1 /*! \file */
2 /*
3 * kmp.h -- KPTS runtime header file.
4 */
5
6 //===----------------------------------------------------------------------===//
7 //
8 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
9 // See https://llvm.org/LICENSE.txt for license information.
10 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
11 //
12 //===----------------------------------------------------------------------===//
13
14 #ifndef KMP_H
15 #define KMP_H
16
17 #include "kmp_config.h"
18
19 /* #define BUILD_PARALLEL_ORDERED 1 */
20
21 /* This fix replaces gettimeofday with clock_gettime for better scalability on
22 the Altix. Requires user code to be linked with -lrt. */
23 //#define FIX_SGI_CLOCK
24
25 /* Defines for OpenMP 3.0 tasking and auto scheduling */
26
27 #ifndef KMP_STATIC_STEAL_ENABLED
28 #define KMP_STATIC_STEAL_ENABLED 1
29 #endif
30
31 #define TASK_CURRENT_NOT_QUEUED 0
32 #define TASK_CURRENT_QUEUED 1
33
34 #ifdef BUILD_TIED_TASK_STACK
35 #define TASK_STACK_EMPTY 0 // entries when the stack is empty
36 #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
37 // Number of entries in each task stack array
38 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
39 // Mask for determining index into stack block
40 #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
41 #endif // BUILD_TIED_TASK_STACK
42
43 #define TASK_NOT_PUSHED 1
44 #define TASK_SUCCESSFULLY_PUSHED 0
45 #define TASK_TIED 1
46 #define TASK_UNTIED 0
47 #define TASK_EXPLICIT 1
48 #define TASK_IMPLICIT 0
49 #define TASK_PROXY 1
50 #define TASK_FULL 0
51 #define TASK_DETACHABLE 1
52 #define TASK_UNDETACHABLE 0
53
54 #define KMP_CANCEL_THREADS
55 #define KMP_THREAD_ATTR
56
57 // Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being
58 // built on Android
59 #if defined(__ANDROID__)
60 #undef KMP_CANCEL_THREADS
61 #endif
62
63 #include <signal.h>
64 #include <stdarg.h>
65 #include <stddef.h>
66 #include <stdio.h>
67 #include <stdlib.h>
68 #include <string.h>
69 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
70 Microsoft library. Some macros provided below to replace these functions */
71 #ifndef __ABSOFT_WIN
72 #include <sys/types.h>
73 #endif
74 #include <limits.h>
75 #include <time.h>
76
77 #include <errno.h>
78
79 #include "kmp_os.h"
80
81 #include "kmp_safe_c_api.h"
82
83 #if KMP_STATS_ENABLED
84 class kmp_stats_list;
85 #endif
86
87 #if KMP_USE_HIER_SCHED
88 // Only include hierarchical scheduling if affinity is supported
89 #undef KMP_USE_HIER_SCHED
90 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
91 #endif
92
93 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
94 #include "hwloc.h"
95 #ifndef HWLOC_OBJ_NUMANODE
96 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
97 #endif
98 #ifndef HWLOC_OBJ_PACKAGE
99 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
100 #endif
101 #if HWLOC_API_VERSION >= 0x00020000
102 // hwloc 2.0 changed type of depth of object from unsigned to int
103 typedef int kmp_hwloc_depth_t;
104 #else
105 typedef unsigned int kmp_hwloc_depth_t;
106 #endif
107 #endif
108
109 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
110 #include <xmmintrin.h>
111 #endif
112
113 #include "kmp_debug.h"
114 #include "kmp_lock.h"
115 #include "kmp_version.h"
116 #if USE_DEBUGGER
117 #include "kmp_debugger.h"
118 #endif
119 #include "kmp_i18n.h"
120
121 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
122
123 #include "kmp_wrapper_malloc.h"
124 #if KMP_OS_UNIX
125 #include <unistd.h>
126 #if !defined NSIG && defined _NSIG
127 #define NSIG _NSIG
128 #endif
129 #endif
130
131 #if KMP_OS_LINUX
132 #pragma weak clock_gettime
133 #endif
134
135 #if OMPT_SUPPORT
136 #include "ompt-internal.h"
137 #endif
138
139 // Affinity format function
140 #include "kmp_str.h"
141
142 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
143 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
144 // free lists of limited size.
145 #ifndef USE_FAST_MEMORY
146 #define USE_FAST_MEMORY 3
147 #endif
148
149 #ifndef KMP_NESTED_HOT_TEAMS
150 #define KMP_NESTED_HOT_TEAMS 0
151 #define USE_NESTED_HOT_ARG(x)
152 #else
153 #if KMP_NESTED_HOT_TEAMS
154 #define USE_NESTED_HOT_ARG(x) , x
155 #else
156 #define USE_NESTED_HOT_ARG(x)
157 #endif
158 #endif
159
160 // Assume using BGET compare_exchange instruction instead of lock by default.
161 #ifndef USE_CMP_XCHG_FOR_BGET
162 #define USE_CMP_XCHG_FOR_BGET 1
163 #endif
164
165 // Test to see if queuing lock is better than bootstrap lock for bget
166 // #ifndef USE_QUEUING_LOCK_FOR_BGET
167 // #define USE_QUEUING_LOCK_FOR_BGET
168 // #endif
169
170 #define KMP_NSEC_PER_SEC 1000000000L
171 #define KMP_USEC_PER_SEC 1000000L
172
173 /*!
174 @ingroup BASIC_TYPES
175 @{
176 */
177
178 /*!
179 Values for bit flags used in the ident_t to describe the fields.
180 */
181 enum {
182 /*! Use trampoline for internal microtasks */
183 KMP_IDENT_IMB = 0x01,
184 /*! Use c-style ident structure */
185 KMP_IDENT_KMPC = 0x02,
186 /* 0x04 is no longer used */
187 /*! Entry point generated by auto-parallelization */
188 KMP_IDENT_AUTOPAR = 0x08,
189 /*! Compiler generates atomic reduction option for kmpc_reduce* */
190 KMP_IDENT_ATOMIC_REDUCE = 0x10,
191 /*! To mark a 'barrier' directive in user code */
192 KMP_IDENT_BARRIER_EXPL = 0x20,
193 /*! To Mark implicit barriers. */
194 KMP_IDENT_BARRIER_IMPL = 0x0040,
195 KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
196 KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
197 KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
198
199 KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
200 KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
201
202 /*! To mark a static loop in OMPT callbacks */
203 KMP_IDENT_WORK_LOOP = 0x200,
204 /*! To mark a sections directive in OMPT callbacks */
205 KMP_IDENT_WORK_SECTIONS = 0x400,
206 /*! To mark a distribute construct in OMPT callbacks */
207 KMP_IDENT_WORK_DISTRIBUTE = 0x800,
208 /*! Atomic hint; bottom four bits as omp_sync_hint_t. Top four reserved and
209 not currently used. If one day we need more bits, then we can use
210 an invalid combination of hints to mean that another, larger field
211 should be used in a different flag. */
212 KMP_IDENT_ATOMIC_HINT_MASK = 0xFF0000,
213 KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
214 KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
215 KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
216 KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
217 };
218
219 /*!
220 * The ident structure that describes a source location.
221 */
222 typedef struct ident {
223 kmp_int32 reserved_1; /**< might be used in Fortran; see above */
224 kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC
225 identifies this union member */
226 kmp_int32 reserved_2; /**< not really used in Fortran any more; see above */
227 #if USE_ITT_BUILD
228 /* but currently used for storing region-specific ITT */
229 /* contextual information. */
230 #endif /* USE_ITT_BUILD */
231 kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++ */
232 char const *psource; /**< String describing the source location.
233 The string is composed of semi-colon separated fields
234 which describe the source file, the function and a pair
235 of line numbers that delimit the construct. */
236 } ident_t;
237 /*!
238 @}
239 */
240
241 // Some forward declarations.
242 typedef union kmp_team kmp_team_t;
243 typedef struct kmp_taskdata kmp_taskdata_t;
244 typedef union kmp_task_team kmp_task_team_t;
245 typedef union kmp_team kmp_team_p;
246 typedef union kmp_info kmp_info_p;
247 typedef union kmp_root kmp_root_p;
248
249 #ifdef __cplusplus
250 extern "C" {
251 #endif
252
253 /* ------------------------------------------------------------------------ */
254
255 /* Pack two 32-bit signed integers into a 64-bit signed integer */
256 /* ToDo: Fix word ordering for big-endian machines. */
257 #define KMP_PACK_64(HIGH_32, LOW_32) \
258 ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
259
260 // Generic string manipulation macros. Assume that _x is of type char *
261 #define SKIP_WS(_x) \
262 { \
263 while (*(_x) == ' ' || *(_x) == '\t') \
264 (_x)++; \
265 }
266 #define SKIP_DIGITS(_x) \
267 { \
268 while (*(_x) >= '0' && *(_x) <= '9') \
269 (_x)++; \
270 }
271 #define SKIP_TOKEN(_x) \
272 { \
273 while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
274 (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
275 (_x)++; \
276 }
277 #define SKIP_TO(_x, _c) \
278 { \
279 while (*(_x) != '\0' && *(_x) != (_c)) \
280 (_x)++; \
281 }
282
283 /* ------------------------------------------------------------------------ */
284
285 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
286 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
287
288 /* ------------------------------------------------------------------------ */
289 /* Enumeration types */
290
291 enum kmp_state_timer {
292 ts_stop,
293 ts_start,
294 ts_pause,
295
296 ts_last_state
297 };
298
299 enum dynamic_mode {
300 dynamic_default,
301 #ifdef USE_LOAD_BALANCE
302 dynamic_load_balance,
303 #endif /* USE_LOAD_BALANCE */
304 dynamic_random,
305 dynamic_thread_limit,
306 dynamic_max
307 };
308
309 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
310 * not include it here */
311 #ifndef KMP_SCHED_TYPE_DEFINED
312 #define KMP_SCHED_TYPE_DEFINED
313 typedef enum kmp_sched {
314 kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
315 // Note: need to adjust __kmp_sch_map global array in case enum is changed
316 kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
317 kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
318 kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
319 kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
320 kmp_sched_upper_std = 5, // upper bound for standard schedules
321 kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
322 kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
323 #if KMP_STATIC_STEAL_ENABLED
324 kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
325 #endif
326 kmp_sched_upper,
327 kmp_sched_default = kmp_sched_static, // default scheduling
328 kmp_sched_monotonic = 0x80000000
329 } kmp_sched_t;
330 #endif
331
332 /*!
333 @ingroup WORK_SHARING
334 * Describes the loop schedule to be used for a parallel for loop.
335 */
336 enum sched_type : kmp_int32 {
337 kmp_sch_lower = 32, /**< lower bound for unordered values */
338 kmp_sch_static_chunked = 33,
339 kmp_sch_static = 34, /**< static unspecialized */
340 kmp_sch_dynamic_chunked = 35,
341 kmp_sch_guided_chunked = 36, /**< guided unspecialized */
342 kmp_sch_runtime = 37,
343 kmp_sch_auto = 38, /**< auto */
344 kmp_sch_trapezoidal = 39,
345
346 /* accessible only through KMP_SCHEDULE environment variable */
347 kmp_sch_static_greedy = 40,
348 kmp_sch_static_balanced = 41,
349 /* accessible only through KMP_SCHEDULE environment variable */
350 kmp_sch_guided_iterative_chunked = 42,
351 kmp_sch_guided_analytical_chunked = 43,
352 /* accessible only through KMP_SCHEDULE environment variable */
353 kmp_sch_static_steal = 44,
354
355 /* static with chunk adjustment (e.g., simd) */
356 kmp_sch_static_balanced_chunked = 45,
357 kmp_sch_guided_simd = 46, /**< guided with chunk adjustment */
358 kmp_sch_runtime_simd = 47, /**< runtime with chunk adjustment */
359
360 /* accessible only through KMP_SCHEDULE environment variable */
361 kmp_sch_upper, /**< upper bound for unordered values */
362
363 kmp_ord_lower = 64, /**< lower bound for ordered values, must be power of 2 */
364 kmp_ord_static_chunked = 65,
365 kmp_ord_static = 66, /**< ordered static unspecialized */
366 kmp_ord_dynamic_chunked = 67,
367 kmp_ord_guided_chunked = 68,
368 kmp_ord_runtime = 69,
369 kmp_ord_auto = 70, /**< ordered auto */
370 kmp_ord_trapezoidal = 71,
371 kmp_ord_upper, /**< upper bound for ordered values */
372
373 /* Schedules for Distribute construct */
374 kmp_distribute_static_chunked = 91, /**< distribute static chunked */
375 kmp_distribute_static = 92, /**< distribute static unspecialized */
376
377 /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
378 single iteration/chunk, even if the loop is serialized. For the schedule
379 types listed above, the entire iteration vector is returned if the loop is
380 serialized. This doesn't work for gcc/gcomp sections. */
381 kmp_nm_lower = 160, /**< lower bound for nomerge values */
382
383 kmp_nm_static_chunked =
384 (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
385 kmp_nm_static = 162, /**< static unspecialized */
386 kmp_nm_dynamic_chunked = 163,
387 kmp_nm_guided_chunked = 164, /**< guided unspecialized */
388 kmp_nm_runtime = 165,
389 kmp_nm_auto = 166, /**< auto */
390 kmp_nm_trapezoidal = 167,
391
392 /* accessible only through KMP_SCHEDULE environment variable */
393 kmp_nm_static_greedy = 168,
394 kmp_nm_static_balanced = 169,
395 /* accessible only through KMP_SCHEDULE environment variable */
396 kmp_nm_guided_iterative_chunked = 170,
397 kmp_nm_guided_analytical_chunked = 171,
398 kmp_nm_static_steal =
399 172, /* accessible only through OMP_SCHEDULE environment variable */
400
401 kmp_nm_ord_static_chunked = 193,
402 kmp_nm_ord_static = 194, /**< ordered static unspecialized */
403 kmp_nm_ord_dynamic_chunked = 195,
404 kmp_nm_ord_guided_chunked = 196,
405 kmp_nm_ord_runtime = 197,
406 kmp_nm_ord_auto = 198, /**< auto */
407 kmp_nm_ord_trapezoidal = 199,
408 kmp_nm_upper, /**< upper bound for nomerge values */
409
410 /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
411 we need to distinguish the three possible cases (no modifier, monotonic
412 modifier, nonmonotonic modifier), we need separate bits for each modifier.
413 The absence of monotonic does not imply nonmonotonic, especially since 4.5
414 says that the behaviour of the "no modifier" case is implementation defined
415 in 4.5, but will become "nonmonotonic" in 5.0.
416
417 Since we're passing a full 32 bit value, we can use a couple of high bits
418 for these flags; out of paranoia we avoid the sign bit.
419
420 These modifiers can be or-ed into non-static schedules by the compiler to
421 pass the additional information. They will be stripped early in the
422 processing in __kmp_dispatch_init when setting up schedules, so most of the
423 code won't ever see schedules with these bits set. */
424 kmp_sch_modifier_monotonic =
425 (1 << 29), /**< Set if the monotonic schedule modifier was present */
426 kmp_sch_modifier_nonmonotonic =
427 (1 << 30), /**< Set if the nonmonotonic schedule modifier was present */
428
429 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
430 (enum sched_type)( \
431 (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))
432 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
433 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
434 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
435 (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
436 #define SCHEDULE_GET_MODIFIERS(s) \
437 ((enum sched_type)( \
438 (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
439 #define SCHEDULE_SET_MODIFIERS(s, m) \
440 (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
441 #define SCHEDULE_NONMONOTONIC 0
442 #define SCHEDULE_MONOTONIC 1
443
444 kmp_sch_default = kmp_sch_static /**< default scheduling algorithm */
445 };
446
447 // Apply modifiers on internal kind to standard kind
448 static inline void
__kmp_sched_apply_mods_stdkind(kmp_sched_t * kind,enum sched_type internal_kind)449 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
450 enum sched_type internal_kind) {
451 if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
452 *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
453 }
454 }
455
456 // Apply modifiers on standard kind to internal kind
457 static inline void
__kmp_sched_apply_mods_intkind(kmp_sched_t kind,enum sched_type * internal_kind)458 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
459 enum sched_type *internal_kind) {
460 if ((int)kind & (int)kmp_sched_monotonic) {
461 *internal_kind = (enum sched_type)((int)*internal_kind |
462 (int)kmp_sch_modifier_monotonic);
463 }
464 }
465
466 // Get standard schedule without modifiers
__kmp_sched_without_mods(kmp_sched_t kind)467 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
468 return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
469 }
470
471 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
472 typedef union kmp_r_sched {
473 struct {
474 enum sched_type r_sched_type;
475 int chunk;
476 };
477 kmp_int64 sched;
478 } kmp_r_sched_t;
479
480 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
481 // internal schedule types
482
483 enum library_type {
484 library_none,
485 library_serial,
486 library_turnaround,
487 library_throughput
488 };
489
490 #if KMP_OS_LINUX
491 enum clock_function_type {
492 clock_function_gettimeofday,
493 clock_function_clock_gettime
494 };
495 #endif /* KMP_OS_LINUX */
496
497 #if KMP_MIC_SUPPORTED
498 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
499 #endif
500
501 /* -- fast reduction stuff ------------------------------------------------ */
502
503 #undef KMP_FAST_REDUCTION_BARRIER
504 #define KMP_FAST_REDUCTION_BARRIER 1
505
506 #undef KMP_FAST_REDUCTION_CORE_DUO
507 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
508 #define KMP_FAST_REDUCTION_CORE_DUO 1
509 #endif
510
511 enum _reduction_method {
512 reduction_method_not_defined = 0,
513 critical_reduce_block = (1 << 8),
514 atomic_reduce_block = (2 << 8),
515 tree_reduce_block = (3 << 8),
516 empty_reduce_block = (4 << 8)
517 };
518
519 // Description of the packed_reduction_method variable:
520 // The packed_reduction_method variable consists of two enum types variables
521 // that are packed together into 0-th byte and 1-st byte:
522 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
523 // barrier that will be used in fast reduction: bs_plain_barrier or
524 // bs_reduction_barrier
525 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
526 // be used in fast reduction;
527 // Reduction method is of 'enum _reduction_method' type and it's defined the way
528 // so that the bits of 0-th byte are empty, so no need to execute a shift
529 // instruction while packing/unpacking
530
531 #if KMP_FAST_REDUCTION_BARRIER
532 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
533 ((reduction_method) | (barrier_type))
534
535 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
536 ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
537
538 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
539 ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
540 #else
541 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
542 (reduction_method)
543
544 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
545 (packed_reduction_method)
546
547 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
548 #endif
549
550 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
551 ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
552 (which_reduction_block))
553
554 #if KMP_FAST_REDUCTION_BARRIER
555 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
556 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
557
558 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
559 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
560 #endif
561
562 typedef int PACKED_REDUCTION_METHOD_T;
563
564 /* -- end of fast reduction stuff ----------------------------------------- */
565
566 #if KMP_OS_WINDOWS
567 #define USE_CBLKDATA
568 #if KMP_MSVC_COMPAT
569 #pragma warning(push)
570 #pragma warning(disable : 271 310)
571 #endif
572 #include <windows.h>
573 #if KMP_MSVC_COMPAT
574 #pragma warning(pop)
575 #endif
576 #endif
577
578 #if KMP_OS_UNIX
579 #include <dlfcn.h>
580 #include <pthread.h>
581 #endif
582
583 /* Only Linux* OS and Windows* OS support thread affinity. */
584 #if KMP_AFFINITY_SUPPORTED
585
586 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
587 #if KMP_OS_WINDOWS
588 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
589 typedef struct GROUP_AFFINITY {
590 KAFFINITY Mask;
591 WORD Group;
592 WORD Reserved[3];
593 } GROUP_AFFINITY;
594 #endif /* _MSC_VER < 1600 */
595 #if KMP_GROUP_AFFINITY
596 extern int __kmp_num_proc_groups;
597 #else
598 static const int __kmp_num_proc_groups = 1;
599 #endif /* KMP_GROUP_AFFINITY */
600 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
601 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
602
603 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
604 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
605
606 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
607 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
608
609 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
610 GROUP_AFFINITY *);
611 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
612 #endif /* KMP_OS_WINDOWS */
613
614 #if KMP_USE_HWLOC
615 extern hwloc_topology_t __kmp_hwloc_topology;
616 extern int __kmp_hwloc_error;
617 extern int __kmp_numa_detected;
618 extern int __kmp_tile_depth;
619 #endif
620
621 extern size_t __kmp_affin_mask_size;
622 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
623 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
624 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
625 #define KMP_CPU_SET_ITERATE(i, mask) \
626 for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
627 #define KMP_CPU_SET(i, mask) (mask)->set(i)
628 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
629 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
630 #define KMP_CPU_ZERO(mask) (mask)->zero()
631 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
632 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
633 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
634 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
635 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
636 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
637 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
638 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
639 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
640 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
641 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
642 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
643 (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
644 #define KMP_CPU_FREE_ARRAY(arr, n) \
645 __kmp_affinity_dispatch->deallocate_mask_array(arr)
646 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
647 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
648 #define __kmp_get_system_affinity(mask, abort_bool) \
649 (mask)->get_system_affinity(abort_bool)
650 #define __kmp_set_system_affinity(mask, abort_bool) \
651 (mask)->set_system_affinity(abort_bool)
652 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
653
654 class KMPAffinity {
655 public:
656 class Mask {
657 public:
658 void *operator new(size_t n);
659 void operator delete(void *p);
660 void *operator new[](size_t n);
661 void operator delete[](void *p);
~Mask()662 virtual ~Mask() {}
663 // Set bit i to 1
set(int i)664 virtual void set(int i) {}
665 // Return bit i
is_set(int i)666 virtual bool is_set(int i) const { return false; }
667 // Set bit i to 0
clear(int i)668 virtual void clear(int i) {}
669 // Zero out entire mask
zero()670 virtual void zero() {}
671 // Copy src into this mask
copy(const Mask * src)672 virtual void copy(const Mask *src) {}
673 // this &= rhs
bitwise_and(const Mask * rhs)674 virtual void bitwise_and(const Mask *rhs) {}
675 // this |= rhs
bitwise_or(const Mask * rhs)676 virtual void bitwise_or(const Mask *rhs) {}
677 // this = ~this
bitwise_not()678 virtual void bitwise_not() {}
679 // API for iterating over an affinity mask
680 // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
begin()681 virtual int begin() const { return 0; }
end()682 virtual int end() const { return 0; }
next(int previous)683 virtual int next(int previous) const { return 0; }
684 // Set the system's affinity to this affinity mask's value
set_system_affinity(bool abort_on_error)685 virtual int set_system_affinity(bool abort_on_error) const { return -1; }
686 // Set this affinity mask to the current system affinity
get_system_affinity(bool abort_on_error)687 virtual int get_system_affinity(bool abort_on_error) { return -1; }
688 // Only 1 DWORD in the mask should have any procs set.
689 // Return the appropriate index, or -1 for an invalid mask.
get_proc_group()690 virtual int get_proc_group() const { return -1; }
691 };
692 void *operator new(size_t n);
693 void operator delete(void *p);
694 // Need virtual destructor
695 virtual ~KMPAffinity() = default;
696 // Determine if affinity is capable
determine_capable(const char * env_var)697 virtual void determine_capable(const char *env_var) {}
698 // Bind the current thread to os proc
bind_thread(int proc)699 virtual void bind_thread(int proc) {}
700 // Factory functions to allocate/deallocate a mask
allocate_mask()701 virtual Mask *allocate_mask() { return nullptr; }
deallocate_mask(Mask * m)702 virtual void deallocate_mask(Mask *m) {}
allocate_mask_array(int num)703 virtual Mask *allocate_mask_array(int num) { return nullptr; }
deallocate_mask_array(Mask * m)704 virtual void deallocate_mask_array(Mask *m) {}
index_mask_array(Mask * m,int index)705 virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
706 static void pick_api();
707 static void destroy_api();
708 enum api_type {
709 NATIVE_OS
710 #if KMP_USE_HWLOC
711 ,
712 HWLOC
713 #endif
714 };
get_api_type()715 virtual api_type get_api_type() const {
716 KMP_ASSERT(0);
717 return NATIVE_OS;
718 }
719
720 private:
721 static bool picked_api;
722 };
723
724 typedef KMPAffinity::Mask kmp_affin_mask_t;
725 extern KMPAffinity *__kmp_affinity_dispatch;
726
727 // Declare local char buffers with this size for printing debug and info
728 // messages, using __kmp_affinity_print_mask().
729 #define KMP_AFFIN_MASK_PRINT_LEN 1024
730
731 enum affinity_type {
732 affinity_none = 0,
733 affinity_physical,
734 affinity_logical,
735 affinity_compact,
736 affinity_scatter,
737 affinity_explicit,
738 affinity_balanced,
739 affinity_disabled, // not used outsize the env var parser
740 affinity_default
741 };
742
743 enum affinity_gran {
744 affinity_gran_fine = 0,
745 affinity_gran_thread,
746 affinity_gran_core,
747 affinity_gran_tile,
748 affinity_gran_numa,
749 affinity_gran_package,
750 affinity_gran_node,
751 #if KMP_GROUP_AFFINITY
752 // The "group" granularity isn't necesssarily coarser than all of the
753 // other levels, but we put it last in the enum.
754 affinity_gran_group,
755 #endif /* KMP_GROUP_AFFINITY */
756 affinity_gran_default
757 };
758
759 enum affinity_top_method {
760 affinity_top_method_all = 0, // try all (supported) methods, in order
761 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
762 affinity_top_method_apicid,
763 affinity_top_method_x2apicid,
764 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
765 affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
766 #if KMP_GROUP_AFFINITY
767 affinity_top_method_group,
768 #endif /* KMP_GROUP_AFFINITY */
769 affinity_top_method_flat,
770 #if KMP_USE_HWLOC
771 affinity_top_method_hwloc,
772 #endif
773 affinity_top_method_default
774 };
775
776 #define affinity_respect_mask_default (-1)
777
778 extern enum affinity_type __kmp_affinity_type; /* Affinity type */
779 extern enum affinity_gran __kmp_affinity_gran; /* Affinity granularity */
780 extern int __kmp_affinity_gran_levels; /* corresponding int value */
781 extern int __kmp_affinity_dups; /* Affinity duplicate masks */
782 extern enum affinity_top_method __kmp_affinity_top_method;
783 extern int __kmp_affinity_compact; /* Affinity 'compact' value */
784 extern int __kmp_affinity_offset; /* Affinity offset value */
785 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */
786 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */
787 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask?
788 extern char *__kmp_affinity_proclist; /* proc ID list */
789 extern kmp_affin_mask_t *__kmp_affinity_masks;
790 extern unsigned __kmp_affinity_num_masks;
791 extern void __kmp_affinity_bind_thread(int which);
792
793 extern kmp_affin_mask_t *__kmp_affin_fullMask;
794 extern char *__kmp_cpuinfo_file;
795
796 #endif /* KMP_AFFINITY_SUPPORTED */
797
798 // This needs to be kept in sync with the values in omp.h !!!
799 typedef enum kmp_proc_bind_t {
800 proc_bind_false = 0,
801 proc_bind_true,
802 proc_bind_master,
803 proc_bind_close,
804 proc_bind_spread,
805 proc_bind_intel, // use KMP_AFFINITY interface
806 proc_bind_default
807 } kmp_proc_bind_t;
808
809 typedef struct kmp_nested_proc_bind_t {
810 kmp_proc_bind_t *bind_types;
811 int size;
812 int used;
813 } kmp_nested_proc_bind_t;
814
815 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
816
817 extern int __kmp_display_affinity;
818 extern char *__kmp_affinity_format;
819 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
820
821 #if KMP_AFFINITY_SUPPORTED
822 #define KMP_PLACE_ALL (-1)
823 #define KMP_PLACE_UNDEFINED (-2)
824 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
825 #define KMP_AFFINITY_NON_PROC_BIND \
826 ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
827 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
828 (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
829 #endif /* KMP_AFFINITY_SUPPORTED */
830
831 extern int __kmp_affinity_num_places;
832
833 typedef enum kmp_cancel_kind_t {
834 cancel_noreq = 0,
835 cancel_parallel = 1,
836 cancel_loop = 2,
837 cancel_sections = 3,
838 cancel_taskgroup = 4
839 } kmp_cancel_kind_t;
840
841 // KMP_HW_SUBSET support:
842 typedef struct kmp_hws_item {
843 int num;
844 int offset;
845 } kmp_hws_item_t;
846
847 extern kmp_hws_item_t __kmp_hws_socket;
848 extern kmp_hws_item_t __kmp_hws_node;
849 extern kmp_hws_item_t __kmp_hws_tile;
850 extern kmp_hws_item_t __kmp_hws_core;
851 extern kmp_hws_item_t __kmp_hws_proc;
852 extern int __kmp_hws_requested;
853 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
854
855 /* ------------------------------------------------------------------------ */
856
857 #define KMP_PAD(type, sz) \
858 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
859
860 // We need to avoid using -1 as a GTID as +1 is added to the gtid
861 // when storing it in a lock, and the value 0 is reserved.
862 #define KMP_GTID_DNE (-2) /* Does not exist */
863 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
864 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
865 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
866 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
867
868 /* OpenMP 5.0 Memory Management support */
869
870 #ifndef __OMP_H
871 // Duplicate type definitions from omp.h
872 typedef uintptr_t omp_uintptr_t;
873
874 typedef enum {
875 OMP_ATK_THREADMODEL = 1,
876 OMP_ATK_ALIGNMENT = 2,
877 OMP_ATK_ACCESS = 3,
878 OMP_ATK_POOL_SIZE = 4,
879 OMP_ATK_FALLBACK = 5,
880 OMP_ATK_FB_DATA = 6,
881 OMP_ATK_PINNED = 7,
882 OMP_ATK_PARTITION = 8
883 } omp_alloctrait_key_t;
884
885 typedef enum {
886 OMP_ATV_FALSE = 0,
887 OMP_ATV_TRUE = 1,
888 OMP_ATV_DEFAULT = 2,
889 OMP_ATV_CONTENDED = 3,
890 OMP_ATV_UNCONTENDED = 4,
891 OMP_ATV_SEQUENTIAL = 5,
892 OMP_ATV_PRIVATE = 6,
893 OMP_ATV_ALL = 7,
894 OMP_ATV_THREAD = 8,
895 OMP_ATV_PTEAM = 9,
896 OMP_ATV_CGROUP = 10,
897 OMP_ATV_DEFAULT_MEM_FB = 11,
898 OMP_ATV_NULL_FB = 12,
899 OMP_ATV_ABORT_FB = 13,
900 OMP_ATV_ALLOCATOR_FB = 14,
901 OMP_ATV_ENVIRONMENT = 15,
902 OMP_ATV_NEAREST = 16,
903 OMP_ATV_BLOCKED = 17,
904 OMP_ATV_INTERLEAVED = 18
905 } omp_alloctrait_value_t;
906
907 typedef void *omp_memspace_handle_t;
908 extern omp_memspace_handle_t const omp_default_mem_space;
909 extern omp_memspace_handle_t const omp_large_cap_mem_space;
910 extern omp_memspace_handle_t const omp_const_mem_space;
911 extern omp_memspace_handle_t const omp_high_bw_mem_space;
912 extern omp_memspace_handle_t const omp_low_lat_mem_space;
913
914 typedef struct {
915 omp_alloctrait_key_t key;
916 omp_uintptr_t value;
917 } omp_alloctrait_t;
918
919 typedef void *omp_allocator_handle_t;
920 extern omp_allocator_handle_t const omp_null_allocator;
921 extern omp_allocator_handle_t const omp_default_mem_alloc;
922 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
923 extern omp_allocator_handle_t const omp_const_mem_alloc;
924 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
925 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
926 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
927 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
928 extern omp_allocator_handle_t const omp_thread_mem_alloc;
929 extern omp_allocator_handle_t const kmp_max_mem_alloc;
930 extern omp_allocator_handle_t __kmp_def_allocator;
931
932 // end of duplicate type definitions from omp.h
933 #endif
934
935 extern int __kmp_memkind_available;
936
937 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
938
939 typedef struct kmp_allocator_t {
940 omp_memspace_handle_t memspace;
941 void **memkind; // pointer to memkind
942 int alignment;
943 omp_alloctrait_value_t fb;
944 kmp_allocator_t *fb_data;
945 kmp_uint64 pool_size;
946 kmp_uint64 pool_used;
947 } kmp_allocator_t;
948
949 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
950 omp_memspace_handle_t,
951 int ntraits,
952 omp_alloctrait_t traits[]);
953 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
954 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
955 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
956 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
957 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
958
959 extern void __kmp_init_memkind();
960 extern void __kmp_fini_memkind();
961
962 /* ------------------------------------------------------------------------ */
963
964 #define KMP_UINT64_MAX \
965 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
966
967 #define KMP_MIN_NTH 1
968
969 #ifndef KMP_MAX_NTH
970 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
971 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
972 #else
973 #define KMP_MAX_NTH INT_MAX
974 #endif
975 #endif /* KMP_MAX_NTH */
976
977 #ifdef PTHREAD_STACK_MIN
978 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
979 #else
980 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
981 #endif
982
983 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
984
985 #if KMP_ARCH_X86
986 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
987 #elif KMP_ARCH_X86_64
988 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
989 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
990 #else
991 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
992 #endif
993
994 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
995 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
996 #define KMP_MAX_MALLOC_POOL_INCR \
997 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
998
999 #define KMP_MIN_STKOFFSET (0)
1000 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1001 #if KMP_OS_DARWIN
1002 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1003 #else
1004 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1005 #endif
1006
1007 #define KMP_MIN_STKPADDING (0)
1008 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1009
1010 #define KMP_BLOCKTIME_MULTIPLIER \
1011 (1000) /* number of blocktime units per second */
1012 #define KMP_MIN_BLOCKTIME (0)
1013 #define KMP_MAX_BLOCKTIME \
1014 (INT_MAX) /* Must be this for "infinite" setting the work */
1015 #define KMP_DEFAULT_BLOCKTIME (200) /* __kmp_blocktime is in milliseconds */
1016
1017 #if KMP_USE_MONITOR
1018 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1019 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1020 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1021
1022 /* Calculate new number of monitor wakeups for a specific block time based on
1023 previous monitor_wakeups. Only allow increasing number of wakeups */
1024 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1025 (((blocktime) == KMP_MAX_BLOCKTIME) \
1026 ? (monitor_wakeups) \
1027 : ((blocktime) == KMP_MIN_BLOCKTIME) \
1028 ? KMP_MAX_MONITOR_WAKEUPS \
1029 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1030 ? (monitor_wakeups) \
1031 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1032
1033 /* Calculate number of intervals for a specific block time based on
1034 monitor_wakeups */
1035 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1036 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1037 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1038 #else
1039 #define KMP_BLOCKTIME(team, tid) \
1040 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1041 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1042 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1043 extern kmp_uint64 __kmp_ticks_per_msec;
1044 #if KMP_COMPILER_ICC
1045 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1046 #else
1047 #define KMP_NOW() __kmp_hardware_timestamp()
1048 #endif
1049 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1050 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1051 (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1052 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1053 #else
1054 // System time is retrieved sporadically while blocking.
1055 extern kmp_uint64 __kmp_now_nsec();
1056 #define KMP_NOW() __kmp_now_nsec()
1057 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1058 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1059 (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1060 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1061 #endif
1062 #endif // KMP_USE_MONITOR
1063
1064 #define KMP_MIN_STATSCOLS 40
1065 #define KMP_MAX_STATSCOLS 4096
1066 #define KMP_DEFAULT_STATSCOLS 80
1067
1068 #define KMP_MIN_INTERVAL 0
1069 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1070 #define KMP_DEFAULT_INTERVAL 0
1071
1072 #define KMP_MIN_CHUNK 1
1073 #define KMP_MAX_CHUNK (INT_MAX - 1)
1074 #define KMP_DEFAULT_CHUNK 1
1075
1076 #define KMP_DFLT_DISP_NUM_BUFF 7
1077 #define KMP_MAX_ORDERED 8
1078
1079 #define KMP_MAX_FIELDS 32
1080
1081 #define KMP_MAX_BRANCH_BITS 31
1082
1083 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1084
1085 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1086
1087 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1088
1089 /* Minimum number of threads before switch to TLS gtid (experimentally
1090 determined) */
1091 /* josh TODO: what about OS X* tuning? */
1092 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1093 #define KMP_TLS_GTID_MIN 5
1094 #else
1095 #define KMP_TLS_GTID_MIN INT_MAX
1096 #endif
1097
1098 #define KMP_MASTER_TID(tid) ((tid) == 0)
1099 #define KMP_WORKER_TID(tid) ((tid) != 0)
1100
1101 #define KMP_MASTER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) == 0)
1102 #define KMP_WORKER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) != 0)
1103 #define KMP_INITIAL_GTID(gtid) ((gtid) == 0)
1104
1105 #ifndef TRUE
1106 #define FALSE 0
1107 #define TRUE (!FALSE)
1108 #endif
1109
1110 /* NOTE: all of the following constants must be even */
1111
1112 #if KMP_OS_WINDOWS
1113 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1114 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1115 #elif KMP_OS_CNK
1116 #define KMP_INIT_WAIT 16U /* initial number of spin-tests */
1117 #define KMP_NEXT_WAIT 8U /* susequent number of spin-tests */
1118 #elif KMP_OS_LINUX
1119 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1120 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1121 #elif KMP_OS_DARWIN
1122 /* TODO: tune for KMP_OS_DARWIN */
1123 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1124 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1125 #elif KMP_OS_DRAGONFLY
1126 /* TODO: tune for KMP_OS_DRAGONFLY */
1127 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1128 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1129 #elif KMP_OS_FREEBSD
1130 /* TODO: tune for KMP_OS_FREEBSD */
1131 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1132 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1133 #elif KMP_OS_NETBSD
1134 /* TODO: tune for KMP_OS_NETBSD */
1135 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1136 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1137 #elif KMP_OS_HURD
1138 /* TODO: tune for KMP_OS_HURD */
1139 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1140 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1141 #elif KMP_OS_OPENBSD
1142 /* TODO: tune for KMP_OS_OPENBSD */
1143 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1144 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1145 #endif
1146
1147 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1148 typedef struct kmp_cpuid {
1149 kmp_uint32 eax;
1150 kmp_uint32 ebx;
1151 kmp_uint32 ecx;
1152 kmp_uint32 edx;
1153 } kmp_cpuid_t;
1154
1155 typedef struct kmp_cpuinfo {
1156 int initialized; // If 0, other fields are not initialized.
1157 int signature; // CPUID(1).EAX
1158 int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1159 int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1160 // Model << 4 ) + Model)
1161 int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1162 int sse2; // 0 if SSE2 instructions are not supported, 1 otherwise.
1163 int rtm; // 0 if RTM instructions are not supported, 1 otherwise.
1164 int cpu_stackoffset;
1165 int apic_id;
1166 int physical_id;
1167 int logical_id;
1168 kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1169 char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1170 } kmp_cpuinfo_t;
1171
1172 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1173
1174 #if KMP_OS_UNIX
1175 // subleaf is only needed for cache and topology discovery and can be set to
1176 // zero in most cases
__kmp_x86_cpuid(int leaf,int subleaf,struct kmp_cpuid * p)1177 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1178 __asm__ __volatile__("cpuid"
1179 : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1180 : "a"(leaf), "c"(subleaf));
1181 }
1182 // Load p into FPU control word
__kmp_load_x87_fpu_control_word(const kmp_int16 * p)1183 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1184 __asm__ __volatile__("fldcw %0" : : "m"(*p));
1185 }
1186 // Store FPU control word into p
__kmp_store_x87_fpu_control_word(kmp_int16 * p)1187 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1188 __asm__ __volatile__("fstcw %0" : "=m"(*p));
1189 }
__kmp_clear_x87_fpu_status_word()1190 static inline void __kmp_clear_x87_fpu_status_word() {
1191 #if KMP_MIC
1192 // 32-bit protected mode x87 FPU state
1193 struct x87_fpu_state {
1194 unsigned cw;
1195 unsigned sw;
1196 unsigned tw;
1197 unsigned fip;
1198 unsigned fips;
1199 unsigned fdp;
1200 unsigned fds;
1201 };
1202 struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1203 __asm__ __volatile__("fstenv %0\n\t" // store FP env
1204 "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1205 "fldenv %0\n\t" // load FP env back
1206 : "+m"(fpu_state), "+m"(fpu_state.sw));
1207 #else
1208 __asm__ __volatile__("fnclex");
1209 #endif // KMP_MIC
1210 }
1211 #if __SSE__
__kmp_load_mxcsr(const kmp_uint32 * p)1212 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
__kmp_store_mxcsr(kmp_uint32 * p)1213 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1214 #else
__kmp_load_mxcsr(const kmp_uint32 * p)1215 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
__kmp_store_mxcsr(kmp_uint32 * p)1216 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1217 #endif
1218 #else
1219 // Windows still has these as external functions in assembly file
1220 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1221 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1222 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1223 extern void __kmp_clear_x87_fpu_status_word();
__kmp_load_mxcsr(const kmp_uint32 * p)1224 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
__kmp_store_mxcsr(kmp_uint32 * p)1225 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1226 #endif // KMP_OS_UNIX
1227
1228 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1229
1230 #if KMP_ARCH_X86
1231 extern void __kmp_x86_pause(void);
1232 #elif KMP_MIC
1233 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1234 // regression after removal of extra PAUSE from spin loops. Changing
1235 // the delay from 100 to 300 showed even better performance than double PAUSE
1236 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
__kmp_x86_pause(void)1237 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1238 #else
__kmp_x86_pause(void)1239 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1240 #endif
1241 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1242 #elif KMP_ARCH_PPC64
1243 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1244 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1245 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1246 #define KMP_CPU_PAUSE() \
1247 do { \
1248 KMP_PPC64_PRI_LOW(); \
1249 KMP_PPC64_PRI_MED(); \
1250 KMP_PPC64_PRI_LOC_MB(); \
1251 } while (0)
1252 #else
1253 #define KMP_CPU_PAUSE() /* nothing to do */
1254 #endif
1255
1256 #define KMP_INIT_YIELD(count) \
1257 { (count) = __kmp_yield_init; }
1258
1259 #define KMP_OVERSUBSCRIBED \
1260 (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1261
1262 #define KMP_TRY_YIELD \
1263 ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1264
1265 #define KMP_TRY_YIELD_OVERSUB \
1266 ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1267
1268 #define KMP_YIELD(cond) \
1269 { \
1270 KMP_CPU_PAUSE(); \
1271 if ((cond) && (KMP_TRY_YIELD)) \
1272 __kmp_yield(); \
1273 }
1274
1275 #define KMP_YIELD_OVERSUB() \
1276 { \
1277 KMP_CPU_PAUSE(); \
1278 if ((KMP_TRY_YIELD_OVERSUB)) \
1279 __kmp_yield(); \
1280 }
1281
1282 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1283 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1284 #define KMP_YIELD_SPIN(count) \
1285 { \
1286 KMP_CPU_PAUSE(); \
1287 if (KMP_TRY_YIELD) { \
1288 (count) -= 2; \
1289 if (!(count)) { \
1290 __kmp_yield(); \
1291 (count) = __kmp_yield_next; \
1292 } \
1293 } \
1294 }
1295
1296 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count) \
1297 { \
1298 KMP_CPU_PAUSE(); \
1299 if ((KMP_TRY_YIELD_OVERSUB)) \
1300 __kmp_yield(); \
1301 else if (__kmp_use_yield == 1) { \
1302 (count) -= 2; \
1303 if (!(count)) { \
1304 __kmp_yield(); \
1305 (count) = __kmp_yield_next; \
1306 } \
1307 } \
1308 }
1309
1310 /* ------------------------------------------------------------------------ */
1311 /* Support datatypes for the orphaned construct nesting checks. */
1312 /* ------------------------------------------------------------------------ */
1313
1314 enum cons_type {
1315 ct_none,
1316 ct_parallel,
1317 ct_pdo,
1318 ct_pdo_ordered,
1319 ct_psections,
1320 ct_psingle,
1321 ct_critical,
1322 ct_ordered_in_parallel,
1323 ct_ordered_in_pdo,
1324 ct_master,
1325 ct_reduce,
1326 ct_barrier
1327 };
1328
1329 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1330
1331 struct cons_data {
1332 ident_t const *ident;
1333 enum cons_type type;
1334 int prev;
1335 kmp_user_lock_p
1336 name; /* address exclusively for critical section name comparison */
1337 };
1338
1339 struct cons_header {
1340 int p_top, w_top, s_top;
1341 int stack_size, stack_top;
1342 struct cons_data *stack_data;
1343 };
1344
1345 struct kmp_region_info {
1346 char *text;
1347 int offset[KMP_MAX_FIELDS];
1348 int length[KMP_MAX_FIELDS];
1349 };
1350
1351 /* ---------------------------------------------------------------------- */
1352 /* ---------------------------------------------------------------------- */
1353
1354 #if KMP_OS_WINDOWS
1355 typedef HANDLE kmp_thread_t;
1356 typedef DWORD kmp_key_t;
1357 #endif /* KMP_OS_WINDOWS */
1358
1359 #if KMP_OS_UNIX
1360 typedef pthread_t kmp_thread_t;
1361 typedef pthread_key_t kmp_key_t;
1362 #endif
1363
1364 extern kmp_key_t __kmp_gtid_threadprivate_key;
1365
1366 typedef struct kmp_sys_info {
1367 long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1368 long minflt; /* the number of page faults serviced without any I/O */
1369 long majflt; /* the number of page faults serviced that required I/O */
1370 long nswap; /* the number of times a process was "swapped" out of memory */
1371 long inblock; /* the number of times the file system had to perform input */
1372 long oublock; /* the number of times the file system had to perform output */
1373 long nvcsw; /* the number of times a context switch was voluntarily */
1374 long nivcsw; /* the number of times a context switch was forced */
1375 } kmp_sys_info_t;
1376
1377 #if USE_ITT_BUILD
1378 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1379 // required type here. Later we will check the type meets requirements.
1380 typedef int kmp_itt_mark_t;
1381 #define KMP_ITT_DEBUG 0
1382 #endif /* USE_ITT_BUILD */
1383
1384 typedef kmp_int32 kmp_critical_name[8];
1385
1386 /*!
1387 @ingroup PARALLEL
1388 The type for a microtask which gets passed to @ref __kmpc_fork_call().
1389 The arguments to the outlined function are
1390 @param global_tid the global thread identity of the thread executing the
1391 function.
1392 @param bound_tid the local identity of the thread executing the function
1393 @param ... pointers to shared variables accessed by the function.
1394 */
1395 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1396 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1397 ...);
1398
1399 /*!
1400 @ingroup THREADPRIVATE
1401 @{
1402 */
1403 /* ---------------------------------------------------------------------------
1404 */
1405 /* Threadprivate initialization/finalization function declarations */
1406
1407 /* for non-array objects: __kmpc_threadprivate_register() */
1408
1409 /*!
1410 Pointer to the constructor function.
1411 The first argument is the <tt>this</tt> pointer
1412 */
1413 typedef void *(*kmpc_ctor)(void *);
1414
1415 /*!
1416 Pointer to the destructor function.
1417 The first argument is the <tt>this</tt> pointer
1418 */
1419 typedef void (*kmpc_dtor)(
1420 void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1421 compiler */
1422 /*!
1423 Pointer to an alternate constructor.
1424 The first argument is the <tt>this</tt> pointer.
1425 */
1426 typedef void *(*kmpc_cctor)(void *, void *);
1427
1428 /* for array objects: __kmpc_threadprivate_register_vec() */
1429 /* First arg: "this" pointer */
1430 /* Last arg: number of array elements */
1431 /*!
1432 Array constructor.
1433 First argument is the <tt>this</tt> pointer
1434 Second argument the number of array elements.
1435 */
1436 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1437 /*!
1438 Pointer to the array destructor function.
1439 The first argument is the <tt>this</tt> pointer
1440 Second argument the number of array elements.
1441 */
1442 typedef void (*kmpc_dtor_vec)(void *, size_t);
1443 /*!
1444 Array constructor.
1445 First argument is the <tt>this</tt> pointer
1446 Third argument the number of array elements.
1447 */
1448 typedef void *(*kmpc_cctor_vec)(void *, void *,
1449 size_t); /* function unused by compiler */
1450
1451 /*!
1452 @}
1453 */
1454
1455 /* keeps tracked of threadprivate cache allocations for cleanup later */
1456 typedef struct kmp_cached_addr {
1457 void **addr; /* address of allocated cache */
1458 void ***compiler_cache; /* pointer to compiler's cache */
1459 void *data; /* pointer to global data */
1460 struct kmp_cached_addr *next; /* pointer to next cached address */
1461 } kmp_cached_addr_t;
1462
1463 struct private_data {
1464 struct private_data *next; /* The next descriptor in the list */
1465 void *data; /* The data buffer for this descriptor */
1466 int more; /* The repeat count for this descriptor */
1467 size_t size; /* The data size for this descriptor */
1468 };
1469
1470 struct private_common {
1471 struct private_common *next;
1472 struct private_common *link;
1473 void *gbl_addr;
1474 void *par_addr; /* par_addr == gbl_addr for MASTER thread */
1475 size_t cmn_size;
1476 };
1477
1478 struct shared_common {
1479 struct shared_common *next;
1480 struct private_data *pod_init;
1481 void *obj_init;
1482 void *gbl_addr;
1483 union {
1484 kmpc_ctor ctor;
1485 kmpc_ctor_vec ctorv;
1486 } ct;
1487 union {
1488 kmpc_cctor cctor;
1489 kmpc_cctor_vec cctorv;
1490 } cct;
1491 union {
1492 kmpc_dtor dtor;
1493 kmpc_dtor_vec dtorv;
1494 } dt;
1495 size_t vec_len;
1496 int is_vec;
1497 size_t cmn_size;
1498 };
1499
1500 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1501 #define KMP_HASH_TABLE_SIZE \
1502 (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1503 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1504 #define KMP_HASH(x) \
1505 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1506
1507 struct common_table {
1508 struct private_common *data[KMP_HASH_TABLE_SIZE];
1509 };
1510
1511 struct shared_table {
1512 struct shared_common *data[KMP_HASH_TABLE_SIZE];
1513 };
1514
1515 /* ------------------------------------------------------------------------ */
1516
1517 #if KMP_USE_HIER_SCHED
1518 // Shared barrier data that exists inside a single unit of the scheduling
1519 // hierarchy
1520 typedef struct kmp_hier_private_bdata_t {
1521 kmp_int32 num_active;
1522 kmp_uint64 index;
1523 kmp_uint64 wait_val[2];
1524 } kmp_hier_private_bdata_t;
1525 #endif
1526
1527 typedef struct kmp_sched_flags {
1528 unsigned ordered : 1;
1529 unsigned nomerge : 1;
1530 unsigned contains_last : 1;
1531 #if KMP_USE_HIER_SCHED
1532 unsigned use_hier : 1;
1533 unsigned unused : 28;
1534 #else
1535 unsigned unused : 29;
1536 #endif
1537 } kmp_sched_flags_t;
1538
1539 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1540
1541 #if KMP_STATIC_STEAL_ENABLED
1542 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1543 kmp_int32 count;
1544 kmp_int32 ub;
1545 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1546 kmp_int32 lb;
1547 kmp_int32 st;
1548 kmp_int32 tc;
1549 kmp_int32 static_steal_counter; /* for static_steal only; maybe better to put
1550 after ub */
1551
1552 // KMP_ALIGN( 16 ) ensures ( if the KMP_ALIGN macro is turned on )
1553 // a) parm3 is properly aligned and
1554 // b) all parm1-4 are in the same cache line.
1555 // Because of parm1-4 are used together, performance seems to be better
1556 // if they are in the same line (not measured though).
1557
1558 struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1559 kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1560 kmp_int32 parm2; // make no real change at least while padding is off.
1561 kmp_int32 parm3;
1562 kmp_int32 parm4;
1563 };
1564
1565 kmp_uint32 ordered_lower;
1566 kmp_uint32 ordered_upper;
1567 #if KMP_OS_WINDOWS
1568 // This var can be placed in the hole between 'tc' and 'parm1', instead of
1569 // 'static_steal_counter'. It would be nice to measure execution times.
1570 // Conditional if/endif can be removed at all.
1571 kmp_int32 last_upper;
1572 #endif /* KMP_OS_WINDOWS */
1573 } dispatch_private_info32_t;
1574
1575 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1576 kmp_int64 count; // current chunk number for static & static-steal scheduling
1577 kmp_int64 ub; /* upper-bound */
1578 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1579 kmp_int64 lb; /* lower-bound */
1580 kmp_int64 st; /* stride */
1581 kmp_int64 tc; /* trip count (number of iterations) */
1582 kmp_int64 static_steal_counter; /* for static_steal only; maybe better to put
1583 after ub */
1584
1585 /* parm[1-4] are used in different ways by different scheduling algorithms */
1586
1587 // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1588 // a) parm3 is properly aligned and
1589 // b) all parm1-4 are in the same cache line.
1590 // Because of parm1-4 are used together, performance seems to be better
1591 // if they are in the same line (not measured though).
1592
1593 struct KMP_ALIGN(32) {
1594 kmp_int64 parm1;
1595 kmp_int64 parm2;
1596 kmp_int64 parm3;
1597 kmp_int64 parm4;
1598 };
1599
1600 kmp_uint64 ordered_lower;
1601 kmp_uint64 ordered_upper;
1602 #if KMP_OS_WINDOWS
1603 // This var can be placed in the hole between 'tc' and 'parm1', instead of
1604 // 'static_steal_counter'. It would be nice to measure execution times.
1605 // Conditional if/endif can be removed at all.
1606 kmp_int64 last_upper;
1607 #endif /* KMP_OS_WINDOWS */
1608 } dispatch_private_info64_t;
1609 #else /* KMP_STATIC_STEAL_ENABLED */
1610 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1611 kmp_int32 lb;
1612 kmp_int32 ub;
1613 kmp_int32 st;
1614 kmp_int32 tc;
1615
1616 kmp_int32 parm1;
1617 kmp_int32 parm2;
1618 kmp_int32 parm3;
1619 kmp_int32 parm4;
1620
1621 kmp_int32 count;
1622
1623 kmp_uint32 ordered_lower;
1624 kmp_uint32 ordered_upper;
1625 #if KMP_OS_WINDOWS
1626 kmp_int32 last_upper;
1627 #endif /* KMP_OS_WINDOWS */
1628 } dispatch_private_info32_t;
1629
1630 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1631 kmp_int64 lb; /* lower-bound */
1632 kmp_int64 ub; /* upper-bound */
1633 kmp_int64 st; /* stride */
1634 kmp_int64 tc; /* trip count (number of iterations) */
1635
1636 /* parm[1-4] are used in different ways by different scheduling algorithms */
1637 kmp_int64 parm1;
1638 kmp_int64 parm2;
1639 kmp_int64 parm3;
1640 kmp_int64 parm4;
1641
1642 kmp_int64 count; /* current chunk number for static scheduling */
1643
1644 kmp_uint64 ordered_lower;
1645 kmp_uint64 ordered_upper;
1646 #if KMP_OS_WINDOWS
1647 kmp_int64 last_upper;
1648 #endif /* KMP_OS_WINDOWS */
1649 } dispatch_private_info64_t;
1650 #endif /* KMP_STATIC_STEAL_ENABLED */
1651
1652 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1653 union private_info {
1654 dispatch_private_info32_t p32;
1655 dispatch_private_info64_t p64;
1656 } u;
1657 enum sched_type schedule; /* scheduling algorithm */
1658 kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1659 kmp_int32 ordered_bumped;
1660 // To retain the structure size after making ordered_iteration scalar
1661 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 3];
1662 // Stack of buffers for nest of serial regions
1663 struct dispatch_private_info *next;
1664 kmp_int32 type_size; /* the size of types in private_info */
1665 #if KMP_USE_HIER_SCHED
1666 kmp_int32 hier_id;
1667 void *parent; /* hierarchical scheduling parent pointer */
1668 #endif
1669 enum cons_type pushed_ws;
1670 } dispatch_private_info_t;
1671
1672 typedef struct dispatch_shared_info32 {
1673 /* chunk index under dynamic, number of idle threads under static-steal;
1674 iteration index otherwise */
1675 volatile kmp_uint32 iteration;
1676 volatile kmp_uint32 num_done;
1677 volatile kmp_uint32 ordered_iteration;
1678 // Dummy to retain the structure size after making ordered_iteration scalar
1679 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1680 } dispatch_shared_info32_t;
1681
1682 typedef struct dispatch_shared_info64 {
1683 /* chunk index under dynamic, number of idle threads under static-steal;
1684 iteration index otherwise */
1685 volatile kmp_uint64 iteration;
1686 volatile kmp_uint64 num_done;
1687 volatile kmp_uint64 ordered_iteration;
1688 // Dummy to retain the structure size after making ordered_iteration scalar
1689 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1690 } dispatch_shared_info64_t;
1691
1692 typedef struct dispatch_shared_info {
1693 union shared_info {
1694 dispatch_shared_info32_t s32;
1695 dispatch_shared_info64_t s64;
1696 } u;
1697 volatile kmp_uint32 buffer_index;
1698 volatile kmp_int32 doacross_buf_idx; // teamwise index
1699 volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1700 kmp_int32 doacross_num_done; // count finished threads
1701 #if KMP_USE_HIER_SCHED
1702 void *hier;
1703 #endif
1704 #if KMP_USE_HWLOC
1705 // When linking with libhwloc, the ORDERED EPCC test slows down on big
1706 // machines (> 48 cores). Performance analysis showed that a cache thrash
1707 // was occurring and this padding helps alleviate the problem.
1708 char padding[64];
1709 #endif
1710 } dispatch_shared_info_t;
1711
1712 typedef struct kmp_disp {
1713 /* Vector for ORDERED SECTION */
1714 void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1715 /* Vector for END ORDERED SECTION */
1716 void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1717
1718 dispatch_shared_info_t *th_dispatch_sh_current;
1719 dispatch_private_info_t *th_dispatch_pr_current;
1720
1721 dispatch_private_info_t *th_disp_buffer;
1722 kmp_int32 th_disp_index;
1723 kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1724 volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1725 union { // we can use union here because doacross cannot be used in
1726 // nonmonotonic loops
1727 kmp_int64 *th_doacross_info; // info on loop bounds
1728 kmp_lock_t *th_steal_lock; // lock used for chunk stealing (8-byte variable)
1729 };
1730 #if KMP_USE_INTERNODE_ALIGNMENT
1731 char more_padding[INTERNODE_CACHE_LINE];
1732 #endif
1733 } kmp_disp_t;
1734
1735 /* ------------------------------------------------------------------------ */
1736 /* Barrier stuff */
1737
1738 /* constants for barrier state update */
1739 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1740 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1741 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1742 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1743
1744 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1745 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1746 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1747
1748 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1749 #error "Barrier sleep bit must be smaller than barrier bump bit"
1750 #endif
1751 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1752 #error "Barrier unused bit must be smaller than barrier bump bit"
1753 #endif
1754
1755 // Constants for release barrier wait state: currently, hierarchical only
1756 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1757 #define KMP_BARRIER_OWN_FLAG \
1758 1 // Normal state; worker waiting on own b_go flag in release
1759 #define KMP_BARRIER_PARENT_FLAG \
1760 2 // Special state; worker waiting on parent's b_go flag in release
1761 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1762 3 // Special state; tells worker to shift from parent to own b_go
1763 #define KMP_BARRIER_SWITCHING \
1764 4 // Special state; worker resets appropriate flag on wake-up
1765
1766 #define KMP_NOT_SAFE_TO_REAP \
1767 0 // Thread th_reap_state: not safe to reap (tasking)
1768 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1769
1770 enum barrier_type {
1771 bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1772 barriers if enabled) */
1773 bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1774 #if KMP_FAST_REDUCTION_BARRIER
1775 bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1776 #endif // KMP_FAST_REDUCTION_BARRIER
1777 bs_last_barrier /* Just a placeholder to mark the end */
1778 };
1779
1780 // to work with reduction barriers just like with plain barriers
1781 #if !KMP_FAST_REDUCTION_BARRIER
1782 #define bs_reduction_barrier bs_plain_barrier
1783 #endif // KMP_FAST_REDUCTION_BARRIER
1784
1785 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1786 bp_linear_bar =
1787 0, /* Single level (degenerate) tree */
1788 bp_tree_bar =
1789 1, /* Balanced tree with branching factor 2^n */
1790 bp_hyper_bar =
1791 2, /* Hypercube-embedded tree with min branching
1792 factor 2^n */
1793 bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1794 bp_last_bar /* Placeholder to mark the end */
1795 } kmp_bar_pat_e;
1796
1797 #define KMP_BARRIER_ICV_PUSH 1
1798
1799 /* Record for holding the values of the internal controls stack records */
1800 typedef struct kmp_internal_control {
1801 int serial_nesting_level; /* corresponds to the value of the
1802 th_team_serialized field */
1803 kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1804 thread) */
1805 kmp_int8
1806 bt_set; /* internal control for whether blocktime is explicitly set */
1807 int blocktime; /* internal control for blocktime */
1808 #if KMP_USE_MONITOR
1809 int bt_intervals; /* internal control for blocktime intervals */
1810 #endif
1811 int nproc; /* internal control for #threads for next parallel region (per
1812 thread) */
1813 int thread_limit; /* internal control for thread-limit-var */
1814 int max_active_levels; /* internal control for max_active_levels */
1815 kmp_r_sched_t
1816 sched; /* internal control for runtime schedule {sched,chunk} pair */
1817 kmp_proc_bind_t proc_bind; /* internal control for affinity */
1818 kmp_int32 default_device; /* internal control for default device */
1819 struct kmp_internal_control *next;
1820 } kmp_internal_control_t;
1821
copy_icvs(kmp_internal_control_t * dst,kmp_internal_control_t * src)1822 static inline void copy_icvs(kmp_internal_control_t *dst,
1823 kmp_internal_control_t *src) {
1824 *dst = *src;
1825 }
1826
1827 /* Thread barrier needs volatile barrier fields */
1828 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1829 // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1830 // uses of it). It is not explicitly aligned below, because we *don't* want
1831 // it to be padded -- instead, we fit b_go into the same cache line with
1832 // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1833 kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1834 // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1835 // same NGO store
1836 volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1837 KMP_ALIGN_CACHE volatile kmp_uint64
1838 b_arrived; // STATE => task reached synch point.
1839 kmp_uint32 *skip_per_level;
1840 kmp_uint32 my_level;
1841 kmp_int32 parent_tid;
1842 kmp_int32 old_tid;
1843 kmp_uint32 depth;
1844 struct kmp_bstate *parent_bar;
1845 kmp_team_t *team;
1846 kmp_uint64 leaf_state;
1847 kmp_uint32 nproc;
1848 kmp_uint8 base_leaf_kids;
1849 kmp_uint8 leaf_kids;
1850 kmp_uint8 offset;
1851 kmp_uint8 wait_flag;
1852 kmp_uint8 use_oncore_barrier;
1853 #if USE_DEBUGGER
1854 // The following field is intended for the debugger solely. Only the worker
1855 // thread itself accesses this field: the worker increases it by 1 when it
1856 // arrives to a barrier.
1857 KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1858 #endif /* USE_DEBUGGER */
1859 } kmp_bstate_t;
1860
1861 union KMP_ALIGN_CACHE kmp_barrier_union {
1862 double b_align; /* use worst case alignment */
1863 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1864 kmp_bstate_t bb;
1865 };
1866
1867 typedef union kmp_barrier_union kmp_balign_t;
1868
1869 /* Team barrier needs only non-volatile arrived counter */
1870 union KMP_ALIGN_CACHE kmp_barrier_team_union {
1871 double b_align; /* use worst case alignment */
1872 char b_pad[CACHE_LINE];
1873 struct {
1874 kmp_uint64 b_arrived; /* STATE => task reached synch point. */
1875 #if USE_DEBUGGER
1876 // The following two fields are indended for the debugger solely. Only
1877 // master of the team accesses these fields: the first one is increased by
1878 // 1 when master arrives to a barrier, the second one is increased by one
1879 // when all the threads arrived.
1880 kmp_uint b_master_arrived;
1881 kmp_uint b_team_arrived;
1882 #endif
1883 };
1884 };
1885
1886 typedef union kmp_barrier_team_union kmp_balign_team_t;
1887
1888 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
1889 threads when a condition changes. This is to workaround an NPTL bug where
1890 padding was added to pthread_cond_t which caused the initialization routine
1891 to write outside of the structure if compiled on pre-NPTL threads. */
1892 #if KMP_OS_WINDOWS
1893 typedef struct kmp_win32_mutex {
1894 /* The Lock */
1895 CRITICAL_SECTION cs;
1896 } kmp_win32_mutex_t;
1897
1898 typedef struct kmp_win32_cond {
1899 /* Count of the number of waiters. */
1900 int waiters_count_;
1901
1902 /* Serialize access to <waiters_count_> */
1903 kmp_win32_mutex_t waiters_count_lock_;
1904
1905 /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
1906 int release_count_;
1907
1908 /* Keeps track of the current "generation" so that we don't allow */
1909 /* one thread to steal all the "releases" from the broadcast. */
1910 int wait_generation_count_;
1911
1912 /* A manual-reset event that's used to block and release waiting threads. */
1913 HANDLE event_;
1914 } kmp_win32_cond_t;
1915 #endif
1916
1917 #if KMP_OS_UNIX
1918
1919 union KMP_ALIGN_CACHE kmp_cond_union {
1920 double c_align;
1921 char c_pad[CACHE_LINE];
1922 pthread_cond_t c_cond;
1923 };
1924
1925 typedef union kmp_cond_union kmp_cond_align_t;
1926
1927 union KMP_ALIGN_CACHE kmp_mutex_union {
1928 double m_align;
1929 char m_pad[CACHE_LINE];
1930 pthread_mutex_t m_mutex;
1931 };
1932
1933 typedef union kmp_mutex_union kmp_mutex_align_t;
1934
1935 #endif /* KMP_OS_UNIX */
1936
1937 typedef struct kmp_desc_base {
1938 void *ds_stackbase;
1939 size_t ds_stacksize;
1940 int ds_stackgrow;
1941 kmp_thread_t ds_thread;
1942 volatile int ds_tid;
1943 int ds_gtid;
1944 #if KMP_OS_WINDOWS
1945 volatile int ds_alive;
1946 DWORD ds_thread_id;
1947 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
1948 However, debugger support (libomp_db) cannot work with handles, because they
1949 uncomparable. For example, debugger requests info about thread with handle h.
1950 h is valid within debugger process, and meaningless within debugee process.
1951 Even if h is duped by call to DuplicateHandle(), so the result h' is valid
1952 within debugee process, but it is a *new* handle which does *not* equal to
1953 any other handle in debugee... The only way to compare handles is convert
1954 them to system-wide ids. GetThreadId() function is available only in
1955 Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
1956 on all Windows* OS flavours (including Windows* 95). Thus, we have to get
1957 thread id by call to GetCurrentThreadId() from within the thread and save it
1958 to let libomp_db identify threads. */
1959 #endif /* KMP_OS_WINDOWS */
1960 } kmp_desc_base_t;
1961
1962 typedef union KMP_ALIGN_CACHE kmp_desc {
1963 double ds_align; /* use worst case alignment */
1964 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
1965 kmp_desc_base_t ds;
1966 } kmp_desc_t;
1967
1968 typedef struct kmp_local {
1969 volatile int this_construct; /* count of single's encountered by thread */
1970 void *reduce_data;
1971 #if KMP_USE_BGET
1972 void *bget_data;
1973 void *bget_list;
1974 #if !USE_CMP_XCHG_FOR_BGET
1975 #ifdef USE_QUEUING_LOCK_FOR_BGET
1976 kmp_lock_t bget_lock; /* Lock for accessing bget free list */
1977 #else
1978 kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
1979 // bootstrap lock so we can use it at library
1980 // shutdown.
1981 #endif /* USE_LOCK_FOR_BGET */
1982 #endif /* ! USE_CMP_XCHG_FOR_BGET */
1983 #endif /* KMP_USE_BGET */
1984
1985 PACKED_REDUCTION_METHOD_T
1986 packed_reduction_method; /* stored by __kmpc_reduce*(), used by
1987 __kmpc_end_reduce*() */
1988
1989 } kmp_local_t;
1990
1991 #define KMP_CHECK_UPDATE(a, b) \
1992 if ((a) != (b)) \
1993 (a) = (b)
1994 #define KMP_CHECK_UPDATE_SYNC(a, b) \
1995 if ((a) != (b)) \
1996 TCW_SYNC_PTR((a), (b))
1997
1998 #define get__blocktime(xteam, xtid) \
1999 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2000 #define get__bt_set(xteam, xtid) \
2001 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2002 #if KMP_USE_MONITOR
2003 #define get__bt_intervals(xteam, xtid) \
2004 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2005 #endif
2006
2007 #define get__dynamic_2(xteam, xtid) \
2008 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2009 #define get__nproc_2(xteam, xtid) \
2010 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2011 #define get__sched_2(xteam, xtid) \
2012 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2013
2014 #define set__blocktime_team(xteam, xtid, xval) \
2015 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2016 (xval))
2017
2018 #if KMP_USE_MONITOR
2019 #define set__bt_intervals_team(xteam, xtid, xval) \
2020 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2021 (xval))
2022 #endif
2023
2024 #define set__bt_set_team(xteam, xtid, xval) \
2025 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2026
2027 #define set__dynamic(xthread, xval) \
2028 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2029 #define get__dynamic(xthread) \
2030 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2031
2032 #define set__nproc(xthread, xval) \
2033 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2034
2035 #define set__thread_limit(xthread, xval) \
2036 (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2037
2038 #define set__max_active_levels(xthread, xval) \
2039 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2040
2041 #define get__max_active_levels(xthread) \
2042 ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2043
2044 #define set__sched(xthread, xval) \
2045 (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2046
2047 #define set__proc_bind(xthread, xval) \
2048 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2049 #define get__proc_bind(xthread) \
2050 ((xthread)->th.th_current_task->td_icvs.proc_bind)
2051
2052 // OpenMP tasking data structures
2053
2054 typedef enum kmp_tasking_mode {
2055 tskm_immediate_exec = 0,
2056 tskm_extra_barrier = 1,
2057 tskm_task_teams = 2,
2058 tskm_max = 2
2059 } kmp_tasking_mode_t;
2060
2061 extern kmp_tasking_mode_t
2062 __kmp_tasking_mode; /* determines how/when to execute tasks */
2063 extern int __kmp_task_stealing_constraint;
2064 extern int __kmp_enable_task_throttling;
2065 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2066 // specified, defaults to 0 otherwise
2067 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2068 extern kmp_int32 __kmp_max_task_priority;
2069 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2070 extern kmp_uint64 __kmp_taskloop_min_tasks;
2071
2072 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2073 taskdata first */
2074 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2075 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2076
2077 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2078 // were spawned and queued since the previous barrier release.
2079 #define KMP_TASKING_ENABLED(task_team) \
2080 (TCR_SYNC_4((task_team)->tt.tt_found_tasks) == TRUE)
2081 /*!
2082 @ingroup BASIC_TYPES
2083 @{
2084 */
2085
2086 /*!
2087 */
2088 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2089
2090 typedef union kmp_cmplrdata {
2091 kmp_int32 priority; /**< priority specified by user for the task */
2092 kmp_routine_entry_t
2093 destructors; /* pointer to function to invoke deconstructors of
2094 firstprivate C++ objects */
2095 /* future data */
2096 } kmp_cmplrdata_t;
2097
2098 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2099 /*!
2100 */
2101 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2102 void *shareds; /**< pointer to block of pointers to shared vars */
2103 kmp_routine_entry_t
2104 routine; /**< pointer to routine to call for executing task */
2105 kmp_int32 part_id; /**< part id for the task */
2106 kmp_cmplrdata_t
2107 data1; /* Two known optional additions: destructors and priority */
2108 kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2109 /* future data */
2110 /* private vars */
2111 } kmp_task_t;
2112
2113 /*!
2114 @}
2115 */
2116
2117 typedef struct kmp_taskgroup {
2118 std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2119 std::atomic<kmp_int32>
2120 cancel_request; // request for cancellation of this taskgroup
2121 struct kmp_taskgroup *parent; // parent taskgroup
2122 // Block of data to perform task reduction
2123 void *reduce_data; // reduction related info
2124 kmp_int32 reduce_num_data; // number of data items to reduce
2125 } kmp_taskgroup_t;
2126
2127 // forward declarations
2128 typedef union kmp_depnode kmp_depnode_t;
2129 typedef struct kmp_depnode_list kmp_depnode_list_t;
2130 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2131
2132 // Compiler sends us this info:
2133 typedef struct kmp_depend_info {
2134 kmp_intptr_t base_addr;
2135 size_t len;
2136 struct {
2137 bool in : 1;
2138 bool out : 1;
2139 bool mtx : 1;
2140 } flags;
2141 } kmp_depend_info_t;
2142
2143 // Internal structures to work with task dependencies:
2144 struct kmp_depnode_list {
2145 kmp_depnode_t *node;
2146 kmp_depnode_list_t *next;
2147 };
2148
2149 // Max number of mutexinoutset dependencies per node
2150 #define MAX_MTX_DEPS 4
2151
2152 typedef struct kmp_base_depnode {
2153 kmp_depnode_list_t *successors; /* used under lock */
2154 kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2155 kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2156 kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2157 kmp_lock_t lock; /* guards shared fields: task, successors */
2158 #if KMP_SUPPORT_GRAPH_OUTPUT
2159 kmp_uint32 id;
2160 #endif
2161 std::atomic<kmp_int32> npredecessors;
2162 std::atomic<kmp_int32> nrefs;
2163 } kmp_base_depnode_t;
2164
2165 union KMP_ALIGN_CACHE kmp_depnode {
2166 double dn_align; /* use worst case alignment */
2167 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2168 kmp_base_depnode_t dn;
2169 };
2170
2171 struct kmp_dephash_entry {
2172 kmp_intptr_t addr;
2173 kmp_depnode_t *last_out;
2174 kmp_depnode_list_t *last_ins;
2175 kmp_depnode_list_t *last_mtxs;
2176 kmp_int32 last_flag;
2177 kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2178 kmp_dephash_entry_t *next_in_bucket;
2179 };
2180
2181 typedef struct kmp_dephash {
2182 kmp_dephash_entry_t **buckets;
2183 size_t size;
2184 size_t generation;
2185 kmp_uint32 nelements;
2186 kmp_uint32 nconflicts;
2187 } kmp_dephash_t;
2188
2189 typedef struct kmp_task_affinity_info {
2190 kmp_intptr_t base_addr;
2191 size_t len;
2192 struct {
2193 bool flag1 : 1;
2194 bool flag2 : 1;
2195 kmp_int32 reserved : 30;
2196 } flags;
2197 } kmp_task_affinity_info_t;
2198
2199 typedef enum kmp_event_type_t {
2200 KMP_EVENT_UNINITIALIZED = 0,
2201 KMP_EVENT_ALLOW_COMPLETION = 1
2202 } kmp_event_type_t;
2203
2204 typedef struct {
2205 kmp_event_type_t type;
2206 kmp_tas_lock_t lock;
2207 union {
2208 kmp_task_t *task;
2209 } ed;
2210 } kmp_event_t;
2211
2212 #ifdef BUILD_TIED_TASK_STACK
2213
2214 /* Tied Task stack definitions */
2215 typedef struct kmp_stack_block {
2216 kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2217 struct kmp_stack_block *sb_next;
2218 struct kmp_stack_block *sb_prev;
2219 } kmp_stack_block_t;
2220
2221 typedef struct kmp_task_stack {
2222 kmp_stack_block_t ts_first_block; // first block of stack entries
2223 kmp_taskdata_t **ts_top; // pointer to the top of stack
2224 kmp_int32 ts_entries; // number of entries on the stack
2225 } kmp_task_stack_t;
2226
2227 #endif // BUILD_TIED_TASK_STACK
2228
2229 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2230 /* Compiler flags */ /* Total compiler flags must be 16 bits */
2231 unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2232 unsigned final : 1; /* task is final(1) so execute immediately */
2233 unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2234 code path */
2235 unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2236 invoke destructors from the runtime */
2237 unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2238 context of the RTL) */
2239 unsigned priority_specified : 1; /* set if the compiler provides priority
2240 setting for the task */
2241 unsigned detachable : 1; /* 1 == can detach */
2242 unsigned reserved : 9; /* reserved for compiler use */
2243
2244 /* Library flags */ /* Total library flags must be 16 bits */
2245 unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2246 unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2247 unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2248 // (1) or may be deferred (0)
2249 unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2250 // (0) [>= 2 threads]
2251 /* If either team_serial or tasking_ser is set, task team may be NULL */
2252 /* Task State Flags: */
2253 unsigned started : 1; /* 1==started, 0==not started */
2254 unsigned executing : 1; /* 1==executing, 0==not executing */
2255 unsigned complete : 1; /* 1==complete, 0==not complete */
2256 unsigned freed : 1; /* 1==freed, 0==allocated */
2257 unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2258 unsigned reserved31 : 7; /* reserved for library use */
2259
2260 } kmp_tasking_flags_t;
2261
2262 struct kmp_taskdata { /* aligned during dynamic allocation */
2263 kmp_int32 td_task_id; /* id, assigned by debugger */
2264 kmp_tasking_flags_t td_flags; /* task flags */
2265 kmp_team_t *td_team; /* team for this task */
2266 kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2267 /* Currently not used except for perhaps IDB */
2268 kmp_taskdata_t *td_parent; /* parent task */
2269 kmp_int32 td_level; /* task nesting level */
2270 std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2271 ident_t *td_ident; /* task identifier */
2272 // Taskwait data.
2273 ident_t *td_taskwait_ident;
2274 kmp_uint32 td_taskwait_counter;
2275 kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2276 KMP_ALIGN_CACHE kmp_internal_control_t
2277 td_icvs; /* Internal control variables for the task */
2278 KMP_ALIGN_CACHE std::atomic<kmp_int32>
2279 td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2280 deallocated */
2281 std::atomic<kmp_int32>
2282 td_incomplete_child_tasks; /* Child tasks not yet complete */
2283 kmp_taskgroup_t
2284 *td_taskgroup; // Each task keeps pointer to its current taskgroup
2285 kmp_dephash_t
2286 *td_dephash; // Dependencies for children tasks are tracked from here
2287 kmp_depnode_t
2288 *td_depnode; // Pointer to graph node if this task has dependencies
2289 kmp_task_team_t *td_task_team;
2290 kmp_int32 td_size_alloc; // The size of task structure, including shareds etc.
2291 #if defined(KMP_GOMP_COMPAT)
2292 // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2293 kmp_int32 td_size_loop_bounds;
2294 #endif
2295 kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2296 #if defined(KMP_GOMP_COMPAT)
2297 // GOMP sends in a copy function for copy constructors
2298 void (*td_copy_func)(void *, void *);
2299 #endif
2300 kmp_event_t td_allow_completion_event;
2301 #if OMPT_SUPPORT
2302 ompt_task_info_t ompt_task_info;
2303 #endif
2304 }; // struct kmp_taskdata
2305
2306 // Make sure padding above worked
2307 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2308
2309 // Data for task team but per thread
2310 typedef struct kmp_base_thread_data {
2311 kmp_info_p *td_thr; // Pointer back to thread info
2312 // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2313 // queued?
2314 kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2315 kmp_taskdata_t *
2316 *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2317 kmp_int32 td_deque_size; // Size of deck
2318 kmp_uint32 td_deque_head; // Head of deque (will wrap)
2319 kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2320 kmp_int32 td_deque_ntasks; // Number of tasks in deque
2321 // GEH: shouldn't this be volatile since used in while-spin?
2322 kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2323 #ifdef BUILD_TIED_TASK_STACK
2324 kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2325 // scheduling constraint
2326 #endif // BUILD_TIED_TASK_STACK
2327 } kmp_base_thread_data_t;
2328
2329 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2330 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2331
2332 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2333 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2334
2335 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2336 kmp_base_thread_data_t td;
2337 double td_align; /* use worst case alignment */
2338 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2339 } kmp_thread_data_t;
2340
2341 // Data for task teams which are used when tasking is enabled for the team
2342 typedef struct kmp_base_task_team {
2343 kmp_bootstrap_lock_t
2344 tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2345 /* must be bootstrap lock since used at library shutdown*/
2346 kmp_task_team_t *tt_next; /* For linking the task team free list */
2347 kmp_thread_data_t
2348 *tt_threads_data; /* Array of per-thread structures for task team */
2349 /* Data survives task team deallocation */
2350 kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2351 executing this team? */
2352 /* TRUE means tt_threads_data is set up and initialized */
2353 kmp_int32 tt_nproc; /* #threads in team */
2354 kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2355 kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2356 kmp_int32 tt_untied_task_encountered;
2357
2358 KMP_ALIGN_CACHE
2359 std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2360
2361 KMP_ALIGN_CACHE
2362 volatile kmp_uint32
2363 tt_active; /* is the team still actively executing tasks */
2364 } kmp_base_task_team_t;
2365
2366 union KMP_ALIGN_CACHE kmp_task_team {
2367 kmp_base_task_team_t tt;
2368 double tt_align; /* use worst case alignment */
2369 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2370 };
2371
2372 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2373 // Free lists keep same-size free memory slots for fast memory allocation
2374 // routines
2375 typedef struct kmp_free_list {
2376 void *th_free_list_self; // Self-allocated tasks free list
2377 void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2378 // threads
2379 void *th_free_list_other; // Non-self free list (to be returned to owner's
2380 // sync list)
2381 } kmp_free_list_t;
2382 #endif
2383 #if KMP_NESTED_HOT_TEAMS
2384 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2385 // are not put in teams pool, and they don't put threads in threads pool.
2386 typedef struct kmp_hot_team_ptr {
2387 kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2388 kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2389 } kmp_hot_team_ptr_t;
2390 #endif
2391 typedef struct kmp_teams_size {
2392 kmp_int32 nteams; // number of teams in a league
2393 kmp_int32 nth; // number of threads in each team of the league
2394 } kmp_teams_size_t;
2395
2396 // This struct stores a thread that acts as a "root" for a contention
2397 // group. Contention groups are rooted at kmp_root threads, but also at
2398 // each master thread of each team created in the teams construct.
2399 // This struct therefore also stores a thread_limit associated with
2400 // that contention group, and a counter to track the number of threads
2401 // active in that contention group. Each thread has a list of these: CG
2402 // root threads have an entry in their list in which cg_root refers to
2403 // the thread itself, whereas other workers in the CG will have a
2404 // single entry where cg_root is same as the entry containing their CG
2405 // root. When a thread encounters a teams construct, it will add a new
2406 // entry to the front of its list, because it now roots a new CG.
2407 typedef struct kmp_cg_root {
2408 kmp_info_p *cg_root; // "root" thread for a contention group
2409 // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2410 // thread_limit clause for teams masters
2411 kmp_int32 cg_thread_limit;
2412 kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2413 struct kmp_cg_root *up; // pointer to higher level CG root in list
2414 } kmp_cg_root_t;
2415
2416 // OpenMP thread data structures
2417
2418 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2419 /* Start with the readonly data which is cache aligned and padded. This is
2420 written before the thread starts working by the master. Uber masters may
2421 update themselves later. Usage does not consider serialized regions. */
2422 kmp_desc_t th_info;
2423 kmp_team_p *th_team; /* team we belong to */
2424 kmp_root_p *th_root; /* pointer to root of task hierarchy */
2425 kmp_info_p *th_next_pool; /* next available thread in the pool */
2426 kmp_disp_t *th_dispatch; /* thread's dispatch data */
2427 int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2428
2429 /* The following are cached from the team info structure */
2430 /* TODO use these in more places as determined to be needed via profiling */
2431 int th_team_nproc; /* number of threads in a team */
2432 kmp_info_p *th_team_master; /* the team's master thread */
2433 int th_team_serialized; /* team is serialized */
2434 microtask_t th_teams_microtask; /* save entry address for teams construct */
2435 int th_teams_level; /* save initial level of teams construct */
2436 /* it is 0 on device but may be any on host */
2437
2438 /* The blocktime info is copied from the team struct to the thread sruct */
2439 /* at the start of a barrier, and the values stored in the team are used */
2440 /* at points in the code where the team struct is no longer guaranteed */
2441 /* to exist (from the POV of worker threads). */
2442 #if KMP_USE_MONITOR
2443 int th_team_bt_intervals;
2444 int th_team_bt_set;
2445 #else
2446 kmp_uint64 th_team_bt_intervals;
2447 #endif
2448
2449 #if KMP_AFFINITY_SUPPORTED
2450 kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2451 #endif
2452 omp_allocator_handle_t th_def_allocator; /* default allocator */
2453 /* The data set by the master at reinit, then R/W by the worker */
2454 KMP_ALIGN_CACHE int
2455 th_set_nproc; /* if > 0, then only use this request for the next fork */
2456 #if KMP_NESTED_HOT_TEAMS
2457 kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2458 #endif
2459 kmp_proc_bind_t
2460 th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2461 kmp_teams_size_t
2462 th_teams_size; /* number of teams/threads in teams construct */
2463 #if KMP_AFFINITY_SUPPORTED
2464 int th_current_place; /* place currently bound to */
2465 int th_new_place; /* place to bind to in par reg */
2466 int th_first_place; /* first place in partition */
2467 int th_last_place; /* last place in partition */
2468 #endif
2469 int th_prev_level; /* previous level for affinity format */
2470 int th_prev_num_threads; /* previous num_threads for affinity format */
2471 #if USE_ITT_BUILD
2472 kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2473 kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2474 kmp_uint64 th_frame_time; /* frame timestamp */
2475 #endif /* USE_ITT_BUILD */
2476 kmp_local_t th_local;
2477 struct private_common *th_pri_head;
2478
2479 /* Now the data only used by the worker (after initial allocation) */
2480 /* TODO the first serial team should actually be stored in the info_t
2481 structure. this will help reduce initial allocation overhead */
2482 KMP_ALIGN_CACHE kmp_team_p
2483 *th_serial_team; /*serialized team held in reserve*/
2484
2485 #if OMPT_SUPPORT
2486 ompt_thread_info_t ompt_thread_info;
2487 #endif
2488
2489 /* The following are also read by the master during reinit */
2490 struct common_table *th_pri_common;
2491
2492 volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2493 /* while awaiting queuing lock acquire */
2494
2495 volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2496
2497 ident_t *th_ident;
2498 unsigned th_x; // Random number generator data
2499 unsigned th_a; // Random number generator data
2500
2501 /* Tasking-related data for the thread */
2502 kmp_task_team_t *th_task_team; // Task team struct
2503 kmp_taskdata_t *th_current_task; // Innermost Task being executed
2504 kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2505 kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2506 // at nested levels
2507 kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2508 kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2509 kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2510 // tasking, thus safe to reap
2511
2512 /* More stuff for keeping track of active/sleeping threads (this part is
2513 written by the worker thread) */
2514 kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2515 int th_active; // ! sleeping; 32 bits for TCR/TCW
2516 struct cons_header *th_cons; // used for consistency check
2517 #if KMP_USE_HIER_SCHED
2518 // used for hierarchical scheduling
2519 kmp_hier_private_bdata_t *th_hier_bar_data;
2520 #endif
2521
2522 /* Add the syncronizing data which is cache aligned and padded. */
2523 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2524
2525 KMP_ALIGN_CACHE volatile kmp_int32
2526 th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2527
2528 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2529 #define NUM_LISTS 4
2530 kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2531 // allocation routines
2532 #endif
2533
2534 #if KMP_OS_WINDOWS
2535 kmp_win32_cond_t th_suspend_cv;
2536 kmp_win32_mutex_t th_suspend_mx;
2537 std::atomic<int> th_suspend_init;
2538 #endif
2539 #if KMP_OS_UNIX
2540 kmp_cond_align_t th_suspend_cv;
2541 kmp_mutex_align_t th_suspend_mx;
2542 std::atomic<int> th_suspend_init_count;
2543 #endif
2544
2545 #if USE_ITT_BUILD
2546 kmp_itt_mark_t th_itt_mark_single;
2547 // alignment ???
2548 #endif /* USE_ITT_BUILD */
2549 #if KMP_STATS_ENABLED
2550 kmp_stats_list *th_stats;
2551 #endif
2552 #if KMP_OS_UNIX
2553 std::atomic<bool> th_blocking;
2554 #endif
2555 kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2556 } kmp_base_info_t;
2557
2558 typedef union KMP_ALIGN_CACHE kmp_info {
2559 double th_align; /* use worst case alignment */
2560 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2561 kmp_base_info_t th;
2562 } kmp_info_t;
2563
2564 // OpenMP thread team data structures
2565
2566 typedef struct kmp_base_data { volatile kmp_uint32 t_value; } kmp_base_data_t;
2567
2568 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2569 double dt_align; /* use worst case alignment */
2570 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2571 kmp_base_data_t dt;
2572 } kmp_sleep_team_t;
2573
2574 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2575 double dt_align; /* use worst case alignment */
2576 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2577 kmp_base_data_t dt;
2578 } kmp_ordered_team_t;
2579
2580 typedef int (*launch_t)(int gtid);
2581
2582 /* Minimum number of ARGV entries to malloc if necessary */
2583 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2584
2585 // Set up how many argv pointers will fit in cache lines containing
2586 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2587 // larger value for more space between the master write/worker read section and
2588 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2589 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2590 #define KMP_INLINE_ARGV_BYTES \
2591 (4 * CACHE_LINE - \
2592 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2593 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2594 CACHE_LINE))
2595 #else
2596 #define KMP_INLINE_ARGV_BYTES \
2597 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2598 #endif
2599 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2600
2601 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2602 // Synchronization Data
2603 // ---------------------------------------------------------------------------
2604 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2605 kmp_balign_team_t t_bar[bs_last_barrier];
2606 std::atomic<int> t_construct; // count of single directive encountered by team
2607 char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2608
2609 // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2610 std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2611 std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2612
2613 // Master only
2614 // ---------------------------------------------------------------------------
2615 KMP_ALIGN_CACHE int t_master_tid; // tid of master in parent team
2616 int t_master_this_cons; // "this_construct" single counter of master in parent
2617 // team
2618 ident_t *t_ident; // if volatile, have to change too much other crud to
2619 // volatile too
2620 kmp_team_p *t_parent; // parent team
2621 kmp_team_p *t_next_pool; // next free team in the team pool
2622 kmp_disp_t *t_dispatch; // thread's dispatch data
2623 kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2624 kmp_proc_bind_t t_proc_bind; // bind type for par region
2625 #if USE_ITT_BUILD
2626 kmp_uint64 t_region_time; // region begin timestamp
2627 #endif /* USE_ITT_BUILD */
2628
2629 // Master write, workers read
2630 // --------------------------------------------------------------------------
2631 KMP_ALIGN_CACHE void **t_argv;
2632 int t_argc;
2633 int t_nproc; // number of threads in team
2634 microtask_t t_pkfn;
2635 launch_t t_invoke; // procedure to launch the microtask
2636
2637 #if OMPT_SUPPORT
2638 ompt_team_info_t ompt_team_info;
2639 ompt_lw_taskteam_t *ompt_serialized_team_info;
2640 #endif
2641
2642 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2643 kmp_int8 t_fp_control_saved;
2644 kmp_int8 t_pad2b;
2645 kmp_int16 t_x87_fpu_control_word; // FP control regs
2646 kmp_uint32 t_mxcsr;
2647 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2648
2649 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2650
2651 KMP_ALIGN_CACHE kmp_info_t **t_threads;
2652 kmp_taskdata_t
2653 *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2654 int t_level; // nested parallel level
2655
2656 KMP_ALIGN_CACHE int t_max_argc;
2657 int t_max_nproc; // max threads this team can handle (dynamically expandable)
2658 int t_serialized; // levels deep of serialized teams
2659 dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2660 int t_id; // team's id, assigned by debugger.
2661 int t_active_level; // nested active parallel level
2662 kmp_r_sched_t t_sched; // run-time schedule for the team
2663 #if KMP_AFFINITY_SUPPORTED
2664 int t_first_place; // first & last place in parent thread's partition.
2665 int t_last_place; // Restore these values to master after par region.
2666 #endif // KMP_AFFINITY_SUPPORTED
2667 int t_display_affinity;
2668 int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2669 // omp_set_num_threads() call
2670 omp_allocator_handle_t t_def_allocator; /* default allocator */
2671
2672 // Read/write by workers as well
2673 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2674 // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2675 // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2676 // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2677 // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2678 char dummy_padding[1024];
2679 #endif
2680 // Internal control stack for additional nested teams.
2681 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2682 // for SERIALIZED teams nested 2 or more levels deep
2683 // typed flag to store request state of cancellation
2684 std::atomic<kmp_int32> t_cancel_request;
2685 int t_master_active; // save on fork, restore on join
2686 void *t_copypriv_data; // team specific pointer to copyprivate data array
2687 #if KMP_OS_WINDOWS
2688 std::atomic<kmp_uint32> t_copyin_counter;
2689 #endif
2690 #if USE_ITT_BUILD
2691 void *t_stack_id; // team specific stack stitching id (for ittnotify)
2692 #endif /* USE_ITT_BUILD */
2693 } kmp_base_team_t;
2694
2695 union KMP_ALIGN_CACHE kmp_team {
2696 kmp_base_team_t t;
2697 double t_align; /* use worst case alignment */
2698 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2699 };
2700
2701 typedef union KMP_ALIGN_CACHE kmp_time_global {
2702 double dt_align; /* use worst case alignment */
2703 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2704 kmp_base_data_t dt;
2705 } kmp_time_global_t;
2706
2707 typedef struct kmp_base_global {
2708 /* cache-aligned */
2709 kmp_time_global_t g_time;
2710
2711 /* non cache-aligned */
2712 volatile int g_abort;
2713 volatile int g_done;
2714
2715 int g_dynamic;
2716 enum dynamic_mode g_dynamic_mode;
2717 } kmp_base_global_t;
2718
2719 typedef union KMP_ALIGN_CACHE kmp_global {
2720 kmp_base_global_t g;
2721 double g_align; /* use worst case alignment */
2722 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2723 } kmp_global_t;
2724
2725 typedef struct kmp_base_root {
2726 // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2727 // (r_in_parallel>= 0)
2728 // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2729 // the synch overhead or keeping r_active
2730 volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2731 // keeps a count of active parallel regions per root
2732 std::atomic<int> r_in_parallel;
2733 // GEH: This is misnamed, should be r_active_levels
2734 kmp_team_t *r_root_team;
2735 kmp_team_t *r_hot_team;
2736 kmp_info_t *r_uber_thread;
2737 kmp_lock_t r_begin_lock;
2738 volatile int r_begin;
2739 int r_blocktime; /* blocktime for this root and descendants */
2740 } kmp_base_root_t;
2741
2742 typedef union KMP_ALIGN_CACHE kmp_root {
2743 kmp_base_root_t r;
2744 double r_align; /* use worst case alignment */
2745 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2746 } kmp_root_t;
2747
2748 struct fortran_inx_info {
2749 kmp_int32 data;
2750 };
2751
2752 /* ------------------------------------------------------------------------ */
2753
2754 extern int __kmp_settings;
2755 extern int __kmp_duplicate_library_ok;
2756 #if USE_ITT_BUILD
2757 extern int __kmp_forkjoin_frames;
2758 extern int __kmp_forkjoin_frames_mode;
2759 #endif
2760 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2761 extern int __kmp_determ_red;
2762
2763 #ifdef KMP_DEBUG
2764 extern int kmp_a_debug;
2765 extern int kmp_b_debug;
2766 extern int kmp_c_debug;
2767 extern int kmp_d_debug;
2768 extern int kmp_e_debug;
2769 extern int kmp_f_debug;
2770 #endif /* KMP_DEBUG */
2771
2772 /* For debug information logging using rotating buffer */
2773 #define KMP_DEBUG_BUF_LINES_INIT 512
2774 #define KMP_DEBUG_BUF_LINES_MIN 1
2775
2776 #define KMP_DEBUG_BUF_CHARS_INIT 128
2777 #define KMP_DEBUG_BUF_CHARS_MIN 2
2778
2779 extern int
2780 __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2781 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2782 extern int
2783 __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2784 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2785 entry pointer */
2786
2787 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2788 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2789 printed in buffer so far */
2790 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2791 recommended in warnings */
2792 /* end rotating debug buffer */
2793
2794 #ifdef KMP_DEBUG
2795 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2796
2797 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2798 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2799 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2800 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2801 extern int __kmp_par_range_lb;
2802 extern int __kmp_par_range_ub;
2803 #endif
2804
2805 /* For printing out dynamic storage map for threads and teams */
2806 extern int
2807 __kmp_storage_map; /* True means print storage map for threads and teams */
2808 extern int __kmp_storage_map_verbose; /* True means storage map includes
2809 placement info */
2810 extern int __kmp_storage_map_verbose_specified;
2811
2812 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2813 extern kmp_cpuinfo_t __kmp_cpuinfo;
2814 #endif
2815
2816 extern volatile int __kmp_init_serial;
2817 extern volatile int __kmp_init_gtid;
2818 extern volatile int __kmp_init_common;
2819 extern volatile int __kmp_init_middle;
2820 extern volatile int __kmp_init_parallel;
2821 #if KMP_USE_MONITOR
2822 extern volatile int __kmp_init_monitor;
2823 #endif
2824 extern volatile int __kmp_init_user_locks;
2825 extern int __kmp_init_counter;
2826 extern int __kmp_root_counter;
2827 extern int __kmp_version;
2828
2829 /* list of address of allocated caches for commons */
2830 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2831
2832 /* Barrier algorithm types and options */
2833 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
2834 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
2835 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
2836 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
2837 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
2838 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
2839 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
2840 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
2841 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
2842 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
2843 extern char const *__kmp_barrier_type_name[bs_last_barrier];
2844 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
2845
2846 /* Global Locks */
2847 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
2848 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
2849 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
2850 extern kmp_bootstrap_lock_t
2851 __kmp_exit_lock; /* exit() is not always thread-safe */
2852 #if KMP_USE_MONITOR
2853 extern kmp_bootstrap_lock_t
2854 __kmp_monitor_lock; /* control monitor thread creation */
2855 #endif
2856 extern kmp_bootstrap_lock_t
2857 __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
2858 __kmp_threads expansion to co-exist */
2859
2860 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
2861 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
2862 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
2863
2864 extern enum library_type __kmp_library;
2865
2866 extern enum sched_type __kmp_sched; /* default runtime scheduling */
2867 extern enum sched_type __kmp_static; /* default static scheduling method */
2868 extern enum sched_type __kmp_guided; /* default guided scheduling method */
2869 extern enum sched_type __kmp_auto; /* default auto scheduling method */
2870 extern int __kmp_chunk; /* default runtime chunk size */
2871
2872 extern size_t __kmp_stksize; /* stack size per thread */
2873 #if KMP_USE_MONITOR
2874 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
2875 #endif
2876 extern size_t __kmp_stkoffset; /* stack offset per thread */
2877 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
2878
2879 extern size_t
2880 __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
2881 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
2882 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
2883 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
2884 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
2885 extern int __kmp_generate_warnings; /* should we issue warnings? */
2886 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
2887
2888 #ifdef DEBUG_SUSPEND
2889 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
2890 #endif
2891
2892 extern kmp_int32 __kmp_use_yield;
2893 extern kmp_int32 __kmp_use_yield_exp_set;
2894 extern kmp_uint32 __kmp_yield_init;
2895 extern kmp_uint32 __kmp_yield_next;
2896
2897 /* ------------------------------------------------------------------------- */
2898 extern int __kmp_allThreadsSpecified;
2899
2900 extern size_t __kmp_align_alloc;
2901 /* following data protected by initialization routines */
2902 extern int __kmp_xproc; /* number of processors in the system */
2903 extern int __kmp_avail_proc; /* number of processors available to the process */
2904 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
2905 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
2906 // maximum total number of concurrently-existing threads on device
2907 extern int __kmp_max_nth;
2908 // maximum total number of concurrently-existing threads in a contention group
2909 extern int __kmp_cg_max_nth;
2910 extern int __kmp_teams_max_nth; // max threads used in a teams construct
2911 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
2912 __kmp_root */
2913 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
2914 region a la OMP_NUM_THREADS */
2915 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
2916 initialization */
2917 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
2918 used (fixed) */
2919 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
2920 (__kmpc_threadprivate_cached()) */
2921 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
2922 blocking (env setting) */
2923 #if KMP_USE_MONITOR
2924 extern int
2925 __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
2926 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
2927 blocking */
2928 #endif
2929 #ifdef KMP_ADJUST_BLOCKTIME
2930 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
2931 #endif /* KMP_ADJUST_BLOCKTIME */
2932 #ifdef KMP_DFLT_NTH_CORES
2933 extern int __kmp_ncores; /* Total number of cores for threads placement */
2934 #endif
2935 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
2936 extern int __kmp_abort_delay;
2937
2938 extern int __kmp_need_register_atfork_specified;
2939 extern int
2940 __kmp_need_register_atfork; /* At initialization, call pthread_atfork to
2941 install fork handler */
2942 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
2943 0 - not set, will be set at runtime
2944 1 - using stack search
2945 2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
2946 X*) or TlsGetValue(Windows* OS))
2947 3 - static TLS (__declspec(thread) __kmp_gtid),
2948 Linux* OS .so only. */
2949 extern int
2950 __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
2951 #ifdef KMP_TDATA_GTID
2952 extern KMP_THREAD_LOCAL int __kmp_gtid;
2953 #endif
2954 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
2955 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
2956 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2957 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
2958 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
2959 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
2960 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2961
2962 // max_active_levels for nested parallelism enabled by default via
2963 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
2964 extern int __kmp_dflt_max_active_levels;
2965 // Indicates whether value of __kmp_dflt_max_active_levels was already
2966 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
2967 extern bool __kmp_dflt_max_active_levels_set;
2968 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
2969 concurrent execution per team */
2970 #if KMP_NESTED_HOT_TEAMS
2971 extern int __kmp_hot_teams_mode;
2972 extern int __kmp_hot_teams_max_level;
2973 #endif
2974
2975 #if KMP_OS_LINUX
2976 extern enum clock_function_type __kmp_clock_function;
2977 extern int __kmp_clock_function_param;
2978 #endif /* KMP_OS_LINUX */
2979
2980 #if KMP_MIC_SUPPORTED
2981 extern enum mic_type __kmp_mic_type;
2982 #endif
2983
2984 #ifdef USE_LOAD_BALANCE
2985 extern double __kmp_load_balance_interval; // load balance algorithm interval
2986 #endif /* USE_LOAD_BALANCE */
2987
2988 // OpenMP 3.1 - Nested num threads array
2989 typedef struct kmp_nested_nthreads_t {
2990 int *nth;
2991 int size;
2992 int used;
2993 } kmp_nested_nthreads_t;
2994
2995 extern kmp_nested_nthreads_t __kmp_nested_nth;
2996
2997 #if KMP_USE_ADAPTIVE_LOCKS
2998
2999 // Parameters for the speculative lock backoff system.
3000 struct kmp_adaptive_backoff_params_t {
3001 // Number of soft retries before it counts as a hard retry.
3002 kmp_uint32 max_soft_retries;
3003 // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3004 // the right
3005 kmp_uint32 max_badness;
3006 };
3007
3008 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3009
3010 #if KMP_DEBUG_ADAPTIVE_LOCKS
3011 extern const char *__kmp_speculative_statsfile;
3012 #endif
3013
3014 #endif // KMP_USE_ADAPTIVE_LOCKS
3015
3016 extern int __kmp_display_env; /* TRUE or FALSE */
3017 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3018 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3019
3020 /* ------------------------------------------------------------------------- */
3021
3022 /* the following are protected by the fork/join lock */
3023 /* write: lock read: anytime */
3024 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3025 /* read/write: lock */
3026 extern volatile kmp_team_t *__kmp_team_pool;
3027 extern volatile kmp_info_t *__kmp_thread_pool;
3028 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3029
3030 // total num threads reachable from some root thread including all root threads
3031 extern volatile int __kmp_nth;
3032 /* total number of threads reachable from some root thread including all root
3033 threads, and those in the thread pool */
3034 extern volatile int __kmp_all_nth;
3035 extern std::atomic<int> __kmp_thread_pool_active_nth;
3036
3037 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3038 /* end data protected by fork/join lock */
3039 /* ------------------------------------------------------------------------- */
3040
3041 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3042 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3043 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3044 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3045 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3046
3047 // AT: Which way is correct?
3048 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3049 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3050 #define __kmp_get_team_num_threads(gtid) \
3051 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3052
KMP_UBER_GTID(int gtid)3053 static inline bool KMP_UBER_GTID(int gtid) {
3054 KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3055 KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3056 return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3057 __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3058 }
3059
__kmp_tid_from_gtid(int gtid)3060 static inline int __kmp_tid_from_gtid(int gtid) {
3061 KMP_DEBUG_ASSERT(gtid >= 0);
3062 return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3063 }
3064
__kmp_gtid_from_tid(int tid,const kmp_team_t * team)3065 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3066 KMP_DEBUG_ASSERT(tid >= 0 && team);
3067 return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3068 }
3069
__kmp_gtid_from_thread(const kmp_info_t * thr)3070 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3071 KMP_DEBUG_ASSERT(thr);
3072 return thr->th.th_info.ds.ds_gtid;
3073 }
3074
__kmp_thread_from_gtid(int gtid)3075 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3076 KMP_DEBUG_ASSERT(gtid >= 0);
3077 return __kmp_threads[gtid];
3078 }
3079
__kmp_team_from_gtid(int gtid)3080 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3081 KMP_DEBUG_ASSERT(gtid >= 0);
3082 return __kmp_threads[gtid]->th.th_team;
3083 }
3084
3085 /* ------------------------------------------------------------------------- */
3086
3087 extern kmp_global_t __kmp_global; /* global status */
3088
3089 extern kmp_info_t __kmp_monitor;
3090 // For Debugging Support Library
3091 extern std::atomic<kmp_int32> __kmp_team_counter;
3092 // For Debugging Support Library
3093 extern std::atomic<kmp_int32> __kmp_task_counter;
3094
3095 #if USE_DEBUGGER
3096 #define _KMP_GEN_ID(counter) \
3097 (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3098 #else
3099 #define _KMP_GEN_ID(counter) (~0)
3100 #endif /* USE_DEBUGGER */
3101
3102 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3103 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3104
3105 /* ------------------------------------------------------------------------ */
3106
3107 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3108 size_t size, char const *format, ...);
3109
3110 extern void __kmp_serial_initialize(void);
3111 extern void __kmp_middle_initialize(void);
3112 extern void __kmp_parallel_initialize(void);
3113
3114 extern void __kmp_internal_begin(void);
3115 extern void __kmp_internal_end_library(int gtid);
3116 extern void __kmp_internal_end_thread(int gtid);
3117 extern void __kmp_internal_end_atexit(void);
3118 extern void __kmp_internal_end_dtor(void);
3119 extern void __kmp_internal_end_dest(void *);
3120
3121 extern int __kmp_register_root(int initial_thread);
3122 extern void __kmp_unregister_root(int gtid);
3123
3124 extern int __kmp_ignore_mppbeg(void);
3125 extern int __kmp_ignore_mppend(void);
3126
3127 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3128 extern void __kmp_exit_single(int gtid);
3129
3130 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3131 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3132
3133 #ifdef USE_LOAD_BALANCE
3134 extern int __kmp_get_load_balance(int);
3135 #endif
3136
3137 extern int __kmp_get_global_thread_id(void);
3138 extern int __kmp_get_global_thread_id_reg(void);
3139 extern void __kmp_exit_thread(int exit_status);
3140 extern void __kmp_abort(char const *format, ...);
3141 extern void __kmp_abort_thread(void);
3142 KMP_NORETURN extern void __kmp_abort_process(void);
3143 extern void __kmp_warn(char const *format, ...);
3144
3145 extern void __kmp_set_num_threads(int new_nth, int gtid);
3146
3147 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3148 // registered.
__kmp_entry_thread()3149 static inline kmp_info_t *__kmp_entry_thread() {
3150 int gtid = __kmp_entry_gtid();
3151
3152 return __kmp_threads[gtid];
3153 }
3154
3155 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3156 extern int __kmp_get_max_active_levels(int gtid);
3157 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3158 extern int __kmp_get_team_size(int gtid, int level);
3159 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3160 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3161
3162 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3163 extern void __kmp_init_random(kmp_info_t *thread);
3164
3165 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3166 extern void __kmp_adjust_num_threads(int new_nproc);
3167 extern void __kmp_check_stksize(size_t *val);
3168
3169 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3170 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3171 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3172 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3173 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3174 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3175
3176 #if USE_FAST_MEMORY
3177 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3178 size_t size KMP_SRC_LOC_DECL);
3179 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3180 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3181 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3182 #define __kmp_fast_allocate(this_thr, size) \
3183 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3184 #define __kmp_fast_free(this_thr, ptr) \
3185 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3186 #endif
3187
3188 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3189 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3190 size_t elsize KMP_SRC_LOC_DECL);
3191 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3192 size_t size KMP_SRC_LOC_DECL);
3193 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3194 #define __kmp_thread_malloc(th, size) \
3195 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3196 #define __kmp_thread_calloc(th, nelem, elsize) \
3197 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3198 #define __kmp_thread_realloc(th, ptr, size) \
3199 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3200 #define __kmp_thread_free(th, ptr) \
3201 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3202
3203 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3204 #define KMP_INTERNAL_FREE(p) free(p)
3205 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3206 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3207
3208 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3209
3210 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3211 kmp_proc_bind_t proc_bind);
3212 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3213 int num_threads);
3214
3215 extern void __kmp_yield();
3216
3217 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3218 enum sched_type schedule, kmp_int32 lb,
3219 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3220 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3221 enum sched_type schedule, kmp_uint32 lb,
3222 kmp_uint32 ub, kmp_int32 st,
3223 kmp_int32 chunk);
3224 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3225 enum sched_type schedule, kmp_int64 lb,
3226 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3227 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3228 enum sched_type schedule, kmp_uint64 lb,
3229 kmp_uint64 ub, kmp_int64 st,
3230 kmp_int64 chunk);
3231
3232 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3233 kmp_int32 *p_last, kmp_int32 *p_lb,
3234 kmp_int32 *p_ub, kmp_int32 *p_st);
3235 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3236 kmp_int32 *p_last, kmp_uint32 *p_lb,
3237 kmp_uint32 *p_ub, kmp_int32 *p_st);
3238 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3239 kmp_int32 *p_last, kmp_int64 *p_lb,
3240 kmp_int64 *p_ub, kmp_int64 *p_st);
3241 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3242 kmp_int32 *p_last, kmp_uint64 *p_lb,
3243 kmp_uint64 *p_ub, kmp_int64 *p_st);
3244
3245 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3246 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3247 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3248 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3249
3250 #ifdef KMP_GOMP_COMPAT
3251
3252 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3253 enum sched_type schedule, kmp_int32 lb,
3254 kmp_int32 ub, kmp_int32 st,
3255 kmp_int32 chunk, int push_ws);
3256 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3257 enum sched_type schedule, kmp_uint32 lb,
3258 kmp_uint32 ub, kmp_int32 st,
3259 kmp_int32 chunk, int push_ws);
3260 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3261 enum sched_type schedule, kmp_int64 lb,
3262 kmp_int64 ub, kmp_int64 st,
3263 kmp_int64 chunk, int push_ws);
3264 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3265 enum sched_type schedule, kmp_uint64 lb,
3266 kmp_uint64 ub, kmp_int64 st,
3267 kmp_int64 chunk, int push_ws);
3268 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3269 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3270 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3271 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3272
3273 #endif /* KMP_GOMP_COMPAT */
3274
3275 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3276 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3277 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3278 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3279 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3280 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3281 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3282 void *obj);
3283 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3284 kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3285
3286 class kmp_flag_32;
3287 class kmp_flag_64;
3288 class kmp_flag_oncore;
3289 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64 *flag,
3290 int final_spin
3291 #if USE_ITT_BUILD
3292 ,
3293 void *itt_sync_obj
3294 #endif
3295 );
3296 extern void __kmp_release_64(kmp_flag_64 *flag);
3297
3298 extern void __kmp_infinite_loop(void);
3299
3300 extern void __kmp_cleanup(void);
3301
3302 #if KMP_HANDLE_SIGNALS
3303 extern int __kmp_handle_signals;
3304 extern void __kmp_install_signals(int parallel_init);
3305 extern void __kmp_remove_signals(void);
3306 #endif
3307
3308 extern void __kmp_clear_system_time(void);
3309 extern void __kmp_read_system_time(double *delta);
3310
3311 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3312
3313 extern void __kmp_expand_host_name(char *buffer, size_t size);
3314 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3315
3316 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3317 extern void
3318 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3319 #endif
3320
3321 extern void
3322 __kmp_runtime_initialize(void); /* machine specific initialization */
3323 extern void __kmp_runtime_destroy(void);
3324
3325 #if KMP_AFFINITY_SUPPORTED
3326 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3327 kmp_affin_mask_t *mask);
3328 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3329 kmp_affin_mask_t *mask);
3330 extern void __kmp_affinity_initialize(void);
3331 extern void __kmp_affinity_uninitialize(void);
3332 extern void __kmp_affinity_set_init_mask(
3333 int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3334 extern void __kmp_affinity_set_place(int gtid);
3335 extern void __kmp_affinity_determine_capable(const char *env_var);
3336 extern int __kmp_aux_set_affinity(void **mask);
3337 extern int __kmp_aux_get_affinity(void **mask);
3338 extern int __kmp_aux_get_affinity_max_proc();
3339 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3340 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3341 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3342 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3343 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3344 extern int kmp_set_thread_affinity_mask_initial(void);
3345 #endif
3346 #endif /* KMP_AFFINITY_SUPPORTED */
3347 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3348 // format string is for affinity, so platforms that do not support
3349 // affinity can still use the other fields, e.g., %n for num_threads
3350 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3351 kmp_str_buf_t *buffer);
3352 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3353
3354 extern void __kmp_cleanup_hierarchy();
3355 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3356
3357 #if KMP_USE_FUTEX
3358
3359 extern int __kmp_futex_determine_capable(void);
3360
3361 #endif // KMP_USE_FUTEX
3362
3363 extern void __kmp_gtid_set_specific(int gtid);
3364 extern int __kmp_gtid_get_specific(void);
3365
3366 extern double __kmp_read_cpu_time(void);
3367
3368 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3369
3370 #if KMP_USE_MONITOR
3371 extern void __kmp_create_monitor(kmp_info_t *th);
3372 #endif
3373
3374 extern void *__kmp_launch_thread(kmp_info_t *thr);
3375
3376 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3377
3378 #if KMP_OS_WINDOWS
3379 extern int __kmp_still_running(kmp_info_t *th);
3380 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3381 extern void __kmp_free_handle(kmp_thread_t tHandle);
3382 #endif
3383
3384 #if KMP_USE_MONITOR
3385 extern void __kmp_reap_monitor(kmp_info_t *th);
3386 #endif
3387 extern void __kmp_reap_worker(kmp_info_t *th);
3388 extern void __kmp_terminate_thread(int gtid);
3389
3390 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3391 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3392 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3393
3394 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag);
3395 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag);
3396 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
3397 extern void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag);
3398 extern void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag);
3399 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
3400
3401 extern void __kmp_elapsed(double *);
3402 extern void __kmp_elapsed_tick(double *);
3403
3404 extern void __kmp_enable(int old_state);
3405 extern void __kmp_disable(int *old_state);
3406
3407 extern void __kmp_thread_sleep(int millis);
3408
3409 extern void __kmp_common_initialize(void);
3410 extern void __kmp_common_destroy(void);
3411 extern void __kmp_common_destroy_gtid(int gtid);
3412
3413 #if KMP_OS_UNIX
3414 extern void __kmp_register_atfork(void);
3415 #endif
3416 extern void __kmp_suspend_initialize(void);
3417 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3418 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3419
3420 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3421 int tid);
3422 extern kmp_team_t *
3423 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3424 #if OMPT_SUPPORT
3425 ompt_data_t ompt_parallel_data,
3426 #endif
3427 kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3428 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3429 extern void __kmp_free_thread(kmp_info_t *);
3430 extern void __kmp_free_team(kmp_root_t *,
3431 kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3432 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3433
3434 /* ------------------------------------------------------------------------ */
3435
3436 extern void __kmp_initialize_bget(kmp_info_t *th);
3437 extern void __kmp_finalize_bget(kmp_info_t *th);
3438
3439 KMP_EXPORT void *kmpc_malloc(size_t size);
3440 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3441 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3442 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3443 KMP_EXPORT void kmpc_free(void *ptr);
3444
3445 /* declarations for internal use */
3446
3447 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3448 size_t reduce_size, void *reduce_data,
3449 void (*reduce)(void *, void *));
3450 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3451 extern int __kmp_barrier_gomp_cancel(int gtid);
3452
3453 /*!
3454 * Tell the fork call which compiler generated the fork call, and therefore how
3455 * to deal with the call.
3456 */
3457 enum fork_context_e {
3458 fork_context_gnu, /**< Called from GNU generated code, so must not invoke the
3459 microtask internally. */
3460 fork_context_intel, /**< Called from Intel generated code. */
3461 fork_context_last
3462 };
3463 extern int __kmp_fork_call(ident_t *loc, int gtid,
3464 enum fork_context_e fork_context, kmp_int32 argc,
3465 microtask_t microtask, launch_t invoker,
3466 kmp_va_list ap);
3467
3468 extern void __kmp_join_call(ident_t *loc, int gtid
3469 #if OMPT_SUPPORT
3470 ,
3471 enum fork_context_e fork_context
3472 #endif
3473 ,
3474 int exit_teams = 0);
3475
3476 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3477 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3478 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3479 extern int __kmp_invoke_task_func(int gtid);
3480 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3481 kmp_info_t *this_thr,
3482 kmp_team_t *team);
3483 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3484 kmp_info_t *this_thr,
3485 kmp_team_t *team);
3486
3487 // should never have been exported
3488 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3489 extern int __kmp_invoke_teams_master(int gtid);
3490 extern void __kmp_teams_master(int gtid);
3491 extern int __kmp_aux_get_team_num();
3492 extern int __kmp_aux_get_num_teams();
3493 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3494 extern void __kmp_user_set_library(enum library_type arg);
3495 extern void __kmp_aux_set_library(enum library_type arg);
3496 extern void __kmp_aux_set_stacksize(size_t arg);
3497 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3498 extern void __kmp_aux_set_defaults(char const *str, int len);
3499
3500 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3501 void kmpc_set_blocktime(int arg);
3502 void ompc_set_nested(int flag);
3503 void ompc_set_dynamic(int flag);
3504 void ompc_set_num_threads(int arg);
3505
3506 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3507 kmp_team_t *team, int tid);
3508 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3509 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3510 kmp_tasking_flags_t *flags,
3511 size_t sizeof_kmp_task_t,
3512 size_t sizeof_shareds,
3513 kmp_routine_entry_t task_entry);
3514 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3515 kmp_team_t *team, int tid,
3516 int set_curr_task);
3517 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3518 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3519
3520 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3521 int gtid,
3522 kmp_task_t *task);
3523 extern void __kmp_fulfill_event(kmp_event_t *event);
3524
3525 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
3526 kmp_flag_32 *flag, int final_spin,
3527 int *thread_finished,
3528 #if USE_ITT_BUILD
3529 void *itt_sync_obj,
3530 #endif /* USE_ITT_BUILD */
3531 kmp_int32 is_constrained);
3532 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
3533 kmp_flag_64 *flag, int final_spin,
3534 int *thread_finished,
3535 #if USE_ITT_BUILD
3536 void *itt_sync_obj,
3537 #endif /* USE_ITT_BUILD */
3538 kmp_int32 is_constrained);
3539 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
3540 kmp_flag_oncore *flag, int final_spin,
3541 int *thread_finished,
3542 #if USE_ITT_BUILD
3543 void *itt_sync_obj,
3544 #endif /* USE_ITT_BUILD */
3545 kmp_int32 is_constrained);
3546
3547 extern void __kmp_free_task_team(kmp_info_t *thread,
3548 kmp_task_team_t *task_team);
3549 extern void __kmp_reap_task_teams(void);
3550 extern void __kmp_wait_to_unref_task_teams(void);
3551 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3552 int always);
3553 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3554 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3555 #if USE_ITT_BUILD
3556 ,
3557 void *itt_sync_obj
3558 #endif /* USE_ITT_BUILD */
3559 ,
3560 int wait = 1);
3561 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3562 int gtid);
3563
3564 extern int __kmp_is_address_mapped(void *addr);
3565 extern kmp_uint64 __kmp_hardware_timestamp(void);
3566
3567 #if KMP_OS_UNIX
3568 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3569 #endif
3570
3571 /* ------------------------------------------------------------------------ */
3572 //
3573 // Assembly routines that have no compiler intrinsic replacement
3574 //
3575
3576 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3577 void *argv[]
3578 #if OMPT_SUPPORT
3579 ,
3580 void **exit_frame_ptr
3581 #endif
3582 );
3583
3584 /* ------------------------------------------------------------------------ */
3585
3586 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3587 KMP_EXPORT void __kmpc_end(ident_t *);
3588
3589 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3590 kmpc_ctor_vec ctor,
3591 kmpc_cctor_vec cctor,
3592 kmpc_dtor_vec dtor,
3593 size_t vector_length);
3594 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3595 kmpc_ctor ctor, kmpc_cctor cctor,
3596 kmpc_dtor dtor);
3597 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3598 void *data, size_t size);
3599
3600 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3601 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3602 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3603 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3604
3605 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3606 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3607 kmpc_micro microtask, ...);
3608
3609 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3610 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3611
3612 KMP_EXPORT void __kmpc_flush(ident_t *);
3613 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3614 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3615 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3616 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3617 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3618 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3619 kmp_critical_name *);
3620 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3621 kmp_critical_name *);
3622 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3623 kmp_critical_name *, uint32_t hint);
3624
3625 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3626 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3627
3628 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3629 kmp_int32 global_tid);
3630
3631 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3632 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3633
3634 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3635 kmp_int32 schedtype, kmp_int32 *plastiter,
3636 kmp_int *plower, kmp_int *pupper,
3637 kmp_int *pstride, kmp_int incr,
3638 kmp_int chunk);
3639
3640 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3641
3642 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3643 size_t cpy_size, void *cpy_data,
3644 void (*cpy_func)(void *, void *),
3645 kmp_int32 didit);
3646
3647 extern void KMPC_SET_NUM_THREADS(int arg);
3648 extern void KMPC_SET_DYNAMIC(int flag);
3649 extern void KMPC_SET_NESTED(int flag);
3650
3651 /* OMP 3.0 tasking interface routines */
3652 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3653 kmp_task_t *new_task);
3654 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3655 kmp_int32 flags,
3656 size_t sizeof_kmp_task_t,
3657 size_t sizeof_shareds,
3658 kmp_routine_entry_t task_entry);
3659 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3660 kmp_int32 flags,
3661 size_t sizeof_kmp_task_t,
3662 size_t sizeof_shareds,
3663 kmp_routine_entry_t task_entry,
3664 kmp_int64 device_id);
3665 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3666 kmp_task_t *task);
3667 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3668 kmp_task_t *task);
3669 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3670 kmp_task_t *new_task);
3671 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3672
3673 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3674 int end_part);
3675
3676 #if TASK_UNUSED
3677 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3678 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3679 kmp_task_t *task);
3680 #endif // TASK_UNUSED
3681
3682 /* ------------------------------------------------------------------------ */
3683
3684 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3685 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3686
3687 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3688 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3689 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3690 kmp_depend_info_t *noalias_dep_list);
3691 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3692 kmp_int32 ndeps,
3693 kmp_depend_info_t *dep_list,
3694 kmp_int32 ndeps_noalias,
3695 kmp_depend_info_t *noalias_dep_list);
3696 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3697 bool serialize_immediate);
3698
3699 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3700 kmp_int32 cncl_kind);
3701 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3702 kmp_int32 cncl_kind);
3703 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3704 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3705
3706 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3707 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3708 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3709 kmp_int32 if_val, kmp_uint64 *lb,
3710 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3711 kmp_int32 sched, kmp_uint64 grainsize,
3712 void *task_dup);
3713 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3714 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
3715 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3716 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
3717 int is_ws, int num,
3718 void *data);
3719 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
3720 int num, void *data);
3721 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
3722 int is_ws);
3723 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
3724 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3725 kmp_task_affinity_info_t *affin_list);
3726
3727 /* Lock interface routines (fast versions with gtid passed in) */
3728 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3729 void **user_lock);
3730 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3731 void **user_lock);
3732 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3733 void **user_lock);
3734 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3735 void **user_lock);
3736 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3737 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3738 void **user_lock);
3739 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3740 void **user_lock);
3741 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3742 void **user_lock);
3743 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3744 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3745 void **user_lock);
3746
3747 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3748 void **user_lock, uintptr_t hint);
3749 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3750 void **user_lock,
3751 uintptr_t hint);
3752
3753 /* Interface to fast scalable reduce methods routines */
3754
3755 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3756 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3757 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3758 kmp_critical_name *lck);
3759 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3760 kmp_critical_name *lck);
3761 KMP_EXPORT kmp_int32 __kmpc_reduce(
3762 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3763 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3764 kmp_critical_name *lck);
3765 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3766 kmp_critical_name *lck);
3767
3768 /* Internal fast reduction routines */
3769
3770 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3771 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3772 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3773 kmp_critical_name *lck);
3774
3775 // this function is for testing set/get/determine reduce method
3776 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3777
3778 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3779 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3780
3781 // C++ port
3782 // missing 'extern "C"' declarations
3783
3784 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3785 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3786 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3787 kmp_int32 num_threads);
3788
3789 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3790 int proc_bind);
3791 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3792 kmp_int32 num_teams,
3793 kmp_int32 num_threads);
3794 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3795 kmpc_micro microtask, ...);
3796 struct kmp_dim { // loop bounds info casted to kmp_int64
3797 kmp_int64 lo; // lower
3798 kmp_int64 up; // upper
3799 kmp_int64 st; // stride
3800 };
3801 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
3802 kmp_int32 num_dims,
3803 const struct kmp_dim *dims);
3804 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
3805 const kmp_int64 *vec);
3806 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
3807 const kmp_int64 *vec);
3808 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
3809
3810 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
3811 void *data, size_t size,
3812 void ***cache);
3813
3814 // Symbols for MS mutual detection.
3815 extern int _You_must_link_with_exactly_one_OpenMP_library;
3816 extern int _You_must_link_with_Intel_OpenMP_library;
3817 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3818 extern int _You_must_link_with_Microsoft_OpenMP_library;
3819 #endif
3820
3821 // The routines below are not exported.
3822 // Consider making them 'static' in corresponding source files.
3823 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
3824 void *data_addr, size_t pc_size);
3825 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
3826 void *data_addr,
3827 size_t pc_size);
3828 void __kmp_threadprivate_resize_cache(int newCapacity);
3829 void __kmp_cleanup_threadprivate_caches();
3830
3831 // ompc_, kmpc_ entries moved from omp.h.
3832 #if KMP_OS_WINDOWS
3833 #define KMPC_CONVENTION __cdecl
3834 #else
3835 #define KMPC_CONVENTION
3836 #endif
3837
3838 #ifndef __OMP_H
3839 typedef enum omp_sched_t {
3840 omp_sched_static = 1,
3841 omp_sched_dynamic = 2,
3842 omp_sched_guided = 3,
3843 omp_sched_auto = 4
3844 } omp_sched_t;
3845 typedef void *kmp_affinity_mask_t;
3846 #endif
3847
3848 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
3849 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
3850 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
3851 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
3852 KMP_EXPORT int KMPC_CONVENTION
3853 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
3854 KMP_EXPORT int KMPC_CONVENTION
3855 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
3856 KMP_EXPORT int KMPC_CONVENTION
3857 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
3858
3859 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
3860 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
3861 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
3862 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
3863 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
3864
3865 enum kmp_target_offload_kind {
3866 tgt_disabled = 0,
3867 tgt_default = 1,
3868 tgt_mandatory = 2
3869 };
3870 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
3871 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
3872 extern kmp_target_offload_kind_t __kmp_target_offload;
3873 extern int __kmpc_get_target_offload();
3874
3875 // Constants used in libomptarget
3876 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
3877 #define KMP_HOST_DEVICE -10 // This is what it is in libomptarget, go figure.
3878 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
3879
3880 // OMP Pause Resource
3881
3882 // The following enum is used both to set the status in __kmp_pause_status, and
3883 // as the internal equivalent of the externally-visible omp_pause_resource_t.
3884 typedef enum kmp_pause_status_t {
3885 kmp_not_paused = 0, // status is not paused, or, requesting resume
3886 kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
3887 kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
3888 } kmp_pause_status_t;
3889
3890 // This stores the pause state of the runtime
3891 extern kmp_pause_status_t __kmp_pause_status;
3892 extern int __kmpc_pause_resource(kmp_pause_status_t level);
3893 extern int __kmp_pause_resource(kmp_pause_status_t level);
3894 // Soft resume sets __kmp_pause_status, and wakes up all threads.
3895 extern void __kmp_resume_if_soft_paused();
3896 // Hard resume simply resets the status to not paused. Library will appear to
3897 // be uninitialized after hard pause. Let OMP constructs trigger required
3898 // initializations.
__kmp_resume_if_hard_paused()3899 static inline void __kmp_resume_if_hard_paused() {
3900 if (__kmp_pause_status == kmp_hard_paused) {
3901 __kmp_pause_status = kmp_not_paused;
3902 }
3903 }
3904
3905 #ifdef __cplusplus
3906 }
3907 #endif
3908
3909 #endif /* KMP_H */
3910