3 * kmp.h -- KPTS runtime header file.
6 //===----------------------------------------------------------------------===//
8 // The LLVM Compiler Infrastructure
10 // This file is dual licensed under the MIT and the University of Illinois Open
11 // Source Licenses. See LICENSE.txt for details.
13 //===----------------------------------------------------------------------===//
18 #include "kmp_config.h"
20 /* #define BUILD_PARALLEL_ORDERED 1 */
22 /* This fix replaces gettimeofday with clock_gettime for better scalability on
23 the Altix. Requires user code to be linked with -lrt. */
24 //#define FIX_SGI_CLOCK
26 /* Defines for OpenMP 3.0 tasking and auto scheduling */
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
38 // Number of entries in each task stack array
39 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
40 // Mask for determining index into stack block
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
56 // Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being
58 #if defined(__ANDROID__)
59 #undef KMP_CANCEL_THREADS
68 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
69 Microsoft library. Some macros provided below to replace these functions */
71 #include <sys/types.h>
80 #include "kmp_safe_c_api.h"
86 #if KMP_USE_HIER_SCHED
87 // Only include hierarchical scheduling if affinity is supported
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"
133 // Affinity format function
137 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
138 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
139 // free lists of limited size.
140 #ifndef USE_FAST_MEMORY
141 #define USE_FAST_MEMORY 3
144 #ifndef KMP_NESTED_HOT_TEAMS
145 #define KMP_NESTED_HOT_TEAMS 0
146 #define USE_NESTED_HOT_ARG(x)
148 #if KMP_NESTED_HOT_TEAMS
150 #define USE_NESTED_HOT_ARG(x) , x
152 // Nested hot teams feature depends on omp 4.0, disable it for earlier versions
153 #undef KMP_NESTED_HOT_TEAMS
154 #define KMP_NESTED_HOT_TEAMS 0
155 #define USE_NESTED_HOT_ARG(x)
158 #define USE_NESTED_HOT_ARG(x)
162 // Assume using BGET compare_exchange instruction instead of lock by default.
163 #ifndef USE_CMP_XCHG_FOR_BGET
164 #define USE_CMP_XCHG_FOR_BGET 1
167 // Test to see if queuing lock is better than bootstrap lock for bget
168 // #ifndef USE_QUEUING_LOCK_FOR_BGET
169 // #define USE_QUEUING_LOCK_FOR_BGET
172 #define KMP_NSEC_PER_SEC 1000000000L
173 #define KMP_USEC_PER_SEC 1000000L
181 Values for bit flags used in the ident_t to describe the fields.
184 /*! Use trampoline for internal microtasks */
185 KMP_IDENT_IMB = 0x01,
186 /*! Use c-style ident structure */
187 KMP_IDENT_KMPC = 0x02,
188 /* 0x04 is no longer used */
189 /*! Entry point generated by auto-parallelization */
190 KMP_IDENT_AUTOPAR = 0x08,
191 /*! Compiler generates atomic reduction option for kmpc_reduce* */
192 KMP_IDENT_ATOMIC_REDUCE = 0x10,
193 /*! To mark a 'barrier' directive in user code */
194 KMP_IDENT_BARRIER_EXPL = 0x20,
195 /*! To Mark implicit barriers. */
196 KMP_IDENT_BARRIER_IMPL = 0x0040,
197 KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
198 KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
199 KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
201 KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
202 KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
204 /*! To mark a static loop in OMPT callbacks */
205 KMP_IDENT_WORK_LOOP = 0x200,
206 /*! To mark a sections directive in OMPT callbacks */
207 KMP_IDENT_WORK_SECTIONS = 0x400,
208 /*! To mark a distirbute construct in OMPT callbacks */
209 KMP_IDENT_WORK_DISTRIBUTE = 0x800,
210 /*! Atomic hint; bottom four bits as omp_sync_hint_t. Top four reserved and
211 not currently used. If one day we need more bits, then we can use
212 an invalid combination of hints to mean that another, larger field
213 should be used in a different flag. */
214 KMP_IDENT_ATOMIC_HINT_MASK = 0xFF0000,
215 KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
216 KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
217 KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
218 KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
222 * The ident structure that describes a source location.
224 typedef struct ident {
225 kmp_int32 reserved_1; /**< might be used in Fortran; see above */
226 kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC
227 identifies this union member */
228 kmp_int32 reserved_2; /**< not really used in Fortran any more; see above */
230 /* but currently used for storing region-specific ITT */
231 /* contextual information. */
232 #endif /* USE_ITT_BUILD */
233 kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++ */
234 char const *psource; /**< String describing the source location.
235 The string is composed of semi-colon separated fields
236 which describe the source file, the function and a pair
237 of line numbers that delimit the construct. */
243 // Some forward declarations.
244 typedef union kmp_team kmp_team_t;
245 typedef struct kmp_taskdata kmp_taskdata_t;
246 typedef union kmp_task_team kmp_task_team_t;
247 typedef union kmp_team kmp_team_p;
248 typedef union kmp_info kmp_info_p;
249 typedef union kmp_root kmp_root_p;
255 /* ------------------------------------------------------------------------ */
257 /* Pack two 32-bit signed integers into a 64-bit signed integer */
258 /* ToDo: Fix word ordering for big-endian machines. */
259 #define KMP_PACK_64(HIGH_32, LOW_32) \
260 ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
262 // Generic string manipulation macros. Assume that _x is of type char *
263 #define SKIP_WS(_x) \
265 while (*(_x) == ' ' || *(_x) == '\t') \
268 #define SKIP_DIGITS(_x) \
270 while (*(_x) >= '0' && *(_x) <= '9') \
273 #define SKIP_TOKEN(_x) \
275 while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
276 (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
279 #define SKIP_TO(_x, _c) \
281 while (*(_x) != '\0' && *(_x) != (_c)) \
285 /* ------------------------------------------------------------------------ */
287 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
288 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
290 /* ------------------------------------------------------------------------ */
291 /* Enumeration types */
293 enum kmp_state_timer {
303 #ifdef USE_LOAD_BALANCE
304 dynamic_load_balance,
305 #endif /* USE_LOAD_BALANCE */
307 dynamic_thread_limit,
311 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
312 * not include it here */
313 #ifndef KMP_SCHED_TYPE_DEFINED
314 #define KMP_SCHED_TYPE_DEFINED
315 typedef enum kmp_sched {
316 kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
317 // Note: need to adjust __kmp_sch_map global array in case enum is changed
318 kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
319 kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
320 kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
321 kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
322 kmp_sched_upper_std = 5, // upper bound for standard schedules
323 kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
324 kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
325 #if KMP_STATIC_STEAL_ENABLED
326 kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
329 kmp_sched_default = kmp_sched_static // default scheduling
334 @ingroup WORK_SHARING
335 * Describes the loop schedule to be used for a parallel for loop.
337 enum sched_type : kmp_int32 {
338 kmp_sch_lower = 32, /**< lower bound for unordered values */
339 kmp_sch_static_chunked = 33,
340 kmp_sch_static = 34, /**< static unspecialized */
341 kmp_sch_dynamic_chunked = 35,
342 kmp_sch_guided_chunked = 36, /**< guided unspecialized */
343 kmp_sch_runtime = 37,
344 kmp_sch_auto = 38, /**< auto */
345 kmp_sch_trapezoidal = 39,
347 /* accessible only through KMP_SCHEDULE environment variable */
348 kmp_sch_static_greedy = 40,
349 kmp_sch_static_balanced = 41,
350 /* accessible only through KMP_SCHEDULE environment variable */
351 kmp_sch_guided_iterative_chunked = 42,
352 kmp_sch_guided_analytical_chunked = 43,
353 /* accessible only through KMP_SCHEDULE environment variable */
354 kmp_sch_static_steal = 44,
357 /* static with chunk adjustment (e.g., simd) */
358 kmp_sch_static_balanced_chunked = 45,
359 kmp_sch_guided_simd = 46, /**< guided with chunk adjustment */
360 kmp_sch_runtime_simd = 47, /**< runtime with chunk adjustment */
363 /* accessible only through KMP_SCHEDULE environment variable */
364 kmp_sch_upper, /**< upper bound for unordered values */
366 kmp_ord_lower = 64, /**< lower bound for ordered values, must be power of 2 */
367 kmp_ord_static_chunked = 65,
368 kmp_ord_static = 66, /**< ordered static unspecialized */
369 kmp_ord_dynamic_chunked = 67,
370 kmp_ord_guided_chunked = 68,
371 kmp_ord_runtime = 69,
372 kmp_ord_auto = 70, /**< ordered auto */
373 kmp_ord_trapezoidal = 71,
374 kmp_ord_upper, /**< upper bound for ordered values */
377 /* Schedules for Distribute construct */
378 kmp_distribute_static_chunked = 91, /**< distribute static chunked */
379 kmp_distribute_static = 92, /**< distribute static unspecialized */
382 /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
383 single iteration/chunk, even if the loop is serialized. For the schedule
384 types listed above, the entire iteration vector is returned if the loop is
385 serialized. This doesn't work for gcc/gcomp sections. */
386 kmp_nm_lower = 160, /**< lower bound for nomerge values */
388 kmp_nm_static_chunked =
389 (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
390 kmp_nm_static = 162, /**< static unspecialized */
391 kmp_nm_dynamic_chunked = 163,
392 kmp_nm_guided_chunked = 164, /**< guided unspecialized */
393 kmp_nm_runtime = 165,
394 kmp_nm_auto = 166, /**< auto */
395 kmp_nm_trapezoidal = 167,
397 /* accessible only through KMP_SCHEDULE environment variable */
398 kmp_nm_static_greedy = 168,
399 kmp_nm_static_balanced = 169,
400 /* accessible only through KMP_SCHEDULE environment variable */
401 kmp_nm_guided_iterative_chunked = 170,
402 kmp_nm_guided_analytical_chunked = 171,
403 kmp_nm_static_steal =
404 172, /* accessible only through OMP_SCHEDULE environment variable */
406 kmp_nm_ord_static_chunked = 193,
407 kmp_nm_ord_static = 194, /**< ordered static unspecialized */
408 kmp_nm_ord_dynamic_chunked = 195,
409 kmp_nm_ord_guided_chunked = 196,
410 kmp_nm_ord_runtime = 197,
411 kmp_nm_ord_auto = 198, /**< auto */
412 kmp_nm_ord_trapezoidal = 199,
413 kmp_nm_upper, /**< upper bound for nomerge values */
416 /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
417 we need to distinguish the three possible cases (no modifier, monotonic
418 modifier, nonmonotonic modifier), we need separate bits for each modifier.
419 The absence of monotonic does not imply nonmonotonic, especially since 4.5
420 says that the behaviour of the "no modifier" case is implementation defined
421 in 4.5, but will become "nonmonotonic" in 5.0.
423 Since we're passing a full 32 bit value, we can use a couple of high bits
424 for these flags; out of paranoia we avoid the sign bit.
426 These modifiers can be or-ed into non-static schedules by the compiler to
427 pass the additional information. They will be stripped early in the
428 processing in __kmp_dispatch_init when setting up schedules, so most of the
429 code won't ever see schedules with these bits set. */
430 kmp_sch_modifier_monotonic =
431 (1 << 29), /**< Set if the monotonic schedule modifier was present */
432 kmp_sch_modifier_nonmonotonic =
433 (1 << 30), /**< Set if the nonmonotonic schedule modifier was present */
435 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
437 (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))
438 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
439 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
440 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
441 (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
443 /* By doing this we hope to avoid multiple tests on OMP_45_ENABLED. Compilers
444 can now eliminate tests on compile time constants and dead code that results
445 from them, so we can leave code guarded by such an if in place. */
446 #define SCHEDULE_WITHOUT_MODIFIERS(s) (s)
447 #define SCHEDULE_HAS_MONOTONIC(s) false
448 #define SCHEDULE_HAS_NONMONOTONIC(s) false
449 #define SCHEDULE_HAS_NO_MODIFIERS(s) true
452 kmp_sch_default = kmp_sch_static /**< default scheduling algorithm */
455 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
456 typedef union kmp_r_sched {
458 enum sched_type r_sched_type;
464 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
465 // internal schedule types
475 enum clock_function_type {
476 clock_function_gettimeofday,
477 clock_function_clock_gettime
479 #endif /* KMP_OS_LINUX */
481 #if KMP_MIC_SUPPORTED
482 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
485 /* -- fast reduction stuff ------------------------------------------------ */
487 #undef KMP_FAST_REDUCTION_BARRIER
488 #define KMP_FAST_REDUCTION_BARRIER 1
490 #undef KMP_FAST_REDUCTION_CORE_DUO
491 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
492 #define KMP_FAST_REDUCTION_CORE_DUO 1
495 enum _reduction_method {
496 reduction_method_not_defined = 0,
497 critical_reduce_block = (1 << 8),
498 atomic_reduce_block = (2 << 8),
499 tree_reduce_block = (3 << 8),
500 empty_reduce_block = (4 << 8)
503 // Description of the packed_reduction_method variable:
504 // The packed_reduction_method variable consists of two enum types variables
505 // that are packed together into 0-th byte and 1-st byte:
506 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
507 // barrier that will be used in fast reduction: bs_plain_barrier or
508 // bs_reduction_barrier
509 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
510 // be used in fast reduction;
511 // Reduction method is of 'enum _reduction_method' type and it's defined the way
512 // so that the bits of 0-th byte are empty, so no need to execute a shift
513 // instruction while packing/unpacking
515 #if KMP_FAST_REDUCTION_BARRIER
516 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
517 ((reduction_method) | (barrier_type))
519 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
520 ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
522 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
523 ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
525 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
528 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
529 (packed_reduction_method)
531 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
534 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
535 ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
536 (which_reduction_block))
538 #if KMP_FAST_REDUCTION_BARRIER
539 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
540 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
542 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
543 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
546 typedef int PACKED_REDUCTION_METHOD_T;
548 /* -- end of fast reduction stuff ----------------------------------------- */
553 #pragma warning(push)
554 #pragma warning(disable : 271 310)
567 /* Only Linux* OS and Windows* OS support thread affinity. */
568 #if KMP_AFFINITY_SUPPORTED
570 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
572 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
573 typedef struct GROUP_AFFINITY {
578 #endif /* _MSC_VER < 1600 */
579 #if KMP_GROUP_AFFINITY
580 extern int __kmp_num_proc_groups;
582 static const int __kmp_num_proc_groups = 1;
583 #endif /* KMP_GROUP_AFFINITY */
584 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
585 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
587 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
588 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
590 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
591 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
593 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
595 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
596 #endif /* KMP_OS_WINDOWS */
599 extern hwloc_topology_t __kmp_hwloc_topology;
600 extern int __kmp_hwloc_error;
601 extern int __kmp_numa_detected;
602 extern int __kmp_tile_depth;
605 extern size_t __kmp_affin_mask_size;
606 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
607 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
608 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
609 #define KMP_CPU_SET_ITERATE(i, mask) \
610 for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
611 #define KMP_CPU_SET(i, mask) (mask)->set(i)
612 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
613 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
614 #define KMP_CPU_ZERO(mask) (mask)->zero()
615 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
616 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
617 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
618 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
619 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
620 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
621 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
622 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
623 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
624 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
625 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
626 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
627 (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
628 #define KMP_CPU_FREE_ARRAY(arr, n) \
629 __kmp_affinity_dispatch->deallocate_mask_array(arr)
630 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
631 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
632 #define __kmp_get_system_affinity(mask, abort_bool) \
633 (mask)->get_system_affinity(abort_bool)
634 #define __kmp_set_system_affinity(mask, abort_bool) \
635 (mask)->set_system_affinity(abort_bool)
636 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
642 void *operator new(size_t n);
643 void operator delete(void *p);
644 void *operator new[](size_t n);
645 void operator delete[](void *p);
648 virtual void set(int i) {}
650 virtual bool is_set(int i) const { return false; }
652 virtual void clear(int i) {}
653 // Zero out entire mask
654 virtual void zero() {}
655 // Copy src into this mask
656 virtual void copy(const Mask *src) {}
658 virtual void bitwise_and(const Mask *rhs) {}
660 virtual void bitwise_or(const Mask *rhs) {}
662 virtual void bitwise_not() {}
663 // API for iterating over an affinity mask
664 // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
665 virtual int begin() const { return 0; }
666 virtual int end() const { return 0; }
667 virtual int next(int previous) const { return 0; }
668 // Set the system's affinity to this affinity mask's value
669 virtual int set_system_affinity(bool abort_on_error) const { return -1; }
670 // Set this affinity mask to the current system affinity
671 virtual int get_system_affinity(bool abort_on_error) { return -1; }
672 // Only 1 DWORD in the mask should have any procs set.
673 // Return the appropriate index, or -1 for an invalid mask.
674 virtual int get_proc_group() const { return -1; }
676 void *operator new(size_t n);
677 void operator delete(void *p);
678 // Need virtual destructor
679 virtual ~KMPAffinity() = default;
680 // Determine if affinity is capable
681 virtual void determine_capable(const char *env_var) {}
682 // Bind the current thread to os proc
683 virtual void bind_thread(int proc) {}
684 // Factory functions to allocate/deallocate a mask
685 virtual Mask *allocate_mask() { return nullptr; }
686 virtual void deallocate_mask(Mask *m) {}
687 virtual Mask *allocate_mask_array(int num) { return nullptr; }
688 virtual void deallocate_mask_array(Mask *m) {}
689 virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
690 static void pick_api();
691 static void destroy_api();
699 virtual api_type get_api_type() const {
705 static bool picked_api;
708 typedef KMPAffinity::Mask kmp_affin_mask_t;
709 extern KMPAffinity *__kmp_affinity_dispatch;
711 // Declare local char buffers with this size for printing debug and info
712 // messages, using __kmp_affinity_print_mask().
713 #define KMP_AFFIN_MASK_PRINT_LEN 1024
723 affinity_disabled, // not used outsize the env var parser
728 affinity_gran_fine = 0,
729 affinity_gran_thread,
733 affinity_gran_package,
735 #if KMP_GROUP_AFFINITY
736 // The "group" granularity isn't necesssarily coarser than all of the
737 // other levels, but we put it last in the enum.
739 #endif /* KMP_GROUP_AFFINITY */
740 affinity_gran_default
743 enum affinity_top_method {
744 affinity_top_method_all = 0, // try all (supported) methods, in order
745 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
746 affinity_top_method_apicid,
747 affinity_top_method_x2apicid,
748 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
749 affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
750 #if KMP_GROUP_AFFINITY
751 affinity_top_method_group,
752 #endif /* KMP_GROUP_AFFINITY */
753 affinity_top_method_flat,
755 affinity_top_method_hwloc,
757 affinity_top_method_default
760 #define affinity_respect_mask_default (-1)
762 extern enum affinity_type __kmp_affinity_type; /* Affinity type */
763 extern enum affinity_gran __kmp_affinity_gran; /* Affinity granularity */
764 extern int __kmp_affinity_gran_levels; /* corresponding int value */
765 extern int __kmp_affinity_dups; /* Affinity duplicate masks */
766 extern enum affinity_top_method __kmp_affinity_top_method;
767 extern int __kmp_affinity_compact; /* Affinity 'compact' value */
768 extern int __kmp_affinity_offset; /* Affinity offset value */
769 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */
770 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */
771 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask?
772 extern char *__kmp_affinity_proclist; /* proc ID list */
773 extern kmp_affin_mask_t *__kmp_affinity_masks;
774 extern unsigned __kmp_affinity_num_masks;
775 extern void __kmp_affinity_bind_thread(int which);
777 extern kmp_affin_mask_t *__kmp_affin_fullMask;
778 extern char *__kmp_cpuinfo_file;
780 #endif /* KMP_AFFINITY_SUPPORTED */
784 // This needs to be kept in sync with the values in omp.h !!!
785 typedef enum kmp_proc_bind_t {
791 proc_bind_intel, // use KMP_AFFINITY interface
795 typedef struct kmp_nested_proc_bind_t {
796 kmp_proc_bind_t *bind_types;
799 } kmp_nested_proc_bind_t;
801 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
803 #endif /* OMP_40_ENABLED */
806 extern int __kmp_display_affinity;
807 extern char *__kmp_affinity_format;
808 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
809 #endif // OMP_50_ENABLED
811 #if KMP_AFFINITY_SUPPORTED
812 #define KMP_PLACE_ALL (-1)
813 #define KMP_PLACE_UNDEFINED (-2)
814 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
815 #define KMP_AFFINITY_NON_PROC_BIND \
816 ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
817 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
818 (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
819 #endif /* KMP_AFFINITY_SUPPORTED */
821 extern int __kmp_affinity_num_places;
824 typedef enum kmp_cancel_kind_t {
831 #endif // OMP_40_ENABLED
833 // KMP_HW_SUBSET support:
834 typedef struct kmp_hws_item {
839 extern kmp_hws_item_t __kmp_hws_socket;
840 extern kmp_hws_item_t __kmp_hws_node;
841 extern kmp_hws_item_t __kmp_hws_tile;
842 extern kmp_hws_item_t __kmp_hws_core;
843 extern kmp_hws_item_t __kmp_hws_proc;
844 extern int __kmp_hws_requested;
845 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
847 /* ------------------------------------------------------------------------ */
849 #define KMP_PAD(type, sz) \
850 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
852 // We need to avoid using -1 as a GTID as +1 is added to the gtid
853 // when storing it in a lock, and the value 0 is reserved.
854 #define KMP_GTID_DNE (-2) /* Does not exist */
855 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
856 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
857 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
858 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
861 /* OpenMP 5.0 Memory Management support */
862 extern int __kmp_memkind_available;
863 extern int __kmp_hbw_mem_available;
864 typedef void *omp_allocator_t;
865 extern const omp_allocator_t *OMP_NULL_ALLOCATOR;
866 extern const omp_allocator_t *omp_default_mem_alloc;
867 extern const omp_allocator_t *omp_large_cap_mem_alloc;
868 extern const omp_allocator_t *omp_const_mem_alloc;
869 extern const omp_allocator_t *omp_high_bw_mem_alloc;
870 extern const omp_allocator_t *omp_low_lat_mem_alloc;
871 extern const omp_allocator_t *omp_cgroup_mem_alloc;
872 extern const omp_allocator_t *omp_pteam_mem_alloc;
873 extern const omp_allocator_t *omp_thread_mem_alloc;
874 extern const omp_allocator_t *__kmp_def_allocator;
876 extern void __kmpc_set_default_allocator(int gtid, const omp_allocator_t *al);
877 extern const omp_allocator_t *__kmpc_get_default_allocator(int gtid);
878 extern void *__kmpc_alloc(int gtid, size_t sz, const omp_allocator_t *al);
879 extern void __kmpc_free(int gtid, void *ptr, const omp_allocator_t *al);
881 extern void __kmp_init_memkind();
882 extern void __kmp_fini_memkind();
883 #endif // OMP_50_ENABLED
885 /* ------------------------------------------------------------------------ */
887 #define KMP_UINT64_MAX \
888 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
890 #define KMP_MIN_NTH 1
893 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
894 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
896 #define KMP_MAX_NTH INT_MAX
898 #endif /* KMP_MAX_NTH */
900 #ifdef PTHREAD_STACK_MIN
901 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
903 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
906 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
909 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
910 #elif KMP_ARCH_X86_64
911 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
912 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
914 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
917 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
918 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
919 #define KMP_MAX_MALLOC_POOL_INCR \
920 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
922 #define KMP_MIN_STKOFFSET (0)
923 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
925 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
927 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
930 #define KMP_MIN_STKPADDING (0)
931 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
933 #define KMP_BLOCKTIME_MULTIPLIER \
934 (1000) /* number of blocktime units per second */
935 #define KMP_MIN_BLOCKTIME (0)
936 #define KMP_MAX_BLOCKTIME \
937 (INT_MAX) /* Must be this for "infinite" setting the work */
938 #define KMP_DEFAULT_BLOCKTIME (200) /* __kmp_blocktime is in milliseconds */
941 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
942 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
943 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
945 /* Calculate new number of monitor wakeups for a specific block time based on
946 previous monitor_wakeups. Only allow increasing number of wakeups */
947 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
948 (((blocktime) == KMP_MAX_BLOCKTIME) \
949 ? (monitor_wakeups) \
950 : ((blocktime) == KMP_MIN_BLOCKTIME) \
951 ? KMP_MAX_MONITOR_WAKEUPS \
952 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
953 ? (monitor_wakeups) \
954 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
956 /* Calculate number of intervals for a specific block time based on
958 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
959 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
960 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
962 #define KMP_BLOCKTIME(team, tid) \
963 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
964 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
965 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
966 extern kmp_uint64 __kmp_ticks_per_msec;
968 #define KMP_NOW() ((kmp_uint64)_rdtsc())
970 #define KMP_NOW() __kmp_hardware_timestamp()
972 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
973 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
974 (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
975 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
977 // System time is retrieved sporadically while blocking.
978 extern kmp_uint64 __kmp_now_nsec();
979 #define KMP_NOW() __kmp_now_nsec()
980 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
981 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
982 (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
983 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
985 #define KMP_YIELD_NOW() \
986 (KMP_NOW_MSEC() / KMP_MAX(__kmp_dflt_blocktime, 1) % \
987 (__kmp_yield_on_count + __kmp_yield_off_count) < \
988 (kmp_uint32)__kmp_yield_on_count)
989 #endif // KMP_USE_MONITOR
991 #define KMP_MIN_STATSCOLS 40
992 #define KMP_MAX_STATSCOLS 4096
993 #define KMP_DEFAULT_STATSCOLS 80
995 #define KMP_MIN_INTERVAL 0
996 #define KMP_MAX_INTERVAL (INT_MAX - 1)
997 #define KMP_DEFAULT_INTERVAL 0
999 #define KMP_MIN_CHUNK 1
1000 #define KMP_MAX_CHUNK (INT_MAX - 1)
1001 #define KMP_DEFAULT_CHUNK 1
1003 #define KMP_MIN_INIT_WAIT 1
1004 #define KMP_MAX_INIT_WAIT (INT_MAX / 2)
1005 #define KMP_DEFAULT_INIT_WAIT 2048U
1007 #define KMP_MIN_NEXT_WAIT 1
1008 #define KMP_MAX_NEXT_WAIT (INT_MAX / 2)
1009 #define KMP_DEFAULT_NEXT_WAIT 1024U
1011 #define KMP_DFLT_DISP_NUM_BUFF 7
1012 #define KMP_MAX_ORDERED 8
1014 #define KMP_MAX_FIELDS 32
1016 #define KMP_MAX_BRANCH_BITS 31
1018 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1020 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1022 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1024 /* Minimum number of threads before switch to TLS gtid (experimentally
1026 /* josh TODO: what about OS X* tuning? */
1027 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1028 #define KMP_TLS_GTID_MIN 5
1030 #define KMP_TLS_GTID_MIN INT_MAX
1033 #define KMP_MASTER_TID(tid) ((tid) == 0)
1034 #define KMP_WORKER_TID(tid) ((tid) != 0)
1036 #define KMP_MASTER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) == 0)
1037 #define KMP_WORKER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) != 0)
1038 #define KMP_INITIAL_GTID(gtid) ((gtid) == 0)
1042 #define TRUE (!FALSE)
1045 /* NOTE: all of the following constants must be even */
1048 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1049 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1051 #define KMP_INIT_WAIT 16U /* initial number of spin-tests */
1052 #define KMP_NEXT_WAIT 8U /* susequent number of spin-tests */
1054 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1055 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1057 /* TODO: tune for KMP_OS_DARWIN */
1058 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1059 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1060 #elif KMP_OS_DRAGONFLY
1061 /* TODO: tune for KMP_OS_DRAGONFLY */
1062 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1063 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1064 #elif KMP_OS_FREEBSD
1065 /* TODO: tune for KMP_OS_FREEBSD */
1066 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1067 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1069 /* TODO: tune for KMP_OS_NETBSD */
1070 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1071 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1073 /* TODO: tune for KMP_OS_HURD */
1074 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1075 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1076 #elif KMP_OS_OPENBSD
1077 /* TODO: tune for KMP_OS_OPENBSD */
1078 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1079 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1082 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1083 typedef struct kmp_cpuid {
1089 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1091 extern void __kmp_x86_pause(void);
1093 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1094 // regression after removal of extra PAUSE from KMP_YIELD_SPIN(). Changing
1095 // the delay from 100 to 300 showed even better performance than double PAUSE
1096 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1097 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1099 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1101 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1102 #elif KMP_ARCH_PPC64
1103 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1104 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1105 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1106 #define KMP_CPU_PAUSE() \
1108 KMP_PPC64_PRI_LOW(); \
1109 KMP_PPC64_PRI_MED(); \
1110 KMP_PPC64_PRI_LOC_MB(); \
1113 #define KMP_CPU_PAUSE() /* nothing to do */
1116 #define KMP_INIT_YIELD(count) \
1117 { (count) = __kmp_yield_init; }
1119 #define KMP_YIELD(cond) \
1122 __kmp_yield((cond)); \
1125 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1126 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1128 #define KMP_YIELD_WHEN(cond, count) \
1133 __kmp_yield(cond); \
1134 (count) = __kmp_yield_next; \
1137 #define KMP_YIELD_SPIN(count) \
1143 (count) = __kmp_yield_next; \
1147 /* ------------------------------------------------------------------------ */
1148 /* Support datatypes for the orphaned construct nesting checks. */
1149 /* ------------------------------------------------------------------------ */
1159 /* the following must be left in order and not split up */
1161 ct_task, // really task inside non-ordered taskq, considered worksharing type
1162 ct_task_ordered, /* really task inside ordered taskq, considered a worksharing
1164 /* the preceding must be left in order and not split up */
1167 ct_ordered_in_parallel,
1169 ct_ordered_in_taskq,
1175 /* test to see if we are in a taskq construct */
1176 #define IS_CONS_TYPE_TASKQ(ct) \
1177 (((int)(ct)) >= ((int)ct_taskq) && ((int)(ct)) <= ((int)ct_task_ordered))
1178 #define IS_CONS_TYPE_ORDERED(ct) \
1179 ((ct) == ct_pdo_ordered || (ct) == ct_task_ordered)
1182 ident_t const *ident;
1183 enum cons_type type;
1186 name; /* address exclusively for critical section name comparison */
1189 struct cons_header {
1190 int p_top, w_top, s_top;
1191 int stack_size, stack_top;
1192 struct cons_data *stack_data;
1195 struct kmp_region_info {
1197 int offset[KMP_MAX_FIELDS];
1198 int length[KMP_MAX_FIELDS];
1201 /* ---------------------------------------------------------------------- */
1202 /* ---------------------------------------------------------------------- */
1205 typedef HANDLE kmp_thread_t;
1206 typedef DWORD kmp_key_t;
1207 #endif /* KMP_OS_WINDOWS */
1210 typedef pthread_t kmp_thread_t;
1211 typedef pthread_key_t kmp_key_t;
1214 extern kmp_key_t __kmp_gtid_threadprivate_key;
1216 typedef struct kmp_sys_info {
1217 long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1218 long minflt; /* the number of page faults serviced without any I/O */
1219 long majflt; /* the number of page faults serviced that required I/O */
1220 long nswap; /* the number of times a process was "swapped" out of memory */
1221 long inblock; /* the number of times the file system had to perform input */
1222 long oublock; /* the number of times the file system had to perform output */
1223 long nvcsw; /* the number of times a context switch was voluntarily */
1224 long nivcsw; /* the number of times a context switch was forced */
1227 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1228 typedef struct kmp_cpuinfo {
1229 int initialized; // If 0, other fields are not initialized.
1230 int signature; // CPUID(1).EAX
1231 int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1232 int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1233 // Model << 4 ) + Model)
1234 int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1235 int sse2; // 0 if SSE2 instructions are not supported, 1 otherwise.
1236 int rtm; // 0 if RTM instructions are not supported, 1 otherwise.
1237 int cpu_stackoffset;
1241 kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1242 char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1247 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1248 // required type here. Later we will check the type meets requirements.
1249 typedef int kmp_itt_mark_t;
1250 #define KMP_ITT_DEBUG 0
1251 #endif /* USE_ITT_BUILD */
1253 /* Taskq data structures */
1255 #define HIGH_WATER_MARK(nslots) (((nslots)*3) / 4)
1256 // num thunks that each thread can simultaneously execute from a task queue
1257 #define __KMP_TASKQ_THUNKS_PER_TH 1
1259 /* flags for taskq_global_flags, kmp_task_queue_t tq_flags, kmpc_thunk_t
1262 #define TQF_IS_ORDERED 0x0001 // __kmpc_taskq interface, taskq ordered
1263 // __kmpc_taskq interface, taskq with lastprivate list
1264 #define TQF_IS_LASTPRIVATE 0x0002
1265 #define TQF_IS_NOWAIT 0x0004 // __kmpc_taskq interface, end taskq nowait
1266 // __kmpc_taskq interface, use heuristics to decide task queue size
1267 #define TQF_HEURISTICS 0x0008
1269 // __kmpc_taskq interface, reserved for future use
1270 #define TQF_INTERFACE_RESERVED1 0x0010
1271 // __kmpc_taskq interface, reserved for future use
1272 #define TQF_INTERFACE_RESERVED2 0x0020
1273 // __kmpc_taskq interface, reserved for future use
1274 #define TQF_INTERFACE_RESERVED3 0x0040
1275 // __kmpc_taskq interface, reserved for future use
1276 #define TQF_INTERFACE_RESERVED4 0x0080
1278 #define TQF_INTERFACE_FLAGS 0x00ff // all the __kmpc_taskq interface flags
1279 // internal/read by instrumentation; only used with TQF_IS_LASTPRIVATE
1280 #define TQF_IS_LAST_TASK 0x0100
1281 // internal use only; this thunk->th_task is the taskq_task
1282 #define TQF_TASKQ_TASK 0x0200
1283 // internal use only; must release worker threads once ANY queued task
1285 #define TQF_RELEASE_WORKERS 0x0400
1286 // internal use only; notify workers that master has finished enqueuing tasks
1287 #define TQF_ALL_TASKS_QUEUED 0x0800
1288 // internal use only: this queue encountered in parallel context: not serialized
1289 #define TQF_PARALLEL_CONTEXT 0x1000
1290 // internal use only; this queue is on the freelist and not in use
1291 #define TQF_DEALLOCATED 0x2000
1293 #define TQF_INTERNAL_FLAGS 0x3f00 // all the internal use only flags
1295 typedef struct KMP_ALIGN_CACHE kmpc_aligned_int32_t {
1297 } kmpc_aligned_int32_t;
1299 typedef struct KMP_ALIGN_CACHE kmpc_aligned_queue_slot_t {
1300 struct kmpc_thunk_t *qs_thunk;
1301 } kmpc_aligned_queue_slot_t;
1303 typedef struct kmpc_task_queue_t {
1304 /* task queue linkage fields for n-ary tree of queues (locked with global
1306 kmp_lock_t tq_link_lck; /* lock for child link, child next/prev links and
1309 struct kmpc_task_queue_t *tq_parent; // pointer to parent taskq, not locked
1310 // for taskq internal freelists, locked with global taskq_freelist_lck
1311 struct kmpc_task_queue_t *tq_next_free;
1313 // pointer to linked-list of children, locked by tq's tq_link_lck
1314 volatile struct kmpc_task_queue_t *tq_first_child;
1315 // next child in linked-list, locked by parent tq's tq_link_lck
1316 struct kmpc_task_queue_t *tq_next_child;
1317 // previous child in linked-list, locked by parent tq's tq_link_lck
1318 struct kmpc_task_queue_t *tq_prev_child;
1319 // reference count of threads with access to this task queue
1320 volatile kmp_int32 tq_ref_count;
1321 /* (other than the thread executing the kmpc_end_taskq call) */
1322 /* locked by parent tq's tq_link_lck */
1324 /* shared data for task queue */
1325 /* per-thread array of pointers to shared variable structures */
1326 struct kmpc_aligned_shared_vars_t *tq_shareds;
1327 /* only one array element exists for all but outermost taskq */
1329 /* bookkeeping for ordered task queue */
1330 kmp_uint32 tq_tasknum_queuing; // ordered task # assigned while queuing tasks
1331 // ordered number of next task to be served (executed)
1332 volatile kmp_uint32 tq_tasknum_serving;
1334 /* thunk storage management for task queue */
1335 kmp_lock_t tq_free_thunks_lck; /* lock for thunk freelist manipulation */
1336 // thunk freelist, chained via th.th_next_free
1337 struct kmpc_thunk_t *tq_free_thunks;
1338 // space allocated for thunks for this task queue
1339 struct kmpc_thunk_t *tq_thunk_space;
1341 /* data fields for queue itself */
1342 kmp_lock_t tq_queue_lck; /* lock for [de]enqueue operations: tq_queue,
1343 tq_head, tq_tail, tq_nfull */
1344 /* array of queue slots to hold thunks for tasks */
1345 kmpc_aligned_queue_slot_t *tq_queue;
1346 volatile struct kmpc_thunk_t *tq_taskq_slot; /* special slot for taskq task
1347 thunk, occupied if not NULL */
1348 kmp_int32 tq_nslots; /* # of tq_thunk_space thunks alloc'd (not incl.
1349 tq_taskq_slot space) */
1350 kmp_int32 tq_head; // enqueue puts item here (index into tq_queue array)
1351 kmp_int32 tq_tail; // dequeue takes item from here (index into tq_queue array)
1352 volatile kmp_int32 tq_nfull; // # of occupied entries in task queue right now
1353 kmp_int32 tq_hiwat; /* high-water mark for tq_nfull and queue scheduling */
1354 volatile kmp_int32 tq_flags; /* TQF_xxx */
1356 /* bookkeeping for outstanding thunks */
1358 /* per-thread array for # of regular thunks currently being executed */
1359 struct kmpc_aligned_int32_t *tq_th_thunks;
1360 kmp_int32 tq_nproc; /* number of thunks in the th_thunks array */
1362 /* statistics library bookkeeping */
1363 ident_t *tq_loc; /* source location information for taskq directive */
1364 } kmpc_task_queue_t;
1366 typedef void (*kmpc_task_t)(kmp_int32 global_tid, struct kmpc_thunk_t *thunk);
1368 /* sizeof_shareds passed as arg to __kmpc_taskq call */
1369 typedef struct kmpc_shared_vars_t { /* aligned during dynamic allocation */
1370 kmpc_task_queue_t *sv_queue; /* (pointers to) shared vars */
1371 } kmpc_shared_vars_t;
1373 typedef struct KMP_ALIGN_CACHE kmpc_aligned_shared_vars_t {
1374 volatile struct kmpc_shared_vars_t *ai_data;
1375 } kmpc_aligned_shared_vars_t;
1377 /* sizeof_thunk passed as arg to kmpc_taskq call */
1378 typedef struct kmpc_thunk_t { /* aligned during dynamic allocation */
1379 union { /* field used for internal freelists too */
1380 kmpc_shared_vars_t *th_shareds;
1381 struct kmpc_thunk_t *th_next_free; /* freelist of individual thunks within
1382 queue, head at tq_free_thunks */
1384 kmpc_task_t th_task; /* taskq_task if flags & TQF_TASKQ_TASK */
1385 struct kmpc_thunk_t *th_encl_thunk; /* pointer to dynamically enclosing thunk
1386 on this thread's call stack */
1387 // TQF_xxx(tq_flags interface plus possible internal flags)
1390 kmp_int32 th_status;
1391 kmp_uint32 th_tasknum; /* task number assigned in order of queuing, used for
1396 typedef struct KMP_ALIGN_CACHE kmp_taskq {
1397 int tq_curr_thunk_capacity;
1399 kmpc_task_queue_t *tq_root;
1400 kmp_int32 tq_global_flags;
1402 kmp_lock_t tq_freelist_lck;
1403 kmpc_task_queue_t *tq_freelist;
1405 kmpc_thunk_t **tq_curr_thunk;
1408 /* END Taskq data structures */
1410 typedef kmp_int32 kmp_critical_name[8];
1414 The type for a microtask which gets passed to @ref __kmpc_fork_call().
1415 The arguments to the outlined function are
1416 @param global_tid the global thread identity of the thread executing the
1418 @param bound_tid the local identitiy of the thread executing the function
1419 @param ... pointers to shared variables accessed by the function.
1421 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1422 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1426 @ingroup THREADPRIVATE
1429 /* ---------------------------------------------------------------------------
1431 /* Threadprivate initialization/finalization function declarations */
1433 /* for non-array objects: __kmpc_threadprivate_register() */
1436 Pointer to the constructor function.
1437 The first argument is the <tt>this</tt> pointer
1439 typedef void *(*kmpc_ctor)(void *);
1442 Pointer to the destructor function.
1443 The first argument is the <tt>this</tt> pointer
1445 typedef void (*kmpc_dtor)(
1446 void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1449 Pointer to an alternate constructor.
1450 The first argument is the <tt>this</tt> pointer.
1452 typedef void *(*kmpc_cctor)(void *, void *);
1454 /* for array objects: __kmpc_threadprivate_register_vec() */
1455 /* First arg: "this" pointer */
1456 /* Last arg: number of array elements */
1459 First argument is the <tt>this</tt> pointer
1460 Second argument the number of array elements.
1462 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1464 Pointer to the array destructor function.
1465 The first argument is the <tt>this</tt> pointer
1466 Second argument the number of array elements.
1468 typedef void (*kmpc_dtor_vec)(void *, size_t);
1471 First argument is the <tt>this</tt> pointer
1472 Third argument the number of array elements.
1474 typedef void *(*kmpc_cctor_vec)(void *, void *,
1475 size_t); /* function unused by compiler */
1481 /* keeps tracked of threadprivate cache allocations for cleanup later */
1482 typedef struct kmp_cached_addr {
1483 void **addr; /* address of allocated cache */
1484 void ***compiler_cache; /* pointer to compiler's cache */
1485 void *data; /* pointer to global data */
1486 struct kmp_cached_addr *next; /* pointer to next cached address */
1487 } kmp_cached_addr_t;
1489 struct private_data {
1490 struct private_data *next; /* The next descriptor in the list */
1491 void *data; /* The data buffer for this descriptor */
1492 int more; /* The repeat count for this descriptor */
1493 size_t size; /* The data size for this descriptor */
1496 struct private_common {
1497 struct private_common *next;
1498 struct private_common *link;
1500 void *par_addr; /* par_addr == gbl_addr for MASTER thread */
1504 struct shared_common {
1505 struct shared_common *next;
1506 struct private_data *pod_init;
1511 kmpc_ctor_vec ctorv;
1515 kmpc_cctor_vec cctorv;
1519 kmpc_dtor_vec dtorv;
1526 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1527 #define KMP_HASH_TABLE_SIZE \
1528 (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1529 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1530 #define KMP_HASH(x) \
1531 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1533 struct common_table {
1534 struct private_common *data[KMP_HASH_TABLE_SIZE];
1537 struct shared_table {
1538 struct shared_common *data[KMP_HASH_TABLE_SIZE];
1541 /* ------------------------------------------------------------------------ */
1543 #if KMP_USE_HIER_SCHED
1544 // Shared barrier data that exists inside a single unit of the scheduling
1546 typedef struct kmp_hier_private_bdata_t {
1547 kmp_int32 num_active;
1549 kmp_uint64 wait_val[2];
1550 } kmp_hier_private_bdata_t;
1553 typedef struct kmp_sched_flags {
1554 unsigned ordered : 1;
1555 unsigned nomerge : 1;
1556 unsigned contains_last : 1;
1557 #if KMP_USE_HIER_SCHED
1558 unsigned use_hier : 1;
1559 unsigned unused : 28;
1561 unsigned unused : 29;
1563 } kmp_sched_flags_t;
1565 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1567 #if KMP_STATIC_STEAL_ENABLED
1568 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1571 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1575 kmp_int32 static_steal_counter; /* for static_steal only; maybe better to put
1578 // KMP_ALIGN( 16 ) ensures ( if the KMP_ALIGN macro is turned on )
1579 // a) parm3 is properly aligned and
1580 // b) all parm1-4 are in the same cache line.
1581 // Because of parm1-4 are used together, performance seems to be better
1582 // if they are in the same line (not measured though).
1584 struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1585 kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1586 kmp_int32 parm2; // make no real change at least while padding is off.
1591 kmp_uint32 ordered_lower;
1592 kmp_uint32 ordered_upper;
1594 // This var can be placed in the hole between 'tc' and 'parm1', instead of
1595 // 'static_steal_counter'. It would be nice to measure execution times.
1596 // Conditional if/endif can be removed at all.
1597 kmp_int32 last_upper;
1598 #endif /* KMP_OS_WINDOWS */
1599 } dispatch_private_info32_t;
1601 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1602 kmp_int64 count; // current chunk number for static & static-steal scheduling
1603 kmp_int64 ub; /* upper-bound */
1604 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1605 kmp_int64 lb; /* lower-bound */
1606 kmp_int64 st; /* stride */
1607 kmp_int64 tc; /* trip count (number of iterations) */
1608 kmp_int64 static_steal_counter; /* for static_steal only; maybe better to put
1611 /* parm[1-4] are used in different ways by different scheduling algorithms */
1613 // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1614 // a) parm3 is properly aligned and
1615 // b) all parm1-4 are in the same cache line.
1616 // Because of parm1-4 are used together, performance seems to be better
1617 // if they are in the same line (not measured though).
1619 struct KMP_ALIGN(32) {
1626 kmp_uint64 ordered_lower;
1627 kmp_uint64 ordered_upper;
1629 // This var can be placed in the hole between 'tc' and 'parm1', instead of
1630 // 'static_steal_counter'. It would be nice to measure execution times.
1631 // Conditional if/endif can be removed at all.
1632 kmp_int64 last_upper;
1633 #endif /* KMP_OS_WINDOWS */
1634 } dispatch_private_info64_t;
1635 #else /* KMP_STATIC_STEAL_ENABLED */
1636 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1649 kmp_uint32 ordered_lower;
1650 kmp_uint32 ordered_upper;
1652 kmp_int32 last_upper;
1653 #endif /* KMP_OS_WINDOWS */
1654 } dispatch_private_info32_t;
1656 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1657 kmp_int64 lb; /* lower-bound */
1658 kmp_int64 ub; /* upper-bound */
1659 kmp_int64 st; /* stride */
1660 kmp_int64 tc; /* trip count (number of iterations) */
1662 /* parm[1-4] are used in different ways by different scheduling algorithms */
1668 kmp_int64 count; /* current chunk number for static scheduling */
1670 kmp_uint64 ordered_lower;
1671 kmp_uint64 ordered_upper;
1673 kmp_int64 last_upper;
1674 #endif /* KMP_OS_WINDOWS */
1675 } dispatch_private_info64_t;
1676 #endif /* KMP_STATIC_STEAL_ENABLED */
1678 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1679 union private_info {
1680 dispatch_private_info32_t p32;
1681 dispatch_private_info64_t p64;
1683 enum sched_type schedule; /* scheduling algorithm */
1684 kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1685 kmp_int32 ordered_bumped;
1686 // To retain the structure size after making ordered_iteration scalar
1687 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 3];
1688 // Stack of buffers for nest of serial regions
1689 struct dispatch_private_info *next;
1690 kmp_int32 type_size; /* the size of types in private_info */
1691 #if KMP_USE_HIER_SCHED
1693 void *parent; /* hierarchical scheduling parent pointer */
1695 enum cons_type pushed_ws;
1696 } dispatch_private_info_t;
1698 typedef struct dispatch_shared_info32 {
1699 /* chunk index under dynamic, number of idle threads under static-steal;
1700 iteration index otherwise */
1701 volatile kmp_uint32 iteration;
1702 volatile kmp_uint32 num_done;
1703 volatile kmp_uint32 ordered_iteration;
1704 // Dummy to retain the structure size after making ordered_iteration scalar
1705 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1706 } dispatch_shared_info32_t;
1708 typedef struct dispatch_shared_info64 {
1709 /* chunk index under dynamic, number of idle threads under static-steal;
1710 iteration index otherwise */
1711 volatile kmp_uint64 iteration;
1712 volatile kmp_uint64 num_done;
1713 volatile kmp_uint64 ordered_iteration;
1714 // Dummy to retain the structure size after making ordered_iteration scalar
1715 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1716 } dispatch_shared_info64_t;
1718 typedef struct dispatch_shared_info {
1720 dispatch_shared_info32_t s32;
1721 dispatch_shared_info64_t s64;
1723 volatile kmp_uint32 buffer_index;
1725 volatile kmp_int32 doacross_buf_idx; // teamwise index
1726 volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1727 kmp_int32 doacross_num_done; // count finished threads
1729 #if KMP_USE_HIER_SCHED
1733 // When linking with libhwloc, the ORDERED EPCC test slows down on big
1734 // machines (> 48 cores). Performance analysis showed that a cache thrash
1735 // was occurring and this padding helps alleviate the problem.
1738 } dispatch_shared_info_t;
1740 typedef struct kmp_disp {
1741 /* Vector for ORDERED SECTION */
1742 void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1743 /* Vector for END ORDERED SECTION */
1744 void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1746 dispatch_shared_info_t *th_dispatch_sh_current;
1747 dispatch_private_info_t *th_dispatch_pr_current;
1749 dispatch_private_info_t *th_disp_buffer;
1750 kmp_int32 th_disp_index;
1752 kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1753 volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1754 union { // we can use union here because doacross cannot be used in
1755 // nonmonotonic loops
1756 kmp_int64 *th_doacross_info; // info on loop bounds
1757 kmp_lock_t *th_steal_lock; // lock used for chunk stealing (8-byte variable)
1760 #if KMP_STATIC_STEAL_ENABLED
1761 kmp_lock_t *th_steal_lock; // lock used for chunk stealing (8-byte variable)
1762 void *dummy_padding[1]; // make it 64 bytes on Intel(R) 64
1764 void *dummy_padding[2]; // make it 64 bytes on Intel(R) 64
1767 #if KMP_USE_INTERNODE_ALIGNMENT
1768 char more_padding[INTERNODE_CACHE_LINE];
1772 /* ------------------------------------------------------------------------ */
1775 /* constants for barrier state update */
1776 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1777 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1778 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1779 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1781 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1782 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1783 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1785 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1786 #error "Barrier sleep bit must be smaller than barrier bump bit"
1788 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1789 #error "Barrier unused bit must be smaller than barrier bump bit"
1792 // Constants for release barrier wait state: currently, hierarchical only
1793 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1794 #define KMP_BARRIER_OWN_FLAG \
1795 1 // Normal state; worker waiting on own b_go flag in release
1796 #define KMP_BARRIER_PARENT_FLAG \
1797 2 // Special state; worker waiting on parent's b_go flag in release
1798 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1799 3 // Special state; tells worker to shift from parent to own b_go
1800 #define KMP_BARRIER_SWITCHING \
1801 4 // Special state; worker resets appropriate flag on wake-up
1803 #define KMP_NOT_SAFE_TO_REAP \
1804 0 // Thread th_reap_state: not safe to reap (tasking)
1805 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1808 bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1809 barriers if enabled) */
1810 bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1811 #if KMP_FAST_REDUCTION_BARRIER
1812 bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1813 #endif // KMP_FAST_REDUCTION_BARRIER
1814 bs_last_barrier /* Just a placeholder to mark the end */
1817 // to work with reduction barriers just like with plain barriers
1818 #if !KMP_FAST_REDUCTION_BARRIER
1819 #define bs_reduction_barrier bs_plain_barrier
1820 #endif // KMP_FAST_REDUCTION_BARRIER
1822 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1824 0, /* Single level (degenerate) tree */
1826 1, /* Balanced tree with branching factor 2^n */
1828 2, /* Hypercube-embedded tree with min branching
1830 bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1831 bp_last_bar /* Placeholder to mark the end */
1834 #define KMP_BARRIER_ICV_PUSH 1
1836 /* Record for holding the values of the internal controls stack records */
1837 typedef struct kmp_internal_control {
1838 int serial_nesting_level; /* corresponds to the value of the
1839 th_team_serialized field */
1840 kmp_int8 nested; /* internal control for nested parallelism (per thread) */
1841 kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1844 bt_set; /* internal control for whether blocktime is explicitly set */
1845 int blocktime; /* internal control for blocktime */
1847 int bt_intervals; /* internal control for blocktime intervals */
1849 int nproc; /* internal control for #threads for next parallel region (per
1851 int max_active_levels; /* internal control for max_active_levels */
1853 sched; /* internal control for runtime schedule {sched,chunk} pair */
1855 kmp_proc_bind_t proc_bind; /* internal control for affinity */
1856 kmp_int32 default_device; /* internal control for default device */
1857 #endif // OMP_40_ENABLED
1858 struct kmp_internal_control *next;
1859 } kmp_internal_control_t;
1861 static inline void copy_icvs(kmp_internal_control_t *dst,
1862 kmp_internal_control_t *src) {
1866 /* Thread barrier needs volatile barrier fields */
1867 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1868 // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1869 // uses of it). It is not explicitly aligned below, because we *don't* want
1870 // it to be padded -- instead, we fit b_go into the same cache line with
1871 // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1872 kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1873 // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1875 volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1876 KMP_ALIGN_CACHE volatile kmp_uint64
1877 b_arrived; // STATE => task reached synch point.
1878 kmp_uint32 *skip_per_level;
1879 kmp_uint32 my_level;
1880 kmp_int32 parent_tid;
1883 struct kmp_bstate *parent_bar;
1885 kmp_uint64 leaf_state;
1887 kmp_uint8 base_leaf_kids;
1888 kmp_uint8 leaf_kids;
1890 kmp_uint8 wait_flag;
1891 kmp_uint8 use_oncore_barrier;
1893 // The following field is intended for the debugger solely. Only the worker
1894 // thread itself accesses this field: the worker increases it by 1 when it
1895 // arrives to a barrier.
1896 KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1897 #endif /* USE_DEBUGGER */
1900 union KMP_ALIGN_CACHE kmp_barrier_union {
1901 double b_align; /* use worst case alignment */
1902 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1906 typedef union kmp_barrier_union kmp_balign_t;
1908 /* Team barrier needs only non-volatile arrived counter */
1909 union KMP_ALIGN_CACHE kmp_barrier_team_union {
1910 double b_align; /* use worst case alignment */
1911 char b_pad[CACHE_LINE];
1913 kmp_uint64 b_arrived; /* STATE => task reached synch point. */
1915 // The following two fields are indended for the debugger solely. Only
1916 // master of the team accesses these fields: the first one is increased by
1917 // 1 when master arrives to a barrier, the second one is increased by one
1918 // when all the threads arrived.
1919 kmp_uint b_master_arrived;
1920 kmp_uint b_team_arrived;
1925 typedef union kmp_barrier_team_union kmp_balign_team_t;
1927 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
1928 threads when a condition changes. This is to workaround an NPTL bug where
1929 padding was added to pthread_cond_t which caused the initialization routine
1930 to write outside of the structure if compiled on pre-NPTL threads. */
1932 typedef struct kmp_win32_mutex {
1934 CRITICAL_SECTION cs;
1935 } kmp_win32_mutex_t;
1937 typedef struct kmp_win32_cond {
1938 /* Count of the number of waiters. */
1941 /* Serialize access to <waiters_count_> */
1942 kmp_win32_mutex_t waiters_count_lock_;
1944 /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
1947 /* Keeps track of the current "generation" so that we don't allow */
1948 /* one thread to steal all the "releases" from the broadcast. */
1949 int wait_generation_count_;
1951 /* A manual-reset event that's used to block and release waiting threads. */
1958 union KMP_ALIGN_CACHE kmp_cond_union {
1960 char c_pad[CACHE_LINE];
1961 pthread_cond_t c_cond;
1964 typedef union kmp_cond_union kmp_cond_align_t;
1966 union KMP_ALIGN_CACHE kmp_mutex_union {
1968 char m_pad[CACHE_LINE];
1969 pthread_mutex_t m_mutex;
1972 typedef union kmp_mutex_union kmp_mutex_align_t;
1974 #endif /* KMP_OS_UNIX */
1976 typedef struct kmp_desc_base {
1978 size_t ds_stacksize;
1980 kmp_thread_t ds_thread;
1981 volatile int ds_tid;
1984 volatile int ds_alive;
1986 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
1987 However, debugger support (libomp_db) cannot work with handles, because they
1988 uncomparable. For example, debugger requests info about thread with handle h.
1989 h is valid within debugger process, and meaningless within debugee process.
1990 Even if h is duped by call to DuplicateHandle(), so the result h' is valid
1991 within debugee process, but it is a *new* handle which does *not* equal to
1992 any other handle in debugee... The only way to compare handles is convert
1993 them to system-wide ids. GetThreadId() function is available only in
1994 Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
1995 on all Windows* OS flavours (including Windows* 95). Thus, we have to get
1996 thread id by call to GetCurrentThreadId() from within the thread and save it
1997 to let libomp_db identify threads. */
1998 #endif /* KMP_OS_WINDOWS */
2001 typedef union KMP_ALIGN_CACHE kmp_desc {
2002 double ds_align; /* use worst case alignment */
2003 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2007 typedef struct kmp_local {
2008 volatile int this_construct; /* count of single's encountered by thread */
2013 #if !USE_CMP_XCHG_FOR_BGET
2014 #ifdef USE_QUEUING_LOCK_FOR_BGET
2015 kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2017 kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2018 // bootstrap lock so we can use it at library
2020 #endif /* USE_LOCK_FOR_BGET */
2021 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2022 #endif /* KMP_USE_BGET */
2024 PACKED_REDUCTION_METHOD_T
2025 packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2026 __kmpc_end_reduce*() */
2030 #define KMP_CHECK_UPDATE(a, b) \
2033 #define KMP_CHECK_UPDATE_SYNC(a, b) \
2035 TCW_SYNC_PTR((a), (b))
2037 #define get__blocktime(xteam, xtid) \
2038 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2039 #define get__bt_set(xteam, xtid) \
2040 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2042 #define get__bt_intervals(xteam, xtid) \
2043 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2046 #define get__nested_2(xteam, xtid) \
2047 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nested)
2048 #define get__dynamic_2(xteam, xtid) \
2049 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2050 #define get__nproc_2(xteam, xtid) \
2051 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2052 #define get__sched_2(xteam, xtid) \
2053 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2055 #define set__blocktime_team(xteam, xtid, xval) \
2056 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2060 #define set__bt_intervals_team(xteam, xtid, xval) \
2061 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2065 #define set__bt_set_team(xteam, xtid, xval) \
2066 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2068 #define set__nested(xthread, xval) \
2069 (((xthread)->th.th_current_task->td_icvs.nested) = (xval))
2070 #define get__nested(xthread) \
2071 (((xthread)->th.th_current_task->td_icvs.nested) ? (FTN_TRUE) : (FTN_FALSE))
2073 #define set__dynamic(xthread, xval) \
2074 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2075 #define get__dynamic(xthread) \
2076 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2078 #define set__nproc(xthread, xval) \
2079 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2081 #define set__max_active_levels(xthread, xval) \
2082 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2084 #define set__sched(xthread, xval) \
2085 (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2089 #define set__proc_bind(xthread, xval) \
2090 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2091 #define get__proc_bind(xthread) \
2092 ((xthread)->th.th_current_task->td_icvs.proc_bind)
2094 #endif /* OMP_40_ENABLED */
2096 // OpenMP tasking data structures
2098 typedef enum kmp_tasking_mode {
2099 tskm_immediate_exec = 0,
2100 tskm_extra_barrier = 1,
2101 tskm_task_teams = 2,
2103 } kmp_tasking_mode_t;
2105 extern kmp_tasking_mode_t
2106 __kmp_tasking_mode; /* determines how/when to execute tasks */
2107 extern int __kmp_task_stealing_constraint;
2109 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2110 // specified, defaults to 0 otherwise
2113 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2114 extern kmp_int32 __kmp_max_task_priority;
2115 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2116 extern kmp_uint64 __kmp_taskloop_min_tasks;
2119 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2121 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2122 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2124 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2125 // were spawned and queued since the previous barrier release.
2126 #define KMP_TASKING_ENABLED(task_team) \
2127 (TCR_SYNC_4((task_team)->tt.tt_found_tasks) == TRUE)
2129 @ingroup BASIC_TYPES
2135 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2137 #if OMP_40_ENABLED || OMP_45_ENABLED
2138 typedef union kmp_cmplrdata {
2140 kmp_int32 priority; /**< priority specified by user for the task */
2141 #endif // OMP_45_ENABLED
2144 destructors; /* pointer to function to invoke deconstructors of
2145 firstprivate C++ objects */
2146 #endif // OMP_40_ENABLED
2151 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2154 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2155 void *shareds; /**< pointer to block of pointers to shared vars */
2157 routine; /**< pointer to routine to call for executing task */
2158 kmp_int32 part_id; /**< part id for the task */
2159 #if OMP_40_ENABLED || OMP_45_ENABLED
2161 data1; /* Two known optional additions: destructors and priority */
2162 kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2173 typedef struct kmp_taskgroup {
2174 std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2175 std::atomic<kmp_int32>
2176 cancel_request; // request for cancellation of this taskgroup
2177 struct kmp_taskgroup *parent; // parent taskgroup
2179 // Block of data to perform task reduction
2180 void *reduce_data; // reduction related info
2181 kmp_int32 reduce_num_data; // number of data items to reduce
2185 // forward declarations
2186 typedef union kmp_depnode kmp_depnode_t;
2187 typedef struct kmp_depnode_list kmp_depnode_list_t;
2188 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2190 // Compiler sends us this info:
2191 typedef struct kmp_depend_info {
2192 kmp_intptr_t base_addr;
2199 } kmp_depend_info_t;
2201 // Internal structures to work with task dependencies:
2202 struct kmp_depnode_list {
2203 kmp_depnode_t *node;
2204 kmp_depnode_list_t *next;
2207 // Max number of mutexinoutset dependencies per node
2208 #define MAX_MTX_DEPS 4
2210 typedef struct kmp_base_depnode {
2211 kmp_depnode_list_t *successors; /* used under lock */
2212 kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2213 kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2214 kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2215 kmp_lock_t lock; /* guards shared fields: task, successors */
2216 #if KMP_SUPPORT_GRAPH_OUTPUT
2219 std::atomic<kmp_int32> npredecessors;
2220 std::atomic<kmp_int32> nrefs;
2221 } kmp_base_depnode_t;
2223 union KMP_ALIGN_CACHE kmp_depnode {
2224 double dn_align; /* use worst case alignment */
2225 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2226 kmp_base_depnode_t dn;
2229 struct kmp_dephash_entry {
2231 kmp_depnode_t *last_out;
2232 kmp_depnode_list_t *last_ins;
2233 kmp_depnode_list_t *last_mtxs;
2234 kmp_int32 last_flag;
2235 kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2236 kmp_dephash_entry_t *next_in_bucket;
2239 typedef struct kmp_dephash {
2240 kmp_dephash_entry_t **buckets;
2243 kmp_uint32 nelements;
2244 kmp_uint32 nconflicts;
2249 typedef struct kmp_task_affinity_info {
2250 kmp_intptr_t base_addr;
2255 kmp_int32 reserved : 30;
2257 } kmp_task_affinity_info_t;
2262 #ifdef BUILD_TIED_TASK_STACK
2264 /* Tied Task stack definitions */
2265 typedef struct kmp_stack_block {
2266 kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2267 struct kmp_stack_block *sb_next;
2268 struct kmp_stack_block *sb_prev;
2269 } kmp_stack_block_t;
2271 typedef struct kmp_task_stack {
2272 kmp_stack_block_t ts_first_block; // first block of stack entries
2273 kmp_taskdata_t **ts_top; // pointer to the top of stack
2274 kmp_int32 ts_entries; // number of entries on the stack
2277 #endif // BUILD_TIED_TASK_STACK
2279 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2280 /* Compiler flags */ /* Total compiler flags must be 16 bits */
2281 unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2282 unsigned final : 1; /* task is final(1) so execute immediately */
2283 unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2286 unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2287 invoke destructors from the runtime */
2289 unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2290 context of the RTL) */
2291 unsigned priority_specified : 1; /* set if the compiler provides priority
2292 setting for the task */
2293 unsigned reserved : 10; /* reserved for compiler use */
2295 unsigned reserved : 12; /* reserved for compiler use */
2297 #else // OMP_40_ENABLED
2298 unsigned reserved : 13; /* reserved for compiler use */
2299 #endif // OMP_40_ENABLED
2301 /* Library flags */ /* Total library flags must be 16 bits */
2302 unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2303 unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2304 unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2305 // (1) or may be deferred (0)
2306 unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2307 // (0) [>= 2 threads]
2308 /* If either team_serial or tasking_ser is set, task team may be NULL */
2309 /* Task State Flags: */
2310 unsigned started : 1; /* 1==started, 0==not started */
2311 unsigned executing : 1; /* 1==executing, 0==not executing */
2312 unsigned complete : 1; /* 1==complete, 0==not complete */
2313 unsigned freed : 1; /* 1==freed, 0==allocateed */
2314 unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2315 unsigned reserved31 : 7; /* reserved for library use */
2317 } kmp_tasking_flags_t;
2319 struct kmp_taskdata { /* aligned during dynamic allocation */
2320 kmp_int32 td_task_id; /* id, assigned by debugger */
2321 kmp_tasking_flags_t td_flags; /* task flags */
2322 kmp_team_t *td_team; /* team for this task */
2323 kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2324 /* Currently not used except for perhaps IDB */
2325 kmp_taskdata_t *td_parent; /* parent task */
2326 kmp_int32 td_level; /* task nesting level */
2327 std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2328 ident_t *td_ident; /* task identifier */
2330 ident_t *td_taskwait_ident;
2331 kmp_uint32 td_taskwait_counter;
2332 kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2333 KMP_ALIGN_CACHE kmp_internal_control_t
2334 td_icvs; /* Internal control variables for the task */
2335 KMP_ALIGN_CACHE std::atomic<kmp_int32>
2336 td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2338 std::atomic<kmp_int32>
2339 td_incomplete_child_tasks; /* Child tasks not yet complete */
2342 *td_taskgroup; // Each task keeps pointer to its current taskgroup
2344 *td_dephash; // Dependencies for children tasks are tracked from here
2346 *td_depnode; // Pointer to graph node if this task has dependencies
2347 #endif // OMP_40_ENABLED
2349 kmp_task_team_t *td_task_team;
2350 kmp_int32 td_size_alloc; // The size of task structure, including shareds etc.
2351 #if defined(KMP_GOMP_COMPAT)
2352 // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2353 kmp_int32 td_size_loop_bounds;
2355 #endif // OMP_45_ENABLED
2356 kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2357 #if defined(KMP_GOMP_COMPAT) && OMP_45_ENABLED
2358 // GOMP sends in a copy function for copy constructors
2359 void (*td_copy_func)(void *, void *);
2362 ompt_task_info_t ompt_task_info;
2364 }; // struct kmp_taskdata
2366 // Make sure padding above worked
2367 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2369 // Data for task team but per thread
2370 typedef struct kmp_base_thread_data {
2371 kmp_info_p *td_thr; // Pointer back to thread info
2372 // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2374 kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2376 *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2377 kmp_int32 td_deque_size; // Size of deck
2378 kmp_uint32 td_deque_head; // Head of deque (will wrap)
2379 kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2380 kmp_int32 td_deque_ntasks; // Number of tasks in deque
2381 // GEH: shouldn't this be volatile since used in while-spin?
2382 kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2383 #ifdef BUILD_TIED_TASK_STACK
2384 kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2385 // scheduling constraint
2386 #endif // BUILD_TIED_TASK_STACK
2387 } kmp_base_thread_data_t;
2389 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2390 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2392 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2393 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2395 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2396 kmp_base_thread_data_t td;
2397 double td_align; /* use worst case alignment */
2398 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2399 } kmp_thread_data_t;
2401 // Data for task teams which are used when tasking is enabled for the team
2402 typedef struct kmp_base_task_team {
2403 kmp_bootstrap_lock_t
2404 tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2405 /* must be bootstrap lock since used at library shutdown*/
2406 kmp_task_team_t *tt_next; /* For linking the task team free list */
2408 *tt_threads_data; /* Array of per-thread structures for task team */
2409 /* Data survives task team deallocation */
2410 kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2411 executing this team? */
2412 /* TRUE means tt_threads_data is set up and initialized */
2413 kmp_int32 tt_nproc; /* #threads in team */
2415 tt_max_threads; /* number of entries allocated for threads_data array */
2418 tt_found_proxy_tasks; /* Have we found proxy tasks since last barrier */
2420 kmp_int32 tt_untied_task_encountered;
2423 std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2427 tt_active; /* is the team still actively executing tasks */
2428 } kmp_base_task_team_t;
2430 union KMP_ALIGN_CACHE kmp_task_team {
2431 kmp_base_task_team_t tt;
2432 double tt_align; /* use worst case alignment */
2433 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2436 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2437 // Free lists keep same-size free memory slots for fast memory allocation
2439 typedef struct kmp_free_list {
2440 void *th_free_list_self; // Self-allocated tasks free list
2441 void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2443 void *th_free_list_other; // Non-self free list (to be returned to owner's
2447 #if KMP_NESTED_HOT_TEAMS
2448 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2449 // are not put in teams pool, and they don't put threads in threads pool.
2450 typedef struct kmp_hot_team_ptr {
2451 kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2452 kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2453 } kmp_hot_team_ptr_t;
2456 typedef struct kmp_teams_size {
2457 kmp_int32 nteams; // number of teams in a league
2458 kmp_int32 nth; // number of threads in each team of the league
2462 // OpenMP thread data structures
2464 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2465 /* Start with the readonly data which is cache aligned and padded. This is
2466 written before the thread starts working by the master. Uber masters may
2467 update themselves later. Usage does not consider serialized regions. */
2469 kmp_team_p *th_team; /* team we belong to */
2470 kmp_root_p *th_root; /* pointer to root of task hierarchy */
2471 kmp_info_p *th_next_pool; /* next available thread in the pool */
2472 kmp_disp_t *th_dispatch; /* thread's dispatch data */
2473 int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2475 /* The following are cached from the team info structure */
2476 /* TODO use these in more places as determined to be needed via profiling */
2477 int th_team_nproc; /* number of threads in a team */
2478 kmp_info_p *th_team_master; /* the team's master thread */
2479 int th_team_serialized; /* team is serialized */
2481 microtask_t th_teams_microtask; /* save entry address for teams construct */
2482 int th_teams_level; /* save initial level of teams construct */
2483 /* it is 0 on device but may be any on host */
2486 /* The blocktime info is copied from the team struct to the thread sruct */
2487 /* at the start of a barrier, and the values stored in the team are used */
2488 /* at points in the code where the team struct is no longer guaranteed */
2489 /* to exist (from the POV of worker threads). */
2491 int th_team_bt_intervals;
2494 kmp_uint64 th_team_bt_intervals;
2497 #if KMP_AFFINITY_SUPPORTED
2498 kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2501 void *const *th_def_allocator; /* per implicit task default allocator */
2503 /* The data set by the master at reinit, then R/W by the worker */
2505 th_set_nproc; /* if > 0, then only use this request for the next fork */
2506 #if KMP_NESTED_HOT_TEAMS
2507 kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2511 th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2513 th_teams_size; /* number of teams/threads in teams construct */
2514 #if KMP_AFFINITY_SUPPORTED
2515 int th_current_place; /* place currently bound to */
2516 int th_new_place; /* place to bind to in par reg */
2517 int th_first_place; /* first place in partition */
2518 int th_last_place; /* last place in partition */
2522 int th_prev_level; /* previous level for affinity format */
2523 int th_prev_num_threads; /* previous num_threads for affinity format */
2526 kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2527 kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2528 kmp_uint64 th_frame_time; /* frame timestamp */
2529 #endif /* USE_ITT_BUILD */
2530 kmp_local_t th_local;
2531 struct private_common *th_pri_head;
2533 /* Now the data only used by the worker (after initial allocation) */
2534 /* TODO the first serial team should actually be stored in the info_t
2535 structure. this will help reduce initial allocation overhead */
2536 KMP_ALIGN_CACHE kmp_team_p
2537 *th_serial_team; /*serialized team held in reserve*/
2540 ompt_thread_info_t ompt_thread_info;
2543 /* The following are also read by the master during reinit */
2544 struct common_table *th_pri_common;
2546 volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2547 /* while awaiting queuing lock acquire */
2549 volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2552 unsigned th_x; // Random number generator data
2553 unsigned th_a; // Random number generator data
2555 /* Tasking-related data for the thread */
2556 kmp_task_team_t *th_task_team; // Task team struct
2557 kmp_taskdata_t *th_current_task; // Innermost Task being executed
2558 kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2559 kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2561 kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2562 kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2563 kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2564 // tasking, thus safe to reap
2566 /* More stuff for keeping track of active/sleeping threads (this part is
2567 written by the worker thread) */
2568 kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2569 int th_active; // ! sleeping; 32 bits for TCR/TCW
2570 struct cons_header *th_cons; // used for consistency check
2571 #if KMP_USE_HIER_SCHED
2572 // used for hierarchical scheduling
2573 kmp_hier_private_bdata_t *th_hier_bar_data;
2576 /* Add the syncronizing data which is cache aligned and padded. */
2577 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2579 KMP_ALIGN_CACHE volatile kmp_int32
2580 th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2582 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2584 kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2585 // allocation routines
2589 kmp_win32_cond_t th_suspend_cv;
2590 kmp_win32_mutex_t th_suspend_mx;
2591 int th_suspend_init;
2594 kmp_cond_align_t th_suspend_cv;
2595 kmp_mutex_align_t th_suspend_mx;
2596 int th_suspend_init_count;
2600 kmp_itt_mark_t th_itt_mark_single;
2602 #endif /* USE_ITT_BUILD */
2603 #if KMP_STATS_ENABLED
2604 kmp_stats_list *th_stats;
2607 std::atomic<bool> th_blocking;
2611 typedef union KMP_ALIGN_CACHE kmp_info {
2612 double th_align; /* use worst case alignment */
2613 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2617 // OpenMP thread team data structures
2619 typedef struct kmp_base_data { volatile kmp_uint32 t_value; } kmp_base_data_t;
2621 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2622 double dt_align; /* use worst case alignment */
2623 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2627 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2628 double dt_align; /* use worst case alignment */
2629 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2631 } kmp_ordered_team_t;
2633 typedef int (*launch_t)(int gtid);
2635 /* Minimum number of ARGV entries to malloc if necessary */
2636 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2638 // Set up how many argv pointers will fit in cache lines containing
2639 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2640 // larger value for more space between the master write/worker read section and
2641 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2642 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2643 #define KMP_INLINE_ARGV_BYTES \
2645 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2646 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2649 #define KMP_INLINE_ARGV_BYTES \
2650 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2652 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2654 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2655 // Synchronization Data
2656 // ---------------------------------------------------------------------------
2657 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2658 kmp_balign_team_t t_bar[bs_last_barrier];
2659 std::atomic<int> t_construct; // count of single directive encountered by team
2660 char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2663 // ---------------------------------------------------------------------------
2664 KMP_ALIGN_CACHE int t_master_tid; // tid of master in parent team
2665 int t_master_this_cons; // "this_construct" single counter of master in parent
2667 ident_t *t_ident; // if volatile, have to change too much other crud to
2669 kmp_team_p *t_parent; // parent team
2670 kmp_team_p *t_next_pool; // next free team in the team pool
2671 kmp_disp_t *t_dispatch; // thread's dispatch data
2672 kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2674 kmp_proc_bind_t t_proc_bind; // bind type for par region
2675 #endif // OMP_40_ENABLED
2677 kmp_uint64 t_region_time; // region begin timestamp
2678 #endif /* USE_ITT_BUILD */
2680 // Master write, workers read
2681 // --------------------------------------------------------------------------
2682 KMP_ALIGN_CACHE void **t_argv;
2684 int t_nproc; // number of threads in team
2686 launch_t t_invoke; // procedure to launch the microtask
2689 ompt_team_info_t ompt_team_info;
2690 ompt_lw_taskteam_t *ompt_serialized_team_info;
2693 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2694 kmp_int8 t_fp_control_saved;
2696 kmp_int16 t_x87_fpu_control_word; // FP control regs
2698 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2700 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2702 KMP_ALIGN_CACHE kmp_info_t **t_threads;
2704 *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2705 int t_level; // nested parallel level
2707 KMP_ALIGN_CACHE int t_max_argc;
2708 int t_max_nproc; // max threads this team can handle (dynamicly expandable)
2709 int t_serialized; // levels deep of serialized teams
2710 dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2711 int t_id; // team's id, assigned by debugger.
2712 int t_active_level; // nested active parallel level
2713 kmp_r_sched_t t_sched; // run-time schedule for the team
2714 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2715 int t_first_place; // first & last place in parent thread's partition.
2716 int t_last_place; // Restore these values to master after par region.
2717 #endif // OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2719 int t_display_affinity;
2721 int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2722 // omp_set_num_threads() call
2724 void *const *t_def_allocator; /* per implicit task default allocator */
2727 // Read/write by workers as well
2728 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2729 // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2730 // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2731 // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2732 // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2733 char dummy_padding[1024];
2735 // Internal control stack for additional nested teams.
2736 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2737 // for SERIALIZED teams nested 2 or more levels deep
2739 // typed flag to store request state of cancellation
2740 std::atomic<kmp_int32> t_cancel_request;
2742 int t_master_active; // save on fork, restore on join
2743 kmp_taskq_t t_taskq; // this team's task queue
2744 void *t_copypriv_data; // team specific pointer to copyprivate data array
2746 std::atomic<kmp_uint32> t_copyin_counter;
2749 void *t_stack_id; // team specific stack stitching id (for ittnotify)
2750 #endif /* USE_ITT_BUILD */
2753 union KMP_ALIGN_CACHE kmp_team {
2755 double t_align; /* use worst case alignment */
2756 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2759 typedef union KMP_ALIGN_CACHE kmp_time_global {
2760 double dt_align; /* use worst case alignment */
2761 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2763 } kmp_time_global_t;
2765 typedef struct kmp_base_global {
2767 kmp_time_global_t g_time;
2769 /* non cache-aligned */
2770 volatile int g_abort;
2771 volatile int g_done;
2774 enum dynamic_mode g_dynamic_mode;
2775 } kmp_base_global_t;
2777 typedef union KMP_ALIGN_CACHE kmp_global {
2778 kmp_base_global_t g;
2779 double g_align; /* use worst case alignment */
2780 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2783 typedef struct kmp_base_root {
2784 // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2785 // (r_in_parallel>= 0)
2786 // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2787 // the synch overhead or keeping r_active
2788 volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2789 // GEH: This is misnamed, should be r_in_parallel
2790 volatile int r_nested; // TODO: GEH - This is unused, just remove it entirely.
2791 // keeps a count of active parallel regions per root
2792 std::atomic<int> r_in_parallel;
2793 // GEH: This is misnamed, should be r_active_levels
2794 kmp_team_t *r_root_team;
2795 kmp_team_t *r_hot_team;
2796 kmp_info_t *r_uber_thread;
2797 kmp_lock_t r_begin_lock;
2798 volatile int r_begin;
2799 int r_blocktime; /* blocktime for this root and descendants */
2800 int r_cg_nthreads; // count of active threads in a contention group
2803 typedef union KMP_ALIGN_CACHE kmp_root {
2805 double r_align; /* use worst case alignment */
2806 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2809 struct fortran_inx_info {
2813 /* ------------------------------------------------------------------------ */
2815 extern int __kmp_settings;
2816 extern int __kmp_duplicate_library_ok;
2818 extern int __kmp_forkjoin_frames;
2819 extern int __kmp_forkjoin_frames_mode;
2821 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2822 extern int __kmp_determ_red;
2825 extern int kmp_a_debug;
2826 extern int kmp_b_debug;
2827 extern int kmp_c_debug;
2828 extern int kmp_d_debug;
2829 extern int kmp_e_debug;
2830 extern int kmp_f_debug;
2831 #endif /* KMP_DEBUG */
2833 /* For debug information logging using rotating buffer */
2834 #define KMP_DEBUG_BUF_LINES_INIT 512
2835 #define KMP_DEBUG_BUF_LINES_MIN 1
2837 #define KMP_DEBUG_BUF_CHARS_INIT 128
2838 #define KMP_DEBUG_BUF_CHARS_MIN 2
2841 __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2842 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2844 __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2845 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2848 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2849 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2850 printed in buffer so far */
2851 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2852 recommended in warnings */
2853 /* end rotating debug buffer */
2856 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2858 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2859 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2860 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2861 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2862 extern int __kmp_par_range_lb;
2863 extern int __kmp_par_range_ub;
2866 /* For printing out dynamic storage map for threads and teams */
2868 __kmp_storage_map; /* True means print storage map for threads and teams */
2869 extern int __kmp_storage_map_verbose; /* True means storage map includes
2871 extern int __kmp_storage_map_verbose_specified;
2873 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2874 extern kmp_cpuinfo_t __kmp_cpuinfo;
2877 extern volatile int __kmp_init_serial;
2878 extern volatile int __kmp_init_gtid;
2879 extern volatile int __kmp_init_common;
2880 extern volatile int __kmp_init_middle;
2881 extern volatile int __kmp_init_parallel;
2883 extern volatile int __kmp_init_monitor;
2885 extern volatile int __kmp_init_user_locks;
2886 extern int __kmp_init_counter;
2887 extern int __kmp_root_counter;
2888 extern int __kmp_version;
2890 /* list of address of allocated caches for commons */
2891 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2893 /* Barrier algorithm types and options */
2894 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
2895 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
2896 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
2897 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
2898 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
2899 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
2900 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
2901 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
2902 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
2903 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
2904 extern char const *__kmp_barrier_type_name[bs_last_barrier];
2905 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
2908 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
2909 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
2910 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
2911 extern kmp_bootstrap_lock_t
2912 __kmp_exit_lock; /* exit() is not always thread-safe */
2914 extern kmp_bootstrap_lock_t
2915 __kmp_monitor_lock; /* control monitor thread creation */
2917 extern kmp_bootstrap_lock_t
2918 __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
2919 __kmp_threads expansion to co-exist */
2921 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
2922 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
2923 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
2925 /* used for yielding spin-waits */
2926 extern unsigned int __kmp_init_wait; /* initial number of spin-tests */
2927 extern unsigned int __kmp_next_wait; /* susequent number of spin-tests */
2929 extern enum library_type __kmp_library;
2931 extern enum sched_type __kmp_sched; /* default runtime scheduling */
2932 extern enum sched_type __kmp_static; /* default static scheduling method */
2933 extern enum sched_type __kmp_guided; /* default guided scheduling method */
2934 extern enum sched_type __kmp_auto; /* default auto scheduling method */
2935 extern int __kmp_chunk; /* default runtime chunk size */
2937 extern size_t __kmp_stksize; /* stack size per thread */
2939 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
2941 extern size_t __kmp_stkoffset; /* stack offset per thread */
2942 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
2945 __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
2946 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
2947 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
2948 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
2949 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
2950 extern int __kmp_generate_warnings; /* should we issue warnings? */
2951 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
2953 #ifdef DEBUG_SUSPEND
2954 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
2957 extern kmp_uint32 __kmp_yield_init;
2958 extern kmp_uint32 __kmp_yield_next;
2961 extern kmp_uint32 __kmp_yielding_on;
2963 extern kmp_uint32 __kmp_yield_cycle;
2964 extern kmp_int32 __kmp_yield_on_count;
2965 extern kmp_int32 __kmp_yield_off_count;
2967 /* ------------------------------------------------------------------------- */
2968 extern int __kmp_allThreadsSpecified;
2970 extern size_t __kmp_align_alloc;
2971 /* following data protected by initialization routines */
2972 extern int __kmp_xproc; /* number of processors in the system */
2973 extern int __kmp_avail_proc; /* number of processors available to the process */
2974 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
2975 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
2976 // maximum total number of concurrently-existing threads on device
2977 extern int __kmp_max_nth;
2978 // maximum total number of concurrently-existing threads in a contention group
2979 extern int __kmp_cg_max_nth;
2980 extern int __kmp_teams_max_nth; // max threads used in a teams construct
2981 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
2983 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
2984 region a la OMP_NUM_THREADS */
2985 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
2987 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
2989 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
2990 (__kmpc_threadprivate_cached()) */
2991 extern int __kmp_dflt_nested; /* nested parallelism enabled by default a la
2993 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
2994 blocking (env setting) */
2997 __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
2998 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3001 #ifdef KMP_ADJUST_BLOCKTIME
3002 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3003 #endif /* KMP_ADJUST_BLOCKTIME */
3004 #ifdef KMP_DFLT_NTH_CORES
3005 extern int __kmp_ncores; /* Total number of cores for threads placement */
3007 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3008 extern int __kmp_abort_delay;
3010 extern int __kmp_need_register_atfork_specified;
3012 __kmp_need_register_atfork; /* At initialization, call pthread_atfork to
3013 install fork handler */
3014 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3015 0 - not set, will be set at runtime
3016 1 - using stack search
3017 2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3018 X*) or TlsGetValue(Windows* OS))
3019 3 - static TLS (__declspec(thread) __kmp_gtid),
3020 Linux* OS .so only. */
3022 __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3023 #ifdef KMP_TDATA_GTID
3024 extern KMP_THREAD_LOCAL int __kmp_gtid;
3026 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3027 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3028 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3029 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3030 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3031 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3032 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3034 extern int __kmp_dflt_max_active_levels; /* max_active_levels for nested
3035 parallelism enabled by default via
3036 OMP_MAX_ACTIVE_LEVELS */
3037 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3038 concurrent execution per team */
3039 #if KMP_NESTED_HOT_TEAMS
3040 extern int __kmp_hot_teams_mode;
3041 extern int __kmp_hot_teams_max_level;
3045 extern enum clock_function_type __kmp_clock_function;
3046 extern int __kmp_clock_function_param;
3047 #endif /* KMP_OS_LINUX */
3049 #if KMP_MIC_SUPPORTED
3050 extern enum mic_type __kmp_mic_type;
3053 #ifdef USE_LOAD_BALANCE
3054 extern double __kmp_load_balance_interval; // load balance algorithm interval
3055 #endif /* USE_LOAD_BALANCE */
3057 // OpenMP 3.1 - Nested num threads array
3058 typedef struct kmp_nested_nthreads_t {
3062 } kmp_nested_nthreads_t;
3064 extern kmp_nested_nthreads_t __kmp_nested_nth;
3066 #if KMP_USE_ADAPTIVE_LOCKS
3068 // Parameters for the speculative lock backoff system.
3069 struct kmp_adaptive_backoff_params_t {
3070 // Number of soft retries before it counts as a hard retry.
3071 kmp_uint32 max_soft_retries;
3072 // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3074 kmp_uint32 max_badness;
3077 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3079 #if KMP_DEBUG_ADAPTIVE_LOCKS
3080 extern const char *__kmp_speculative_statsfile;
3083 #endif // KMP_USE_ADAPTIVE_LOCKS
3086 extern int __kmp_display_env; /* TRUE or FALSE */
3087 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3088 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3091 /* ------------------------------------------------------------------------- */
3093 /* the following are protected by the fork/join lock */
3094 /* write: lock read: anytime */
3095 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3096 /* read/write: lock */
3097 extern volatile kmp_team_t *__kmp_team_pool;
3098 extern volatile kmp_info_t *__kmp_thread_pool;
3099 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3101 // total num threads reachable from some root thread including all root threads
3102 extern volatile int __kmp_nth;
3103 /* total number of threads reachable from some root thread including all root
3104 threads, and those in the thread pool */
3105 extern volatile int __kmp_all_nth;
3106 extern int __kmp_thread_pool_nth;
3107 extern std::atomic<int> __kmp_thread_pool_active_nth;
3109 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3110 /* end data protected by fork/join lock */
3111 /* ------------------------------------------------------------------------- */
3113 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3114 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3115 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3116 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3117 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3119 // AT: Which way is correct?
3120 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3121 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3122 #define __kmp_get_team_num_threads(gtid) \
3123 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3125 static inline bool KMP_UBER_GTID(int gtid) {
3126 KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3127 KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3128 return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3129 __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3132 static inline int __kmp_tid_from_gtid(int gtid) {
3133 KMP_DEBUG_ASSERT(gtid >= 0);
3134 return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3137 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3138 KMP_DEBUG_ASSERT(tid >= 0 && team);
3139 return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3142 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3143 KMP_DEBUG_ASSERT(thr);
3144 return thr->th.th_info.ds.ds_gtid;
3147 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3148 KMP_DEBUG_ASSERT(gtid >= 0);
3149 return __kmp_threads[gtid];
3152 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3153 KMP_DEBUG_ASSERT(gtid >= 0);
3154 return __kmp_threads[gtid]->th.th_team;
3157 /* ------------------------------------------------------------------------- */
3159 extern kmp_global_t __kmp_global; /* global status */
3161 extern kmp_info_t __kmp_monitor;
3162 // For Debugging Support Library
3163 extern std::atomic<kmp_int32> __kmp_team_counter;
3164 // For Debugging Support Library
3165 extern std::atomic<kmp_int32> __kmp_task_counter;
3168 #define _KMP_GEN_ID(counter) \
3169 (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3171 #define _KMP_GEN_ID(counter) (~0)
3172 #endif /* USE_DEBUGGER */
3174 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3175 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3177 /* ------------------------------------------------------------------------ */
3179 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3180 size_t size, char const *format, ...);
3182 extern void __kmp_serial_initialize(void);
3183 extern void __kmp_middle_initialize(void);
3184 extern void __kmp_parallel_initialize(void);
3186 extern void __kmp_internal_begin(void);
3187 extern void __kmp_internal_end_library(int gtid);
3188 extern void __kmp_internal_end_thread(int gtid);
3189 extern void __kmp_internal_end_atexit(void);
3190 extern void __kmp_internal_end_fini(void);
3191 extern void __kmp_internal_end_dtor(void);
3192 extern void __kmp_internal_end_dest(void *);
3194 extern int __kmp_register_root(int initial_thread);
3195 extern void __kmp_unregister_root(int gtid);
3197 extern int __kmp_ignore_mppbeg(void);
3198 extern int __kmp_ignore_mppend(void);
3200 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3201 extern void __kmp_exit_single(int gtid);
3203 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3204 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3206 #ifdef USE_LOAD_BALANCE
3207 extern int __kmp_get_load_balance(int);
3210 extern int __kmp_get_global_thread_id(void);
3211 extern int __kmp_get_global_thread_id_reg(void);
3212 extern void __kmp_exit_thread(int exit_status);
3213 extern void __kmp_abort(char const *format, ...);
3214 extern void __kmp_abort_thread(void);
3215 KMP_NORETURN extern void __kmp_abort_process(void);
3216 extern void __kmp_warn(char const *format, ...);
3218 extern void __kmp_set_num_threads(int new_nth, int gtid);
3220 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3222 static inline kmp_info_t *__kmp_entry_thread() {
3223 int gtid = __kmp_entry_gtid();
3225 return __kmp_threads[gtid];
3228 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3229 extern int __kmp_get_max_active_levels(int gtid);
3230 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3231 extern int __kmp_get_team_size(int gtid, int level);
3232 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3233 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3235 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3236 extern void __kmp_init_random(kmp_info_t *thread);
3238 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3239 extern void __kmp_adjust_num_threads(int new_nproc);
3241 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3242 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3243 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3244 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3245 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3246 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3249 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3250 size_t size KMP_SRC_LOC_DECL);
3251 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3252 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3253 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3254 #define __kmp_fast_allocate(this_thr, size) \
3255 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3256 #define __kmp_fast_free(this_thr, ptr) \
3257 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3260 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3261 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3262 size_t elsize KMP_SRC_LOC_DECL);
3263 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3264 size_t size KMP_SRC_LOC_DECL);
3265 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3266 #define __kmp_thread_malloc(th, size) \
3267 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3268 #define __kmp_thread_calloc(th, nelem, elsize) \
3269 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3270 #define __kmp_thread_realloc(th, ptr, size) \
3271 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3272 #define __kmp_thread_free(th, ptr) \
3273 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3275 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3276 #define KMP_INTERNAL_FREE(p) free(p)
3277 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3278 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3280 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3283 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3284 kmp_proc_bind_t proc_bind);
3285 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3289 extern void __kmp_yield(int cond);
3291 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3292 enum sched_type schedule, kmp_int32 lb,
3293 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3294 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3295 enum sched_type schedule, kmp_uint32 lb,
3296 kmp_uint32 ub, kmp_int32 st,
3298 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3299 enum sched_type schedule, kmp_int64 lb,
3300 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3301 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3302 enum sched_type schedule, kmp_uint64 lb,
3303 kmp_uint64 ub, kmp_int64 st,
3306 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3307 kmp_int32 *p_last, kmp_int32 *p_lb,
3308 kmp_int32 *p_ub, kmp_int32 *p_st);
3309 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3310 kmp_int32 *p_last, kmp_uint32 *p_lb,
3311 kmp_uint32 *p_ub, kmp_int32 *p_st);
3312 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3313 kmp_int32 *p_last, kmp_int64 *p_lb,
3314 kmp_int64 *p_ub, kmp_int64 *p_st);
3315 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3316 kmp_int32 *p_last, kmp_uint64 *p_lb,
3317 kmp_uint64 *p_ub, kmp_int64 *p_st);
3319 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3320 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3321 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3322 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3324 #ifdef KMP_GOMP_COMPAT
3326 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3327 enum sched_type schedule, kmp_int32 lb,
3328 kmp_int32 ub, kmp_int32 st,
3329 kmp_int32 chunk, int push_ws);
3330 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3331 enum sched_type schedule, kmp_uint32 lb,
3332 kmp_uint32 ub, kmp_int32 st,
3333 kmp_int32 chunk, int push_ws);
3334 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3335 enum sched_type schedule, kmp_int64 lb,
3336 kmp_int64 ub, kmp_int64 st,
3337 kmp_int64 chunk, int push_ws);
3338 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3339 enum sched_type schedule, kmp_uint64 lb,
3340 kmp_uint64 ub, kmp_int64 st,
3341 kmp_int64 chunk, int push_ws);
3342 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3343 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3344 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3345 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3347 #endif /* KMP_GOMP_COMPAT */
3349 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3350 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3351 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3352 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3353 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3354 extern kmp_uint32 __kmp_wait_yield_4(kmp_uint32 volatile *spinner,
3356 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3358 extern void __kmp_wait_yield_4_ptr(void *spinner, kmp_uint32 checker,
3359 kmp_uint32 (*pred)(void *, kmp_uint32),
3364 class kmp_flag_oncore;
3365 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64 *flag,
3372 extern void __kmp_release_64(kmp_flag_64 *flag);
3374 extern void __kmp_infinite_loop(void);
3376 extern void __kmp_cleanup(void);
3378 #if KMP_HANDLE_SIGNALS
3379 extern int __kmp_handle_signals;
3380 extern void __kmp_install_signals(int parallel_init);
3381 extern void __kmp_remove_signals(void);
3384 extern void __kmp_clear_system_time(void);
3385 extern void __kmp_read_system_time(double *delta);
3387 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3389 extern void __kmp_expand_host_name(char *buffer, size_t size);
3390 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3392 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3394 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3398 __kmp_runtime_initialize(void); /* machine specific initialization */
3399 extern void __kmp_runtime_destroy(void);
3401 #if KMP_AFFINITY_SUPPORTED
3402 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3403 kmp_affin_mask_t *mask);
3404 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3405 kmp_affin_mask_t *mask);
3406 extern void __kmp_affinity_initialize(void);
3407 extern void __kmp_affinity_uninitialize(void);
3408 extern void __kmp_affinity_set_init_mask(
3409 int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3411 extern void __kmp_affinity_set_place(int gtid);
3413 extern void __kmp_affinity_determine_capable(const char *env_var);
3414 extern int __kmp_aux_set_affinity(void **mask);
3415 extern int __kmp_aux_get_affinity(void **mask);
3416 extern int __kmp_aux_get_affinity_max_proc();
3417 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3418 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3419 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3420 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3422 extern int kmp_set_thread_affinity_mask_initial(void);
3424 #endif /* KMP_AFFINITY_SUPPORTED */
3426 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3427 // format string is for affinity, so platforms that do not support
3428 // affinity can still use the other fields, e.g., %n for num_threads
3429 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3430 kmp_str_buf_t *buffer);
3431 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3434 extern void __kmp_cleanup_hierarchy();
3435 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3439 extern int __kmp_futex_determine_capable(void);
3441 #endif // KMP_USE_FUTEX
3443 extern void __kmp_gtid_set_specific(int gtid);
3444 extern int __kmp_gtid_get_specific(void);
3446 extern double __kmp_read_cpu_time(void);
3448 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3451 extern void __kmp_create_monitor(kmp_info_t *th);
3454 extern void *__kmp_launch_thread(kmp_info_t *thr);
3456 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3459 extern int __kmp_still_running(kmp_info_t *th);
3460 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3461 extern void __kmp_free_handle(kmp_thread_t tHandle);
3465 extern void __kmp_reap_monitor(kmp_info_t *th);
3467 extern void __kmp_reap_worker(kmp_info_t *th);
3468 extern void __kmp_terminate_thread(int gtid);
3470 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag);
3471 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag);
3472 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
3473 extern void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag);
3474 extern void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag);
3475 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
3477 extern void __kmp_elapsed(double *);
3478 extern void __kmp_elapsed_tick(double *);
3480 extern void __kmp_enable(int old_state);
3481 extern void __kmp_disable(int *old_state);
3483 extern void __kmp_thread_sleep(int millis);
3485 extern void __kmp_common_initialize(void);
3486 extern void __kmp_common_destroy(void);
3487 extern void __kmp_common_destroy_gtid(int gtid);
3490 extern void __kmp_register_atfork(void);
3492 extern void __kmp_suspend_initialize(void);
3493 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3495 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3499 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3501 ompt_data_t ompt_parallel_data,
3503 kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3504 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3507 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3509 ompt_id_t ompt_parallel_id,
3511 kmp_internal_control_t *new_icvs,
3512 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3513 #endif // OMP_40_ENABLED
3514 extern void __kmp_free_thread(kmp_info_t *);
3515 extern void __kmp_free_team(kmp_root_t *,
3516 kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3517 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3519 /* ------------------------------------------------------------------------ */
3521 extern void __kmp_initialize_bget(kmp_info_t *th);
3522 extern void __kmp_finalize_bget(kmp_info_t *th);
3524 KMP_EXPORT void *kmpc_malloc(size_t size);
3525 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3526 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3527 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3528 KMP_EXPORT void kmpc_free(void *ptr);
3530 /* declarations for internal use */
3532 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3533 size_t reduce_size, void *reduce_data,
3534 void (*reduce)(void *, void *));
3535 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3538 * Tell the fork call which compiler generated the fork call, and therefore how
3539 * to deal with the call.
3541 enum fork_context_e {
3542 fork_context_gnu, /**< Called from GNU generated code, so must not invoke the
3543 microtask internally. */
3544 fork_context_intel, /**< Called from Intel generated code. */
3547 extern int __kmp_fork_call(ident_t *loc, int gtid,
3548 enum fork_context_e fork_context, kmp_int32 argc,
3549 microtask_t microtask, launch_t invoker,
3550 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
3551 #if (KMP_ARCH_ARM || KMP_ARCH_X86_64 || KMP_ARCH_AARCH64) && KMP_OS_LINUX
3558 extern void __kmp_join_call(ident_t *loc, int gtid
3561 enum fork_context_e fork_context
3569 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3570 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3571 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3572 extern int __kmp_invoke_task_func(int gtid);
3573 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3574 kmp_info_t *this_thr,
3576 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3577 kmp_info_t *this_thr,
3580 // should never have been exported
3581 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3583 extern int __kmp_invoke_teams_master(int gtid);
3584 extern void __kmp_teams_master(int gtid);
3585 extern int __kmp_aux_get_team_num();
3586 extern int __kmp_aux_get_num_teams();
3588 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3589 extern void __kmp_user_set_library(enum library_type arg);
3590 extern void __kmp_aux_set_library(enum library_type arg);
3591 extern void __kmp_aux_set_stacksize(size_t arg);
3592 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3593 extern void __kmp_aux_set_defaults(char const *str, int len);
3595 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3596 void kmpc_set_blocktime(int arg);
3597 void ompc_set_nested(int flag);
3598 void ompc_set_dynamic(int flag);
3599 void ompc_set_num_threads(int arg);
3601 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3602 kmp_team_t *team, int tid);
3603 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3604 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3605 kmp_tasking_flags_t *flags,
3606 size_t sizeof_kmp_task_t,
3607 size_t sizeof_shareds,
3608 kmp_routine_entry_t task_entry);
3609 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3610 kmp_team_t *team, int tid,
3612 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3613 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3614 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
3615 kmp_flag_32 *flag, int final_spin,
3616 int *thread_finished,
3619 #endif /* USE_ITT_BUILD */
3620 kmp_int32 is_constrained);
3621 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
3622 kmp_flag_64 *flag, int final_spin,
3623 int *thread_finished,
3626 #endif /* USE_ITT_BUILD */
3627 kmp_int32 is_constrained);
3628 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
3629 kmp_flag_oncore *flag, int final_spin,
3630 int *thread_finished,
3633 #endif /* USE_ITT_BUILD */
3634 kmp_int32 is_constrained);
3636 extern void __kmp_free_task_team(kmp_info_t *thread,
3637 kmp_task_team_t *task_team);
3638 extern void __kmp_reap_task_teams(void);
3639 extern void __kmp_wait_to_unref_task_teams(void);
3640 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3642 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3643 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3647 #endif /* USE_ITT_BUILD */
3650 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3653 extern int __kmp_is_address_mapped(void *addr);
3654 extern kmp_uint64 __kmp_hardware_timestamp(void);
3657 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3660 /* ------------------------------------------------------------------------ */
3662 // Assembly routines that have no compiler intrinsic replacement
3665 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3667 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
3670 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*(p)); }
3671 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
3673 static inline void __kmp_load_mxcsr(const kmp_uint32 *) {}
3674 static inline void __kmp_store_mxcsr(kmp_uint32 *) {}
3677 extern void __kmp_load_x87_fpu_control_word(kmp_int16 *p);
3678 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
3679 extern void __kmp_clear_x87_fpu_status_word();
3680 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
3682 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3684 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3688 void **exit_frame_ptr
3692 /* ------------------------------------------------------------------------ */
3694 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3695 KMP_EXPORT void __kmpc_end(ident_t *);
3697 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3699 kmpc_cctor_vec cctor,
3701 size_t vector_length);
3702 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3703 kmpc_ctor ctor, kmpc_cctor cctor,
3705 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3706 void *data, size_t size);
3708 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3709 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3710 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3711 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3713 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3714 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3715 kmpc_micro microtask, ...);
3717 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3718 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3720 KMP_EXPORT void __kmpc_flush(ident_t *);
3721 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3722 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3723 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3724 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3725 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3726 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3727 kmp_critical_name *);
3728 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3729 kmp_critical_name *);
3732 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3733 kmp_critical_name *, uint32_t hint);
3736 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3737 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3739 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3740 kmp_int32 global_tid);
3742 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3743 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3745 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3746 kmp_int32 schedtype, kmp_int32 *plastiter,
3747 kmp_int *plower, kmp_int *pupper,
3748 kmp_int *pstride, kmp_int incr,
3751 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3753 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3754 size_t cpy_size, void *cpy_data,
3755 void (*cpy_func)(void *, void *),
3758 extern void KMPC_SET_NUM_THREADS(int arg);
3759 extern void KMPC_SET_DYNAMIC(int flag);
3760 extern void KMPC_SET_NESTED(int flag);
3762 /* Taskq interface routines */
3763 KMP_EXPORT kmpc_thunk_t *__kmpc_taskq(ident_t *loc, kmp_int32 global_tid,
3764 kmpc_task_t taskq_task,
3765 size_t sizeof_thunk,
3766 size_t sizeof_shareds, kmp_int32 flags,
3767 kmpc_shared_vars_t **shareds);
3768 KMP_EXPORT void __kmpc_end_taskq(ident_t *loc, kmp_int32 global_tid,
3769 kmpc_thunk_t *thunk);
3770 KMP_EXPORT kmp_int32 __kmpc_task(ident_t *loc, kmp_int32 global_tid,
3771 kmpc_thunk_t *thunk);
3772 KMP_EXPORT void __kmpc_taskq_task(ident_t *loc, kmp_int32 global_tid,
3773 kmpc_thunk_t *thunk, kmp_int32 status);
3774 KMP_EXPORT void __kmpc_end_taskq_task(ident_t *loc, kmp_int32 global_tid,
3775 kmpc_thunk_t *thunk);
3776 KMP_EXPORT kmpc_thunk_t *__kmpc_task_buffer(ident_t *loc, kmp_int32 global_tid,
3777 kmpc_thunk_t *taskq_thunk,
3780 /* OMP 3.0 tasking interface routines */
3781 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3782 kmp_task_t *new_task);
3783 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3785 size_t sizeof_kmp_task_t,
3786 size_t sizeof_shareds,
3787 kmp_routine_entry_t task_entry);
3788 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3790 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3792 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3793 kmp_task_t *new_task);
3794 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3796 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3800 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3801 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3803 #endif // TASK_UNUSED
3805 /* ------------------------------------------------------------------------ */
3809 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3810 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3812 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3813 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3814 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3815 kmp_depend_info_t *noalias_dep_list);
3816 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3818 kmp_depend_info_t *dep_list,
3819 kmp_int32 ndeps_noalias,
3820 kmp_depend_info_t *noalias_dep_list);
3822 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3823 bool serialize_immediate);
3825 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3826 kmp_int32 cncl_kind);
3827 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3828 kmp_int32 cncl_kind);
3829 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3830 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3834 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3835 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3836 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3837 kmp_int32 if_val, kmp_uint64 *lb,
3838 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3839 kmp_int32 sched, kmp_uint64 grainsize,
3843 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3844 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3845 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
3846 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3847 kmp_task_affinity_info_t *affin_list);
3852 /* Lock interface routines (fast versions with gtid passed in) */
3853 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3855 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3857 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3859 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3861 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3862 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3864 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3866 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3868 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3869 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3873 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3874 void **user_lock, uintptr_t hint);
3875 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3880 /* Interface to fast scalable reduce methods routines */
3882 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3883 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3884 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3885 kmp_critical_name *lck);
3886 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3887 kmp_critical_name *lck);
3888 KMP_EXPORT kmp_int32 __kmpc_reduce(
3889 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3890 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3891 kmp_critical_name *lck);
3892 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3893 kmp_critical_name *lck);
3895 /* Internal fast reduction routines */
3897 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3898 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3899 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3900 kmp_critical_name *lck);
3902 // this function is for testing set/get/determine reduce method
3903 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3905 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3906 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3909 // missing 'extern "C"' declarations
3911 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3912 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3913 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3914 kmp_int32 num_threads);
3917 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3919 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3920 kmp_int32 num_teams,
3921 kmp_int32 num_threads);
3922 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3923 kmpc_micro microtask, ...);
3926 struct kmp_dim { // loop bounds info casted to kmp_int64
3927 kmp_int64 lo; // lower
3928 kmp_int64 up; // upper
3929 kmp_int64 st; // stride
3931 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
3933 const struct kmp_dim *dims);
3934 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
3935 const kmp_int64 *vec);
3936 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
3937 const kmp_int64 *vec);
3938 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
3941 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
3942 void *data, size_t size,
3945 // Symbols for MS mutual detection.
3946 extern int _You_must_link_with_exactly_one_OpenMP_library;
3947 extern int _You_must_link_with_Intel_OpenMP_library;
3948 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3949 extern int _You_must_link_with_Microsoft_OpenMP_library;
3952 // The routines below are not exported.
3953 // Consider making them 'static' in corresponding source files.
3954 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
3955 void *data_addr, size_t pc_size);
3956 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
3959 void __kmp_threadprivate_resize_cache(int newCapacity);
3960 void __kmp_cleanup_threadprivate_caches();
3962 // ompc_, kmpc_ entries moved from omp.h.
3964 #define KMPC_CONVENTION __cdecl
3966 #define KMPC_CONVENTION
3970 typedef enum omp_sched_t {
3971 omp_sched_static = 1,
3972 omp_sched_dynamic = 2,
3973 omp_sched_guided = 3,
3976 typedef void *kmp_affinity_mask_t;
3979 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
3980 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
3981 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
3982 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
3983 KMP_EXPORT int KMPC_CONVENTION
3984 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
3985 KMP_EXPORT int KMPC_CONVENTION
3986 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
3987 KMP_EXPORT int KMPC_CONVENTION
3988 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
3990 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
3991 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
3992 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
3993 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
3994 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
3997 enum kmp_target_offload_kind {
4002 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4003 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4004 extern kmp_target_offload_kind_t __kmp_target_offload;
4005 extern int __kmpc_get_target_offload();
4009 // Constants used in libomptarget
4010 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4011 #define KMP_HOST_DEVICE -10 // This is what it is in libomptarget, go figure.
4012 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4013 #endif // OMP_40_ENABLED