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 <limits>
70 #include <type_traits>
71 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
72 Microsoft library. Some macros provided below to replace these functions */
73 #ifndef __ABSOFT_WIN
74 #include <sys/types.h>
75 #endif
76 #include <limits.h>
77 #include <time.h>
78
79 #include <errno.h>
80
81 #include "kmp_os.h"
82
83 #include "kmp_safe_c_api.h"
84
85 #if KMP_STATS_ENABLED
86 class kmp_stats_list;
87 #endif
88
89 #if KMP_USE_HIER_SCHED
90 // Only include hierarchical scheduling if affinity is supported
91 #undef KMP_USE_HIER_SCHED
92 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
93 #endif
94
95 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
96 #include "hwloc.h"
97 #ifndef HWLOC_OBJ_NUMANODE
98 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
99 #endif
100 #ifndef HWLOC_OBJ_PACKAGE
101 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
102 #endif
103 #if HWLOC_API_VERSION >= 0x00020000
104 // hwloc 2.0 changed type of depth of object from unsigned to int
105 typedef int kmp_hwloc_depth_t;
106 #else
107 typedef unsigned int kmp_hwloc_depth_t;
108 #endif
109 #endif
110
111 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
112 #include <xmmintrin.h>
113 #endif
114
115 #include "kmp_debug.h"
116 #include "kmp_lock.h"
117 #include "kmp_version.h"
118 #include "kmp_barrier.h"
119 #if USE_DEBUGGER
120 #include "kmp_debugger.h"
121 #endif
122 #include "kmp_i18n.h"
123
124 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
125
126 #include "kmp_wrapper_malloc.h"
127 #if KMP_OS_UNIX
128 #include <unistd.h>
129 #if !defined NSIG && defined _NSIG
130 #define NSIG _NSIG
131 #endif
132 #endif
133
134 #if KMP_OS_LINUX
135 #pragma weak clock_gettime
136 #endif
137
138 #if OMPT_SUPPORT
139 #include "ompt-internal.h"
140 #endif
141
142 #if OMPD_SUPPORT
143 #include "ompd-specific.h"
144 #endif
145
146 #ifndef UNLIKELY
147 #define UNLIKELY(x) (x)
148 #endif
149
150 // Affinity format function
151 #include "kmp_str.h"
152
153 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
154 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
155 // free lists of limited size.
156 #ifndef USE_FAST_MEMORY
157 #define USE_FAST_MEMORY 3
158 #endif
159
160 #ifndef KMP_NESTED_HOT_TEAMS
161 #define KMP_NESTED_HOT_TEAMS 0
162 #define USE_NESTED_HOT_ARG(x)
163 #else
164 #if KMP_NESTED_HOT_TEAMS
165 #define USE_NESTED_HOT_ARG(x) , x
166 #else
167 #define USE_NESTED_HOT_ARG(x)
168 #endif
169 #endif
170
171 // Assume using BGET compare_exchange instruction instead of lock by default.
172 #ifndef USE_CMP_XCHG_FOR_BGET
173 #define USE_CMP_XCHG_FOR_BGET 1
174 #endif
175
176 // Test to see if queuing lock is better than bootstrap lock for bget
177 // #ifndef USE_QUEUING_LOCK_FOR_BGET
178 // #define USE_QUEUING_LOCK_FOR_BGET
179 // #endif
180
181 #define KMP_NSEC_PER_SEC 1000000000L
182 #define KMP_USEC_PER_SEC 1000000L
183
184 /*!
185 @ingroup BASIC_TYPES
186 @{
187 */
188
189 /*!
190 Values for bit flags used in the ident_t to describe the fields.
191 */
192 enum {
193 /*! Use trampoline for internal microtasks */
194 KMP_IDENT_IMB = 0x01,
195 /*! Use c-style ident structure */
196 KMP_IDENT_KMPC = 0x02,
197 /* 0x04 is no longer used */
198 /*! Entry point generated by auto-parallelization */
199 KMP_IDENT_AUTOPAR = 0x08,
200 /*! Compiler generates atomic reduction option for kmpc_reduce* */
201 KMP_IDENT_ATOMIC_REDUCE = 0x10,
202 /*! To mark a 'barrier' directive in user code */
203 KMP_IDENT_BARRIER_EXPL = 0x20,
204 /*! To Mark implicit barriers. */
205 KMP_IDENT_BARRIER_IMPL = 0x0040,
206 KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
207 KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
208 KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
209
210 KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
211 KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
212
213 /*! To mark a static loop in OMPT callbacks */
214 KMP_IDENT_WORK_LOOP = 0x200,
215 /*! To mark a sections directive in OMPT callbacks */
216 KMP_IDENT_WORK_SECTIONS = 0x400,
217 /*! To mark a distribute construct in OMPT callbacks */
218 KMP_IDENT_WORK_DISTRIBUTE = 0x800,
219 /*! Atomic hint; bottom four bits as omp_sync_hint_t. Top four reserved and
220 not currently used. If one day we need more bits, then we can use
221 an invalid combination of hints to mean that another, larger field
222 should be used in a different flag. */
223 KMP_IDENT_ATOMIC_HINT_MASK = 0xFF0000,
224 KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
225 KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
226 KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
227 KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
228 KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
229 };
230
231 /*!
232 * The ident structure that describes a source location.
233 */
234 typedef struct ident {
235 kmp_int32 reserved_1; /**< might be used in Fortran; see above */
236 kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC
237 identifies this union member */
238 kmp_int32 reserved_2; /**< not really used in Fortran any more; see above */
239 #if USE_ITT_BUILD
240 /* but currently used for storing region-specific ITT */
241 /* contextual information. */
242 #endif /* USE_ITT_BUILD */
243 kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++ */
244 char const *psource; /**< String describing the source location.
245 The string is composed of semi-colon separated fields
246 which describe the source file, the function and a pair
247 of line numbers that delimit the construct. */
248 // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
get_openmp_versionident249 kmp_int32 get_openmp_version() {
250 return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
251 }
252 } ident_t;
253 /*!
254 @}
255 */
256
257 // Some forward declarations.
258 typedef union kmp_team kmp_team_t;
259 typedef struct kmp_taskdata kmp_taskdata_t;
260 typedef union kmp_task_team kmp_task_team_t;
261 typedef union kmp_team kmp_team_p;
262 typedef union kmp_info kmp_info_p;
263 typedef union kmp_root kmp_root_p;
264
265 template <bool C = false, bool S = true> class kmp_flag_32;
266 template <bool C = false, bool S = true> class kmp_flag_64;
267 template <bool C = false, bool S = true> class kmp_atomic_flag_64;
268 class kmp_flag_oncore;
269
270 #ifdef __cplusplus
271 extern "C" {
272 #endif
273
274 /* ------------------------------------------------------------------------ */
275
276 /* Pack two 32-bit signed integers into a 64-bit signed integer */
277 /* ToDo: Fix word ordering for big-endian machines. */
278 #define KMP_PACK_64(HIGH_32, LOW_32) \
279 ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
280
281 // Generic string manipulation macros. Assume that _x is of type char *
282 #define SKIP_WS(_x) \
283 { \
284 while (*(_x) == ' ' || *(_x) == '\t') \
285 (_x)++; \
286 }
287 #define SKIP_DIGITS(_x) \
288 { \
289 while (*(_x) >= '0' && *(_x) <= '9') \
290 (_x)++; \
291 }
292 #define SKIP_TOKEN(_x) \
293 { \
294 while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
295 (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
296 (_x)++; \
297 }
298 #define SKIP_TO(_x, _c) \
299 { \
300 while (*(_x) != '\0' && *(_x) != (_c)) \
301 (_x)++; \
302 }
303
304 /* ------------------------------------------------------------------------ */
305
306 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
307 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
308
309 /* ------------------------------------------------------------------------ */
310 /* Enumeration types */
311
312 enum kmp_state_timer {
313 ts_stop,
314 ts_start,
315 ts_pause,
316
317 ts_last_state
318 };
319
320 enum dynamic_mode {
321 dynamic_default,
322 #ifdef USE_LOAD_BALANCE
323 dynamic_load_balance,
324 #endif /* USE_LOAD_BALANCE */
325 dynamic_random,
326 dynamic_thread_limit,
327 dynamic_max
328 };
329
330 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
331 * not include it here */
332 #ifndef KMP_SCHED_TYPE_DEFINED
333 #define KMP_SCHED_TYPE_DEFINED
334 typedef enum kmp_sched {
335 kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
336 // Note: need to adjust __kmp_sch_map global array in case enum is changed
337 kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
338 kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
339 kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
340 kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
341 kmp_sched_upper_std = 5, // upper bound for standard schedules
342 kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
343 kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
344 #if KMP_STATIC_STEAL_ENABLED
345 kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
346 #endif
347 kmp_sched_upper,
348 kmp_sched_default = kmp_sched_static, // default scheduling
349 kmp_sched_monotonic = 0x80000000
350 } kmp_sched_t;
351 #endif
352
353 /*!
354 @ingroup WORK_SHARING
355 * Describes the loop schedule to be used for a parallel for loop.
356 */
357 enum sched_type : kmp_int32 {
358 kmp_sch_lower = 32, /**< lower bound for unordered values */
359 kmp_sch_static_chunked = 33,
360 kmp_sch_static = 34, /**< static unspecialized */
361 kmp_sch_dynamic_chunked = 35,
362 kmp_sch_guided_chunked = 36, /**< guided unspecialized */
363 kmp_sch_runtime = 37,
364 kmp_sch_auto = 38, /**< auto */
365 kmp_sch_trapezoidal = 39,
366
367 /* accessible only through KMP_SCHEDULE environment variable */
368 kmp_sch_static_greedy = 40,
369 kmp_sch_static_balanced = 41,
370 /* accessible only through KMP_SCHEDULE environment variable */
371 kmp_sch_guided_iterative_chunked = 42,
372 kmp_sch_guided_analytical_chunked = 43,
373 /* accessible only through KMP_SCHEDULE environment variable */
374 kmp_sch_static_steal = 44,
375
376 /* static with chunk adjustment (e.g., simd) */
377 kmp_sch_static_balanced_chunked = 45,
378 kmp_sch_guided_simd = 46, /**< guided with chunk adjustment */
379 kmp_sch_runtime_simd = 47, /**< runtime with chunk adjustment */
380
381 /* accessible only through KMP_SCHEDULE environment variable */
382 kmp_sch_upper, /**< upper bound for unordered values */
383
384 kmp_ord_lower = 64, /**< lower bound for ordered values, must be power of 2 */
385 kmp_ord_static_chunked = 65,
386 kmp_ord_static = 66, /**< ordered static unspecialized */
387 kmp_ord_dynamic_chunked = 67,
388 kmp_ord_guided_chunked = 68,
389 kmp_ord_runtime = 69,
390 kmp_ord_auto = 70, /**< ordered auto */
391 kmp_ord_trapezoidal = 71,
392 kmp_ord_upper, /**< upper bound for ordered values */
393
394 /* Schedules for Distribute construct */
395 kmp_distribute_static_chunked = 91, /**< distribute static chunked */
396 kmp_distribute_static = 92, /**< distribute static unspecialized */
397
398 /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
399 single iteration/chunk, even if the loop is serialized. For the schedule
400 types listed above, the entire iteration vector is returned if the loop is
401 serialized. This doesn't work for gcc/gcomp sections. */
402 kmp_nm_lower = 160, /**< lower bound for nomerge values */
403
404 kmp_nm_static_chunked =
405 (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
406 kmp_nm_static = 162, /**< static unspecialized */
407 kmp_nm_dynamic_chunked = 163,
408 kmp_nm_guided_chunked = 164, /**< guided unspecialized */
409 kmp_nm_runtime = 165,
410 kmp_nm_auto = 166, /**< auto */
411 kmp_nm_trapezoidal = 167,
412
413 /* accessible only through KMP_SCHEDULE environment variable */
414 kmp_nm_static_greedy = 168,
415 kmp_nm_static_balanced = 169,
416 /* accessible only through KMP_SCHEDULE environment variable */
417 kmp_nm_guided_iterative_chunked = 170,
418 kmp_nm_guided_analytical_chunked = 171,
419 kmp_nm_static_steal =
420 172, /* accessible only through OMP_SCHEDULE environment variable */
421
422 kmp_nm_ord_static_chunked = 193,
423 kmp_nm_ord_static = 194, /**< ordered static unspecialized */
424 kmp_nm_ord_dynamic_chunked = 195,
425 kmp_nm_ord_guided_chunked = 196,
426 kmp_nm_ord_runtime = 197,
427 kmp_nm_ord_auto = 198, /**< auto */
428 kmp_nm_ord_trapezoidal = 199,
429 kmp_nm_upper, /**< upper bound for nomerge values */
430
431 /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
432 we need to distinguish the three possible cases (no modifier, monotonic
433 modifier, nonmonotonic modifier), we need separate bits for each modifier.
434 The absence of monotonic does not imply nonmonotonic, especially since 4.5
435 says that the behaviour of the "no modifier" case is implementation defined
436 in 4.5, but will become "nonmonotonic" in 5.0.
437
438 Since we're passing a full 32 bit value, we can use a couple of high bits
439 for these flags; out of paranoia we avoid the sign bit.
440
441 These modifiers can be or-ed into non-static schedules by the compiler to
442 pass the additional information. They will be stripped early in the
443 processing in __kmp_dispatch_init when setting up schedules, so most of the
444 code won't ever see schedules with these bits set. */
445 kmp_sch_modifier_monotonic =
446 (1 << 29), /**< Set if the monotonic schedule modifier was present */
447 kmp_sch_modifier_nonmonotonic =
448 (1 << 30), /**< Set if the nonmonotonic schedule modifier was present */
449
450 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
451 (enum sched_type)( \
452 (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))
453 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
454 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
455 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
456 (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
457 #define SCHEDULE_GET_MODIFIERS(s) \
458 ((enum sched_type)( \
459 (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
460 #define SCHEDULE_SET_MODIFIERS(s, m) \
461 (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
462 #define SCHEDULE_NONMONOTONIC 0
463 #define SCHEDULE_MONOTONIC 1
464
465 kmp_sch_default = kmp_sch_static /**< default scheduling algorithm */
466 };
467
468 // Apply modifiers on internal kind to standard kind
469 static inline void
__kmp_sched_apply_mods_stdkind(kmp_sched_t * kind,enum sched_type internal_kind)470 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
471 enum sched_type internal_kind) {
472 if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
473 *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
474 }
475 }
476
477 // Apply modifiers on standard kind to internal kind
478 static inline void
__kmp_sched_apply_mods_intkind(kmp_sched_t kind,enum sched_type * internal_kind)479 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
480 enum sched_type *internal_kind) {
481 if ((int)kind & (int)kmp_sched_monotonic) {
482 *internal_kind = (enum sched_type)((int)*internal_kind |
483 (int)kmp_sch_modifier_monotonic);
484 }
485 }
486
487 // Get standard schedule without modifiers
__kmp_sched_without_mods(kmp_sched_t kind)488 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
489 return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
490 }
491
492 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
493 typedef union kmp_r_sched {
494 struct {
495 enum sched_type r_sched_type;
496 int chunk;
497 };
498 kmp_int64 sched;
499 } kmp_r_sched_t;
500
501 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
502 // internal schedule types
503
504 enum library_type {
505 library_none,
506 library_serial,
507 library_turnaround,
508 library_throughput
509 };
510
511 #if KMP_OS_LINUX
512 enum clock_function_type {
513 clock_function_gettimeofday,
514 clock_function_clock_gettime
515 };
516 #endif /* KMP_OS_LINUX */
517
518 #if KMP_MIC_SUPPORTED
519 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
520 #endif
521
522 /* -- fast reduction stuff ------------------------------------------------ */
523
524 #undef KMP_FAST_REDUCTION_BARRIER
525 #define KMP_FAST_REDUCTION_BARRIER 1
526
527 #undef KMP_FAST_REDUCTION_CORE_DUO
528 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
529 #define KMP_FAST_REDUCTION_CORE_DUO 1
530 #endif
531
532 enum _reduction_method {
533 reduction_method_not_defined = 0,
534 critical_reduce_block = (1 << 8),
535 atomic_reduce_block = (2 << 8),
536 tree_reduce_block = (3 << 8),
537 empty_reduce_block = (4 << 8)
538 };
539
540 // Description of the packed_reduction_method variable:
541 // The packed_reduction_method variable consists of two enum types variables
542 // that are packed together into 0-th byte and 1-st byte:
543 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
544 // barrier that will be used in fast reduction: bs_plain_barrier or
545 // bs_reduction_barrier
546 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
547 // be used in fast reduction;
548 // Reduction method is of 'enum _reduction_method' type and it's defined the way
549 // so that the bits of 0-th byte are empty, so no need to execute a shift
550 // instruction while packing/unpacking
551
552 #if KMP_FAST_REDUCTION_BARRIER
553 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
554 ((reduction_method) | (barrier_type))
555
556 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
557 ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
558
559 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
560 ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
561 #else
562 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
563 (reduction_method)
564
565 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
566 (packed_reduction_method)
567
568 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
569 #endif
570
571 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
572 ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
573 (which_reduction_block))
574
575 #if KMP_FAST_REDUCTION_BARRIER
576 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
577 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
578
579 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
580 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
581 #endif
582
583 typedef int PACKED_REDUCTION_METHOD_T;
584
585 /* -- end of fast reduction stuff ----------------------------------------- */
586
587 #if KMP_OS_WINDOWS
588 #define USE_CBLKDATA
589 #if KMP_MSVC_COMPAT
590 #pragma warning(push)
591 #pragma warning(disable : 271 310)
592 #endif
593 #include <windows.h>
594 #if KMP_MSVC_COMPAT
595 #pragma warning(pop)
596 #endif
597 #endif
598
599 #if KMP_OS_UNIX
600 #include <dlfcn.h>
601 #include <pthread.h>
602 #endif
603
604 enum kmp_hw_t : int {
605 KMP_HW_UNKNOWN = -1,
606 KMP_HW_SOCKET = 0,
607 KMP_HW_PROC_GROUP,
608 KMP_HW_NUMA,
609 KMP_HW_DIE,
610 KMP_HW_LLC,
611 KMP_HW_L3,
612 KMP_HW_TILE,
613 KMP_HW_MODULE,
614 KMP_HW_L2,
615 KMP_HW_L1,
616 KMP_HW_CORE,
617 KMP_HW_THREAD,
618 KMP_HW_LAST
619 };
620
621 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \
622 KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
623 #define KMP_ASSERT_VALID_HW_TYPE(type) \
624 KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
625
626 #define KMP_FOREACH_HW_TYPE(type) \
627 for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \
628 type = (kmp_hw_t)((int)type + 1))
629
630 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
631 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
632
633 /* Only Linux* OS and Windows* OS support thread affinity. */
634 #if KMP_AFFINITY_SUPPORTED
635
636 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
637 #if KMP_OS_WINDOWS
638 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
639 typedef struct GROUP_AFFINITY {
640 KAFFINITY Mask;
641 WORD Group;
642 WORD Reserved[3];
643 } GROUP_AFFINITY;
644 #endif /* _MSC_VER < 1600 */
645 #if KMP_GROUP_AFFINITY
646 extern int __kmp_num_proc_groups;
647 #else
648 static const int __kmp_num_proc_groups = 1;
649 #endif /* KMP_GROUP_AFFINITY */
650 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
651 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
652
653 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
654 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
655
656 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
657 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
658
659 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
660 GROUP_AFFINITY *);
661 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
662 #endif /* KMP_OS_WINDOWS */
663
664 #if KMP_USE_HWLOC
665 extern hwloc_topology_t __kmp_hwloc_topology;
666 extern int __kmp_hwloc_error;
667 #endif
668
669 extern size_t __kmp_affin_mask_size;
670 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
671 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
672 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
673 #define KMP_CPU_SET_ITERATE(i, mask) \
674 for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
675 #define KMP_CPU_SET(i, mask) (mask)->set(i)
676 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
677 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
678 #define KMP_CPU_ZERO(mask) (mask)->zero()
679 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
680 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
681 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
682 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
683 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
684 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
685 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
686 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
687 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
688 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
689 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
690 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
691 (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
692 #define KMP_CPU_FREE_ARRAY(arr, n) \
693 __kmp_affinity_dispatch->deallocate_mask_array(arr)
694 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
695 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
696 #define __kmp_get_system_affinity(mask, abort_bool) \
697 (mask)->get_system_affinity(abort_bool)
698 #define __kmp_set_system_affinity(mask, abort_bool) \
699 (mask)->set_system_affinity(abort_bool)
700 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
701
702 class KMPAffinity {
703 public:
704 class Mask {
705 public:
706 void *operator new(size_t n);
707 void operator delete(void *p);
708 void *operator new[](size_t n);
709 void operator delete[](void *p);
~Mask()710 virtual ~Mask() {}
711 // Set bit i to 1
set(int i)712 virtual void set(int i) {}
713 // Return bit i
is_set(int i)714 virtual bool is_set(int i) const { return false; }
715 // Set bit i to 0
clear(int i)716 virtual void clear(int i) {}
717 // Zero out entire mask
zero()718 virtual void zero() {}
719 // Copy src into this mask
copy(const Mask * src)720 virtual void copy(const Mask *src) {}
721 // this &= rhs
bitwise_and(const Mask * rhs)722 virtual void bitwise_and(const Mask *rhs) {}
723 // this |= rhs
bitwise_or(const Mask * rhs)724 virtual void bitwise_or(const Mask *rhs) {}
725 // this = ~this
bitwise_not()726 virtual void bitwise_not() {}
727 // API for iterating over an affinity mask
728 // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
begin()729 virtual int begin() const { return 0; }
end()730 virtual int end() const { return 0; }
next(int previous)731 virtual int next(int previous) const { return 0; }
732 #if KMP_OS_WINDOWS
set_process_affinity(bool abort_on_error)733 virtual int set_process_affinity(bool abort_on_error) const { return -1; }
734 #endif
735 // Set the system's affinity to this affinity mask's value
set_system_affinity(bool abort_on_error)736 virtual int set_system_affinity(bool abort_on_error) const { return -1; }
737 // Set this affinity mask to the current system affinity
get_system_affinity(bool abort_on_error)738 virtual int get_system_affinity(bool abort_on_error) { return -1; }
739 // Only 1 DWORD in the mask should have any procs set.
740 // Return the appropriate index, or -1 for an invalid mask.
get_proc_group()741 virtual int get_proc_group() const { return -1; }
742 };
743 void *operator new(size_t n);
744 void operator delete(void *p);
745 // Need virtual destructor
746 virtual ~KMPAffinity() = default;
747 // Determine if affinity is capable
determine_capable(const char * env_var)748 virtual void determine_capable(const char *env_var) {}
749 // Bind the current thread to os proc
bind_thread(int proc)750 virtual void bind_thread(int proc) {}
751 // Factory functions to allocate/deallocate a mask
allocate_mask()752 virtual Mask *allocate_mask() { return nullptr; }
deallocate_mask(Mask * m)753 virtual void deallocate_mask(Mask *m) {}
allocate_mask_array(int num)754 virtual Mask *allocate_mask_array(int num) { return nullptr; }
deallocate_mask_array(Mask * m)755 virtual void deallocate_mask_array(Mask *m) {}
index_mask_array(Mask * m,int index)756 virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
757 static void pick_api();
758 static void destroy_api();
759 enum api_type {
760 NATIVE_OS
761 #if KMP_USE_HWLOC
762 ,
763 HWLOC
764 #endif
765 };
get_api_type()766 virtual api_type get_api_type() const {
767 KMP_ASSERT(0);
768 return NATIVE_OS;
769 }
770
771 private:
772 static bool picked_api;
773 };
774
775 typedef KMPAffinity::Mask kmp_affin_mask_t;
776 extern KMPAffinity *__kmp_affinity_dispatch;
777
778 // Declare local char buffers with this size for printing debug and info
779 // messages, using __kmp_affinity_print_mask().
780 #define KMP_AFFIN_MASK_PRINT_LEN 1024
781
782 enum affinity_type {
783 affinity_none = 0,
784 affinity_physical,
785 affinity_logical,
786 affinity_compact,
787 affinity_scatter,
788 affinity_explicit,
789 affinity_balanced,
790 affinity_disabled, // not used outsize the env var parser
791 affinity_default
792 };
793
794 enum affinity_top_method {
795 affinity_top_method_all = 0, // try all (supported) methods, in order
796 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
797 affinity_top_method_apicid,
798 affinity_top_method_x2apicid,
799 affinity_top_method_x2apicid_1f,
800 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
801 affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
802 #if KMP_GROUP_AFFINITY
803 affinity_top_method_group,
804 #endif /* KMP_GROUP_AFFINITY */
805 affinity_top_method_flat,
806 #if KMP_USE_HWLOC
807 affinity_top_method_hwloc,
808 #endif
809 affinity_top_method_default
810 };
811
812 #define affinity_respect_mask_default (-1)
813
814 extern enum affinity_type __kmp_affinity_type; /* Affinity type */
815 extern kmp_hw_t __kmp_affinity_gran; /* Affinity granularity */
816 extern int __kmp_affinity_gran_levels; /* corresponding int value */
817 extern int __kmp_affinity_dups; /* Affinity duplicate masks */
818 extern enum affinity_top_method __kmp_affinity_top_method;
819 extern int __kmp_affinity_compact; /* Affinity 'compact' value */
820 extern int __kmp_affinity_offset; /* Affinity offset value */
821 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */
822 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */
823 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask?
824 extern char *__kmp_affinity_proclist; /* proc ID list */
825 extern kmp_affin_mask_t *__kmp_affinity_masks;
826 extern unsigned __kmp_affinity_num_masks;
827 extern void __kmp_affinity_bind_thread(int which);
828
829 extern kmp_affin_mask_t *__kmp_affin_fullMask;
830 extern char *__kmp_cpuinfo_file;
831
832 #endif /* KMP_AFFINITY_SUPPORTED */
833
834 // This needs to be kept in sync with the values in omp.h !!!
835 typedef enum kmp_proc_bind_t {
836 proc_bind_false = 0,
837 proc_bind_true,
838 proc_bind_primary,
839 proc_bind_close,
840 proc_bind_spread,
841 proc_bind_intel, // use KMP_AFFINITY interface
842 proc_bind_default
843 } kmp_proc_bind_t;
844
845 typedef struct kmp_nested_proc_bind_t {
846 kmp_proc_bind_t *bind_types;
847 int size;
848 int used;
849 } kmp_nested_proc_bind_t;
850
851 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
852
853 extern int __kmp_display_affinity;
854 extern char *__kmp_affinity_format;
855 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
856 #if OMPT_SUPPORT
857 extern int __kmp_tool;
858 extern char *__kmp_tool_libraries;
859 #endif // OMPT_SUPPORT
860
861 #if KMP_AFFINITY_SUPPORTED
862 #define KMP_PLACE_ALL (-1)
863 #define KMP_PLACE_UNDEFINED (-2)
864 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
865 #define KMP_AFFINITY_NON_PROC_BIND \
866 ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
867 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
868 (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
869 #endif /* KMP_AFFINITY_SUPPORTED */
870
871 extern int __kmp_affinity_num_places;
872
873 typedef enum kmp_cancel_kind_t {
874 cancel_noreq = 0,
875 cancel_parallel = 1,
876 cancel_loop = 2,
877 cancel_sections = 3,
878 cancel_taskgroup = 4
879 } kmp_cancel_kind_t;
880
881 // KMP_HW_SUBSET support:
882 typedef struct kmp_hws_item {
883 int num;
884 int offset;
885 } kmp_hws_item_t;
886
887 extern kmp_hws_item_t __kmp_hws_socket;
888 extern kmp_hws_item_t __kmp_hws_die;
889 extern kmp_hws_item_t __kmp_hws_node;
890 extern kmp_hws_item_t __kmp_hws_tile;
891 extern kmp_hws_item_t __kmp_hws_core;
892 extern kmp_hws_item_t __kmp_hws_proc;
893 extern int __kmp_hws_requested;
894 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
895
896 /* ------------------------------------------------------------------------ */
897
898 #define KMP_PAD(type, sz) \
899 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
900
901 // We need to avoid using -1 as a GTID as +1 is added to the gtid
902 // when storing it in a lock, and the value 0 is reserved.
903 #define KMP_GTID_DNE (-2) /* Does not exist */
904 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
905 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
906 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
907 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
908
909 /* OpenMP 5.0 Memory Management support */
910
911 #ifndef __OMP_H
912 // Duplicate type definitions from omp.h
913 typedef uintptr_t omp_uintptr_t;
914
915 typedef enum {
916 omp_atk_sync_hint = 1,
917 omp_atk_alignment = 2,
918 omp_atk_access = 3,
919 omp_atk_pool_size = 4,
920 omp_atk_fallback = 5,
921 omp_atk_fb_data = 6,
922 omp_atk_pinned = 7,
923 omp_atk_partition = 8
924 } omp_alloctrait_key_t;
925
926 typedef enum {
927 omp_atv_false = 0,
928 omp_atv_true = 1,
929 omp_atv_contended = 3,
930 omp_atv_uncontended = 4,
931 omp_atv_serialized = 5,
932 omp_atv_sequential = omp_atv_serialized, // (deprecated)
933 omp_atv_private = 6,
934 omp_atv_all = 7,
935 omp_atv_thread = 8,
936 omp_atv_pteam = 9,
937 omp_atv_cgroup = 10,
938 omp_atv_default_mem_fb = 11,
939 omp_atv_null_fb = 12,
940 omp_atv_abort_fb = 13,
941 omp_atv_allocator_fb = 14,
942 omp_atv_environment = 15,
943 omp_atv_nearest = 16,
944 omp_atv_blocked = 17,
945 omp_atv_interleaved = 18
946 } omp_alloctrait_value_t;
947 #define omp_atv_default ((omp_uintptr_t)-1)
948
949 typedef void *omp_memspace_handle_t;
950 extern omp_memspace_handle_t const omp_default_mem_space;
951 extern omp_memspace_handle_t const omp_large_cap_mem_space;
952 extern omp_memspace_handle_t const omp_const_mem_space;
953 extern omp_memspace_handle_t const omp_high_bw_mem_space;
954 extern omp_memspace_handle_t const omp_low_lat_mem_space;
955 // Preview of target memory support
956 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
957 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
958 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
959
960 typedef struct {
961 omp_alloctrait_key_t key;
962 omp_uintptr_t value;
963 } omp_alloctrait_t;
964
965 typedef void *omp_allocator_handle_t;
966 extern omp_allocator_handle_t const omp_null_allocator;
967 extern omp_allocator_handle_t const omp_default_mem_alloc;
968 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
969 extern omp_allocator_handle_t const omp_const_mem_alloc;
970 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
971 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
972 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
973 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
974 extern omp_allocator_handle_t const omp_thread_mem_alloc;
975 // Preview of target memory support
976 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
977 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
978 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
979 extern omp_allocator_handle_t const kmp_max_mem_alloc;
980 extern omp_allocator_handle_t __kmp_def_allocator;
981
982 // end of duplicate type definitions from omp.h
983 #endif
984
985 extern int __kmp_memkind_available;
986
987 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
988
989 typedef struct kmp_allocator_t {
990 omp_memspace_handle_t memspace;
991 void **memkind; // pointer to memkind
992 size_t alignment;
993 omp_alloctrait_value_t fb;
994 kmp_allocator_t *fb_data;
995 kmp_uint64 pool_size;
996 kmp_uint64 pool_used;
997 } kmp_allocator_t;
998
999 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1000 omp_memspace_handle_t,
1001 int ntraits,
1002 omp_alloctrait_t traits[]);
1003 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1004 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1005 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1006 // external interfaces, may be used by compiler
1007 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1008 extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1009 omp_allocator_handle_t al);
1010 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1011 omp_allocator_handle_t al);
1012 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1013 omp_allocator_handle_t al,
1014 omp_allocator_handle_t free_al);
1015 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1016 // internal interfaces, contain real implementation
1017 extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1018 omp_allocator_handle_t al);
1019 extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1020 omp_allocator_handle_t al);
1021 extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1022 omp_allocator_handle_t al,
1023 omp_allocator_handle_t free_al);
1024 extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1025
1026 extern void __kmp_init_memkind();
1027 extern void __kmp_fini_memkind();
1028 extern void __kmp_init_target_mem();
1029
1030 /* ------------------------------------------------------------------------ */
1031
1032 #define KMP_UINT64_MAX \
1033 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1034
1035 #define KMP_MIN_NTH 1
1036
1037 #ifndef KMP_MAX_NTH
1038 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1039 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
1040 #else
1041 #define KMP_MAX_NTH INT_MAX
1042 #endif
1043 #endif /* KMP_MAX_NTH */
1044
1045 #ifdef PTHREAD_STACK_MIN
1046 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
1047 #else
1048 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1049 #endif
1050
1051 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1052
1053 #if KMP_ARCH_X86
1054 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1055 #elif KMP_ARCH_X86_64
1056 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1057 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1058 #else
1059 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1060 #endif
1061
1062 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1063 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1064 #define KMP_MAX_MALLOC_POOL_INCR \
1065 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1066
1067 #define KMP_MIN_STKOFFSET (0)
1068 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1069 #if KMP_OS_DARWIN
1070 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1071 #else
1072 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1073 #endif
1074
1075 #define KMP_MIN_STKPADDING (0)
1076 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1077
1078 #define KMP_BLOCKTIME_MULTIPLIER \
1079 (1000) /* number of blocktime units per second */
1080 #define KMP_MIN_BLOCKTIME (0)
1081 #define KMP_MAX_BLOCKTIME \
1082 (INT_MAX) /* Must be this for "infinite" setting the work */
1083 #define KMP_DEFAULT_BLOCKTIME (200) /* __kmp_blocktime is in milliseconds */
1084
1085 #if KMP_USE_MONITOR
1086 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1087 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1088 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1089
1090 /* Calculate new number of monitor wakeups for a specific block time based on
1091 previous monitor_wakeups. Only allow increasing number of wakeups */
1092 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1093 (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \
1094 : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \
1095 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1096 ? (monitor_wakeups) \
1097 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1098
1099 /* Calculate number of intervals for a specific block time based on
1100 monitor_wakeups */
1101 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1102 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1103 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1104 #else
1105 #define KMP_BLOCKTIME(team, tid) \
1106 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1107 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1108 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1109 extern kmp_uint64 __kmp_ticks_per_msec;
1110 #if KMP_COMPILER_ICC
1111 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1112 #else
1113 #define KMP_NOW() __kmp_hardware_timestamp()
1114 #endif
1115 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1116 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1117 (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1118 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1119 #else
1120 // System time is retrieved sporadically while blocking.
1121 extern kmp_uint64 __kmp_now_nsec();
1122 #define KMP_NOW() __kmp_now_nsec()
1123 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1124 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1125 (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1126 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1127 #endif
1128 #endif // KMP_USE_MONITOR
1129
1130 #define KMP_MIN_STATSCOLS 40
1131 #define KMP_MAX_STATSCOLS 4096
1132 #define KMP_DEFAULT_STATSCOLS 80
1133
1134 #define KMP_MIN_INTERVAL 0
1135 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1136 #define KMP_DEFAULT_INTERVAL 0
1137
1138 #define KMP_MIN_CHUNK 1
1139 #define KMP_MAX_CHUNK (INT_MAX - 1)
1140 #define KMP_DEFAULT_CHUNK 1
1141
1142 #define KMP_MIN_DISP_NUM_BUFF 1
1143 #define KMP_DFLT_DISP_NUM_BUFF 7
1144 #define KMP_MAX_DISP_NUM_BUFF 4096
1145
1146 #define KMP_MAX_ORDERED 8
1147
1148 #define KMP_MAX_FIELDS 32
1149
1150 #define KMP_MAX_BRANCH_BITS 31
1151
1152 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1153
1154 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1155
1156 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1157
1158 /* Minimum number of threads before switch to TLS gtid (experimentally
1159 determined) */
1160 /* josh TODO: what about OS X* tuning? */
1161 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1162 #define KMP_TLS_GTID_MIN 5
1163 #else
1164 #define KMP_TLS_GTID_MIN INT_MAX
1165 #endif
1166
1167 #define KMP_MASTER_TID(tid) (0 == (tid))
1168 #define KMP_WORKER_TID(tid) (0 != (tid))
1169
1170 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1171 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1172 #define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1173
1174 #ifndef TRUE
1175 #define FALSE 0
1176 #define TRUE (!FALSE)
1177 #endif
1178
1179 /* NOTE: all of the following constants must be even */
1180
1181 #if KMP_OS_WINDOWS
1182 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1183 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1184 #elif KMP_OS_LINUX
1185 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1186 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1187 #elif KMP_OS_DARWIN
1188 /* TODO: tune for KMP_OS_DARWIN */
1189 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1190 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1191 #elif KMP_OS_DRAGONFLY
1192 /* TODO: tune for KMP_OS_DRAGONFLY */
1193 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1194 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1195 #elif KMP_OS_FREEBSD
1196 /* TODO: tune for KMP_OS_FREEBSD */
1197 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1198 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1199 #elif KMP_OS_NETBSD
1200 /* TODO: tune for KMP_OS_NETBSD */
1201 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1202 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1203 #elif KMP_OS_HURD
1204 /* TODO: tune for KMP_OS_HURD */
1205 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1206 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1207 #elif KMP_OS_OPENBSD
1208 /* TODO: tune for KMP_OS_OPENBSD */
1209 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1210 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1211 #endif
1212
1213 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1214 typedef struct kmp_cpuid {
1215 kmp_uint32 eax;
1216 kmp_uint32 ebx;
1217 kmp_uint32 ecx;
1218 kmp_uint32 edx;
1219 } kmp_cpuid_t;
1220
1221 typedef struct kmp_cpuinfo_flags_t {
1222 unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1223 unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1224 unsigned reserved : 30; // Ensure size of 32 bits
1225 } kmp_cpuinfo_flags_t;
1226
1227 typedef struct kmp_cpuinfo {
1228 int initialized; // If 0, other fields are not initialized.
1229 int signature; // CPUID(1).EAX
1230 int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1231 int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1232 // Model << 4 ) + Model)
1233 int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1234 kmp_cpuinfo_flags_t flags;
1235 int apic_id;
1236 int physical_id;
1237 int logical_id;
1238 kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1239 char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1240 } kmp_cpuinfo_t;
1241
1242 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1243
1244 #if KMP_OS_UNIX
1245 // subleaf is only needed for cache and topology discovery and can be set to
1246 // zero in most cases
__kmp_x86_cpuid(int leaf,int subleaf,struct kmp_cpuid * p)1247 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1248 __asm__ __volatile__("cpuid"
1249 : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1250 : "a"(leaf), "c"(subleaf));
1251 }
1252 // Load p into FPU control word
__kmp_load_x87_fpu_control_word(const kmp_int16 * p)1253 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1254 __asm__ __volatile__("fldcw %0" : : "m"(*p));
1255 }
1256 // Store FPU control word into p
__kmp_store_x87_fpu_control_word(kmp_int16 * p)1257 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1258 __asm__ __volatile__("fstcw %0" : "=m"(*p));
1259 }
__kmp_clear_x87_fpu_status_word()1260 static inline void __kmp_clear_x87_fpu_status_word() {
1261 #if KMP_MIC
1262 // 32-bit protected mode x87 FPU state
1263 struct x87_fpu_state {
1264 unsigned cw;
1265 unsigned sw;
1266 unsigned tw;
1267 unsigned fip;
1268 unsigned fips;
1269 unsigned fdp;
1270 unsigned fds;
1271 };
1272 struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1273 __asm__ __volatile__("fstenv %0\n\t" // store FP env
1274 "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1275 "fldenv %0\n\t" // load FP env back
1276 : "+m"(fpu_state), "+m"(fpu_state.sw));
1277 #else
1278 __asm__ __volatile__("fnclex");
1279 #endif // KMP_MIC
1280 }
1281 #if __SSE__
__kmp_load_mxcsr(const kmp_uint32 * p)1282 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
__kmp_store_mxcsr(kmp_uint32 * p)1283 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1284 #else
__kmp_load_mxcsr(const kmp_uint32 * p)1285 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
__kmp_store_mxcsr(kmp_uint32 * p)1286 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1287 #endif
1288 #else
1289 // Windows still has these as external functions in assembly file
1290 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1291 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1292 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1293 extern void __kmp_clear_x87_fpu_status_word();
__kmp_load_mxcsr(const kmp_uint32 * p)1294 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
__kmp_store_mxcsr(kmp_uint32 * p)1295 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1296 #endif // KMP_OS_UNIX
1297
1298 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1299
1300 #if KMP_ARCH_X86
1301 extern void __kmp_x86_pause(void);
1302 #elif KMP_MIC
1303 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1304 // regression after removal of extra PAUSE from spin loops. Changing
1305 // the delay from 100 to 300 showed even better performance than double PAUSE
1306 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
__kmp_x86_pause(void)1307 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1308 #else
__kmp_x86_pause(void)1309 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1310 #endif
1311 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1312 #elif KMP_ARCH_PPC64
1313 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1314 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1315 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1316 #define KMP_CPU_PAUSE() \
1317 do { \
1318 KMP_PPC64_PRI_LOW(); \
1319 KMP_PPC64_PRI_MED(); \
1320 KMP_PPC64_PRI_LOC_MB(); \
1321 } while (0)
1322 #else
1323 #define KMP_CPU_PAUSE() /* nothing to do */
1324 #endif
1325
1326 #define KMP_INIT_YIELD(count) \
1327 { (count) = __kmp_yield_init; }
1328
1329 #define KMP_OVERSUBSCRIBED \
1330 (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1331
1332 #define KMP_TRY_YIELD \
1333 ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1334
1335 #define KMP_TRY_YIELD_OVERSUB \
1336 ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1337
1338 #define KMP_YIELD(cond) \
1339 { \
1340 KMP_CPU_PAUSE(); \
1341 if ((cond) && (KMP_TRY_YIELD)) \
1342 __kmp_yield(); \
1343 }
1344
1345 #define KMP_YIELD_OVERSUB() \
1346 { \
1347 KMP_CPU_PAUSE(); \
1348 if ((KMP_TRY_YIELD_OVERSUB)) \
1349 __kmp_yield(); \
1350 }
1351
1352 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1353 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1354 #define KMP_YIELD_SPIN(count) \
1355 { \
1356 KMP_CPU_PAUSE(); \
1357 if (KMP_TRY_YIELD) { \
1358 (count) -= 2; \
1359 if (!(count)) { \
1360 __kmp_yield(); \
1361 (count) = __kmp_yield_next; \
1362 } \
1363 } \
1364 }
1365
1366 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count) \
1367 { \
1368 KMP_CPU_PAUSE(); \
1369 if ((KMP_TRY_YIELD_OVERSUB)) \
1370 __kmp_yield(); \
1371 else if (__kmp_use_yield == 1) { \
1372 (count) -= 2; \
1373 if (!(count)) { \
1374 __kmp_yield(); \
1375 (count) = __kmp_yield_next; \
1376 } \
1377 } \
1378 }
1379
1380 // User-level Monitor/Mwait
1381 #if KMP_HAVE_UMWAIT
1382 // We always try for UMWAIT first
1383 #if KMP_HAVE_WAITPKG_INTRINSICS
1384 #if KMP_HAVE_IMMINTRIN_H
1385 #include <immintrin.h>
1386 #elif KMP_HAVE_INTRIN_H
1387 #include <intrin.h>
1388 #endif
1389 #endif // KMP_HAVE_WAITPKG_INTRINSICS
1390 KMP_ATTRIBUTE_TARGET_WAITPKG
__kmp_tpause(uint32_t hint,uint64_t counter)1391 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1392 #if !KMP_HAVE_WAITPKG_INTRINSICS
1393 uint32_t timeHi = uint32_t(counter >> 32);
1394 uint32_t timeLo = uint32_t(counter & 0xffffffff);
1395 char flag;
1396 __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1397 "setb %0"
1398 : "=r"(flag)
1399 : "a"(timeLo), "d"(timeHi), "c"(hint)
1400 :);
1401 return flag;
1402 #else
1403 return _tpause(hint, counter);
1404 #endif
1405 }
1406 KMP_ATTRIBUTE_TARGET_WAITPKG
__kmp_umonitor(void * cacheline)1407 static inline void __kmp_umonitor(void *cacheline) {
1408 #if !KMP_HAVE_WAITPKG_INTRINSICS
1409 __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1410 :
1411 : "a"(cacheline)
1412 :);
1413 #else
1414 _umonitor(cacheline);
1415 #endif
1416 }
1417 KMP_ATTRIBUTE_TARGET_WAITPKG
__kmp_umwait(uint32_t hint,uint64_t counter)1418 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1419 #if !KMP_HAVE_WAITPKG_INTRINSICS
1420 uint32_t timeHi = uint32_t(counter >> 32);
1421 uint32_t timeLo = uint32_t(counter & 0xffffffff);
1422 char flag;
1423 __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1424 "setb %0"
1425 : "=r"(flag)
1426 : "a"(timeLo), "d"(timeHi), "c"(hint)
1427 :);
1428 return flag;
1429 #else
1430 return _umwait(hint, counter);
1431 #endif
1432 }
1433 #elif KMP_HAVE_MWAIT
1434 #if KMP_OS_UNIX
1435 #include <pmmintrin.h>
1436 #else
1437 #include <intrin.h>
1438 #endif
1439 #if KMP_OS_UNIX
1440 __attribute__((target("sse3")))
1441 #endif
1442 static inline void
__kmp_mm_monitor(void * cacheline,unsigned extensions,unsigned hints)1443 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1444 _mm_monitor(cacheline, extensions, hints);
1445 }
1446 #if KMP_OS_UNIX
1447 __attribute__((target("sse3")))
1448 #endif
1449 static inline void
__kmp_mm_mwait(unsigned extensions,unsigned hints)1450 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1451 _mm_mwait(extensions, hints);
1452 }
1453 #endif // KMP_HAVE_UMWAIT
1454
1455 /* ------------------------------------------------------------------------ */
1456 /* Support datatypes for the orphaned construct nesting checks. */
1457 /* ------------------------------------------------------------------------ */
1458
1459 /* When adding to this enum, add its corresponding string in cons_text_c[]
1460 * array in kmp_error.cpp */
1461 enum cons_type {
1462 ct_none,
1463 ct_parallel,
1464 ct_pdo,
1465 ct_pdo_ordered,
1466 ct_psections,
1467 ct_psingle,
1468 ct_critical,
1469 ct_ordered_in_parallel,
1470 ct_ordered_in_pdo,
1471 ct_master,
1472 ct_reduce,
1473 ct_barrier,
1474 ct_masked
1475 };
1476
1477 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1478
1479 struct cons_data {
1480 ident_t const *ident;
1481 enum cons_type type;
1482 int prev;
1483 kmp_user_lock_p
1484 name; /* address exclusively for critical section name comparison */
1485 };
1486
1487 struct cons_header {
1488 int p_top, w_top, s_top;
1489 int stack_size, stack_top;
1490 struct cons_data *stack_data;
1491 };
1492
1493 struct kmp_region_info {
1494 char *text;
1495 int offset[KMP_MAX_FIELDS];
1496 int length[KMP_MAX_FIELDS];
1497 };
1498
1499 /* ---------------------------------------------------------------------- */
1500 /* ---------------------------------------------------------------------- */
1501
1502 #if KMP_OS_WINDOWS
1503 typedef HANDLE kmp_thread_t;
1504 typedef DWORD kmp_key_t;
1505 #endif /* KMP_OS_WINDOWS */
1506
1507 #if KMP_OS_UNIX
1508 typedef pthread_t kmp_thread_t;
1509 typedef pthread_key_t kmp_key_t;
1510 #endif
1511
1512 extern kmp_key_t __kmp_gtid_threadprivate_key;
1513
1514 typedef struct kmp_sys_info {
1515 long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1516 long minflt; /* the number of page faults serviced without any I/O */
1517 long majflt; /* the number of page faults serviced that required I/O */
1518 long nswap; /* the number of times a process was "swapped" out of memory */
1519 long inblock; /* the number of times the file system had to perform input */
1520 long oublock; /* the number of times the file system had to perform output */
1521 long nvcsw; /* the number of times a context switch was voluntarily */
1522 long nivcsw; /* the number of times a context switch was forced */
1523 } kmp_sys_info_t;
1524
1525 #if USE_ITT_BUILD
1526 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1527 // required type here. Later we will check the type meets requirements.
1528 typedef int kmp_itt_mark_t;
1529 #define KMP_ITT_DEBUG 0
1530 #endif /* USE_ITT_BUILD */
1531
1532 typedef kmp_int32 kmp_critical_name[8];
1533
1534 /*!
1535 @ingroup PARALLEL
1536 The type for a microtask which gets passed to @ref __kmpc_fork_call().
1537 The arguments to the outlined function are
1538 @param global_tid the global thread identity of the thread executing the
1539 function.
1540 @param bound_tid the local identity of the thread executing the function
1541 @param ... pointers to shared variables accessed by the function.
1542 */
1543 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1544 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1545 ...);
1546
1547 /*!
1548 @ingroup THREADPRIVATE
1549 @{
1550 */
1551 /* ---------------------------------------------------------------------------
1552 */
1553 /* Threadprivate initialization/finalization function declarations */
1554
1555 /* for non-array objects: __kmpc_threadprivate_register() */
1556
1557 /*!
1558 Pointer to the constructor function.
1559 The first argument is the <tt>this</tt> pointer
1560 */
1561 typedef void *(*kmpc_ctor)(void *);
1562
1563 /*!
1564 Pointer to the destructor function.
1565 The first argument is the <tt>this</tt> pointer
1566 */
1567 typedef void (*kmpc_dtor)(
1568 void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1569 compiler */
1570 /*!
1571 Pointer to an alternate constructor.
1572 The first argument is the <tt>this</tt> pointer.
1573 */
1574 typedef void *(*kmpc_cctor)(void *, void *);
1575
1576 /* for array objects: __kmpc_threadprivate_register_vec() */
1577 /* First arg: "this" pointer */
1578 /* Last arg: number of array elements */
1579 /*!
1580 Array constructor.
1581 First argument is the <tt>this</tt> pointer
1582 Second argument the number of array elements.
1583 */
1584 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1585 /*!
1586 Pointer to the array destructor function.
1587 The first argument is the <tt>this</tt> pointer
1588 Second argument the number of array elements.
1589 */
1590 typedef void (*kmpc_dtor_vec)(void *, size_t);
1591 /*!
1592 Array constructor.
1593 First argument is the <tt>this</tt> pointer
1594 Third argument the number of array elements.
1595 */
1596 typedef void *(*kmpc_cctor_vec)(void *, void *,
1597 size_t); /* function unused by compiler */
1598
1599 /*!
1600 @}
1601 */
1602
1603 /* keeps tracked of threadprivate cache allocations for cleanup later */
1604 typedef struct kmp_cached_addr {
1605 void **addr; /* address of allocated cache */
1606 void ***compiler_cache; /* pointer to compiler's cache */
1607 void *data; /* pointer to global data */
1608 struct kmp_cached_addr *next; /* pointer to next cached address */
1609 } kmp_cached_addr_t;
1610
1611 struct private_data {
1612 struct private_data *next; /* The next descriptor in the list */
1613 void *data; /* The data buffer for this descriptor */
1614 int more; /* The repeat count for this descriptor */
1615 size_t size; /* The data size for this descriptor */
1616 };
1617
1618 struct private_common {
1619 struct private_common *next;
1620 struct private_common *link;
1621 void *gbl_addr;
1622 void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1623 size_t cmn_size;
1624 };
1625
1626 struct shared_common {
1627 struct shared_common *next;
1628 struct private_data *pod_init;
1629 void *obj_init;
1630 void *gbl_addr;
1631 union {
1632 kmpc_ctor ctor;
1633 kmpc_ctor_vec ctorv;
1634 } ct;
1635 union {
1636 kmpc_cctor cctor;
1637 kmpc_cctor_vec cctorv;
1638 } cct;
1639 union {
1640 kmpc_dtor dtor;
1641 kmpc_dtor_vec dtorv;
1642 } dt;
1643 size_t vec_len;
1644 int is_vec;
1645 size_t cmn_size;
1646 };
1647
1648 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1649 #define KMP_HASH_TABLE_SIZE \
1650 (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1651 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1652 #define KMP_HASH(x) \
1653 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1654
1655 struct common_table {
1656 struct private_common *data[KMP_HASH_TABLE_SIZE];
1657 };
1658
1659 struct shared_table {
1660 struct shared_common *data[KMP_HASH_TABLE_SIZE];
1661 };
1662
1663 /* ------------------------------------------------------------------------ */
1664
1665 #if KMP_USE_HIER_SCHED
1666 // Shared barrier data that exists inside a single unit of the scheduling
1667 // hierarchy
1668 typedef struct kmp_hier_private_bdata_t {
1669 kmp_int32 num_active;
1670 kmp_uint64 index;
1671 kmp_uint64 wait_val[2];
1672 } kmp_hier_private_bdata_t;
1673 #endif
1674
1675 typedef struct kmp_sched_flags {
1676 unsigned ordered : 1;
1677 unsigned nomerge : 1;
1678 unsigned contains_last : 1;
1679 #if KMP_USE_HIER_SCHED
1680 unsigned use_hier : 1;
1681 unsigned unused : 28;
1682 #else
1683 unsigned unused : 29;
1684 #endif
1685 } kmp_sched_flags_t;
1686
1687 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1688
1689 #if KMP_STATIC_STEAL_ENABLED
1690 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1691 kmp_int32 count;
1692 kmp_int32 ub;
1693 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1694 kmp_int32 lb;
1695 kmp_int32 st;
1696 kmp_int32 tc;
1697 kmp_lock_t *steal_lock; // lock used for chunk stealing
1698 // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1699 // a) parm3 is properly aligned and
1700 // b) all parm1-4 are on the same cache line.
1701 // Because of parm1-4 are used together, performance seems to be better
1702 // if they are on the same cache line (not measured though).
1703
1704 struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1705 kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1706 kmp_int32 parm2; // make no real change at least while padding is off.
1707 kmp_int32 parm3;
1708 kmp_int32 parm4;
1709 };
1710
1711 kmp_uint32 ordered_lower;
1712 kmp_uint32 ordered_upper;
1713 #if KMP_OS_WINDOWS
1714 kmp_int32 last_upper;
1715 #endif /* KMP_OS_WINDOWS */
1716 } dispatch_private_info32_t;
1717
1718 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1719 kmp_int64 count; // current chunk number for static & static-steal scheduling
1720 kmp_int64 ub; /* upper-bound */
1721 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1722 kmp_int64 lb; /* lower-bound */
1723 kmp_int64 st; /* stride */
1724 kmp_int64 tc; /* trip count (number of iterations) */
1725 kmp_lock_t *steal_lock; // lock used for chunk stealing
1726 /* parm[1-4] are used in different ways by different scheduling algorithms */
1727
1728 // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1729 // a) parm3 is properly aligned and
1730 // b) all parm1-4 are in the same cache line.
1731 // Because of parm1-4 are used together, performance seems to be better
1732 // if they are in the same line (not measured though).
1733
1734 struct KMP_ALIGN(32) {
1735 kmp_int64 parm1;
1736 kmp_int64 parm2;
1737 kmp_int64 parm3;
1738 kmp_int64 parm4;
1739 };
1740
1741 kmp_uint64 ordered_lower;
1742 kmp_uint64 ordered_upper;
1743 #if KMP_OS_WINDOWS
1744 kmp_int64 last_upper;
1745 #endif /* KMP_OS_WINDOWS */
1746 } dispatch_private_info64_t;
1747 #else /* KMP_STATIC_STEAL_ENABLED */
1748 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1749 kmp_int32 lb;
1750 kmp_int32 ub;
1751 kmp_int32 st;
1752 kmp_int32 tc;
1753
1754 kmp_int32 parm1;
1755 kmp_int32 parm2;
1756 kmp_int32 parm3;
1757 kmp_int32 parm4;
1758
1759 kmp_int32 count;
1760
1761 kmp_uint32 ordered_lower;
1762 kmp_uint32 ordered_upper;
1763 #if KMP_OS_WINDOWS
1764 kmp_int32 last_upper;
1765 #endif /* KMP_OS_WINDOWS */
1766 } dispatch_private_info32_t;
1767
1768 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1769 kmp_int64 lb; /* lower-bound */
1770 kmp_int64 ub; /* upper-bound */
1771 kmp_int64 st; /* stride */
1772 kmp_int64 tc; /* trip count (number of iterations) */
1773
1774 /* parm[1-4] are used in different ways by different scheduling algorithms */
1775 kmp_int64 parm1;
1776 kmp_int64 parm2;
1777 kmp_int64 parm3;
1778 kmp_int64 parm4;
1779
1780 kmp_int64 count; /* current chunk number for static scheduling */
1781
1782 kmp_uint64 ordered_lower;
1783 kmp_uint64 ordered_upper;
1784 #if KMP_OS_WINDOWS
1785 kmp_int64 last_upper;
1786 #endif /* KMP_OS_WINDOWS */
1787 } dispatch_private_info64_t;
1788 #endif /* KMP_STATIC_STEAL_ENABLED */
1789
1790 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1791 union private_info {
1792 dispatch_private_info32_t p32;
1793 dispatch_private_info64_t p64;
1794 } u;
1795 enum sched_type schedule; /* scheduling algorithm */
1796 kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1797 std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
1798 kmp_int32 ordered_bumped;
1799 // Stack of buffers for nest of serial regions
1800 struct dispatch_private_info *next;
1801 kmp_int32 type_size; /* the size of types in private_info */
1802 #if KMP_USE_HIER_SCHED
1803 kmp_int32 hier_id;
1804 void *parent; /* hierarchical scheduling parent pointer */
1805 #endif
1806 enum cons_type pushed_ws;
1807 } dispatch_private_info_t;
1808
1809 typedef struct dispatch_shared_info32 {
1810 /* chunk index under dynamic, number of idle threads under static-steal;
1811 iteration index otherwise */
1812 volatile kmp_uint32 iteration;
1813 volatile kmp_int32 num_done;
1814 volatile kmp_uint32 ordered_iteration;
1815 // Dummy to retain the structure size after making ordered_iteration scalar
1816 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1817 } dispatch_shared_info32_t;
1818
1819 typedef struct dispatch_shared_info64 {
1820 /* chunk index under dynamic, number of idle threads under static-steal;
1821 iteration index otherwise */
1822 volatile kmp_uint64 iteration;
1823 volatile kmp_int64 num_done;
1824 volatile kmp_uint64 ordered_iteration;
1825 // Dummy to retain the structure size after making ordered_iteration scalar
1826 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1827 } dispatch_shared_info64_t;
1828
1829 typedef struct dispatch_shared_info {
1830 union shared_info {
1831 dispatch_shared_info32_t s32;
1832 dispatch_shared_info64_t s64;
1833 } u;
1834 volatile kmp_uint32 buffer_index;
1835 volatile kmp_int32 doacross_buf_idx; // teamwise index
1836 volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1837 kmp_int32 doacross_num_done; // count finished threads
1838 #if KMP_USE_HIER_SCHED
1839 void *hier;
1840 #endif
1841 #if KMP_USE_HWLOC
1842 // When linking with libhwloc, the ORDERED EPCC test slows down on big
1843 // machines (> 48 cores). Performance analysis showed that a cache thrash
1844 // was occurring and this padding helps alleviate the problem.
1845 char padding[64];
1846 #endif
1847 } dispatch_shared_info_t;
1848
1849 typedef struct kmp_disp {
1850 /* Vector for ORDERED SECTION */
1851 void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1852 /* Vector for END ORDERED SECTION */
1853 void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1854
1855 dispatch_shared_info_t *th_dispatch_sh_current;
1856 dispatch_private_info_t *th_dispatch_pr_current;
1857
1858 dispatch_private_info_t *th_disp_buffer;
1859 kmp_uint32 th_disp_index;
1860 kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1861 volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1862 kmp_int64 *th_doacross_info; // info on loop bounds
1863 #if KMP_USE_INTERNODE_ALIGNMENT
1864 char more_padding[INTERNODE_CACHE_LINE];
1865 #endif
1866 } kmp_disp_t;
1867
1868 /* ------------------------------------------------------------------------ */
1869 /* Barrier stuff */
1870
1871 /* constants for barrier state update */
1872 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1873 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1874 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1875 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1876
1877 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1878 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1879 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1880
1881 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1882 #error "Barrier sleep bit must be smaller than barrier bump bit"
1883 #endif
1884 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1885 #error "Barrier unused bit must be smaller than barrier bump bit"
1886 #endif
1887
1888 // Constants for release barrier wait state: currently, hierarchical only
1889 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1890 #define KMP_BARRIER_OWN_FLAG \
1891 1 // Normal state; worker waiting on own b_go flag in release
1892 #define KMP_BARRIER_PARENT_FLAG \
1893 2 // Special state; worker waiting on parent's b_go flag in release
1894 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1895 3 // Special state; tells worker to shift from parent to own b_go
1896 #define KMP_BARRIER_SWITCHING \
1897 4 // Special state; worker resets appropriate flag on wake-up
1898
1899 #define KMP_NOT_SAFE_TO_REAP \
1900 0 // Thread th_reap_state: not safe to reap (tasking)
1901 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1902
1903 // The flag_type describes the storage used for the flag.
1904 enum flag_type {
1905 flag32, /**< atomic 32 bit flags */
1906 flag64, /**< 64 bit flags */
1907 atomic_flag64, /**< atomic 64 bit flags */
1908 flag_oncore, /**< special 64-bit flag for on-core barrier (hierarchical) */
1909 flag_unset
1910 };
1911
1912 enum barrier_type {
1913 bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1914 barriers if enabled) */
1915 bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1916 #if KMP_FAST_REDUCTION_BARRIER
1917 bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1918 #endif // KMP_FAST_REDUCTION_BARRIER
1919 bs_last_barrier /* Just a placeholder to mark the end */
1920 };
1921
1922 // to work with reduction barriers just like with plain barriers
1923 #if !KMP_FAST_REDUCTION_BARRIER
1924 #define bs_reduction_barrier bs_plain_barrier
1925 #endif // KMP_FAST_REDUCTION_BARRIER
1926
1927 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1928 bp_linear_bar =
1929 0, /* Single level (degenerate) tree */
1930 bp_tree_bar =
1931 1, /* Balanced tree with branching factor 2^n */
1932 bp_hyper_bar = 2, /* Hypercube-embedded tree with min
1933 branching factor 2^n */
1934 bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1935 bp_dist_bar = 4, /* Distributed barrier */
1936 bp_last_bar /* Placeholder to mark the end */
1937 } kmp_bar_pat_e;
1938
1939 #define KMP_BARRIER_ICV_PUSH 1
1940
1941 /* Record for holding the values of the internal controls stack records */
1942 typedef struct kmp_internal_control {
1943 int serial_nesting_level; /* corresponds to the value of the
1944 th_team_serialized field */
1945 kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1946 thread) */
1947 kmp_int8
1948 bt_set; /* internal control for whether blocktime is explicitly set */
1949 int blocktime; /* internal control for blocktime */
1950 #if KMP_USE_MONITOR
1951 int bt_intervals; /* internal control for blocktime intervals */
1952 #endif
1953 int nproc; /* internal control for #threads for next parallel region (per
1954 thread) */
1955 int thread_limit; /* internal control for thread-limit-var */
1956 int max_active_levels; /* internal control for max_active_levels */
1957 kmp_r_sched_t
1958 sched; /* internal control for runtime schedule {sched,chunk} pair */
1959 kmp_proc_bind_t proc_bind; /* internal control for affinity */
1960 kmp_int32 default_device; /* internal control for default device */
1961 struct kmp_internal_control *next;
1962 } kmp_internal_control_t;
1963
copy_icvs(kmp_internal_control_t * dst,kmp_internal_control_t * src)1964 static inline void copy_icvs(kmp_internal_control_t *dst,
1965 kmp_internal_control_t *src) {
1966 *dst = *src;
1967 }
1968
1969 /* Thread barrier needs volatile barrier fields */
1970 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1971 // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1972 // uses of it). It is not explicitly aligned below, because we *don't* want
1973 // it to be padded -- instead, we fit b_go into the same cache line with
1974 // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1975 kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1976 // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1977 // same NGO store
1978 volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1979 KMP_ALIGN_CACHE volatile kmp_uint64
1980 b_arrived; // STATE => task reached synch point.
1981 kmp_uint32 *skip_per_level;
1982 kmp_uint32 my_level;
1983 kmp_int32 parent_tid;
1984 kmp_int32 old_tid;
1985 kmp_uint32 depth;
1986 struct kmp_bstate *parent_bar;
1987 kmp_team_t *team;
1988 kmp_uint64 leaf_state;
1989 kmp_uint32 nproc;
1990 kmp_uint8 base_leaf_kids;
1991 kmp_uint8 leaf_kids;
1992 kmp_uint8 offset;
1993 kmp_uint8 wait_flag;
1994 kmp_uint8 use_oncore_barrier;
1995 #if USE_DEBUGGER
1996 // The following field is intended for the debugger solely. Only the worker
1997 // thread itself accesses this field: the worker increases it by 1 when it
1998 // arrives to a barrier.
1999 KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2000 #endif /* USE_DEBUGGER */
2001 } kmp_bstate_t;
2002
2003 union KMP_ALIGN_CACHE kmp_barrier_union {
2004 double b_align; /* use worst case alignment */
2005 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2006 kmp_bstate_t bb;
2007 };
2008
2009 typedef union kmp_barrier_union kmp_balign_t;
2010
2011 /* Team barrier needs only non-volatile arrived counter */
2012 union KMP_ALIGN_CACHE kmp_barrier_team_union {
2013 double b_align; /* use worst case alignment */
2014 char b_pad[CACHE_LINE];
2015 struct {
2016 kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2017 #if USE_DEBUGGER
2018 // The following two fields are indended for the debugger solely. Only
2019 // primary thread of the team accesses these fields: the first one is
2020 // increased by 1 when the primary thread arrives to a barrier, the second
2021 // one is increased by one when all the threads arrived.
2022 kmp_uint b_master_arrived;
2023 kmp_uint b_team_arrived;
2024 #endif
2025 };
2026 };
2027
2028 typedef union kmp_barrier_team_union kmp_balign_team_t;
2029
2030 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2031 threads when a condition changes. This is to workaround an NPTL bug where
2032 padding was added to pthread_cond_t which caused the initialization routine
2033 to write outside of the structure if compiled on pre-NPTL threads. */
2034 #if KMP_OS_WINDOWS
2035 typedef struct kmp_win32_mutex {
2036 /* The Lock */
2037 CRITICAL_SECTION cs;
2038 } kmp_win32_mutex_t;
2039
2040 typedef struct kmp_win32_cond {
2041 /* Count of the number of waiters. */
2042 int waiters_count_;
2043
2044 /* Serialize access to <waiters_count_> */
2045 kmp_win32_mutex_t waiters_count_lock_;
2046
2047 /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2048 int release_count_;
2049
2050 /* Keeps track of the current "generation" so that we don't allow */
2051 /* one thread to steal all the "releases" from the broadcast. */
2052 int wait_generation_count_;
2053
2054 /* A manual-reset event that's used to block and release waiting threads. */
2055 HANDLE event_;
2056 } kmp_win32_cond_t;
2057 #endif
2058
2059 #if KMP_OS_UNIX
2060
2061 union KMP_ALIGN_CACHE kmp_cond_union {
2062 double c_align;
2063 char c_pad[CACHE_LINE];
2064 pthread_cond_t c_cond;
2065 };
2066
2067 typedef union kmp_cond_union kmp_cond_align_t;
2068
2069 union KMP_ALIGN_CACHE kmp_mutex_union {
2070 double m_align;
2071 char m_pad[CACHE_LINE];
2072 pthread_mutex_t m_mutex;
2073 };
2074
2075 typedef union kmp_mutex_union kmp_mutex_align_t;
2076
2077 #endif /* KMP_OS_UNIX */
2078
2079 typedef struct kmp_desc_base {
2080 void *ds_stackbase;
2081 size_t ds_stacksize;
2082 int ds_stackgrow;
2083 kmp_thread_t ds_thread;
2084 volatile int ds_tid;
2085 int ds_gtid;
2086 #if KMP_OS_WINDOWS
2087 volatile int ds_alive;
2088 DWORD ds_thread_id;
2089 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2090 However, debugger support (libomp_db) cannot work with handles, because they
2091 uncomparable. For example, debugger requests info about thread with handle h.
2092 h is valid within debugger process, and meaningless within debugee process.
2093 Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2094 within debugee process, but it is a *new* handle which does *not* equal to
2095 any other handle in debugee... The only way to compare handles is convert
2096 them to system-wide ids. GetThreadId() function is available only in
2097 Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2098 on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2099 thread id by call to GetCurrentThreadId() from within the thread and save it
2100 to let libomp_db identify threads. */
2101 #endif /* KMP_OS_WINDOWS */
2102 } kmp_desc_base_t;
2103
2104 typedef union KMP_ALIGN_CACHE kmp_desc {
2105 double ds_align; /* use worst case alignment */
2106 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2107 kmp_desc_base_t ds;
2108 } kmp_desc_t;
2109
2110 typedef struct kmp_local {
2111 volatile int this_construct; /* count of single's encountered by thread */
2112 void *reduce_data;
2113 #if KMP_USE_BGET
2114 void *bget_data;
2115 void *bget_list;
2116 #if !USE_CMP_XCHG_FOR_BGET
2117 #ifdef USE_QUEUING_LOCK_FOR_BGET
2118 kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2119 #else
2120 kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2121 // bootstrap lock so we can use it at library
2122 // shutdown.
2123 #endif /* USE_LOCK_FOR_BGET */
2124 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2125 #endif /* KMP_USE_BGET */
2126
2127 PACKED_REDUCTION_METHOD_T
2128 packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2129 __kmpc_end_reduce*() */
2130
2131 } kmp_local_t;
2132
2133 #define KMP_CHECK_UPDATE(a, b) \
2134 if ((a) != (b)) \
2135 (a) = (b)
2136 #define KMP_CHECK_UPDATE_SYNC(a, b) \
2137 if ((a) != (b)) \
2138 TCW_SYNC_PTR((a), (b))
2139
2140 #define get__blocktime(xteam, xtid) \
2141 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2142 #define get__bt_set(xteam, xtid) \
2143 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2144 #if KMP_USE_MONITOR
2145 #define get__bt_intervals(xteam, xtid) \
2146 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2147 #endif
2148
2149 #define get__dynamic_2(xteam, xtid) \
2150 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2151 #define get__nproc_2(xteam, xtid) \
2152 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2153 #define get__sched_2(xteam, xtid) \
2154 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2155
2156 #define set__blocktime_team(xteam, xtid, xval) \
2157 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2158 (xval))
2159
2160 #if KMP_USE_MONITOR
2161 #define set__bt_intervals_team(xteam, xtid, xval) \
2162 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2163 (xval))
2164 #endif
2165
2166 #define set__bt_set_team(xteam, xtid, xval) \
2167 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2168
2169 #define set__dynamic(xthread, xval) \
2170 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2171 #define get__dynamic(xthread) \
2172 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2173
2174 #define set__nproc(xthread, xval) \
2175 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2176
2177 #define set__thread_limit(xthread, xval) \
2178 (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2179
2180 #define set__max_active_levels(xthread, xval) \
2181 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2182
2183 #define get__max_active_levels(xthread) \
2184 ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2185
2186 #define set__sched(xthread, xval) \
2187 (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2188
2189 #define set__proc_bind(xthread, xval) \
2190 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2191 #define get__proc_bind(xthread) \
2192 ((xthread)->th.th_current_task->td_icvs.proc_bind)
2193
2194 // OpenMP tasking data structures
2195
2196 typedef enum kmp_tasking_mode {
2197 tskm_immediate_exec = 0,
2198 tskm_extra_barrier = 1,
2199 tskm_task_teams = 2,
2200 tskm_max = 2
2201 } kmp_tasking_mode_t;
2202
2203 extern kmp_tasking_mode_t
2204 __kmp_tasking_mode; /* determines how/when to execute tasks */
2205 extern int __kmp_task_stealing_constraint;
2206 extern int __kmp_enable_task_throttling;
2207 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2208 // specified, defaults to 0 otherwise
2209 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2210 extern kmp_int32 __kmp_max_task_priority;
2211 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2212 extern kmp_uint64 __kmp_taskloop_min_tasks;
2213
2214 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2215 taskdata first */
2216 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2217 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2218
2219 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2220 // were spawned and queued since the previous barrier release.
2221 #define KMP_TASKING_ENABLED(task_team) \
2222 (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2223 /*!
2224 @ingroup BASIC_TYPES
2225 @{
2226 */
2227
2228 /*!
2229 */
2230 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2231
2232 typedef union kmp_cmplrdata {
2233 kmp_int32 priority; /**< priority specified by user for the task */
2234 kmp_routine_entry_t
2235 destructors; /* pointer to function to invoke deconstructors of
2236 firstprivate C++ objects */
2237 /* future data */
2238 } kmp_cmplrdata_t;
2239
2240 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2241 /*!
2242 */
2243 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2244 void *shareds; /**< pointer to block of pointers to shared vars */
2245 kmp_routine_entry_t
2246 routine; /**< pointer to routine to call for executing task */
2247 kmp_int32 part_id; /**< part id for the task */
2248 kmp_cmplrdata_t
2249 data1; /* Two known optional additions: destructors and priority */
2250 kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2251 /* future data */
2252 /* private vars */
2253 } kmp_task_t;
2254
2255 /*!
2256 @}
2257 */
2258
2259 typedef struct kmp_taskgroup {
2260 std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2261 std::atomic<kmp_int32>
2262 cancel_request; // request for cancellation of this taskgroup
2263 struct kmp_taskgroup *parent; // parent taskgroup
2264 // Block of data to perform task reduction
2265 void *reduce_data; // reduction related info
2266 kmp_int32 reduce_num_data; // number of data items to reduce
2267 uintptr_t *gomp_data; // gomp reduction data
2268 } kmp_taskgroup_t;
2269
2270 // forward declarations
2271 typedef union kmp_depnode kmp_depnode_t;
2272 typedef struct kmp_depnode_list kmp_depnode_list_t;
2273 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2274
2275 // macros for checking dep flag as an integer
2276 #define KMP_DEP_IN 0x1
2277 #define KMP_DEP_OUT 0x2
2278 #define KMP_DEP_INOUT 0x3
2279 #define KMP_DEP_MTX 0x4
2280 #define KMP_DEP_SET 0x8
2281 #define KMP_DEP_ALL 0x80
2282 // Compiler sends us this info:
2283 typedef struct kmp_depend_info {
2284 kmp_intptr_t base_addr;
2285 size_t len;
2286 union {
2287 kmp_uint8 flag; // flag as an unsigned char
2288 struct { // flag as a set of 8 bits
2289 unsigned in : 1;
2290 unsigned out : 1;
2291 unsigned mtx : 1;
2292 unsigned set : 1;
2293 unsigned unused : 3;
2294 unsigned all : 1;
2295 } flags;
2296 };
2297 } kmp_depend_info_t;
2298
2299 // Internal structures to work with task dependencies:
2300 struct kmp_depnode_list {
2301 kmp_depnode_t *node;
2302 kmp_depnode_list_t *next;
2303 };
2304
2305 // Max number of mutexinoutset dependencies per node
2306 #define MAX_MTX_DEPS 4
2307
2308 typedef struct kmp_base_depnode {
2309 kmp_depnode_list_t *successors; /* used under lock */
2310 kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2311 kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2312 kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2313 kmp_lock_t lock; /* guards shared fields: task, successors */
2314 #if KMP_SUPPORT_GRAPH_OUTPUT
2315 kmp_uint32 id;
2316 #endif
2317 std::atomic<kmp_int32> npredecessors;
2318 std::atomic<kmp_int32> nrefs;
2319 } kmp_base_depnode_t;
2320
2321 union KMP_ALIGN_CACHE kmp_depnode {
2322 double dn_align; /* use worst case alignment */
2323 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2324 kmp_base_depnode_t dn;
2325 };
2326
2327 struct kmp_dephash_entry {
2328 kmp_intptr_t addr;
2329 kmp_depnode_t *last_out;
2330 kmp_depnode_list_t *last_set;
2331 kmp_depnode_list_t *prev_set;
2332 kmp_uint8 last_flag;
2333 kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2334 kmp_dephash_entry_t *next_in_bucket;
2335 };
2336
2337 typedef struct kmp_dephash {
2338 kmp_dephash_entry_t **buckets;
2339 size_t size;
2340 kmp_depnode_t *last_all;
2341 size_t generation;
2342 kmp_uint32 nelements;
2343 kmp_uint32 nconflicts;
2344 } kmp_dephash_t;
2345
2346 typedef struct kmp_task_affinity_info {
2347 kmp_intptr_t base_addr;
2348 size_t len;
2349 struct {
2350 bool flag1 : 1;
2351 bool flag2 : 1;
2352 kmp_int32 reserved : 30;
2353 } flags;
2354 } kmp_task_affinity_info_t;
2355
2356 typedef enum kmp_event_type_t {
2357 KMP_EVENT_UNINITIALIZED = 0,
2358 KMP_EVENT_ALLOW_COMPLETION = 1
2359 } kmp_event_type_t;
2360
2361 typedef struct {
2362 kmp_event_type_t type;
2363 kmp_tas_lock_t lock;
2364 union {
2365 kmp_task_t *task;
2366 } ed;
2367 } kmp_event_t;
2368
2369 #ifdef BUILD_TIED_TASK_STACK
2370
2371 /* Tied Task stack definitions */
2372 typedef struct kmp_stack_block {
2373 kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2374 struct kmp_stack_block *sb_next;
2375 struct kmp_stack_block *sb_prev;
2376 } kmp_stack_block_t;
2377
2378 typedef struct kmp_task_stack {
2379 kmp_stack_block_t ts_first_block; // first block of stack entries
2380 kmp_taskdata_t **ts_top; // pointer to the top of stack
2381 kmp_int32 ts_entries; // number of entries on the stack
2382 } kmp_task_stack_t;
2383
2384 #endif // BUILD_TIED_TASK_STACK
2385
2386 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2387 /* Compiler flags */ /* Total compiler flags must be 16 bits */
2388 unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2389 unsigned final : 1; /* task is final(1) so execute immediately */
2390 unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2391 code path */
2392 unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2393 invoke destructors from the runtime */
2394 unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2395 context of the RTL) */
2396 unsigned priority_specified : 1; /* set if the compiler provides priority
2397 setting for the task */
2398 unsigned detachable : 1; /* 1 == can detach */
2399 unsigned hidden_helper : 1; /* 1 == hidden helper task */
2400 unsigned reserved : 8; /* reserved for compiler use */
2401
2402 /* Library flags */ /* Total library flags must be 16 bits */
2403 unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2404 unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2405 unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2406 // (1) or may be deferred (0)
2407 unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2408 // (0) [>= 2 threads]
2409 /* If either team_serial or tasking_ser is set, task team may be NULL */
2410 /* Task State Flags: */
2411 unsigned started : 1; /* 1==started, 0==not started */
2412 unsigned executing : 1; /* 1==executing, 0==not executing */
2413 unsigned complete : 1; /* 1==complete, 0==not complete */
2414 unsigned freed : 1; /* 1==freed, 0==allocated */
2415 unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2416 unsigned reserved31 : 7; /* reserved for library use */
2417
2418 } kmp_tasking_flags_t;
2419
2420 struct kmp_taskdata { /* aligned during dynamic allocation */
2421 kmp_int32 td_task_id; /* id, assigned by debugger */
2422 kmp_tasking_flags_t td_flags; /* task flags */
2423 kmp_team_t *td_team; /* team for this task */
2424 kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2425 /* Currently not used except for perhaps IDB */
2426 kmp_taskdata_t *td_parent; /* parent task */
2427 kmp_int32 td_level; /* task nesting level */
2428 std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2429 ident_t *td_ident; /* task identifier */
2430 // Taskwait data.
2431 ident_t *td_taskwait_ident;
2432 kmp_uint32 td_taskwait_counter;
2433 kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2434 KMP_ALIGN_CACHE kmp_internal_control_t
2435 td_icvs; /* Internal control variables for the task */
2436 KMP_ALIGN_CACHE std::atomic<kmp_int32>
2437 td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2438 deallocated */
2439 std::atomic<kmp_int32>
2440 td_incomplete_child_tasks; /* Child tasks not yet complete */
2441 kmp_taskgroup_t
2442 *td_taskgroup; // Each task keeps pointer to its current taskgroup
2443 kmp_dephash_t
2444 *td_dephash; // Dependencies for children tasks are tracked from here
2445 kmp_depnode_t
2446 *td_depnode; // Pointer to graph node if this task has dependencies
2447 kmp_task_team_t *td_task_team;
2448 size_t td_size_alloc; // Size of task structure, including shareds etc.
2449 #if defined(KMP_GOMP_COMPAT)
2450 // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2451 kmp_int32 td_size_loop_bounds;
2452 #endif
2453 kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2454 #if defined(KMP_GOMP_COMPAT)
2455 // GOMP sends in a copy function for copy constructors
2456 void (*td_copy_func)(void *, void *);
2457 #endif
2458 kmp_event_t td_allow_completion_event;
2459 #if OMPT_SUPPORT
2460 ompt_task_info_t ompt_task_info;
2461 #endif
2462 }; // struct kmp_taskdata
2463
2464 // Make sure padding above worked
2465 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2466
2467 // Data for task team but per thread
2468 typedef struct kmp_base_thread_data {
2469 kmp_info_p *td_thr; // Pointer back to thread info
2470 // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2471 // queued?
2472 kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2473 kmp_taskdata_t *
2474 *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2475 kmp_int32 td_deque_size; // Size of deck
2476 kmp_uint32 td_deque_head; // Head of deque (will wrap)
2477 kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2478 kmp_int32 td_deque_ntasks; // Number of tasks in deque
2479 // GEH: shouldn't this be volatile since used in while-spin?
2480 kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2481 #ifdef BUILD_TIED_TASK_STACK
2482 kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2483 // scheduling constraint
2484 #endif // BUILD_TIED_TASK_STACK
2485 } kmp_base_thread_data_t;
2486
2487 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2488 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2489
2490 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2491 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2492
2493 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2494 kmp_base_thread_data_t td;
2495 double td_align; /* use worst case alignment */
2496 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2497 } kmp_thread_data_t;
2498
2499 // Data for task teams which are used when tasking is enabled for the team
2500 typedef struct kmp_base_task_team {
2501 kmp_bootstrap_lock_t
2502 tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2503 /* must be bootstrap lock since used at library shutdown*/
2504 kmp_task_team_t *tt_next; /* For linking the task team free list */
2505 kmp_thread_data_t
2506 *tt_threads_data; /* Array of per-thread structures for task team */
2507 /* Data survives task team deallocation */
2508 kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2509 executing this team? */
2510 /* TRUE means tt_threads_data is set up and initialized */
2511 kmp_int32 tt_nproc; /* #threads in team */
2512 kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2513 kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2514 kmp_int32 tt_untied_task_encountered;
2515 // There is hidden helper thread encountered in this task team so that we must
2516 // wait when waiting on task team
2517 kmp_int32 tt_hidden_helper_task_encountered;
2518
2519 KMP_ALIGN_CACHE
2520 std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2521
2522 KMP_ALIGN_CACHE
2523 volatile kmp_uint32
2524 tt_active; /* is the team still actively executing tasks */
2525 } kmp_base_task_team_t;
2526
2527 union KMP_ALIGN_CACHE kmp_task_team {
2528 kmp_base_task_team_t tt;
2529 double tt_align; /* use worst case alignment */
2530 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2531 };
2532
2533 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2534 // Free lists keep same-size free memory slots for fast memory allocation
2535 // routines
2536 typedef struct kmp_free_list {
2537 void *th_free_list_self; // Self-allocated tasks free list
2538 void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2539 // threads
2540 void *th_free_list_other; // Non-self free list (to be returned to owner's
2541 // sync list)
2542 } kmp_free_list_t;
2543 #endif
2544 #if KMP_NESTED_HOT_TEAMS
2545 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2546 // are not put in teams pool, and they don't put threads in threads pool.
2547 typedef struct kmp_hot_team_ptr {
2548 kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2549 kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2550 } kmp_hot_team_ptr_t;
2551 #endif
2552 typedef struct kmp_teams_size {
2553 kmp_int32 nteams; // number of teams in a league
2554 kmp_int32 nth; // number of threads in each team of the league
2555 } kmp_teams_size_t;
2556
2557 // This struct stores a thread that acts as a "root" for a contention
2558 // group. Contention groups are rooted at kmp_root threads, but also at
2559 // each primary thread of each team created in the teams construct.
2560 // This struct therefore also stores a thread_limit associated with
2561 // that contention group, and a counter to track the number of threads
2562 // active in that contention group. Each thread has a list of these: CG
2563 // root threads have an entry in their list in which cg_root refers to
2564 // the thread itself, whereas other workers in the CG will have a
2565 // single entry where cg_root is same as the entry containing their CG
2566 // root. When a thread encounters a teams construct, it will add a new
2567 // entry to the front of its list, because it now roots a new CG.
2568 typedef struct kmp_cg_root {
2569 kmp_info_p *cg_root; // "root" thread for a contention group
2570 // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2571 // thread_limit clause for teams primary threads
2572 kmp_int32 cg_thread_limit;
2573 kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2574 struct kmp_cg_root *up; // pointer to higher level CG root in list
2575 } kmp_cg_root_t;
2576
2577 // OpenMP thread data structures
2578
2579 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2580 /* Start with the readonly data which is cache aligned and padded. This is
2581 written before the thread starts working by the primary thread. Uber
2582 masters may update themselves later. Usage does not consider serialized
2583 regions. */
2584 kmp_desc_t th_info;
2585 kmp_team_p *th_team; /* team we belong to */
2586 kmp_root_p *th_root; /* pointer to root of task hierarchy */
2587 kmp_info_p *th_next_pool; /* next available thread in the pool */
2588 kmp_disp_t *th_dispatch; /* thread's dispatch data */
2589 int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2590
2591 /* The following are cached from the team info structure */
2592 /* TODO use these in more places as determined to be needed via profiling */
2593 int th_team_nproc; /* number of threads in a team */
2594 kmp_info_p *th_team_master; /* the team's primary thread */
2595 int th_team_serialized; /* team is serialized */
2596 microtask_t th_teams_microtask; /* save entry address for teams construct */
2597 int th_teams_level; /* save initial level of teams construct */
2598 /* it is 0 on device but may be any on host */
2599
2600 /* The blocktime info is copied from the team struct to the thread struct */
2601 /* at the start of a barrier, and the values stored in the team are used */
2602 /* at points in the code where the team struct is no longer guaranteed */
2603 /* to exist (from the POV of worker threads). */
2604 #if KMP_USE_MONITOR
2605 int th_team_bt_intervals;
2606 int th_team_bt_set;
2607 #else
2608 kmp_uint64 th_team_bt_intervals;
2609 #endif
2610
2611 #if KMP_AFFINITY_SUPPORTED
2612 kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2613 #endif
2614 omp_allocator_handle_t th_def_allocator; /* default allocator */
2615 /* The data set by the primary thread at reinit, then R/W by the worker */
2616 KMP_ALIGN_CACHE int
2617 th_set_nproc; /* if > 0, then only use this request for the next fork */
2618 #if KMP_NESTED_HOT_TEAMS
2619 kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2620 #endif
2621 kmp_proc_bind_t
2622 th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2623 kmp_teams_size_t
2624 th_teams_size; /* number of teams/threads in teams construct */
2625 #if KMP_AFFINITY_SUPPORTED
2626 int th_current_place; /* place currently bound to */
2627 int th_new_place; /* place to bind to in par reg */
2628 int th_first_place; /* first place in partition */
2629 int th_last_place; /* last place in partition */
2630 #endif
2631 int th_prev_level; /* previous level for affinity format */
2632 int th_prev_num_threads; /* previous num_threads for affinity format */
2633 #if USE_ITT_BUILD
2634 kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2635 kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2636 kmp_uint64 th_frame_time; /* frame timestamp */
2637 #endif /* USE_ITT_BUILD */
2638 kmp_local_t th_local;
2639 struct private_common *th_pri_head;
2640
2641 /* Now the data only used by the worker (after initial allocation) */
2642 /* TODO the first serial team should actually be stored in the info_t
2643 structure. this will help reduce initial allocation overhead */
2644 KMP_ALIGN_CACHE kmp_team_p
2645 *th_serial_team; /*serialized team held in reserve*/
2646
2647 #if OMPT_SUPPORT
2648 ompt_thread_info_t ompt_thread_info;
2649 #endif
2650
2651 /* The following are also read by the primary thread during reinit */
2652 struct common_table *th_pri_common;
2653
2654 volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2655 /* while awaiting queuing lock acquire */
2656
2657 volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2658 flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2659
2660 ident_t *th_ident;
2661 unsigned th_x; // Random number generator data
2662 unsigned th_a; // Random number generator data
2663
2664 /* Tasking-related data for the thread */
2665 kmp_task_team_t *th_task_team; // Task team struct
2666 kmp_taskdata_t *th_current_task; // Innermost Task being executed
2667 kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2668 kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2669 // at nested levels
2670 kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2671 kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2672 kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2673 // tasking, thus safe to reap
2674
2675 /* More stuff for keeping track of active/sleeping threads (this part is
2676 written by the worker thread) */
2677 kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2678 int th_active; // ! sleeping; 32 bits for TCR/TCW
2679 std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
2680 // 0 = not used in team; 1 = used in team;
2681 // 2 = transitioning to not used in team; 3 = transitioning to used in team
2682 struct cons_header *th_cons; // used for consistency check
2683 #if KMP_USE_HIER_SCHED
2684 // used for hierarchical scheduling
2685 kmp_hier_private_bdata_t *th_hier_bar_data;
2686 #endif
2687
2688 /* Add the syncronizing data which is cache aligned and padded. */
2689 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2690
2691 KMP_ALIGN_CACHE volatile kmp_int32
2692 th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2693
2694 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2695 #define NUM_LISTS 4
2696 kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2697 // allocation routines
2698 #endif
2699
2700 #if KMP_OS_WINDOWS
2701 kmp_win32_cond_t th_suspend_cv;
2702 kmp_win32_mutex_t th_suspend_mx;
2703 std::atomic<int> th_suspend_init;
2704 #endif
2705 #if KMP_OS_UNIX
2706 kmp_cond_align_t th_suspend_cv;
2707 kmp_mutex_align_t th_suspend_mx;
2708 std::atomic<int> th_suspend_init_count;
2709 #endif
2710
2711 #if USE_ITT_BUILD
2712 kmp_itt_mark_t th_itt_mark_single;
2713 // alignment ???
2714 #endif /* USE_ITT_BUILD */
2715 #if KMP_STATS_ENABLED
2716 kmp_stats_list *th_stats;
2717 #endif
2718 #if KMP_OS_UNIX
2719 std::atomic<bool> th_blocking;
2720 #endif
2721 kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2722 } kmp_base_info_t;
2723
2724 typedef union KMP_ALIGN_CACHE kmp_info {
2725 double th_align; /* use worst case alignment */
2726 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2727 kmp_base_info_t th;
2728 } kmp_info_t;
2729
2730 // OpenMP thread team data structures
2731
2732 typedef struct kmp_base_data {
2733 volatile kmp_uint32 t_value;
2734 } kmp_base_data_t;
2735
2736 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2737 double dt_align; /* use worst case alignment */
2738 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2739 kmp_base_data_t dt;
2740 } kmp_sleep_team_t;
2741
2742 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2743 double dt_align; /* use worst case alignment */
2744 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2745 kmp_base_data_t dt;
2746 } kmp_ordered_team_t;
2747
2748 typedef int (*launch_t)(int gtid);
2749
2750 /* Minimum number of ARGV entries to malloc if necessary */
2751 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2752
2753 // Set up how many argv pointers will fit in cache lines containing
2754 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2755 // larger value for more space between the primary write/worker read section and
2756 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2757 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2758 #define KMP_INLINE_ARGV_BYTES \
2759 (4 * CACHE_LINE - \
2760 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2761 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2762 CACHE_LINE))
2763 #else
2764 #define KMP_INLINE_ARGV_BYTES \
2765 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2766 #endif
2767 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2768
2769 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2770 // Synchronization Data
2771 // ---------------------------------------------------------------------------
2772 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2773 kmp_balign_team_t t_bar[bs_last_barrier];
2774 std::atomic<int> t_construct; // count of single directive encountered by team
2775 char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2776
2777 // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2778 std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2779 std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2780
2781 // Primary thread only
2782 // ---------------------------------------------------------------------------
2783 KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
2784 int t_master_this_cons; // "this_construct" single counter of primary thread
2785 // in parent team
2786 ident_t *t_ident; // if volatile, have to change too much other crud to
2787 // volatile too
2788 kmp_team_p *t_parent; // parent team
2789 kmp_team_p *t_next_pool; // next free team in the team pool
2790 kmp_disp_t *t_dispatch; // thread's dispatch data
2791 kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2792 kmp_proc_bind_t t_proc_bind; // bind type for par region
2793 #if USE_ITT_BUILD
2794 kmp_uint64 t_region_time; // region begin timestamp
2795 #endif /* USE_ITT_BUILD */
2796
2797 // Primary thread write, workers read
2798 // --------------------------------------------------------------------------
2799 KMP_ALIGN_CACHE void **t_argv;
2800 int t_argc;
2801 int t_nproc; // number of threads in team
2802 microtask_t t_pkfn;
2803 launch_t t_invoke; // procedure to launch the microtask
2804
2805 #if OMPT_SUPPORT
2806 ompt_team_info_t ompt_team_info;
2807 ompt_lw_taskteam_t *ompt_serialized_team_info;
2808 #endif
2809
2810 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2811 kmp_int8 t_fp_control_saved;
2812 kmp_int8 t_pad2b;
2813 kmp_int16 t_x87_fpu_control_word; // FP control regs
2814 kmp_uint32 t_mxcsr;
2815 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2816
2817 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2818
2819 KMP_ALIGN_CACHE kmp_info_t **t_threads;
2820 kmp_taskdata_t
2821 *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2822 int t_level; // nested parallel level
2823
2824 KMP_ALIGN_CACHE int t_max_argc;
2825 int t_max_nproc; // max threads this team can handle (dynamically expandable)
2826 int t_serialized; // levels deep of serialized teams
2827 dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2828 int t_id; // team's id, assigned by debugger.
2829 int t_active_level; // nested active parallel level
2830 kmp_r_sched_t t_sched; // run-time schedule for the team
2831 #if KMP_AFFINITY_SUPPORTED
2832 int t_first_place; // first & last place in parent thread's partition.
2833 int t_last_place; // Restore these values to primary thread after par region.
2834 #endif // KMP_AFFINITY_SUPPORTED
2835 int t_display_affinity;
2836 int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2837 // omp_set_num_threads() call
2838 omp_allocator_handle_t t_def_allocator; /* default allocator */
2839
2840 // Read/write by workers as well
2841 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2842 // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2843 // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2844 // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2845 // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2846 char dummy_padding[1024];
2847 #endif
2848 // Internal control stack for additional nested teams.
2849 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2850 // for SERIALIZED teams nested 2 or more levels deep
2851 // typed flag to store request state of cancellation
2852 std::atomic<kmp_int32> t_cancel_request;
2853 int t_master_active; // save on fork, restore on join
2854 void *t_copypriv_data; // team specific pointer to copyprivate data array
2855 #if KMP_OS_WINDOWS
2856 std::atomic<kmp_uint32> t_copyin_counter;
2857 #endif
2858 #if USE_ITT_BUILD
2859 void *t_stack_id; // team specific stack stitching id (for ittnotify)
2860 #endif /* USE_ITT_BUILD */
2861 distributedBarrier *b; // Distributed barrier data associated with team
2862 } kmp_base_team_t;
2863
2864 union KMP_ALIGN_CACHE kmp_team {
2865 kmp_base_team_t t;
2866 double t_align; /* use worst case alignment */
2867 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2868 };
2869
2870 typedef union KMP_ALIGN_CACHE kmp_time_global {
2871 double dt_align; /* use worst case alignment */
2872 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2873 kmp_base_data_t dt;
2874 } kmp_time_global_t;
2875
2876 typedef struct kmp_base_global {
2877 /* cache-aligned */
2878 kmp_time_global_t g_time;
2879
2880 /* non cache-aligned */
2881 volatile int g_abort;
2882 volatile int g_done;
2883
2884 int g_dynamic;
2885 enum dynamic_mode g_dynamic_mode;
2886 } kmp_base_global_t;
2887
2888 typedef union KMP_ALIGN_CACHE kmp_global {
2889 kmp_base_global_t g;
2890 double g_align; /* use worst case alignment */
2891 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2892 } kmp_global_t;
2893
2894 typedef struct kmp_base_root {
2895 // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2896 // (r_in_parallel>= 0)
2897 // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2898 // the synch overhead or keeping r_active
2899 volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2900 // keeps a count of active parallel regions per root
2901 std::atomic<int> r_in_parallel;
2902 // GEH: This is misnamed, should be r_active_levels
2903 kmp_team_t *r_root_team;
2904 kmp_team_t *r_hot_team;
2905 kmp_info_t *r_uber_thread;
2906 kmp_lock_t r_begin_lock;
2907 volatile int r_begin;
2908 int r_blocktime; /* blocktime for this root and descendants */
2909 #if KMP_AFFINITY_SUPPORTED
2910 int r_affinity_assigned;
2911 #endif // KMP_AFFINITY_SUPPORTED
2912 } kmp_base_root_t;
2913
2914 typedef union KMP_ALIGN_CACHE kmp_root {
2915 kmp_base_root_t r;
2916 double r_align; /* use worst case alignment */
2917 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2918 } kmp_root_t;
2919
2920 struct fortran_inx_info {
2921 kmp_int32 data;
2922 };
2923
2924 /* ------------------------------------------------------------------------ */
2925
2926 extern int __kmp_settings;
2927 extern int __kmp_duplicate_library_ok;
2928 #if USE_ITT_BUILD
2929 extern int __kmp_forkjoin_frames;
2930 extern int __kmp_forkjoin_frames_mode;
2931 #endif
2932 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2933 extern int __kmp_determ_red;
2934
2935 #ifdef KMP_DEBUG
2936 extern int kmp_a_debug;
2937 extern int kmp_b_debug;
2938 extern int kmp_c_debug;
2939 extern int kmp_d_debug;
2940 extern int kmp_e_debug;
2941 extern int kmp_f_debug;
2942 #endif /* KMP_DEBUG */
2943
2944 /* For debug information logging using rotating buffer */
2945 #define KMP_DEBUG_BUF_LINES_INIT 512
2946 #define KMP_DEBUG_BUF_LINES_MIN 1
2947
2948 #define KMP_DEBUG_BUF_CHARS_INIT 128
2949 #define KMP_DEBUG_BUF_CHARS_MIN 2
2950
2951 extern int
2952 __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2953 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2954 extern int
2955 __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2956 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2957 entry pointer */
2958
2959 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2960 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2961 printed in buffer so far */
2962 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2963 recommended in warnings */
2964 /* end rotating debug buffer */
2965
2966 #ifdef KMP_DEBUG
2967 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2968
2969 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2970 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2971 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2972 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2973 extern int __kmp_par_range_lb;
2974 extern int __kmp_par_range_ub;
2975 #endif
2976
2977 /* For printing out dynamic storage map for threads and teams */
2978 extern int
2979 __kmp_storage_map; /* True means print storage map for threads and teams */
2980 extern int __kmp_storage_map_verbose; /* True means storage map includes
2981 placement info */
2982 extern int __kmp_storage_map_verbose_specified;
2983
2984 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2985 extern kmp_cpuinfo_t __kmp_cpuinfo;
2986 #endif
2987
2988 extern volatile int __kmp_init_serial;
2989 extern volatile int __kmp_init_gtid;
2990 extern volatile int __kmp_init_common;
2991 extern volatile int __kmp_init_middle;
2992 extern volatile int __kmp_init_parallel;
2993 #if KMP_USE_MONITOR
2994 extern volatile int __kmp_init_monitor;
2995 #endif
2996 extern volatile int __kmp_init_user_locks;
2997 extern volatile int __kmp_init_hidden_helper_threads;
2998 extern int __kmp_init_counter;
2999 extern int __kmp_root_counter;
3000 extern int __kmp_version;
3001
3002 /* list of address of allocated caches for commons */
3003 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3004
3005 /* Barrier algorithm types and options */
3006 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3007 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3008 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3009 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3010 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3011 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3012 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3013 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3014 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3015 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3016 extern char const *__kmp_barrier_type_name[bs_last_barrier];
3017 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3018
3019 /* Global Locks */
3020 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3021 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3022 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3023 extern kmp_bootstrap_lock_t
3024 __kmp_exit_lock; /* exit() is not always thread-safe */
3025 #if KMP_USE_MONITOR
3026 extern kmp_bootstrap_lock_t
3027 __kmp_monitor_lock; /* control monitor thread creation */
3028 #endif
3029 extern kmp_bootstrap_lock_t
3030 __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3031 __kmp_threads expansion to co-exist */
3032
3033 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
3034 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
3035 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3036
3037 extern enum library_type __kmp_library;
3038
3039 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3040 extern enum sched_type __kmp_static; /* default static scheduling method */
3041 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3042 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3043 extern int __kmp_chunk; /* default runtime chunk size */
3044 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3045
3046 extern size_t __kmp_stksize; /* stack size per thread */
3047 #if KMP_USE_MONITOR
3048 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3049 #endif
3050 extern size_t __kmp_stkoffset; /* stack offset per thread */
3051 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3052
3053 extern size_t
3054 __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3055 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3056 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3057 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
3058 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3059 extern int __kmp_generate_warnings; /* should we issue warnings? */
3060 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3061
3062 #ifdef DEBUG_SUSPEND
3063 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3064 #endif
3065
3066 extern kmp_int32 __kmp_use_yield;
3067 extern kmp_int32 __kmp_use_yield_exp_set;
3068 extern kmp_uint32 __kmp_yield_init;
3069 extern kmp_uint32 __kmp_yield_next;
3070
3071 /* ------------------------------------------------------------------------- */
3072 extern int __kmp_allThreadsSpecified;
3073
3074 extern size_t __kmp_align_alloc;
3075 /* following data protected by initialization routines */
3076 extern int __kmp_xproc; /* number of processors in the system */
3077 extern int __kmp_avail_proc; /* number of processors available to the process */
3078 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3079 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3080 // maximum total number of concurrently-existing threads on device
3081 extern int __kmp_max_nth;
3082 // maximum total number of concurrently-existing threads in a contention group
3083 extern int __kmp_cg_max_nth;
3084 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3085 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3086 __kmp_root */
3087 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3088 region a la OMP_NUM_THREADS */
3089 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3090 initialization */
3091 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3092 used (fixed) */
3093 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3094 (__kmpc_threadprivate_cached()) */
3095 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
3096 blocking (env setting) */
3097 #if KMP_USE_MONITOR
3098 extern int
3099 __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3100 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3101 blocking */
3102 #endif
3103 #ifdef KMP_ADJUST_BLOCKTIME
3104 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3105 #endif /* KMP_ADJUST_BLOCKTIME */
3106 #ifdef KMP_DFLT_NTH_CORES
3107 extern int __kmp_ncores; /* Total number of cores for threads placement */
3108 #endif
3109 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3110 extern int __kmp_abort_delay;
3111
3112 extern int __kmp_need_register_atfork_specified;
3113 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3114 to install fork handler */
3115 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3116 0 - not set, will be set at runtime
3117 1 - using stack search
3118 2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3119 X*) or TlsGetValue(Windows* OS))
3120 3 - static TLS (__declspec(thread) __kmp_gtid),
3121 Linux* OS .so only. */
3122 extern int
3123 __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3124 #ifdef KMP_TDATA_GTID
3125 extern KMP_THREAD_LOCAL int __kmp_gtid;
3126 #endif
3127 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3128 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3129 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3130 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3131 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3132 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3133 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3134
3135 // max_active_levels for nested parallelism enabled by default via
3136 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3137 extern int __kmp_dflt_max_active_levels;
3138 // Indicates whether value of __kmp_dflt_max_active_levels was already
3139 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3140 extern bool __kmp_dflt_max_active_levels_set;
3141 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3142 concurrent execution per team */
3143 #if KMP_NESTED_HOT_TEAMS
3144 extern int __kmp_hot_teams_mode;
3145 extern int __kmp_hot_teams_max_level;
3146 #endif
3147
3148 #if KMP_OS_LINUX
3149 extern enum clock_function_type __kmp_clock_function;
3150 extern int __kmp_clock_function_param;
3151 #endif /* KMP_OS_LINUX */
3152
3153 #if KMP_MIC_SUPPORTED
3154 extern enum mic_type __kmp_mic_type;
3155 #endif
3156
3157 #ifdef USE_LOAD_BALANCE
3158 extern double __kmp_load_balance_interval; // load balance algorithm interval
3159 #endif /* USE_LOAD_BALANCE */
3160
3161 // OpenMP 3.1 - Nested num threads array
3162 typedef struct kmp_nested_nthreads_t {
3163 int *nth;
3164 int size;
3165 int used;
3166 } kmp_nested_nthreads_t;
3167
3168 extern kmp_nested_nthreads_t __kmp_nested_nth;
3169
3170 #if KMP_USE_ADAPTIVE_LOCKS
3171
3172 // Parameters for the speculative lock backoff system.
3173 struct kmp_adaptive_backoff_params_t {
3174 // Number of soft retries before it counts as a hard retry.
3175 kmp_uint32 max_soft_retries;
3176 // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3177 // the right
3178 kmp_uint32 max_badness;
3179 };
3180
3181 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3182
3183 #if KMP_DEBUG_ADAPTIVE_LOCKS
3184 extern const char *__kmp_speculative_statsfile;
3185 #endif
3186
3187 #endif // KMP_USE_ADAPTIVE_LOCKS
3188
3189 extern int __kmp_display_env; /* TRUE or FALSE */
3190 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3191 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3192 extern int __kmp_nteams;
3193 extern int __kmp_teams_thread_limit;
3194
3195 /* ------------------------------------------------------------------------- */
3196
3197 /* the following are protected by the fork/join lock */
3198 /* write: lock read: anytime */
3199 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3200 /* read/write: lock */
3201 extern volatile kmp_team_t *__kmp_team_pool;
3202 extern volatile kmp_info_t *__kmp_thread_pool;
3203 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3204
3205 // total num threads reachable from some root thread including all root threads
3206 extern volatile int __kmp_nth;
3207 /* total number of threads reachable from some root thread including all root
3208 threads, and those in the thread pool */
3209 extern volatile int __kmp_all_nth;
3210 extern std::atomic<int> __kmp_thread_pool_active_nth;
3211
3212 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3213 /* end data protected by fork/join lock */
3214 /* ------------------------------------------------------------------------- */
3215
3216 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3217 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3218 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3219 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3220 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3221
3222 // AT: Which way is correct?
3223 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3224 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3225 #define __kmp_get_team_num_threads(gtid) \
3226 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3227
KMP_UBER_GTID(int gtid)3228 static inline bool KMP_UBER_GTID(int gtid) {
3229 KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3230 KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3231 return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3232 __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3233 }
3234
__kmp_tid_from_gtid(int gtid)3235 static inline int __kmp_tid_from_gtid(int gtid) {
3236 KMP_DEBUG_ASSERT(gtid >= 0);
3237 return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3238 }
3239
__kmp_gtid_from_tid(int tid,const kmp_team_t * team)3240 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3241 KMP_DEBUG_ASSERT(tid >= 0 && team);
3242 return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3243 }
3244
__kmp_gtid_from_thread(const kmp_info_t * thr)3245 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3246 KMP_DEBUG_ASSERT(thr);
3247 return thr->th.th_info.ds.ds_gtid;
3248 }
3249
__kmp_thread_from_gtid(int gtid)3250 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3251 KMP_DEBUG_ASSERT(gtid >= 0);
3252 return __kmp_threads[gtid];
3253 }
3254
__kmp_team_from_gtid(int gtid)3255 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3256 KMP_DEBUG_ASSERT(gtid >= 0);
3257 return __kmp_threads[gtid]->th.th_team;
3258 }
3259
__kmp_assert_valid_gtid(kmp_int32 gtid)3260 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3261 if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3262 KMP_FATAL(ThreadIdentInvalid);
3263 }
3264
3265 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3266 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3267 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3268 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3269 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3270 #endif
3271
3272 /* ------------------------------------------------------------------------- */
3273
3274 extern kmp_global_t __kmp_global; /* global status */
3275
3276 extern kmp_info_t __kmp_monitor;
3277 // For Debugging Support Library
3278 extern std::atomic<kmp_int32> __kmp_team_counter;
3279 // For Debugging Support Library
3280 extern std::atomic<kmp_int32> __kmp_task_counter;
3281
3282 #if USE_DEBUGGER
3283 #define _KMP_GEN_ID(counter) \
3284 (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3285 #else
3286 #define _KMP_GEN_ID(counter) (~0)
3287 #endif /* USE_DEBUGGER */
3288
3289 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3290 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3291
3292 /* ------------------------------------------------------------------------ */
3293
3294 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3295 size_t size, char const *format, ...);
3296
3297 extern void __kmp_serial_initialize(void);
3298 extern void __kmp_middle_initialize(void);
3299 extern void __kmp_parallel_initialize(void);
3300
3301 extern void __kmp_internal_begin(void);
3302 extern void __kmp_internal_end_library(int gtid);
3303 extern void __kmp_internal_end_thread(int gtid);
3304 extern void __kmp_internal_end_atexit(void);
3305 extern void __kmp_internal_end_dtor(void);
3306 extern void __kmp_internal_end_dest(void *);
3307
3308 extern int __kmp_register_root(int initial_thread);
3309 extern void __kmp_unregister_root(int gtid);
3310 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3311
3312 extern int __kmp_ignore_mppbeg(void);
3313 extern int __kmp_ignore_mppend(void);
3314
3315 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3316 extern void __kmp_exit_single(int gtid);
3317
3318 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3319 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3320
3321 #ifdef USE_LOAD_BALANCE
3322 extern int __kmp_get_load_balance(int);
3323 #endif
3324
3325 extern int __kmp_get_global_thread_id(void);
3326 extern int __kmp_get_global_thread_id_reg(void);
3327 extern void __kmp_exit_thread(int exit_status);
3328 extern void __kmp_abort(char const *format, ...);
3329 extern void __kmp_abort_thread(void);
3330 KMP_NORETURN extern void __kmp_abort_process(void);
3331 extern void __kmp_warn(char const *format, ...);
3332
3333 extern void __kmp_set_num_threads(int new_nth, int gtid);
3334
3335 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3336 // registered.
__kmp_entry_thread()3337 static inline kmp_info_t *__kmp_entry_thread() {
3338 int gtid = __kmp_entry_gtid();
3339
3340 return __kmp_threads[gtid];
3341 }
3342
3343 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3344 extern int __kmp_get_max_active_levels(int gtid);
3345 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3346 extern int __kmp_get_team_size(int gtid, int level);
3347 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3348 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3349
3350 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3351 extern void __kmp_init_random(kmp_info_t *thread);
3352
3353 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3354 extern void __kmp_adjust_num_threads(int new_nproc);
3355 extern void __kmp_check_stksize(size_t *val);
3356
3357 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3358 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3359 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3360 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3361 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3362 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3363
3364 #if USE_FAST_MEMORY
3365 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3366 size_t size KMP_SRC_LOC_DECL);
3367 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3368 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3369 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3370 #define __kmp_fast_allocate(this_thr, size) \
3371 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3372 #define __kmp_fast_free(this_thr, ptr) \
3373 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3374 #endif
3375
3376 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3377 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3378 size_t elsize KMP_SRC_LOC_DECL);
3379 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3380 size_t size KMP_SRC_LOC_DECL);
3381 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3382 #define __kmp_thread_malloc(th, size) \
3383 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3384 #define __kmp_thread_calloc(th, nelem, elsize) \
3385 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3386 #define __kmp_thread_realloc(th, ptr, size) \
3387 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3388 #define __kmp_thread_free(th, ptr) \
3389 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3390
3391 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3392 #define KMP_INTERNAL_FREE(p) free(p)
3393 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3394 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3395
3396 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3397
3398 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3399 kmp_proc_bind_t proc_bind);
3400 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3401 int num_threads);
3402 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3403 int num_teams_ub, int num_threads);
3404
3405 extern void __kmp_yield();
3406
3407 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3408 enum sched_type schedule, kmp_int32 lb,
3409 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3410 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3411 enum sched_type schedule, kmp_uint32 lb,
3412 kmp_uint32 ub, kmp_int32 st,
3413 kmp_int32 chunk);
3414 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3415 enum sched_type schedule, kmp_int64 lb,
3416 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3417 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3418 enum sched_type schedule, kmp_uint64 lb,
3419 kmp_uint64 ub, kmp_int64 st,
3420 kmp_int64 chunk);
3421
3422 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3423 kmp_int32 *p_last, kmp_int32 *p_lb,
3424 kmp_int32 *p_ub, kmp_int32 *p_st);
3425 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3426 kmp_int32 *p_last, kmp_uint32 *p_lb,
3427 kmp_uint32 *p_ub, kmp_int32 *p_st);
3428 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3429 kmp_int32 *p_last, kmp_int64 *p_lb,
3430 kmp_int64 *p_ub, kmp_int64 *p_st);
3431 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3432 kmp_int32 *p_last, kmp_uint64 *p_lb,
3433 kmp_uint64 *p_ub, kmp_int64 *p_st);
3434
3435 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3436 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3437 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3438 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3439
3440 #ifdef KMP_GOMP_COMPAT
3441
3442 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3443 enum sched_type schedule, kmp_int32 lb,
3444 kmp_int32 ub, kmp_int32 st,
3445 kmp_int32 chunk, int push_ws);
3446 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3447 enum sched_type schedule, kmp_uint32 lb,
3448 kmp_uint32 ub, kmp_int32 st,
3449 kmp_int32 chunk, int push_ws);
3450 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3451 enum sched_type schedule, kmp_int64 lb,
3452 kmp_int64 ub, kmp_int64 st,
3453 kmp_int64 chunk, int push_ws);
3454 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3455 enum sched_type schedule, kmp_uint64 lb,
3456 kmp_uint64 ub, kmp_int64 st,
3457 kmp_int64 chunk, int push_ws);
3458 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3459 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3460 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3461 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3462
3463 #endif /* KMP_GOMP_COMPAT */
3464
3465 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3466 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3467 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3468 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3469 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3470 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3471 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3472 void *obj);
3473 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3474 kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3475
3476 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3477 int final_spin
3478 #if USE_ITT_BUILD
3479 ,
3480 void *itt_sync_obj
3481 #endif
3482 );
3483 extern void __kmp_release_64(kmp_flag_64<> *flag);
3484
3485 extern void __kmp_infinite_loop(void);
3486
3487 extern void __kmp_cleanup(void);
3488
3489 #if KMP_HANDLE_SIGNALS
3490 extern int __kmp_handle_signals;
3491 extern void __kmp_install_signals(int parallel_init);
3492 extern void __kmp_remove_signals(void);
3493 #endif
3494
3495 extern void __kmp_clear_system_time(void);
3496 extern void __kmp_read_system_time(double *delta);
3497
3498 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3499
3500 extern void __kmp_expand_host_name(char *buffer, size_t size);
3501 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3502
3503 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && KMP_ARCH_AARCH64)
3504 extern void
3505 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3506 #endif
3507
3508 extern void
3509 __kmp_runtime_initialize(void); /* machine specific initialization */
3510 extern void __kmp_runtime_destroy(void);
3511
3512 #if KMP_AFFINITY_SUPPORTED
3513 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3514 kmp_affin_mask_t *mask);
3515 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3516 kmp_affin_mask_t *mask);
3517 extern void __kmp_affinity_initialize(void);
3518 extern void __kmp_affinity_uninitialize(void);
3519 extern void __kmp_affinity_set_init_mask(
3520 int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3521 extern void __kmp_affinity_set_place(int gtid);
3522 extern void __kmp_affinity_determine_capable(const char *env_var);
3523 extern int __kmp_aux_set_affinity(void **mask);
3524 extern int __kmp_aux_get_affinity(void **mask);
3525 extern int __kmp_aux_get_affinity_max_proc();
3526 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3527 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3528 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3529 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3530 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3531 extern int kmp_set_thread_affinity_mask_initial(void);
3532 #endif
__kmp_assign_root_init_mask()3533 static inline void __kmp_assign_root_init_mask() {
3534 int gtid = __kmp_entry_gtid();
3535 kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3536 if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3537 __kmp_affinity_set_init_mask(gtid, TRUE);
3538 r->r.r_affinity_assigned = TRUE;
3539 }
3540 }
3541 #else /* KMP_AFFINITY_SUPPORTED */
3542 #define __kmp_assign_root_init_mask() /* Nothing */
3543 #endif /* KMP_AFFINITY_SUPPORTED */
3544 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3545 // format string is for affinity, so platforms that do not support
3546 // affinity can still use the other fields, e.g., %n for num_threads
3547 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3548 kmp_str_buf_t *buffer);
3549 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3550
3551 extern void __kmp_cleanup_hierarchy();
3552 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3553
3554 #if KMP_USE_FUTEX
3555
3556 extern int __kmp_futex_determine_capable(void);
3557
3558 #endif // KMP_USE_FUTEX
3559
3560 extern void __kmp_gtid_set_specific(int gtid);
3561 extern int __kmp_gtid_get_specific(void);
3562
3563 extern double __kmp_read_cpu_time(void);
3564
3565 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3566
3567 #if KMP_USE_MONITOR
3568 extern void __kmp_create_monitor(kmp_info_t *th);
3569 #endif
3570
3571 extern void *__kmp_launch_thread(kmp_info_t *thr);
3572
3573 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3574
3575 #if KMP_OS_WINDOWS
3576 extern int __kmp_still_running(kmp_info_t *th);
3577 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3578 extern void __kmp_free_handle(kmp_thread_t tHandle);
3579 #endif
3580
3581 #if KMP_USE_MONITOR
3582 extern void __kmp_reap_monitor(kmp_info_t *th);
3583 #endif
3584 extern void __kmp_reap_worker(kmp_info_t *th);
3585 extern void __kmp_terminate_thread(int gtid);
3586
3587 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3588 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3589 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3590
3591 extern void __kmp_elapsed(double *);
3592 extern void __kmp_elapsed_tick(double *);
3593
3594 extern void __kmp_enable(int old_state);
3595 extern void __kmp_disable(int *old_state);
3596
3597 extern void __kmp_thread_sleep(int millis);
3598
3599 extern void __kmp_common_initialize(void);
3600 extern void __kmp_common_destroy(void);
3601 extern void __kmp_common_destroy_gtid(int gtid);
3602
3603 #if KMP_OS_UNIX
3604 extern void __kmp_register_atfork(void);
3605 #endif
3606 extern void __kmp_suspend_initialize(void);
3607 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3608 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3609
3610 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3611 int tid);
3612 extern kmp_team_t *
3613 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3614 #if OMPT_SUPPORT
3615 ompt_data_t ompt_parallel_data,
3616 #endif
3617 kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3618 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3619 extern void __kmp_free_thread(kmp_info_t *);
3620 extern void __kmp_free_team(kmp_root_t *,
3621 kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3622 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3623
3624 /* ------------------------------------------------------------------------ */
3625
3626 extern void __kmp_initialize_bget(kmp_info_t *th);
3627 extern void __kmp_finalize_bget(kmp_info_t *th);
3628
3629 KMP_EXPORT void *kmpc_malloc(size_t size);
3630 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3631 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3632 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3633 KMP_EXPORT void kmpc_free(void *ptr);
3634
3635 /* declarations for internal use */
3636
3637 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3638 size_t reduce_size, void *reduce_data,
3639 void (*reduce)(void *, void *));
3640 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3641 extern int __kmp_barrier_gomp_cancel(int gtid);
3642
3643 /*!
3644 * Tell the fork call which compiler generated the fork call, and therefore how
3645 * to deal with the call.
3646 */
3647 enum fork_context_e {
3648 fork_context_gnu, /**< Called from GNU generated code, so must not invoke the
3649 microtask internally. */
3650 fork_context_intel, /**< Called from Intel generated code. */
3651 fork_context_last
3652 };
3653 extern int __kmp_fork_call(ident_t *loc, int gtid,
3654 enum fork_context_e fork_context, kmp_int32 argc,
3655 microtask_t microtask, launch_t invoker,
3656 kmp_va_list ap);
3657
3658 extern void __kmp_join_call(ident_t *loc, int gtid
3659 #if OMPT_SUPPORT
3660 ,
3661 enum fork_context_e fork_context
3662 #endif
3663 ,
3664 int exit_teams = 0);
3665
3666 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3667 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3668 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3669 extern int __kmp_invoke_task_func(int gtid);
3670 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3671 kmp_info_t *this_thr,
3672 kmp_team_t *team);
3673 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3674 kmp_info_t *this_thr,
3675 kmp_team_t *team);
3676
3677 // should never have been exported
3678 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3679 extern int __kmp_invoke_teams_master(int gtid);
3680 extern void __kmp_teams_master(int gtid);
3681 extern int __kmp_aux_get_team_num();
3682 extern int __kmp_aux_get_num_teams();
3683 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3684 extern void __kmp_user_set_library(enum library_type arg);
3685 extern void __kmp_aux_set_library(enum library_type arg);
3686 extern void __kmp_aux_set_stacksize(size_t arg);
3687 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3688 extern void __kmp_aux_set_defaults(char const *str, size_t len);
3689
3690 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3691 void kmpc_set_blocktime(int arg);
3692 void ompc_set_nested(int flag);
3693 void ompc_set_dynamic(int flag);
3694 void ompc_set_num_threads(int arg);
3695
3696 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3697 kmp_team_t *team, int tid);
3698 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3699 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3700 kmp_tasking_flags_t *flags,
3701 size_t sizeof_kmp_task_t,
3702 size_t sizeof_shareds,
3703 kmp_routine_entry_t task_entry);
3704 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3705 kmp_team_t *team, int tid,
3706 int set_curr_task);
3707 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3708 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3709
3710 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3711 int gtid,
3712 kmp_task_t *task);
3713 extern void __kmp_fulfill_event(kmp_event_t *event);
3714
3715 extern void __kmp_free_task_team(kmp_info_t *thread,
3716 kmp_task_team_t *task_team);
3717 extern void __kmp_reap_task_teams(void);
3718 extern void __kmp_wait_to_unref_task_teams(void);
3719 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3720 int always);
3721 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3722 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3723 #if USE_ITT_BUILD
3724 ,
3725 void *itt_sync_obj
3726 #endif /* USE_ITT_BUILD */
3727 ,
3728 int wait = 1);
3729 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3730 int gtid);
3731
3732 extern int __kmp_is_address_mapped(void *addr);
3733 extern kmp_uint64 __kmp_hardware_timestamp(void);
3734
3735 #if KMP_OS_UNIX
3736 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3737 #endif
3738
3739 /* ------------------------------------------------------------------------ */
3740 //
3741 // Assembly routines that have no compiler intrinsic replacement
3742 //
3743
3744 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3745 void *argv[]
3746 #if OMPT_SUPPORT
3747 ,
3748 void **exit_frame_ptr
3749 #endif
3750 );
3751
3752 /* ------------------------------------------------------------------------ */
3753
3754 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3755 KMP_EXPORT void __kmpc_end(ident_t *);
3756
3757 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3758 kmpc_ctor_vec ctor,
3759 kmpc_cctor_vec cctor,
3760 kmpc_dtor_vec dtor,
3761 size_t vector_length);
3762 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3763 kmpc_ctor ctor, kmpc_cctor cctor,
3764 kmpc_dtor dtor);
3765 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3766 void *data, size_t size);
3767
3768 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3769 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3770 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3771 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3772
3773 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3774 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3775 kmpc_micro microtask, ...);
3776
3777 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3778 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3779
3780 KMP_EXPORT void __kmpc_flush(ident_t *);
3781 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3782 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3783 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3784 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
3785 kmp_int32 filter);
3786 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
3787 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3788 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3789 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3790 kmp_critical_name *);
3791 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3792 kmp_critical_name *);
3793 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3794 kmp_critical_name *, uint32_t hint);
3795
3796 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3797 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3798
3799 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3800 kmp_int32 global_tid);
3801
3802 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3803 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3804
3805 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3806 kmp_int32 schedtype, kmp_int32 *plastiter,
3807 kmp_int *plower, kmp_int *pupper,
3808 kmp_int *pstride, kmp_int incr,
3809 kmp_int chunk);
3810
3811 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3812
3813 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3814 size_t cpy_size, void *cpy_data,
3815 void (*cpy_func)(void *, void *),
3816 kmp_int32 didit);
3817
3818 extern void KMPC_SET_NUM_THREADS(int arg);
3819 extern void KMPC_SET_DYNAMIC(int flag);
3820 extern void KMPC_SET_NESTED(int flag);
3821
3822 /* OMP 3.0 tasking interface routines */
3823 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3824 kmp_task_t *new_task);
3825 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3826 kmp_int32 flags,
3827 size_t sizeof_kmp_task_t,
3828 size_t sizeof_shareds,
3829 kmp_routine_entry_t task_entry);
3830 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
3831 ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
3832 size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
3833 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3834 kmp_task_t *task);
3835 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3836 kmp_task_t *task);
3837 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3838 kmp_task_t *new_task);
3839 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3840
3841 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3842 int end_part);
3843
3844 #if TASK_UNUSED
3845 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3846 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3847 kmp_task_t *task);
3848 #endif // TASK_UNUSED
3849
3850 /* ------------------------------------------------------------------------ */
3851
3852 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3853 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3854
3855 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3856 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3857 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3858 kmp_depend_info_t *noalias_dep_list);
3859 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3860 kmp_int32 ndeps,
3861 kmp_depend_info_t *dep_list,
3862 kmp_int32 ndeps_noalias,
3863 kmp_depend_info_t *noalias_dep_list);
3864 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3865 bool serialize_immediate);
3866
3867 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3868 kmp_int32 cncl_kind);
3869 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3870 kmp_int32 cncl_kind);
3871 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3872 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3873
3874 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3875 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3876 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3877 kmp_int32 if_val, kmp_uint64 *lb,
3878 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3879 kmp_int32 sched, kmp_uint64 grainsize,
3880 void *task_dup);
3881 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
3882 kmp_task_t *task, kmp_int32 if_val,
3883 kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
3884 kmp_int32 nogroup, kmp_int32 sched,
3885 kmp_uint64 grainsize, kmp_int32 modifier,
3886 void *task_dup);
3887 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3888 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
3889 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3890 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
3891 int is_ws, int num,
3892 void *data);
3893 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
3894 int num, void *data);
3895 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
3896 int is_ws);
3897 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
3898 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3899 kmp_task_affinity_info_t *affin_list);
3900 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
3901 KMP_EXPORT int __kmp_get_max_teams(void);
3902 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
3903 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
3904
3905 /* Lock interface routines (fast versions with gtid passed in) */
3906 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3907 void **user_lock);
3908 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3909 void **user_lock);
3910 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3911 void **user_lock);
3912 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3913 void **user_lock);
3914 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3915 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3916 void **user_lock);
3917 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3918 void **user_lock);
3919 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3920 void **user_lock);
3921 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3922 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3923 void **user_lock);
3924
3925 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3926 void **user_lock, uintptr_t hint);
3927 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3928 void **user_lock,
3929 uintptr_t hint);
3930
3931 /* Interface to fast scalable reduce methods routines */
3932
3933 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3934 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3935 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3936 kmp_critical_name *lck);
3937 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3938 kmp_critical_name *lck);
3939 KMP_EXPORT kmp_int32 __kmpc_reduce(
3940 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3941 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3942 kmp_critical_name *lck);
3943 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3944 kmp_critical_name *lck);
3945
3946 /* Internal fast reduction routines */
3947
3948 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3949 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3950 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3951 kmp_critical_name *lck);
3952
3953 // this function is for testing set/get/determine reduce method
3954 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3955
3956 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3957 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3958
3959 // C++ port
3960 // missing 'extern "C"' declarations
3961
3962 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3963 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3964 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3965 kmp_int32 num_threads);
3966
3967 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3968 int proc_bind);
3969 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3970 kmp_int32 num_teams,
3971 kmp_int32 num_threads);
3972 /* Function for OpenMP 5.1 num_teams clause */
3973 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
3974 kmp_int32 num_teams_lb,
3975 kmp_int32 num_teams_ub,
3976 kmp_int32 num_threads);
3977 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3978 kmpc_micro microtask, ...);
3979 struct kmp_dim { // loop bounds info casted to kmp_int64
3980 kmp_int64 lo; // lower
3981 kmp_int64 up; // upper
3982 kmp_int64 st; // stride
3983 };
3984 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
3985 kmp_int32 num_dims,
3986 const struct kmp_dim *dims);
3987 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
3988 const kmp_int64 *vec);
3989 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
3990 const kmp_int64 *vec);
3991 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
3992
3993 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
3994 void *data, size_t size,
3995 void ***cache);
3996
3997 // Symbols for MS mutual detection.
3998 extern int _You_must_link_with_exactly_one_OpenMP_library;
3999 extern int _You_must_link_with_Intel_OpenMP_library;
4000 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
4001 extern int _You_must_link_with_Microsoft_OpenMP_library;
4002 #endif
4003
4004 // The routines below are not exported.
4005 // Consider making them 'static' in corresponding source files.
4006 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4007 void *data_addr, size_t pc_size);
4008 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4009 void *data_addr,
4010 size_t pc_size);
4011 void __kmp_threadprivate_resize_cache(int newCapacity);
4012 void __kmp_cleanup_threadprivate_caches();
4013
4014 // ompc_, kmpc_ entries moved from omp.h.
4015 #if KMP_OS_WINDOWS
4016 #define KMPC_CONVENTION __cdecl
4017 #else
4018 #define KMPC_CONVENTION
4019 #endif
4020
4021 #ifndef __OMP_H
4022 typedef enum omp_sched_t {
4023 omp_sched_static = 1,
4024 omp_sched_dynamic = 2,
4025 omp_sched_guided = 3,
4026 omp_sched_auto = 4
4027 } omp_sched_t;
4028 typedef void *kmp_affinity_mask_t;
4029 #endif
4030
4031 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4032 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4033 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4034 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4035 KMP_EXPORT int KMPC_CONVENTION
4036 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4037 KMP_EXPORT int KMPC_CONVENTION
4038 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4039 KMP_EXPORT int KMPC_CONVENTION
4040 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4041
4042 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4043 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4044 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4045 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4046 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4047 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4048 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4049 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4050 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4051 char const *format);
4052
4053 enum kmp_target_offload_kind {
4054 tgt_disabled = 0,
4055 tgt_default = 1,
4056 tgt_mandatory = 2
4057 };
4058 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4059 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4060 extern kmp_target_offload_kind_t __kmp_target_offload;
4061 extern int __kmpc_get_target_offload();
4062
4063 // Constants used in libomptarget
4064 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4065 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4066
4067 // OMP Pause Resource
4068
4069 // The following enum is used both to set the status in __kmp_pause_status, and
4070 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4071 typedef enum kmp_pause_status_t {
4072 kmp_not_paused = 0, // status is not paused, or, requesting resume
4073 kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4074 kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4075 } kmp_pause_status_t;
4076
4077 // This stores the pause state of the runtime
4078 extern kmp_pause_status_t __kmp_pause_status;
4079 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4080 extern int __kmp_pause_resource(kmp_pause_status_t level);
4081 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4082 extern void __kmp_resume_if_soft_paused();
4083 // Hard resume simply resets the status to not paused. Library will appear to
4084 // be uninitialized after hard pause. Let OMP constructs trigger required
4085 // initializations.
__kmp_resume_if_hard_paused()4086 static inline void __kmp_resume_if_hard_paused() {
4087 if (__kmp_pause_status == kmp_hard_paused) {
4088 __kmp_pause_status = kmp_not_paused;
4089 }
4090 }
4091
4092 extern void __kmp_omp_display_env(int verbose);
4093
4094 // 1: it is initializing hidden helper team
4095 extern volatile int __kmp_init_hidden_helper;
4096 // 1: the hidden helper team is done
4097 extern volatile int __kmp_hidden_helper_team_done;
4098 // 1: enable hidden helper task
4099 extern kmp_int32 __kmp_enable_hidden_helper;
4100 // Main thread of hidden helper team
4101 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4102 // Descriptors for the hidden helper threads
4103 extern kmp_info_t **__kmp_hidden_helper_threads;
4104 // Number of hidden helper threads
4105 extern kmp_int32 __kmp_hidden_helper_threads_num;
4106 // Number of hidden helper tasks that have not been executed yet
4107 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4108
4109 extern void __kmp_hidden_helper_initialize();
4110 extern void __kmp_hidden_helper_threads_initz_routine();
4111 extern void __kmp_do_initialize_hidden_helper_threads();
4112 extern void __kmp_hidden_helper_threads_initz_wait();
4113 extern void __kmp_hidden_helper_initz_release();
4114 extern void __kmp_hidden_helper_threads_deinitz_wait();
4115 extern void __kmp_hidden_helper_threads_deinitz_release();
4116 extern void __kmp_hidden_helper_main_thread_wait();
4117 extern void __kmp_hidden_helper_worker_thread_wait();
4118 extern void __kmp_hidden_helper_worker_thread_signal();
4119 extern void __kmp_hidden_helper_main_thread_release();
4120
4121 // Check whether a given thread is a hidden helper thread
4122 #define KMP_HIDDEN_HELPER_THREAD(gtid) \
4123 ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4124
4125 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4126 ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4127
4128 #define KMP_HIDDEN_HELPER_TEAM(team) \
4129 (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4130
4131 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4132 // main thread, is skipped.
4133 #define KMP_GTID_TO_SHADOW_GTID(gtid) \
4134 ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4135
4136 // Return the adjusted gtid value by subtracting from gtid the number
4137 // of hidden helper threads. This adjusted value is the gtid the thread would
4138 // have received if there were no hidden helper threads.
__kmp_adjust_gtid_for_hidden_helpers(int gtid)4139 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4140 int adjusted_gtid = gtid;
4141 if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4142 gtid - __kmp_hidden_helper_threads_num >= 0) {
4143 adjusted_gtid -= __kmp_hidden_helper_threads_num;
4144 }
4145 return adjusted_gtid;
4146 }
4147
4148 // Support for error directive
4149 typedef enum kmp_severity_t {
4150 severity_warning = 1,
4151 severity_fatal = 2
4152 } kmp_severity_t;
4153 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4154
4155 // Support for scope directive
4156 KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4157 KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4158
4159 #ifdef __cplusplus
4160 }
4161 #endif
4162
4163 template <bool C, bool S>
4164 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4165 template <bool C, bool S>
4166 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4167 template <bool C, bool S>
4168 extern void __kmp_atomic_suspend_64(int th_gtid,
4169 kmp_atomic_flag_64<C, S> *flag);
4170 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4171 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4172 template <bool C, bool S>
4173 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4174 template <bool C, bool S>
4175 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4176 template <bool C, bool S>
4177 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4178 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4179 #endif
4180 template <bool C, bool S>
4181 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4182 template <bool C, bool S>
4183 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4184 template <bool C, bool S>
4185 extern void __kmp_atomic_resume_64(int target_gtid,
4186 kmp_atomic_flag_64<C, S> *flag);
4187 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4188
4189 template <bool C, bool S>
4190 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4191 kmp_flag_32<C, S> *flag, int final_spin,
4192 int *thread_finished,
4193 #if USE_ITT_BUILD
4194 void *itt_sync_obj,
4195 #endif /* USE_ITT_BUILD */
4196 kmp_int32 is_constrained);
4197 template <bool C, bool S>
4198 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4199 kmp_flag_64<C, S> *flag, int final_spin,
4200 int *thread_finished,
4201 #if USE_ITT_BUILD
4202 void *itt_sync_obj,
4203 #endif /* USE_ITT_BUILD */
4204 kmp_int32 is_constrained);
4205 template <bool C, bool S>
4206 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4207 kmp_atomic_flag_64<C, S> *flag,
4208 int final_spin, int *thread_finished,
4209 #if USE_ITT_BUILD
4210 void *itt_sync_obj,
4211 #endif /* USE_ITT_BUILD */
4212 kmp_int32 is_constrained);
4213 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4214 kmp_flag_oncore *flag, int final_spin,
4215 int *thread_finished,
4216 #if USE_ITT_BUILD
4217 void *itt_sync_obj,
4218 #endif /* USE_ITT_BUILD */
4219 kmp_int32 is_constrained);
4220
4221 extern int __kmp_nesting_mode;
4222 extern int __kmp_nesting_mode_nlevels;
4223 extern int *__kmp_nesting_nth_level;
4224 extern void __kmp_init_nesting_mode();
4225 extern void __kmp_set_nesting_mode_threads();
4226
4227 /// This class safely opens and closes a C-style FILE* object using RAII
4228 /// semantics. There are also methods which allow using stdout or stderr as
4229 /// the underlying FILE* object. With the implicit conversion operator to
4230 /// FILE*, an object with this type can be used in any function which takes
4231 /// a FILE* object e.g., fprintf().
4232 /// No close method is needed at use sites.
4233 class kmp_safe_raii_file_t {
4234 FILE *f;
4235
close()4236 void close() {
4237 if (f && f != stdout && f != stderr) {
4238 fclose(f);
4239 f = nullptr;
4240 }
4241 }
4242
4243 public:
kmp_safe_raii_file_t()4244 kmp_safe_raii_file_t() : f(nullptr) {}
4245 kmp_safe_raii_file_t(const char *filename, const char *mode,
4246 const char *env_var = nullptr)
f(nullptr)4247 : f(nullptr) {
4248 open(filename, mode, env_var);
4249 }
~kmp_safe_raii_file_t()4250 ~kmp_safe_raii_file_t() { close(); }
4251
4252 /// Open filename using mode. This is automatically closed in the destructor.
4253 /// The env_var parameter indicates the environment variable the filename
4254 /// came from if != nullptr.
4255 void open(const char *filename, const char *mode,
4256 const char *env_var = nullptr) {
4257 KMP_ASSERT(!f);
4258 f = fopen(filename, mode);
4259 if (!f) {
4260 int code = errno;
4261 if (env_var) {
4262 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4263 KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4264 } else {
4265 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4266 __kmp_msg_null);
4267 }
4268 }
4269 }
4270 /// Set the FILE* object to stdout and output there
4271 /// No open call should happen before this call.
set_stdout()4272 void set_stdout() {
4273 KMP_ASSERT(!f);
4274 f = stdout;
4275 }
4276 /// Set the FILE* object to stderr and output there
4277 /// No open call should happen before this call.
set_stderr()4278 void set_stderr() {
4279 KMP_ASSERT(!f);
4280 f = stderr;
4281 }
4282 operator bool() { return bool(f); }
4283 operator FILE *() { return f; }
4284 };
4285
4286 template <typename SourceType, typename TargetType,
4287 bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4288 bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4289 bool isSourceSigned = std::is_signed<SourceType>::value,
4290 bool isTargetSigned = std::is_signed<TargetType>::value>
4291 struct kmp_convert {};
4292
4293 // Both types are signed; Source smaller
4294 template <typename SourceType, typename TargetType>
4295 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4296 static TargetType to(SourceType src) { return (TargetType)src; }
4297 };
4298 // Source equal
4299 template <typename SourceType, typename TargetType>
4300 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4301 static TargetType to(SourceType src) { return src; }
4302 };
4303 // Source bigger
4304 template <typename SourceType, typename TargetType>
4305 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4306 static TargetType to(SourceType src) {
4307 KMP_ASSERT(src <= static_cast<SourceType>(
4308 (std::numeric_limits<TargetType>::max)()));
4309 KMP_ASSERT(src >= static_cast<SourceType>(
4310 (std::numeric_limits<TargetType>::min)()));
4311 return (TargetType)src;
4312 }
4313 };
4314
4315 // Source signed, Target unsigned
4316 // Source smaller
4317 template <typename SourceType, typename TargetType>
4318 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4319 static TargetType to(SourceType src) {
4320 KMP_ASSERT(src >= 0);
4321 return (TargetType)src;
4322 }
4323 };
4324 // Source equal
4325 template <typename SourceType, typename TargetType>
4326 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4327 static TargetType to(SourceType src) {
4328 KMP_ASSERT(src >= 0);
4329 return (TargetType)src;
4330 }
4331 };
4332 // Source bigger
4333 template <typename SourceType, typename TargetType>
4334 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4335 static TargetType to(SourceType src) {
4336 KMP_ASSERT(src >= 0);
4337 KMP_ASSERT(src <= static_cast<SourceType>(
4338 (std::numeric_limits<TargetType>::max)()));
4339 return (TargetType)src;
4340 }
4341 };
4342
4343 // Source unsigned, Target signed
4344 // Source smaller
4345 template <typename SourceType, typename TargetType>
4346 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4347 static TargetType to(SourceType src) { return (TargetType)src; }
4348 };
4349 // Source equal
4350 template <typename SourceType, typename TargetType>
4351 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4352 static TargetType to(SourceType src) {
4353 KMP_ASSERT(src <= static_cast<SourceType>(
4354 (std::numeric_limits<TargetType>::max)()));
4355 return (TargetType)src;
4356 }
4357 };
4358 // Source bigger
4359 template <typename SourceType, typename TargetType>
4360 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4361 static TargetType to(SourceType src) {
4362 KMP_ASSERT(src <= static_cast<SourceType>(
4363 (std::numeric_limits<TargetType>::max)()));
4364 return (TargetType)src;
4365 }
4366 };
4367
4368 // Source unsigned, Target unsigned
4369 // Source smaller
4370 template <typename SourceType, typename TargetType>
4371 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4372 static TargetType to(SourceType src) { return (TargetType)src; }
4373 };
4374 // Source equal
4375 template <typename SourceType, typename TargetType>
4376 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4377 static TargetType to(SourceType src) { return src; }
4378 };
4379 // Source bigger
4380 template <typename SourceType, typename TargetType>
4381 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4382 static TargetType to(SourceType src) {
4383 KMP_ASSERT(src <= static_cast<SourceType>(
4384 (std::numeric_limits<TargetType>::max)()));
4385 return (TargetType)src;
4386 }
4387 };
4388
4389 template <typename T1, typename T2>
4390 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4391 *dest = kmp_convert<T1, T2>::to(src);
4392 }
4393
4394 #endif /* KMP_H */
4395