers;
2987 #if KMP_NESTED_HOT_TEAMS
2988 extern int __kmp_hot_teams_mode;
2989 extern int __kmp_hot_teams_max_level;
2993 extern enum clock_function_type __kmp_clock_function;
2994 extern int __kmp_clock_function_param;
2997 #if KMP_MIC_SUPPORTED
2998 extern enum mic_type __kmp_mic_type;
3001 #ifdef USE_LOAD_BALANCE
3002 extern double __kmp_load_balance_interval;
3006 typedef struct kmp_nested_nthreads_t {
3010 } kmp_nested_nthreads_t;
3012 extern kmp_nested_nthreads_t __kmp_nested_nth;
3014 #if KMP_USE_ADAPTIVE_LOCKS
3017 struct kmp_adaptive_backoff_params_t {
3019 kmp_uint32 max_soft_retries;
3022 kmp_uint32 max_badness;
3025 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3027 #if KMP_DEBUG_ADAPTIVE_LOCKS
3028 extern const char *__kmp_speculative_statsfile;
3031 #endif // KMP_USE_ADAPTIVE_LOCKS
3034 extern int __kmp_display_env;
3035 extern int __kmp_display_env_verbose;
3036 extern int __kmp_omp_cancellation;
3043 extern kmp_info_t **__kmp_threads;
3045 extern volatile kmp_team_t *__kmp_team_pool;
3046 extern volatile kmp_info_t *__kmp_thread_pool;
3047 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3050 extern volatile int __kmp_nth;
3053 extern volatile int __kmp_all_nth;
3054 extern int __kmp_thread_pool_nth;
3055 extern std::atomic<int> __kmp_thread_pool_active_nth;
3057 extern kmp_root_t **__kmp_root;
3061 extern kmp_global_t __kmp_global;
3063 extern kmp_info_t __kmp_monitor;
3065 extern std::atomic<kmp_uint32> __kmp_team_counter;
3067 extern std::atomic<kmp_uint32> __kmp_task_counter;
3070 #define _KMP_GEN_ID(counter) \
3071 (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3073 #define _KMP_GEN_ID(counter) (~0)
3076 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3077 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3081 extern void __kmp_print_storage_map_gtid(
int gtid,
void *p1,
void *p2,
3082 size_t size,
char const *format, ...);
3084 extern void __kmp_serial_initialize(
void);
3085 extern void __kmp_middle_initialize(
void);
3086 extern void __kmp_parallel_initialize(
void);
3088 extern void __kmp_internal_begin(
void);
3089 extern void __kmp_internal_end_library(
int gtid);
3090 extern void __kmp_internal_end_thread(
int gtid);
3091 extern void __kmp_internal_end_atexit(
void);
3092 extern void __kmp_internal_end_fini(
void);
3093 extern void __kmp_internal_end_dtor(
void);
3094 extern void __kmp_internal_end_dest(
void *);
3096 extern int __kmp_register_root(
int initial_thread);
3097 extern void __kmp_unregister_root(
int gtid);
3099 extern int __kmp_ignore_mppbeg(
void);
3100 extern int __kmp_ignore_mppend(
void);
3102 extern int __kmp_enter_single(
int gtid,
ident_t *id_ref,
int push_ws);
3103 extern void __kmp_exit_single(
int gtid);
3105 extern void __kmp_parallel_deo(
int *gtid_ref,
int *cid_ref,
ident_t *loc_ref);
3106 extern void __kmp_parallel_dxo(
int *gtid_ref,
int *cid_ref,
ident_t *loc_ref);
3108 #ifdef USE_LOAD_BALANCE
3109 extern int __kmp_get_load_balance(
int);
3112 extern int __kmp_get_global_thread_id(
void);
3113 extern int __kmp_get_global_thread_id_reg(
void);
3114 extern void __kmp_exit_thread(
int exit_status);
3115 extern void __kmp_abort(
char const *format, ...);
3116 extern void __kmp_abort_thread(
void);
3117 KMP_NORETURN
extern void __kmp_abort_process(
void);
3118 extern void __kmp_warn(
char const *format, ...);
3120 extern void __kmp_set_num_threads(
int new_nth,
int gtid);
3124 static inline kmp_info_t *__kmp_entry_thread() {
3125 int gtid = __kmp_entry_gtid();
3127 return __kmp_threads[gtid];
3130 extern void __kmp_set_max_active_levels(
int gtid,
int new_max_active_levels);
3131 extern int __kmp_get_max_active_levels(
int gtid);
3132 extern int __kmp_get_ancestor_thread_num(
int gtid,
int level);
3133 extern int __kmp_get_team_size(
int gtid,
int level);
3134 extern void __kmp_set_schedule(
int gtid, kmp_sched_t new_sched,
int chunk);
3135 extern void __kmp_get_schedule(
int gtid, kmp_sched_t *sched,
int *chunk);
3137 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3138 extern void __kmp_init_random(kmp_info_t *thread);
3140 extern kmp_r_sched_t __kmp_get_schedule_global(
void);
3141 extern void __kmp_adjust_num_threads(
int new_nproc);
3143 extern void *___kmp_allocate(
size_t size KMP_SRC_LOC_DECL);
3144 extern void *___kmp_page_allocate(
size_t size KMP_SRC_LOC_DECL);
3145 extern void ___kmp_free(
void *ptr KMP_SRC_LOC_DECL);
3146 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3147 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3148 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3151 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3152 size_t size KMP_SRC_LOC_DECL);
3153 extern void ___kmp_fast_free(kmp_info_t *this_thr,
void *ptr KMP_SRC_LOC_DECL);
3154 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3155 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3156 #define __kmp_fast_allocate(this_thr, size) \
3157 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3158 #define __kmp_fast_free(this_thr, ptr) \
3159 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3162 extern void *___kmp_thread_malloc(kmp_info_t *th,
size_t size KMP_SRC_LOC_DECL);
3163 extern void *___kmp_thread_calloc(kmp_info_t *th,
size_t nelem,
3164 size_t elsize KMP_SRC_LOC_DECL);
3165 extern void *___kmp_thread_realloc(kmp_info_t *th,
void *ptr,
3166 size_t size KMP_SRC_LOC_DECL);
3167 extern void ___kmp_thread_free(kmp_info_t *th,
void *ptr KMP_SRC_LOC_DECL);
3168 #define __kmp_thread_malloc(th, size) \
3169 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3170 #define __kmp_thread_calloc(th, nelem, elsize) \
3171 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3172 #define __kmp_thread_realloc(th, ptr, size) \
3173 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3174 #define __kmp_thread_free(th, ptr) \
3175 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3177 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3178 #define KMP_INTERNAL_FREE(p) free(p)
3179 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3180 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3182 extern void __kmp_push_num_threads(
ident_t *loc,
int gtid,
int num_threads);
3185 extern void __kmp_push_proc_bind(
ident_t *loc,
int gtid,
3186 kmp_proc_bind_t proc_bind);
3187 extern void __kmp_push_num_teams(
ident_t *loc,
int gtid,
int num_teams,
3191 extern void __kmp_yield(
int cond);
3195 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3198 kmp_uint32 ub, kmp_int32 st,
3202 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3205 kmp_uint64 ub, kmp_int64 st,
3209 kmp_int32 *p_last, kmp_int32 *p_lb,
3210 kmp_int32 *p_ub, kmp_int32 *p_st);
3212 kmp_int32 *p_last, kmp_uint32 *p_lb,
3213 kmp_uint32 *p_ub, kmp_int32 *p_st);
3215 kmp_int32 *p_last, kmp_int64 *p_lb,
3216 kmp_int64 *p_ub, kmp_int64 *p_st);
3218 kmp_int32 *p_last, kmp_uint64 *p_lb,
3219 kmp_uint64 *p_ub, kmp_int64 *p_st);
3226 #ifdef KMP_GOMP_COMPAT
3228 extern void __kmp_aux_dispatch_init_4(
ident_t *loc, kmp_int32 gtid,
3230 kmp_int32 ub, kmp_int32 st,
3231 kmp_int32 chunk,
int push_ws);
3232 extern void __kmp_aux_dispatch_init_4u(
ident_t *loc, kmp_int32 gtid,
3234 kmp_uint32 ub, kmp_int32 st,
3235 kmp_int32 chunk,
int push_ws);
3236 extern void __kmp_aux_dispatch_init_8(
ident_t *loc, kmp_int32 gtid,
3238 kmp_int64 ub, kmp_int64 st,
3239 kmp_int64 chunk,
int push_ws);
3240 extern void __kmp_aux_dispatch_init_8u(
ident_t *loc, kmp_int32 gtid,
3242 kmp_uint64 ub, kmp_int64 st,
3243 kmp_int64 chunk,
int push_ws);
3244 extern void __kmp_aux_dispatch_fini_chunk_4(
ident_t *loc, kmp_int32 gtid);
3245 extern void __kmp_aux_dispatch_fini_chunk_8(
ident_t *loc, kmp_int32 gtid);
3246 extern void __kmp_aux_dispatch_fini_chunk_4u(
ident_t *loc, kmp_int32 gtid);
3247 extern void __kmp_aux_dispatch_fini_chunk_8u(
ident_t *loc, kmp_int32 gtid);
3251 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3252 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3253 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3254 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3255 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3256 extern kmp_uint32 __kmp_wait_yield_4(kmp_uint32
volatile *spinner,
3258 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3260 extern void __kmp_wait_yield_4_ptr(
void *spinner, kmp_uint32 checker,
3261 kmp_uint32 (*pred)(
void *, kmp_uint32),
3266 class kmp_flag_oncore;
3267 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64 *flag,
3274 extern void __kmp_release_64(kmp_flag_64 *flag);
3276 extern void __kmp_infinite_loop(
void);
3278 extern void __kmp_cleanup(
void);
3280 #if KMP_HANDLE_SIGNALS
3281 extern int __kmp_handle_signals;
3282 extern void __kmp_install_signals(
int parallel_init);
3283 extern void __kmp_remove_signals(
void);
3286 extern void __kmp_clear_system_time(
void);
3287 extern void __kmp_read_system_time(
double *delta);
3289 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3291 extern void __kmp_expand_host_name(
char *buffer,
size_t size);
3292 extern void __kmp_expand_file_name(
char *result,
size_t rlen,
char *pattern);
3294 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3296 __kmp_initialize_system_tick(
void);
3300 __kmp_runtime_initialize(
void);
3301 extern void __kmp_runtime_destroy(
void);
3303 #if KMP_AFFINITY_SUPPORTED
3304 extern char *__kmp_affinity_print_mask(
char *buf,
int buf_len,
3305 kmp_affin_mask_t *mask);
3306 extern void __kmp_affinity_initialize(
void);
3307 extern void __kmp_affinity_uninitialize(
void);
3308 extern void __kmp_affinity_set_init_mask(
3309 int gtid,
int isa_root);
3311 extern void __kmp_affinity_set_place(
int gtid);
3313 extern void __kmp_affinity_determine_capable(
const char *env_var);
3314 extern int __kmp_aux_set_affinity(
void **mask);
3315 extern int __kmp_aux_get_affinity(
void **mask);
3316 extern int __kmp_aux_get_affinity_max_proc();
3317 extern int __kmp_aux_set_affinity_mask_proc(
int proc,
void **mask);
3318 extern int __kmp_aux_unset_affinity_mask_proc(
int proc,
void **mask);
3319 extern int __kmp_aux_get_affinity_mask_proc(
int proc,
void **mask);
3320 extern void __kmp_balanced_affinity(
int tid,
int team_size);
3322 extern int kmp_set_thread_affinity_mask_initial(
void);
3326 extern void __kmp_cleanup_hierarchy();
3327 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3331 extern int __kmp_futex_determine_capable(
void);
3333 #endif // KMP_USE_FUTEX
3335 extern void __kmp_gtid_set_specific(
int gtid);
3336 extern int __kmp_gtid_get_specific(
void);
3338 extern double __kmp_read_cpu_time(
void);
3340 extern int __kmp_read_system_info(
struct kmp_sys_info *info);
3343 extern void __kmp_create_monitor(kmp_info_t *th);
3346 extern void *__kmp_launch_thread(kmp_info_t *thr);
3348 extern void __kmp_create_worker(
int gtid, kmp_info_t *th,
size_t stack_size);
3351 extern int __kmp_still_running(kmp_info_t *th);
3352 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3353 extern void __kmp_free_handle(kmp_thread_t tHandle);
3357 extern void __kmp_reap_monitor(kmp_info_t *th);
3359 extern void __kmp_reap_worker(kmp_info_t *th);
3360 extern void __kmp_terminate_thread(
int gtid);
3362 extern void __kmp_suspend_32(
int th_gtid, kmp_flag_32 *flag);
3363 extern void __kmp_suspend_64(
int th_gtid, kmp_flag_64 *flag);
3364 extern void __kmp_suspend_oncore(
int th_gtid, kmp_flag_oncore *flag);
3365 extern void __kmp_resume_32(
int target_gtid, kmp_flag_32 *flag);
3366 extern void __kmp_resume_64(
int target_gtid, kmp_flag_64 *flag);
3367 extern void __kmp_resume_oncore(
int target_gtid, kmp_flag_oncore *flag);
3369 extern void __kmp_elapsed(
double *);
3370 extern void __kmp_elapsed_tick(
double *);
3372 extern void __kmp_enable(
int old_state);
3373 extern void __kmp_disable(
int *old_state);
3375 extern void __kmp_thread_sleep(
int millis);
3377 extern void __kmp_common_initialize(
void);
3378 extern void __kmp_common_destroy(
void);
3379 extern void __kmp_common_destroy_gtid(
int gtid);
3382 extern void __kmp_register_atfork(
void);
3384 extern void __kmp_suspend_initialize(
void);
3385 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3387 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3391 __kmp_allocate_team(kmp_root_t *root,
int new_nproc,
int max_nproc,
3393 ompt_data_t ompt_parallel_data,
3395 kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3396 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3399 __kmp_allocate_team(kmp_root_t *root,
int new_nproc,
int max_nproc,
3401 ompt_id_t ompt_parallel_id,
3403 kmp_internal_control_t *new_icvs,
3404 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3405 #endif // OMP_40_ENABLED
3406 extern void __kmp_free_thread(kmp_info_t *);
3407 extern void __kmp_free_team(kmp_root_t *,
3408 kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3409 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3413 extern void __kmp_initialize_bget(kmp_info_t *th);
3414 extern void __kmp_finalize_bget(kmp_info_t *th);
3416 KMP_EXPORT
void *kmpc_malloc(
size_t size);
3417 KMP_EXPORT
void *kmpc_aligned_malloc(
size_t size,
size_t alignment);
3418 KMP_EXPORT
void *kmpc_calloc(
size_t nelem,
size_t elsize);
3419 KMP_EXPORT
void *kmpc_realloc(
void *ptr,
size_t size);
3420 KMP_EXPORT
void kmpc_free(
void *ptr);
3424 extern int __kmp_barrier(
enum barrier_type bt,
int gtid,
int is_split,
3425 size_t reduce_size,
void *reduce_data,
3426 void (*reduce)(
void *,
void *));
3427 extern void __kmp_end_split_barrier(
enum barrier_type bt,
int gtid);
3433 enum fork_context_e {
3439 extern int __kmp_fork_call(
ident_t *loc,
int gtid,
3440 enum fork_context_e fork_context, kmp_int32 argc,
3441 microtask_t microtask, launch_t invoker,
3443 #
if (KMP_ARCH_ARM || KMP_ARCH_X86_64 || KMP_ARCH_AARCH64) && KMP_OS_LINUX
3450 extern void __kmp_join_call(
ident_t *loc,
int gtid
3453 enum fork_context_e fork_context
3461 extern void __kmp_serialized_parallel(
ident_t *
id, kmp_int32 gtid);
3462 extern void __kmp_internal_fork(
ident_t *
id,
int gtid, kmp_team_t *team);
3463 extern void __kmp_internal_join(
ident_t *
id,
int gtid, kmp_team_t *team);
3464 extern int __kmp_invoke_task_func(
int gtid);
3465 extern void __kmp_run_before_invoked_task(
int gtid,
int tid,
3466 kmp_info_t *this_thr,
3468 extern void __kmp_run_after_invoked_task(
int gtid,
int tid,
3469 kmp_info_t *this_thr,
3473 KMP_EXPORT
int __kmpc_invoke_task_func(
int gtid);
3475 extern int __kmp_invoke_teams_master(
int gtid);
3476 extern void __kmp_teams_master(
int gtid);
3478 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3479 extern void __kmp_user_set_library(
enum library_type arg);
3480 extern void __kmp_aux_set_library(
enum library_type arg);
3481 extern void __kmp_aux_set_stacksize(
size_t arg);
3482 extern void __kmp_aux_set_blocktime(
int arg, kmp_info_t *thread,
int tid);
3483 extern void __kmp_aux_set_defaults(
char const *str,
int len);
3486 void kmpc_set_blocktime(
int arg);
3487 void ompc_set_nested(
int flag);
3488 void ompc_set_dynamic(
int flag);
3489 void ompc_set_num_threads(
int arg);
3491 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3492 kmp_team_t *team,
int tid);
3493 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3494 extern kmp_task_t *__kmp_task_alloc(
ident_t *loc_ref, kmp_int32 gtid,
3495 kmp_tasking_flags_t *flags,
3496 size_t sizeof_kmp_task_t,
3497 size_t sizeof_shareds,
3498 kmp_routine_entry_t task_entry);
3499 extern void __kmp_init_implicit_task(
ident_t *loc_ref, kmp_info_t *this_thr,
3500 kmp_team_t *team,
int tid,
3502 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3503 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3504 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
3505 kmp_flag_32 *flag,
int final_spin,
3506 int *thread_finished,
3510 kmp_int32 is_constrained);
3511 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
3512 kmp_flag_64 *flag,
int final_spin,
3513 int *thread_finished,
3517 kmp_int32 is_constrained);
3518 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
3519 kmp_flag_oncore *flag,
int final_spin,
3520 int *thread_finished,
3524 kmp_int32 is_constrained);
3526 extern void __kmp_free_task_team(kmp_info_t *thread,
3527 kmp_task_team_t *task_team);
3528 extern void __kmp_reap_task_teams(
void);
3529 extern void __kmp_wait_to_unref_task_teams(
void);
3530 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3532 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3533 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3540 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3543 extern int __kmp_is_address_mapped(
void *addr);
3544 extern kmp_uint64 __kmp_hardware_timestamp(
void);
3547 extern int __kmp_read_from_file(
char const *path,
char const *format, ...);
3555 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3557 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
3559 #define __kmp_load_mxcsr(p) _mm_setcsr(*(p))
3560 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
3562 extern void __kmp_load_x87_fpu_control_word(kmp_int16 *p);
3563 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
3564 extern void __kmp_clear_x87_fpu_status_word();
3565 #define KMP_X86_MXCSR_MASK 0xffffffc0
3569 extern int __kmp_invoke_microtask(microtask_t pkfn,
int gtid,
int npr,
int argc,
3573 void **exit_frame_ptr
3586 size_t vector_length);
3590 KMP_EXPORT
void *__kmpc_threadprivate(
ident_t *, kmp_int32 global_tid,
3591 void *data,
size_t size);
3612 kmp_critical_name *);
3614 kmp_critical_name *);
3617 KMP_EXPORT
void __kmpc_critical_with_hint(
ident_t *, kmp_int32 global_tid,
3618 kmp_critical_name *, uintptr_t hint);
3625 kmp_int32 global_tid);
3630 KMP_EXPORT
void KMPC_FOR_STATIC_INIT(
ident_t *loc, kmp_int32 global_tid,
3631 kmp_int32 schedtype, kmp_int32 *plastiter,
3632 kmp_int *plower, kmp_int *pupper,
3633 kmp_int *pstride, kmp_int incr,
3639 size_t cpy_size,
void *cpy_data,
3640 void (*cpy_func)(
void *,
void *),
3643 extern void KMPC_SET_NUM_THREADS(
int arg);
3644 extern void KMPC_SET_DYNAMIC(
int flag);
3645 extern void KMPC_SET_NESTED(
int flag);
3648 KMP_EXPORT kmpc_thunk_t *__kmpc_taskq(
ident_t *loc, kmp_int32 global_tid,
3649 kmpc_task_t taskq_task,
3650 size_t sizeof_thunk,
3651 size_t sizeof_shareds, kmp_int32 flags,
3652 kmpc_shared_vars_t **shareds);
3653 KMP_EXPORT
void __kmpc_end_taskq(
ident_t *loc, kmp_int32 global_tid,
3654 kmpc_thunk_t *thunk);
3655 KMP_EXPORT kmp_int32 __kmpc_task(
ident_t *loc, kmp_int32 global_tid,
3656 kmpc_thunk_t *thunk);
3657 KMP_EXPORT
void __kmpc_taskq_task(
ident_t *loc, kmp_int32 global_tid,
3658 kmpc_thunk_t *thunk, kmp_int32 status);
3659 KMP_EXPORT
void __kmpc_end_taskq_task(
ident_t *loc, kmp_int32 global_tid,
3660 kmpc_thunk_t *thunk);
3661 KMP_EXPORT kmpc_thunk_t *__kmpc_task_buffer(
ident_t *loc, kmp_int32 global_tid,
3662 kmpc_thunk_t *taskq_thunk,
3666 KMP_EXPORT kmp_int32 __kmpc_omp_task(
ident_t *loc_ref, kmp_int32 gtid,
3667 kmp_task_t *new_task);
3668 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(
ident_t *loc_ref, kmp_int32 gtid,
3670 size_t sizeof_kmp_task_t,
3671 size_t sizeof_shareds,
3672 kmp_routine_entry_t task_entry);
3673 KMP_EXPORT
void __kmpc_omp_task_begin_if0(
ident_t *loc_ref, kmp_int32 gtid,
3675 KMP_EXPORT
void __kmpc_omp_task_complete_if0(
ident_t *loc_ref, kmp_int32 gtid,
3677 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(
ident_t *loc_ref, kmp_int32 gtid,
3678 kmp_task_t *new_task);
3679 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(
ident_t *loc_ref, kmp_int32 gtid);
3681 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(
ident_t *loc_ref, kmp_int32 gtid,
3685 void __kmpc_omp_task_begin(
ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3686 void __kmpc_omp_task_complete(
ident_t *loc_ref, kmp_int32 gtid,
3688 #endif // TASK_UNUSED
3694 KMP_EXPORT
void __kmpc_taskgroup(
ident_t *loc,
int gtid);
3695 KMP_EXPORT
void __kmpc_end_taskgroup(
ident_t *loc,
int gtid);
3698 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3699 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3700 kmp_depend_info_t *noalias_dep_list);
3703 kmp_depend_info_t *dep_list,
3704 kmp_int32 ndeps_noalias,
3705 kmp_depend_info_t *noalias_dep_list);
3706 extern void __kmp_release_deps(kmp_int32 gtid, kmp_taskdata_t *task);
3707 extern void __kmp_dephash_free_entries(kmp_info_t *thread, kmp_dephash_t *h);
3708 extern void __kmp_dephash_free(kmp_info_t *thread, kmp_dephash_t *h);
3710 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3711 bool serialize_immediate);
3713 KMP_EXPORT kmp_int32 __kmpc_cancel(
ident_t *loc_ref, kmp_int32 gtid,
3714 kmp_int32 cncl_kind);
3715 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(
ident_t *loc_ref, kmp_int32 gtid,
3716 kmp_int32 cncl_kind);
3717 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(
ident_t *loc_ref, kmp_int32 gtid);
3718 KMP_EXPORT
int __kmp_get_cancellation_status(
int cancel_kind);
3722 KMP_EXPORT
void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3723 KMP_EXPORT
void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3724 KMP_EXPORT
void __kmpc_taskloop(
ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3725 kmp_int32 if_val, kmp_uint64 *lb,
3726 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3727 kmp_int32 sched, kmp_uint64 grainsize,
3732 KMP_EXPORT
void *__kmpc_task_reduction_init(
int gtid,
int num_data,
void *data);
3733 KMP_EXPORT
void *__kmpc_task_reduction_get_th_data(
int gtid,
void *tg,
void *d);
3739 KMP_EXPORT
void __kmpc_init_lock(
ident_t *loc, kmp_int32 gtid,
3741 KMP_EXPORT
void __kmpc_init_nest_lock(
ident_t *loc, kmp_int32 gtid,
3743 KMP_EXPORT
void __kmpc_destroy_lock(
ident_t *loc, kmp_int32 gtid,
3745 KMP_EXPORT
void __kmpc_destroy_nest_lock(
ident_t *loc, kmp_int32 gtid,
3747 KMP_EXPORT
void __kmpc_set_lock(
ident_t *loc, kmp_int32 gtid,
void **user_lock);
3748 KMP_EXPORT
void __kmpc_set_nest_lock(
ident_t *loc, kmp_int32 gtid,
3750 KMP_EXPORT
void __kmpc_unset_lock(
ident_t *loc, kmp_int32 gtid,
3752 KMP_EXPORT
void __kmpc_unset_nest_lock(
ident_t *loc, kmp_int32 gtid,
3754 KMP_EXPORT
int __kmpc_test_lock(
ident_t *loc, kmp_int32 gtid,
void **user_lock);
3755 KMP_EXPORT
int __kmpc_test_nest_lock(
ident_t *loc, kmp_int32 gtid,
3759 KMP_EXPORT
void __kmpc_init_lock_with_hint(
ident_t *loc, kmp_int32 gtid,
3760 void **user_lock, uintptr_t hint);
3761 KMP_EXPORT
void __kmpc_init_nest_lock_with_hint(
ident_t *loc, kmp_int32 gtid,
3769 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
size_t reduce_size,
3770 void *reduce_data,
void (*reduce_func)(
void *lhs_data,
void *rhs_data),
3771 kmp_critical_name *lck);
3773 kmp_critical_name *lck);
3775 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
size_t reduce_size,
3776 void *reduce_data,
void (*reduce_func)(
void *lhs_data,
void *rhs_data),
3777 kmp_critical_name *lck);
3779 kmp_critical_name *lck);
3783 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3784 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
size_t reduce_size,
3785 void *reduce_data,
void (*reduce_func)(
void *lhs_data,
void *rhs_data),
3786 kmp_critical_name *lck);
3789 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(
void);
3791 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3792 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3798 KMP_EXPORT
void __kmpc_pop_num_threads(
ident_t *loc, kmp_int32 global_tid);
3800 kmp_int32 num_threads);
3803 KMP_EXPORT
void __kmpc_push_proc_bind(
ident_t *loc, kmp_int32 global_tid,
3806 kmp_int32 num_teams,
3807 kmp_int32 num_threads);
3817 KMP_EXPORT
void __kmpc_doacross_init(
ident_t *loc, kmp_int32 gtid,
3819 const struct kmp_dim *dims);
3820 KMP_EXPORT
void __kmpc_doacross_wait(
ident_t *loc, kmp_int32 gtid,
3821 const kmp_int64 *vec);
3822 KMP_EXPORT
void __kmpc_doacross_post(
ident_t *loc, kmp_int32 gtid,
3823 const kmp_int64 *vec);
3824 KMP_EXPORT
void __kmpc_doacross_fini(
ident_t *loc, kmp_int32 gtid);
3828 void *data,
size_t size,
3832 extern int _You_must_link_with_exactly_one_OpenMP_library;
3833 extern int _You_must_link_with_Intel_OpenMP_library;
3834 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3835 extern int _You_must_link_with_Microsoft_OpenMP_library;
3840 void kmp_threadprivate_insert_private_data(
int gtid,
void *pc_addr,
3841 void *data_addr,
size_t pc_size);
3842 struct private_common *kmp_threadprivate_insert(
int gtid,
void *pc_addr,
3845 void __kmp_threadprivate_resize_cache(
int newCapacity);
3846 void __kmp_cleanup_threadprivate_caches();
3850 #define KMPC_CONVENTION __cdecl
3852 #define KMPC_CONVENTION
3856 typedef enum omp_sched_t {
3857 omp_sched_static = 1,
3858 omp_sched_dynamic = 2,
3859 omp_sched_guided = 3,
3862 typedef void *kmp_affinity_mask_t;
3865 KMP_EXPORT
void KMPC_CONVENTION ompc_set_max_active_levels(
int);
3866 KMP_EXPORT
void KMPC_CONVENTION ompc_set_schedule(omp_sched_t,
int);
3867 KMP_EXPORT
int KMPC_CONVENTION ompc_get_ancestor_thread_num(
int);
3868 KMP_EXPORT
int KMPC_CONVENTION ompc_get_team_size(
int);
3869 KMP_EXPORT
int KMPC_CONVENTION
3870 kmpc_set_affinity_mask_proc(
int, kmp_affinity_mask_t *);
3871 KMP_EXPORT
int KMPC_CONVENTION
3872 kmpc_unset_affinity_mask_proc(
int, kmp_affinity_mask_t *);
3873 KMP_EXPORT
int KMPC_CONVENTION
3874 kmpc_get_affinity_mask_proc(
int, kmp_affinity_mask_t *);
3876 KMP_EXPORT
void KMPC_CONVENTION kmpc_set_stacksize(
int);
3877 KMP_EXPORT
void KMPC_CONVENTION kmpc_set_stacksize_s(
size_t);
3878 KMP_EXPORT
void KMPC_CONVENTION kmpc_set_library(
int);
3879 KMP_EXPORT
void KMPC_CONVENTION kmpc_set_defaults(
char const *);
3880 KMP_EXPORT
void KMPC_CONVENTION kmpc_set_disp_num_buffers(
int);
3883 enum kmp_target_offload_kind {
3888 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
3890 extern kmp_target_offload_kind_t __kmp_target_offload;
3891 extern int __kmpc_get_target_offload();