1 /*
2 * kmp_gsupport.cpp
3 */
4
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "kmp.h"
14 #include "kmp_atomic.h"
15 #include "kmp_utils.h"
16
17 #if OMPT_SUPPORT
18 #include "ompt-specific.h"
19 #endif
20
21 enum {
22 KMP_GOMP_TASK_UNTIED_FLAG = 1,
23 KMP_GOMP_TASK_FINAL_FLAG = 2,
24 KMP_GOMP_TASK_DEPENDS_FLAG = 8
25 };
26
27 enum {
28 KMP_GOMP_DEPOBJ_IN = 1,
29 KMP_GOMP_DEPOBJ_OUT = 2,
30 KMP_GOMP_DEPOBJ_INOUT = 3,
31 KMP_GOMP_DEPOBJ_MTXINOUTSET = 4
32 };
33
34 // This class helps convert gomp dependency info into
35 // kmp_depend_info_t structures
36 class kmp_gomp_depends_info_t {
37 void **depend;
38 kmp_int32 num_deps;
39 size_t num_out, num_mutexinout, num_in, num_depobj;
40 size_t offset;
41
42 public:
kmp_gomp_depends_info_t(void ** depend)43 kmp_gomp_depends_info_t(void **depend) : depend(depend) {
44 size_t ndeps = (kmp_intptr_t)depend[0];
45 // GOMP taskdep structure:
46 // if depend[0] != 0:
47 // depend = [ ndeps | nout | &out | ... | &out | &in | ... | &in ]
48 //
49 // if depend[0] == 0:
50 // depend = [ 0 | ndeps | nout | nmtx | nin | &out | ... | &out | &mtx |
51 // ... | &mtx | &in | ... | &in | &depobj | ... | &depobj ]
52 if (ndeps) {
53 num_out = (kmp_intptr_t)depend[1];
54 num_in = ndeps - num_out;
55 num_mutexinout = num_depobj = 0;
56 offset = 2;
57 } else {
58 ndeps = (kmp_intptr_t)depend[1];
59 num_out = (kmp_intptr_t)depend[2];
60 num_mutexinout = (kmp_intptr_t)depend[3];
61 num_in = (kmp_intptr_t)depend[4];
62 num_depobj = ndeps - num_out - num_mutexinout - num_in;
63 KMP_ASSERT(num_depobj <= ndeps);
64 offset = 5;
65 }
66 num_deps = static_cast<kmp_int32>(ndeps);
67 }
get_num_deps() const68 kmp_int32 get_num_deps() const { return num_deps; }
get_kmp_depend(size_t index) const69 kmp_depend_info_t get_kmp_depend(size_t index) const {
70 kmp_depend_info_t retval;
71 memset(&retval, '\0', sizeof(retval));
72 KMP_ASSERT(index < (size_t)num_deps);
73 retval.len = 0;
74 // Because inout and out are logically equivalent,
75 // use inout and in dependency flags. GOMP does not provide a
76 // way to distinguish if user specified out vs. inout.
77 if (index < num_out) {
78 retval.flags.in = 1;
79 retval.flags.out = 1;
80 retval.base_addr = (kmp_intptr_t)depend[offset + index];
81 } else if (index >= num_out && index < (num_out + num_mutexinout)) {
82 retval.flags.mtx = 1;
83 retval.base_addr = (kmp_intptr_t)depend[offset + index];
84 } else if (index >= (num_out + num_mutexinout) &&
85 index < (num_out + num_mutexinout + num_in)) {
86 retval.flags.in = 1;
87 retval.base_addr = (kmp_intptr_t)depend[offset + index];
88 } else {
89 // depobj is a two element array (size of elements are size of pointer)
90 // depobj[0] = base_addr
91 // depobj[1] = type (in, out, inout, mutexinoutset, etc.)
92 kmp_intptr_t *depobj = (kmp_intptr_t *)depend[offset + index];
93 retval.base_addr = depobj[0];
94 switch (depobj[1]) {
95 case KMP_GOMP_DEPOBJ_IN:
96 retval.flags.in = 1;
97 break;
98 case KMP_GOMP_DEPOBJ_OUT:
99 retval.flags.out = 1;
100 break;
101 case KMP_GOMP_DEPOBJ_INOUT:
102 retval.flags.in = 1;
103 retval.flags.out = 1;
104 break;
105 case KMP_GOMP_DEPOBJ_MTXINOUTSET:
106 retval.flags.mtx = 1;
107 break;
108 default:
109 KMP_FATAL(GompFeatureNotSupported, "Unknown depobj type");
110 }
111 }
112 return retval;
113 }
114 };
115
116 #ifdef __cplusplus
117 extern "C" {
118 #endif // __cplusplus
119
120 #define MKLOC(loc, routine) \
121 static ident_t loc = {0, KMP_IDENT_KMPC, 0, 0, ";unknown;unknown;0;0;;"};
122
123 #include "kmp_ftn_os.h"
124
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER)125 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER)(void) {
126 int gtid = __kmp_entry_gtid();
127 MKLOC(loc, "GOMP_barrier");
128 KA_TRACE(20, ("GOMP_barrier: T#%d\n", gtid));
129 #if OMPT_SUPPORT && OMPT_OPTIONAL
130 ompt_frame_t *ompt_frame;
131 if (ompt_enabled.enabled) {
132 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
133 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
134 }
135 OMPT_STORE_RETURN_ADDRESS(gtid);
136 #endif
137 __kmpc_barrier(&loc, gtid);
138 #if OMPT_SUPPORT && OMPT_OPTIONAL
139 if (ompt_enabled.enabled) {
140 ompt_frame->enter_frame = ompt_data_none;
141 }
142 #endif
143 }
144
145 // Mutual exclusion
146
147 // The symbol that icc/ifort generates for unnamed critical sections
148 // - .gomp_critical_user_ - is defined using .comm in any objects reference it.
149 // We can't reference it directly here in C code, as the symbol contains a ".".
150 //
151 // The RTL contains an assembly language definition of .gomp_critical_user_
152 // with another symbol __kmp_unnamed_critical_addr initialized with it's
153 // address.
154 extern kmp_critical_name *__kmp_unnamed_critical_addr;
155
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_START)156 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_START)(void) {
157 int gtid = __kmp_entry_gtid();
158 MKLOC(loc, "GOMP_critical_start");
159 KA_TRACE(20, ("GOMP_critical_start: T#%d\n", gtid));
160 #if OMPT_SUPPORT && OMPT_OPTIONAL
161 OMPT_STORE_RETURN_ADDRESS(gtid);
162 #endif
163 __kmpc_critical(&loc, gtid, __kmp_unnamed_critical_addr);
164 }
165
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_END)166 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_END)(void) {
167 int gtid = __kmp_get_gtid();
168 MKLOC(loc, "GOMP_critical_end");
169 KA_TRACE(20, ("GOMP_critical_end: T#%d\n", gtid));
170 #if OMPT_SUPPORT && OMPT_OPTIONAL
171 OMPT_STORE_RETURN_ADDRESS(gtid);
172 #endif
173 __kmpc_end_critical(&loc, gtid, __kmp_unnamed_critical_addr);
174 }
175
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_START)176 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_START)(void **pptr) {
177 int gtid = __kmp_entry_gtid();
178 MKLOC(loc, "GOMP_critical_name_start");
179 KA_TRACE(20, ("GOMP_critical_name_start: T#%d\n", gtid));
180 __kmpc_critical(&loc, gtid, (kmp_critical_name *)pptr);
181 }
182
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_END)183 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_END)(void **pptr) {
184 int gtid = __kmp_get_gtid();
185 MKLOC(loc, "GOMP_critical_name_end");
186 KA_TRACE(20, ("GOMP_critical_name_end: T#%d\n", gtid));
187 __kmpc_end_critical(&loc, gtid, (kmp_critical_name *)pptr);
188 }
189
190 // The Gnu codegen tries to use locked operations to perform atomic updates
191 // inline. If it can't, then it calls GOMP_atomic_start() before performing
192 // the update and GOMP_atomic_end() afterward, regardless of the data type.
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_START)193 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_START)(void) {
194 int gtid = __kmp_entry_gtid();
195 KA_TRACE(20, ("GOMP_atomic_start: T#%d\n", gtid));
196
197 #if OMPT_SUPPORT
198 __ompt_thread_assign_wait_id(0);
199 #endif
200
201 __kmp_acquire_atomic_lock(&__kmp_atomic_lock, gtid);
202 }
203
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_END)204 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_END)(void) {
205 int gtid = __kmp_get_gtid();
206 KA_TRACE(20, ("GOMP_atomic_end: T#%d\n", gtid));
207 __kmp_release_atomic_lock(&__kmp_atomic_lock, gtid);
208 }
209
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_START)210 int KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_START)(void) {
211 int gtid = __kmp_entry_gtid();
212 MKLOC(loc, "GOMP_single_start");
213 KA_TRACE(20, ("GOMP_single_start: T#%d\n", gtid));
214
215 if (!TCR_4(__kmp_init_parallel))
216 __kmp_parallel_initialize();
217 __kmp_resume_if_soft_paused();
218
219 // 3rd parameter == FALSE prevents kmp_enter_single from pushing a
220 // workshare when USE_CHECKS is defined. We need to avoid the push,
221 // as there is no corresponding GOMP_single_end() call.
222 kmp_int32 rc = __kmp_enter_single(gtid, &loc, FALSE);
223
224 #if OMPT_SUPPORT && OMPT_OPTIONAL
225 kmp_info_t *this_thr = __kmp_threads[gtid];
226 kmp_team_t *team = this_thr->th.th_team;
227 int tid = __kmp_tid_from_gtid(gtid);
228
229 if (ompt_enabled.enabled) {
230 if (rc) {
231 if (ompt_enabled.ompt_callback_work) {
232 ompt_callbacks.ompt_callback(ompt_callback_work)(
233 ompt_work_single_executor, ompt_scope_begin,
234 &(team->t.ompt_team_info.parallel_data),
235 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
236 1, OMPT_GET_RETURN_ADDRESS(0));
237 }
238 } else {
239 if (ompt_enabled.ompt_callback_work) {
240 ompt_callbacks.ompt_callback(ompt_callback_work)(
241 ompt_work_single_other, ompt_scope_begin,
242 &(team->t.ompt_team_info.parallel_data),
243 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
244 1, OMPT_GET_RETURN_ADDRESS(0));
245 ompt_callbacks.ompt_callback(ompt_callback_work)(
246 ompt_work_single_other, ompt_scope_end,
247 &(team->t.ompt_team_info.parallel_data),
248 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
249 1, OMPT_GET_RETURN_ADDRESS(0));
250 }
251 }
252 }
253 #endif
254
255 return rc;
256 }
257
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_START)258 void *KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_START)(void) {
259 void *retval;
260 int gtid = __kmp_entry_gtid();
261 MKLOC(loc, "GOMP_single_copy_start");
262 KA_TRACE(20, ("GOMP_single_copy_start: T#%d\n", gtid));
263
264 if (!TCR_4(__kmp_init_parallel))
265 __kmp_parallel_initialize();
266 __kmp_resume_if_soft_paused();
267
268 // If this is the first thread to enter, return NULL. The generated code will
269 // then call GOMP_single_copy_end() for this thread only, with the
270 // copyprivate data pointer as an argument.
271 if (__kmp_enter_single(gtid, &loc, FALSE))
272 return NULL;
273
274 // Wait for the first thread to set the copyprivate data pointer,
275 // and for all other threads to reach this point.
276
277 #if OMPT_SUPPORT && OMPT_OPTIONAL
278 ompt_frame_t *ompt_frame;
279 if (ompt_enabled.enabled) {
280 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
281 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
282 }
283 OMPT_STORE_RETURN_ADDRESS(gtid);
284 #endif
285 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
286
287 // Retrieve the value of the copyprivate data point, and wait for all
288 // threads to do likewise, then return.
289 retval = __kmp_team_from_gtid(gtid)->t.t_copypriv_data;
290 {
291 #if OMPT_SUPPORT && OMPT_OPTIONAL
292 OMPT_STORE_RETURN_ADDRESS(gtid);
293 #endif
294 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
295 }
296 #if OMPT_SUPPORT && OMPT_OPTIONAL
297 if (ompt_enabled.enabled) {
298 ompt_frame->enter_frame = ompt_data_none;
299 }
300 #endif
301 return retval;
302 }
303
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_END)304 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_END)(void *data) {
305 int gtid = __kmp_get_gtid();
306 KA_TRACE(20, ("GOMP_single_copy_end: T#%d\n", gtid));
307
308 // Set the copyprivate data pointer fo the team, then hit the barrier so that
309 // the other threads will continue on and read it. Hit another barrier before
310 // continuing, so that the know that the copyprivate data pointer has been
311 // propagated to all threads before trying to reuse the t_copypriv_data field.
312 __kmp_team_from_gtid(gtid)->t.t_copypriv_data = data;
313 #if OMPT_SUPPORT && OMPT_OPTIONAL
314 ompt_frame_t *ompt_frame;
315 if (ompt_enabled.enabled) {
316 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
317 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
318 }
319 OMPT_STORE_RETURN_ADDRESS(gtid);
320 #endif
321 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
322 {
323 #if OMPT_SUPPORT && OMPT_OPTIONAL
324 OMPT_STORE_RETURN_ADDRESS(gtid);
325 #endif
326 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
327 }
328 #if OMPT_SUPPORT && OMPT_OPTIONAL
329 if (ompt_enabled.enabled) {
330 ompt_frame->enter_frame = ompt_data_none;
331 }
332 #endif
333 }
334
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_START)335 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_START)(void) {
336 int gtid = __kmp_entry_gtid();
337 MKLOC(loc, "GOMP_ordered_start");
338 KA_TRACE(20, ("GOMP_ordered_start: T#%d\n", gtid));
339 #if OMPT_SUPPORT && OMPT_OPTIONAL
340 OMPT_STORE_RETURN_ADDRESS(gtid);
341 #endif
342 __kmpc_ordered(&loc, gtid);
343 }
344
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_END)345 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_END)(void) {
346 int gtid = __kmp_get_gtid();
347 MKLOC(loc, "GOMP_ordered_end");
348 KA_TRACE(20, ("GOMP_ordered_start: T#%d\n", gtid));
349 #if OMPT_SUPPORT && OMPT_OPTIONAL
350 OMPT_STORE_RETURN_ADDRESS(gtid);
351 #endif
352 __kmpc_end_ordered(&loc, gtid);
353 }
354
355 // Dispatch macro defs
356 //
357 // They come in two flavors: 64-bit unsigned, and either 32-bit signed
358 // (IA-32 architecture) or 64-bit signed (Intel(R) 64).
359
360 #if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_MIPS || KMP_ARCH_WASM || \
361 KMP_ARCH_PPC || KMP_ARCH_AARCH64_32
362 #define KMP_DISPATCH_INIT __kmp_aux_dispatch_init_4
363 #define KMP_DISPATCH_FINI_CHUNK __kmp_aux_dispatch_fini_chunk_4
364 #define KMP_DISPATCH_NEXT __kmpc_dispatch_next_4
365 #else
366 #define KMP_DISPATCH_INIT __kmp_aux_dispatch_init_8
367 #define KMP_DISPATCH_FINI_CHUNK __kmp_aux_dispatch_fini_chunk_8
368 #define KMP_DISPATCH_NEXT __kmpc_dispatch_next_8
369 #endif /* KMP_ARCH_X86 */
370
371 #define KMP_DISPATCH_INIT_ULL __kmp_aux_dispatch_init_8u
372 #define KMP_DISPATCH_FINI_CHUNK_ULL __kmp_aux_dispatch_fini_chunk_8u
373 #define KMP_DISPATCH_NEXT_ULL __kmpc_dispatch_next_8u
374
375 // The parallel construct
376
377 #ifndef KMP_DEBUG
378 static
379 #endif /* KMP_DEBUG */
380 void
__kmp_GOMP_microtask_wrapper(int * gtid,int * npr,void (* task)(void *),void * data)381 __kmp_GOMP_microtask_wrapper(int *gtid, int *npr, void (*task)(void *),
382 void *data) {
383 #if OMPT_SUPPORT
384 kmp_info_t *thr;
385 ompt_frame_t *ompt_frame;
386 ompt_state_t enclosing_state;
387
388 if (ompt_enabled.enabled) {
389 // get pointer to thread data structure
390 thr = __kmp_threads[*gtid];
391
392 // save enclosing task state; set current state for task
393 enclosing_state = thr->th.ompt_thread_info.state;
394 thr->th.ompt_thread_info.state = ompt_state_work_parallel;
395
396 // set task frame
397 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
398 ompt_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
399 }
400 #endif
401
402 task(data);
403
404 #if OMPT_SUPPORT
405 if (ompt_enabled.enabled) {
406 // clear task frame
407 ompt_frame->exit_frame = ompt_data_none;
408
409 // restore enclosing state
410 thr->th.ompt_thread_info.state = enclosing_state;
411 }
412 #endif
413 }
414
415 #ifndef KMP_DEBUG
416 static
417 #endif /* KMP_DEBUG */
418 void
__kmp_GOMP_parallel_microtask_wrapper(int * gtid,int * npr,void (* task)(void *),void * data,unsigned num_threads,ident_t * loc,enum sched_type schedule,long start,long end,long incr,long chunk_size)419 __kmp_GOMP_parallel_microtask_wrapper(int *gtid, int *npr,
420 void (*task)(void *), void *data,
421 unsigned num_threads, ident_t *loc,
422 enum sched_type schedule, long start,
423 long end, long incr,
424 long chunk_size) {
425 // Initialize the loop worksharing construct.
426
427 KMP_DISPATCH_INIT(loc, *gtid, schedule, start, end, incr, chunk_size,
428 schedule != kmp_sch_static);
429
430 #if OMPT_SUPPORT
431 kmp_info_t *thr;
432 ompt_frame_t *ompt_frame;
433 ompt_state_t enclosing_state;
434
435 if (ompt_enabled.enabled) {
436 thr = __kmp_threads[*gtid];
437 // save enclosing task state; set current state for task
438 enclosing_state = thr->th.ompt_thread_info.state;
439 thr->th.ompt_thread_info.state = ompt_state_work_parallel;
440
441 // set task frame
442 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
443 ompt_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
444 }
445 #endif
446
447 // Now invoke the microtask.
448 task(data);
449
450 #if OMPT_SUPPORT
451 if (ompt_enabled.enabled) {
452 // clear task frame
453 ompt_frame->exit_frame = ompt_data_none;
454
455 // reset enclosing state
456 thr->th.ompt_thread_info.state = enclosing_state;
457 }
458 #endif
459 }
460
__kmp_GOMP_fork_call(ident_t * loc,int gtid,unsigned num_threads,unsigned flags,void (* unwrapped_task)(void *),microtask_t wrapper,int argc,...)461 static void __kmp_GOMP_fork_call(ident_t *loc, int gtid, unsigned num_threads,
462 unsigned flags, void (*unwrapped_task)(void *),
463 microtask_t wrapper, int argc, ...) {
464 int rc;
465 kmp_info_t *thr = __kmp_threads[gtid];
466 kmp_team_t *team = thr->th.th_team;
467 int tid = __kmp_tid_from_gtid(gtid);
468
469 va_list ap;
470 va_start(ap, argc);
471
472 if (num_threads != 0)
473 __kmp_push_num_threads(loc, gtid, num_threads);
474 if (flags != 0)
475 __kmp_push_proc_bind(loc, gtid, (kmp_proc_bind_t)flags);
476 rc = __kmp_fork_call(loc, gtid, fork_context_gnu, argc, wrapper,
477 __kmp_invoke_task_func, kmp_va_addr_of(ap));
478
479 va_end(ap);
480
481 if (rc) {
482 __kmp_run_before_invoked_task(gtid, tid, thr, team);
483 }
484
485 #if OMPT_SUPPORT
486 int ompt_team_size;
487 if (ompt_enabled.enabled) {
488 ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
489 ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
490
491 // implicit task callback
492 if (ompt_enabled.ompt_callback_implicit_task) {
493 ompt_team_size = __kmp_team_from_gtid(gtid)->t.t_nproc;
494 ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
495 ompt_scope_begin, &(team_info->parallel_data),
496 &(task_info->task_data), ompt_team_size, __kmp_tid_from_gtid(gtid),
497 ompt_task_implicit); // TODO: Can this be ompt_task_initial?
498 task_info->thread_num = __kmp_tid_from_gtid(gtid);
499 }
500 thr->th.ompt_thread_info.state = ompt_state_work_parallel;
501 }
502 #endif
503 }
504
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_START)505 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_START)(void (*task)(void *),
506 void *data,
507 unsigned num_threads) {
508 int gtid = __kmp_entry_gtid();
509
510 #if OMPT_SUPPORT
511 ompt_frame_t *parent_frame, *frame;
512
513 if (ompt_enabled.enabled) {
514 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL);
515 parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
516 }
517 OMPT_STORE_RETURN_ADDRESS(gtid);
518 #endif
519
520 MKLOC(loc, "GOMP_parallel_start");
521 KA_TRACE(20, ("GOMP_parallel_start: T#%d\n", gtid));
522 __kmp_GOMP_fork_call(&loc, gtid, num_threads, 0u, task,
523 (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task,
524 data);
525 #if OMPT_SUPPORT
526 if (ompt_enabled.enabled) {
527 __ompt_get_task_info_internal(0, NULL, NULL, &frame, NULL, NULL);
528 frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
529 }
530 #endif
531 #if OMPD_SUPPORT
532 if (ompd_state & OMPD_ENABLE_BP)
533 ompd_bp_parallel_begin();
534 #endif
535 }
536
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)537 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(void) {
538 int gtid = __kmp_get_gtid();
539 kmp_info_t *thr;
540
541 thr = __kmp_threads[gtid];
542
543 MKLOC(loc, "GOMP_parallel_end");
544 KA_TRACE(20, ("GOMP_parallel_end: T#%d\n", gtid));
545
546 if (!thr->th.th_team->t.t_serialized) {
547 __kmp_run_after_invoked_task(gtid, __kmp_tid_from_gtid(gtid), thr,
548 thr->th.th_team);
549 }
550 #if OMPT_SUPPORT
551 if (ompt_enabled.enabled) {
552 // Implicit task is finished here, in the barrier we might schedule
553 // deferred tasks,
554 // these don't see the implicit task on the stack
555 OMPT_CUR_TASK_INFO(thr)->frame.exit_frame = ompt_data_none;
556 }
557 #endif
558
559 __kmp_join_call(&loc, gtid
560 #if OMPT_SUPPORT
561 ,
562 fork_context_gnu
563 #endif
564 );
565 #if OMPD_SUPPORT
566 if (ompd_state & OMPD_ENABLE_BP)
567 ompd_bp_parallel_end();
568 #endif
569 }
570
571 // Loop worksharing constructs
572
573 // The Gnu codegen passes in an exclusive upper bound for the overall range,
574 // but the libguide dispatch code expects an inclusive upper bound, hence the
575 // "end - incr" 5th argument to KMP_DISPATCH_INIT (and the " ub - str" 11th
576 // argument to __kmp_GOMP_fork_call).
577 //
578 // Conversely, KMP_DISPATCH_NEXT returns and inclusive upper bound in *p_ub,
579 // but the Gnu codegen expects an exclusive upper bound, so the adjustment
580 // "*p_ub += stride" compensates for the discrepancy.
581 //
582 // Correction: the gnu codegen always adjusts the upper bound by +-1, not the
583 // stride value. We adjust the dispatch parameters accordingly (by +-1), but
584 // we still adjust p_ub by the actual stride value.
585 //
586 // The "runtime" versions do not take a chunk_sz parameter.
587 //
588 // The profile lib cannot support construct checking of unordered loops that
589 // are predetermined by the compiler to be statically scheduled, as the gcc
590 // codegen will not always emit calls to GOMP_loop_static_next() to get the
591 // next iteration. Instead, it emits inline code to call omp_get_thread_num()
592 // num and calculate the iteration space using the result. It doesn't do this
593 // with ordered static loop, so they can be checked.
594
595 #if OMPT_SUPPORT
596 #define IF_OMPT_SUPPORT(code) code
597 #else
598 #define IF_OMPT_SUPPORT(code)
599 #endif
600
601 #define LOOP_START(func, schedule) \
602 int func(long lb, long ub, long str, long chunk_sz, long *p_lb, \
603 long *p_ub) { \
604 int status; \
605 long stride; \
606 int gtid = __kmp_entry_gtid(); \
607 MKLOC(loc, KMP_STR(func)); \
608 KA_TRACE( \
609 20, \
610 (KMP_STR( \
611 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \
612 gtid, lb, ub, str, chunk_sz)); \
613 \
614 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
615 { \
616 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
617 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
618 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
619 (schedule) != kmp_sch_static); \
620 } \
621 { \
622 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
623 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
624 (kmp_int *)p_ub, (kmp_int *)&stride); \
625 } \
626 if (status) { \
627 KMP_DEBUG_ASSERT(stride == str); \
628 *p_ub += (str > 0) ? 1 : -1; \
629 } \
630 } else { \
631 status = 0; \
632 } \
633 \
634 KA_TRACE( \
635 20, \
636 (KMP_STR( \
637 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \
638 gtid, *p_lb, *p_ub, status)); \
639 return status; \
640 }
641
642 #define LOOP_RUNTIME_START(func, schedule) \
643 int func(long lb, long ub, long str, long *p_lb, long *p_ub) { \
644 int status; \
645 long stride; \
646 long chunk_sz = 0; \
647 int gtid = __kmp_entry_gtid(); \
648 MKLOC(loc, KMP_STR(func)); \
649 KA_TRACE( \
650 20, \
651 (KMP_STR(func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \
652 gtid, lb, ub, str, chunk_sz)); \
653 \
654 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
655 { \
656 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
657 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
658 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
659 TRUE); \
660 } \
661 { \
662 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
663 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
664 (kmp_int *)p_ub, (kmp_int *)&stride); \
665 } \
666 if (status) { \
667 KMP_DEBUG_ASSERT(stride == str); \
668 *p_ub += (str > 0) ? 1 : -1; \
669 } \
670 } else { \
671 status = 0; \
672 } \
673 \
674 KA_TRACE( \
675 20, \
676 (KMP_STR( \
677 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \
678 gtid, *p_lb, *p_ub, status)); \
679 return status; \
680 }
681
682 #define KMP_DOACROSS_FINI(status, gtid) \
683 if (!status && __kmp_threads[gtid]->th.th_dispatch->th_doacross_flags) { \
684 __kmpc_doacross_fini(NULL, gtid); \
685 }
686
687 #define LOOP_NEXT(func, fini_code) \
688 int func(long *p_lb, long *p_ub) { \
689 int status; \
690 long stride; \
691 int gtid = __kmp_get_gtid(); \
692 MKLOC(loc, KMP_STR(func)); \
693 KA_TRACE(20, (KMP_STR(func) ": T#%d\n", gtid)); \
694 \
695 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
696 fini_code status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
697 (kmp_int *)p_ub, (kmp_int *)&stride); \
698 if (status) { \
699 *p_ub += (stride > 0) ? 1 : -1; \
700 } \
701 KMP_DOACROSS_FINI(status, gtid) \
702 \
703 KA_TRACE( \
704 20, \
705 (KMP_STR(func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, stride 0x%lx, " \
706 "returning %d\n", \
707 gtid, *p_lb, *p_ub, stride, status)); \
708 return status; \
709 }
710
LOOP_START(KMP_EXPAND_NAME (KMP_API_NAME_GOMP_LOOP_STATIC_START),kmp_sch_static)711 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_START), kmp_sch_static)
712 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT), {})
713 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START),
714 kmp_sch_dynamic_chunked)
715 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_START),
716 kmp_sch_dynamic_chunked)
717 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT), {})
718 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_NEXT), {})
719 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_START),
720 kmp_sch_guided_chunked)
721 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_START),
722 kmp_sch_guided_chunked)
723 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT), {})
724 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_NEXT), {})
725 LOOP_RUNTIME_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_START),
726 kmp_sch_runtime)
727 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT), {})
728 LOOP_RUNTIME_START(
729 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_START),
730 kmp_sch_runtime)
731 LOOP_RUNTIME_START(
732 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_START),
733 kmp_sch_runtime)
734 LOOP_NEXT(
735 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_NEXT), {})
736 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_NEXT), {})
737
738 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START),
739 kmp_ord_static)
740 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT),
741 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); })
742 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START),
743 kmp_ord_dynamic_chunked)
744 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT),
745 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); })
746 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START),
747 kmp_ord_guided_chunked)
748 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT),
749 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); })
750 LOOP_RUNTIME_START(
751 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START),
752 kmp_ord_runtime)
753 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT),
754 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); })
755
756 #define LOOP_DOACROSS_START(func, schedule) \
757 bool func(unsigned ncounts, long *counts, long chunk_sz, long *p_lb, \
758 long *p_ub) { \
759 int status; \
760 long stride, lb, ub, str; \
761 int gtid = __kmp_entry_gtid(); \
762 struct kmp_dim *dims = \
763 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \
764 MKLOC(loc, KMP_STR(func)); \
765 for (unsigned i = 0; i < ncounts; ++i) { \
766 dims[i].lo = 0; \
767 dims[i].up = counts[i] - 1; \
768 dims[i].st = 1; \
769 } \
770 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \
771 lb = 0; \
772 ub = counts[0]; \
773 str = 1; \
774 KA_TRACE(20, (KMP_STR(func) ": T#%d, ncounts %u, lb 0x%lx, ub 0x%lx, str " \
775 "0x%lx, chunk_sz " \
776 "0x%lx\n", \
777 gtid, ncounts, lb, ub, str, chunk_sz)); \
778 \
779 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
780 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
781 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
782 (schedule) != kmp_sch_static); \
783 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
784 (kmp_int *)p_ub, (kmp_int *)&stride); \
785 if (status) { \
786 KMP_DEBUG_ASSERT(stride == str); \
787 *p_ub += (str > 0) ? 1 : -1; \
788 } \
789 } else { \
790 status = 0; \
791 } \
792 KMP_DOACROSS_FINI(status, gtid); \
793 \
794 KA_TRACE( \
795 20, \
796 (KMP_STR( \
797 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \
798 gtid, *p_lb, *p_ub, status)); \
799 __kmp_free(dims); \
800 return status; \
801 }
802
803 #define LOOP_DOACROSS_RUNTIME_START(func, schedule) \
804 int func(unsigned ncounts, long *counts, long *p_lb, long *p_ub) { \
805 int status; \
806 long stride, lb, ub, str; \
807 long chunk_sz = 0; \
808 int gtid = __kmp_entry_gtid(); \
809 struct kmp_dim *dims = \
810 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \
811 MKLOC(loc, KMP_STR(func)); \
812 for (unsigned i = 0; i < ncounts; ++i) { \
813 dims[i].lo = 0; \
814 dims[i].up = counts[i] - 1; \
815 dims[i].st = 1; \
816 } \
817 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \
818 lb = 0; \
819 ub = counts[0]; \
820 str = 1; \
821 KA_TRACE( \
822 20, \
823 (KMP_STR(func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \
824 gtid, lb, ub, str, chunk_sz)); \
825 \
826 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
827 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
828 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, TRUE); \
829 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
830 (kmp_int *)p_ub, (kmp_int *)&stride); \
831 if (status) { \
832 KMP_DEBUG_ASSERT(stride == str); \
833 *p_ub += (str > 0) ? 1 : -1; \
834 } \
835 } else { \
836 status = 0; \
837 } \
838 KMP_DOACROSS_FINI(status, gtid); \
839 \
840 KA_TRACE( \
841 20, \
842 (KMP_STR( \
843 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \
844 gtid, *p_lb, *p_ub, status)); \
845 __kmp_free(dims); \
846 return status; \
847 }
848
849 LOOP_DOACROSS_START(
850 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START),
851 kmp_sch_static)
852 LOOP_DOACROSS_START(
853 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START),
854 kmp_sch_dynamic_chunked)
855 LOOP_DOACROSS_START(
856 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START),
857 kmp_sch_guided_chunked)
858 LOOP_DOACROSS_RUNTIME_START(
859 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START),
860 kmp_sch_runtime)
861
862 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END)(void) {
863 int gtid = __kmp_get_gtid();
864 KA_TRACE(20, ("GOMP_loop_end: T#%d\n", gtid))
865
866 #if OMPT_SUPPORT && OMPT_OPTIONAL
867 ompt_frame_t *ompt_frame;
868 if (ompt_enabled.enabled) {
869 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
870 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
871 OMPT_STORE_RETURN_ADDRESS(gtid);
872 }
873 #endif
874 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
875 #if OMPT_SUPPORT && OMPT_OPTIONAL
876 if (ompt_enabled.enabled) {
877 ompt_frame->enter_frame = ompt_data_none;
878 }
879 #endif
880
881 KA_TRACE(20, ("GOMP_loop_end exit: T#%d\n", gtid))
882 }
883
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_NOWAIT)884 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_NOWAIT)(void) {
885 KA_TRACE(20, ("GOMP_loop_end_nowait: T#%d\n", __kmp_get_gtid()))
886 }
887
888 // Unsigned long long loop worksharing constructs
889 //
890 // These are new with gcc 4.4
891
892 #define LOOP_START_ULL(func, schedule) \
893 int func(int up, unsigned long long lb, unsigned long long ub, \
894 unsigned long long str, unsigned long long chunk_sz, \
895 unsigned long long *p_lb, unsigned long long *p_ub) { \
896 int status; \
897 long long str2 = up ? ((long long)str) : -((long long)str); \
898 long long stride; \
899 int gtid = __kmp_entry_gtid(); \
900 MKLOC(loc, KMP_STR(func)); \
901 \
902 KA_TRACE(20, (KMP_STR(func) ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str " \
903 "0x%llx, chunk_sz 0x%llx\n", \
904 gtid, up, lb, ub, str, chunk_sz)); \
905 \
906 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
907 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \
908 (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \
909 (schedule) != kmp_sch_static); \
910 status = \
911 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
912 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
913 if (status) { \
914 KMP_DEBUG_ASSERT(stride == str2); \
915 *p_ub += (str > 0) ? 1 : -1; \
916 } \
917 } else { \
918 status = 0; \
919 } \
920 \
921 KA_TRACE( \
922 20, \
923 (KMP_STR( \
924 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \
925 gtid, *p_lb, *p_ub, status)); \
926 return status; \
927 }
928
929 #define LOOP_RUNTIME_START_ULL(func, schedule) \
930 int func(int up, unsigned long long lb, unsigned long long ub, \
931 unsigned long long str, unsigned long long *p_lb, \
932 unsigned long long *p_ub) { \
933 int status; \
934 long long str2 = up ? ((long long)str) : -((long long)str); \
935 unsigned long long stride; \
936 unsigned long long chunk_sz = 0; \
937 int gtid = __kmp_entry_gtid(); \
938 MKLOC(loc, KMP_STR(func)); \
939 \
940 KA_TRACE(20, (KMP_STR(func) ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str " \
941 "0x%llx, chunk_sz 0x%llx\n", \
942 gtid, up, lb, ub, str, chunk_sz)); \
943 \
944 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
945 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \
946 (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \
947 TRUE); \
948 status = \
949 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
950 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
951 if (status) { \
952 KMP_DEBUG_ASSERT((long long)stride == str2); \
953 *p_ub += (str > 0) ? 1 : -1; \
954 } \
955 } else { \
956 status = 0; \
957 } \
958 \
959 KA_TRACE( \
960 20, \
961 (KMP_STR( \
962 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \
963 gtid, *p_lb, *p_ub, status)); \
964 return status; \
965 }
966
967 #define LOOP_NEXT_ULL(func, fini_code) \
968 int func(unsigned long long *p_lb, unsigned long long *p_ub) { \
969 int status; \
970 long long stride; \
971 int gtid = __kmp_get_gtid(); \
972 MKLOC(loc, KMP_STR(func)); \
973 KA_TRACE(20, (KMP_STR(func) ": T#%d\n", gtid)); \
974 \
975 fini_code status = \
976 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
977 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
978 if (status) { \
979 *p_ub += (stride > 0) ? 1 : -1; \
980 } \
981 \
982 KA_TRACE( \
983 20, \
984 (KMP_STR( \
985 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, stride 0x%llx, " \
986 "returning %d\n", \
987 gtid, *p_lb, *p_ub, stride, status)); \
988 return status; \
989 }
990
LOOP_START_ULL(KMP_EXPAND_NAME (KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START),kmp_sch_static)991 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START),
992 kmp_sch_static)
993 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT), {})
994 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START),
995 kmp_sch_dynamic_chunked)
996 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT), {})
997 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START),
998 kmp_sch_guided_chunked)
999 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT), {})
1000 LOOP_START_ULL(
1001 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_START),
1002 kmp_sch_dynamic_chunked)
1003 LOOP_NEXT_ULL(
1004 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_NEXT), {})
1005 LOOP_START_ULL(
1006 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_START),
1007 kmp_sch_guided_chunked)
1008 LOOP_NEXT_ULL(
1009 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_NEXT), {})
1010 LOOP_RUNTIME_START_ULL(
1011 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START), kmp_sch_runtime)
1012 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT), {})
1013 LOOP_RUNTIME_START_ULL(
1014 KMP_EXPAND_NAME(
1015 KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_START),
1016 kmp_sch_runtime)
1017 LOOP_RUNTIME_START_ULL(
1018 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_START),
1019 kmp_sch_runtime)
1020 LOOP_NEXT_ULL(
1021 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_NEXT),
1022 {})
1023 LOOP_NEXT_ULL(
1024 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_NEXT), {})
1025
1026 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START),
1027 kmp_ord_static)
1028 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT),
1029 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); })
1030 LOOP_START_ULL(
1031 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START),
1032 kmp_ord_dynamic_chunked)
1033 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT),
1034 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); })
1035 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START),
1036 kmp_ord_guided_chunked)
1037 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT),
1038 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); })
1039 LOOP_RUNTIME_START_ULL(
1040 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START),
1041 kmp_ord_runtime)
1042 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT),
1043 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); })
1044
1045 #define LOOP_DOACROSS_START_ULL(func, schedule) \
1046 int func(unsigned ncounts, unsigned long long *counts, \
1047 unsigned long long chunk_sz, unsigned long long *p_lb, \
1048 unsigned long long *p_ub) { \
1049 int status; \
1050 long long stride, str, lb, ub; \
1051 int gtid = __kmp_entry_gtid(); \
1052 struct kmp_dim *dims = \
1053 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \
1054 MKLOC(loc, KMP_STR(func)); \
1055 for (unsigned i = 0; i < ncounts; ++i) { \
1056 dims[i].lo = 0; \
1057 dims[i].up = counts[i] - 1; \
1058 dims[i].st = 1; \
1059 } \
1060 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \
1061 lb = 0; \
1062 ub = counts[0]; \
1063 str = 1; \
1064 \
1065 KA_TRACE(20, (KMP_STR(func) ": T#%d, lb 0x%llx, ub 0x%llx, str " \
1066 "0x%llx, chunk_sz 0x%llx\n", \
1067 gtid, lb, ub, str, chunk_sz)); \
1068 \
1069 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
1070 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \
1071 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
1072 (schedule) != kmp_sch_static); \
1073 status = \
1074 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
1075 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
1076 if (status) { \
1077 KMP_DEBUG_ASSERT(stride == str); \
1078 *p_ub += (str > 0) ? 1 : -1; \
1079 } \
1080 } else { \
1081 status = 0; \
1082 } \
1083 KMP_DOACROSS_FINI(status, gtid); \
1084 \
1085 KA_TRACE( \
1086 20, \
1087 (KMP_STR( \
1088 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \
1089 gtid, *p_lb, *p_ub, status)); \
1090 __kmp_free(dims); \
1091 return status; \
1092 }
1093
1094 #define LOOP_DOACROSS_RUNTIME_START_ULL(func, schedule) \
1095 int func(unsigned ncounts, unsigned long long *counts, \
1096 unsigned long long *p_lb, unsigned long long *p_ub) { \
1097 int status; \
1098 unsigned long long stride, str, lb, ub; \
1099 unsigned long long chunk_sz = 0; \
1100 int gtid = __kmp_entry_gtid(); \
1101 struct kmp_dim *dims = \
1102 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \
1103 MKLOC(loc, KMP_STR(func)); \
1104 for (unsigned i = 0; i < ncounts; ++i) { \
1105 dims[i].lo = 0; \
1106 dims[i].up = counts[i] - 1; \
1107 dims[i].st = 1; \
1108 } \
1109 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \
1110 lb = 0; \
1111 ub = counts[0]; \
1112 str = 1; \
1113 KA_TRACE(20, (KMP_STR(func) ": T#%d, lb 0x%llx, ub 0x%llx, str " \
1114 "0x%llx, chunk_sz 0x%llx\n", \
1115 gtid, lb, ub, str, chunk_sz)); \
1116 \
1117 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
1118 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \
1119 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
1120 TRUE); \
1121 status = \
1122 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
1123 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
1124 if (status) { \
1125 KMP_DEBUG_ASSERT(stride == str); \
1126 *p_ub += (str > 0) ? 1 : -1; \
1127 } \
1128 } else { \
1129 status = 0; \
1130 } \
1131 KMP_DOACROSS_FINI(status, gtid); \
1132 \
1133 KA_TRACE( \
1134 20, \
1135 (KMP_STR( \
1136 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \
1137 gtid, *p_lb, *p_ub, status)); \
1138 __kmp_free(dims); \
1139 return status; \
1140 }
1141
1142 LOOP_DOACROSS_START_ULL(
1143 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START),
1144 kmp_sch_static)
1145 LOOP_DOACROSS_START_ULL(
1146 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START),
1147 kmp_sch_dynamic_chunked)
1148 LOOP_DOACROSS_START_ULL(
1149 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START),
1150 kmp_sch_guided_chunked)
1151 LOOP_DOACROSS_RUNTIME_START_ULL(
1152 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START),
1153 kmp_sch_runtime)
1154
1155 // Combined parallel / loop worksharing constructs
1156 //
1157 // There are no ull versions (yet).
1158
1159 #define PARALLEL_LOOP_START(func, schedule, ompt_pre, ompt_post) \
1160 void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \
1161 long ub, long str, long chunk_sz) { \
1162 int gtid = __kmp_entry_gtid(); \
1163 MKLOC(loc, KMP_STR(func)); \
1164 KA_TRACE( \
1165 20, \
1166 (KMP_STR( \
1167 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \
1168 gtid, lb, ub, str, chunk_sz)); \
1169 \
1170 ompt_pre(); \
1171 \
1172 __kmp_GOMP_fork_call(&loc, gtid, num_threads, 0u, task, \
1173 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \
1174 9, task, data, num_threads, &loc, (schedule), lb, \
1175 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \
1176 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid)); \
1177 \
1178 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
1179 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
1180 (schedule) != kmp_sch_static); \
1181 \
1182 ompt_post(); \
1183 \
1184 KA_TRACE(20, (KMP_STR(func) " exit: T#%d\n", gtid)); \
1185 }
1186
1187 #if OMPT_SUPPORT && OMPT_OPTIONAL
1188
1189 #define OMPT_LOOP_PRE() \
1190 ompt_frame_t *parent_frame; \
1191 if (ompt_enabled.enabled) { \
1192 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); \
1193 parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); \
1194 OMPT_STORE_RETURN_ADDRESS(gtid); \
1195 }
1196
1197 #define OMPT_LOOP_POST() \
1198 if (ompt_enabled.enabled) { \
1199 parent_frame->enter_frame = ompt_data_none; \
1200 }
1201
1202 #else
1203
1204 #define OMPT_LOOP_PRE()
1205
1206 #define OMPT_LOOP_POST()
1207
1208 #endif
1209
1210 PARALLEL_LOOP_START(
1211 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START),
1212 kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1213 PARALLEL_LOOP_START(
1214 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START),
1215 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1216 PARALLEL_LOOP_START(
1217 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START),
1218 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1219 PARALLEL_LOOP_START(
1220 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START),
1221 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1222
1223 // Tasking constructs
1224
1225 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASK)(void (*func)(void *), void *data,
1226 void (*copy_func)(void *, void *),
1227 long arg_size, long arg_align,
1228 bool if_cond, unsigned gomp_flags,
1229 void **depend) {
1230 MKLOC(loc, "GOMP_task");
1231 int gtid = __kmp_entry_gtid();
1232 kmp_int32 flags = 0;
1233 kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags;
1234
1235 KA_TRACE(20, ("GOMP_task: T#%d\n", gtid));
1236
1237 // The low-order bit is the "untied" flag
1238 if (!(gomp_flags & KMP_GOMP_TASK_UNTIED_FLAG)) {
1239 input_flags->tiedness = TASK_TIED;
1240 }
1241 // The second low-order bit is the "final" flag
1242 if (gomp_flags & KMP_GOMP_TASK_FINAL_FLAG) {
1243 input_flags->final = 1;
1244 }
1245 input_flags->native = 1;
1246 // __kmp_task_alloc() sets up all other flags
1247
1248 if (!if_cond) {
1249 arg_size = 0;
1250 }
1251
1252 kmp_task_t *task = __kmp_task_alloc(
1253 &loc, gtid, input_flags, sizeof(kmp_task_t),
1254 arg_size ? arg_size + arg_align - 1 : 0, (kmp_routine_entry_t)func);
1255
1256 if (arg_size > 0) {
1257 if (arg_align > 0) {
1258 task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) /
1259 arg_align * arg_align);
1260 }
1261 // else error??
1262
1263 if (copy_func) {
1264 (*copy_func)(task->shareds, data);
1265 } else {
1266 KMP_MEMCPY(task->shareds, data, arg_size);
1267 }
1268 }
1269
1270 #if OMPT_SUPPORT
1271 kmp_taskdata_t *current_task;
1272 if (ompt_enabled.enabled) {
1273 current_task = __kmp_threads[gtid]->th.th_current_task;
1274 current_task->ompt_task_info.frame.enter_frame.ptr =
1275 OMPT_GET_FRAME_ADDRESS(0);
1276 }
1277 OMPT_STORE_RETURN_ADDRESS(gtid);
1278 #endif
1279
1280 if (if_cond) {
1281 if (gomp_flags & KMP_GOMP_TASK_DEPENDS_FLAG) {
1282 KMP_ASSERT(depend);
1283 kmp_gomp_depends_info_t gomp_depends(depend);
1284 kmp_int32 ndeps = gomp_depends.get_num_deps();
1285 SimpleVLA<kmp_depend_info_t> dep_list(ndeps);
1286 for (kmp_int32 i = 0; i < ndeps; i++)
1287 dep_list[i] = gomp_depends.get_kmp_depend(i);
1288 kmp_int32 ndeps_cnv;
1289 __kmp_type_convert(ndeps, &ndeps_cnv);
1290 __kmpc_omp_task_with_deps(&loc, gtid, task, ndeps_cnv, dep_list, 0, NULL);
1291 } else {
1292 __kmpc_omp_task(&loc, gtid, task);
1293 }
1294 } else {
1295 #if OMPT_SUPPORT
1296 ompt_thread_info_t oldInfo;
1297 kmp_info_t *thread;
1298 kmp_taskdata_t *taskdata;
1299 if (ompt_enabled.enabled) {
1300 // Store the threads states and restore them after the task
1301 thread = __kmp_threads[gtid];
1302 taskdata = KMP_TASK_TO_TASKDATA(task);
1303 oldInfo = thread->th.ompt_thread_info;
1304 thread->th.ompt_thread_info.wait_id = 0;
1305 thread->th.ompt_thread_info.state = ompt_state_work_parallel;
1306 taskdata->ompt_task_info.frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1307 }
1308 OMPT_STORE_RETURN_ADDRESS(gtid);
1309 #endif
1310 if (gomp_flags & KMP_GOMP_TASK_DEPENDS_FLAG) {
1311 KMP_ASSERT(depend);
1312 kmp_gomp_depends_info_t gomp_depends(depend);
1313 kmp_int32 ndeps = gomp_depends.get_num_deps();
1314 SimpleVLA<kmp_depend_info_t> dep_list(ndeps);
1315 for (kmp_int32 i = 0; i < ndeps; i++)
1316 dep_list[i] = gomp_depends.get_kmp_depend(i);
1317 __kmpc_omp_wait_deps(&loc, gtid, ndeps, dep_list, 0, NULL);
1318 }
1319
1320 __kmpc_omp_task_begin_if0(&loc, gtid, task);
1321 func(data);
1322 __kmpc_omp_task_complete_if0(&loc, gtid, task);
1323
1324 #if OMPT_SUPPORT
1325 if (ompt_enabled.enabled) {
1326 thread->th.ompt_thread_info = oldInfo;
1327 taskdata->ompt_task_info.frame.exit_frame = ompt_data_none;
1328 }
1329 #endif
1330 }
1331 #if OMPT_SUPPORT
1332 if (ompt_enabled.enabled) {
1333 current_task->ompt_task_info.frame.enter_frame = ompt_data_none;
1334 }
1335 #endif
1336
1337 KA_TRACE(20, ("GOMP_task exit: T#%d\n", gtid));
1338 }
1339
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT)1340 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT)(void) {
1341 MKLOC(loc, "GOMP_taskwait");
1342 int gtid = __kmp_entry_gtid();
1343
1344 #if OMPT_SUPPORT
1345 OMPT_STORE_RETURN_ADDRESS(gtid);
1346 #endif
1347
1348 KA_TRACE(20, ("GOMP_taskwait: T#%d\n", gtid));
1349
1350 __kmpc_omp_taskwait(&loc, gtid);
1351
1352 KA_TRACE(20, ("GOMP_taskwait exit: T#%d\n", gtid));
1353 }
1354
1355 // Sections worksharing constructs
1356 //
1357 // For the sections construct, we initialize a dynamically scheduled loop
1358 // worksharing construct with lb 1 and stride 1, and use the iteration #'s
1359 // that its returns as sections ids.
1360 //
1361 // There are no special entry points for ordered sections, so we always use
1362 // the dynamically scheduled workshare, even if the sections aren't ordered.
1363
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_START)1364 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_START)(unsigned count) {
1365 int status;
1366 kmp_int lb, ub, stride;
1367 int gtid = __kmp_entry_gtid();
1368 MKLOC(loc, "GOMP_sections_start");
1369 KA_TRACE(20, ("GOMP_sections_start: T#%d\n", gtid));
1370
1371 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE);
1372
1373 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride);
1374 if (status) {
1375 KMP_DEBUG_ASSERT(stride == 1);
1376 KMP_DEBUG_ASSERT(lb > 0);
1377 KMP_ASSERT(lb == ub);
1378 } else {
1379 lb = 0;
1380 }
1381
1382 KA_TRACE(20, ("GOMP_sections_start exit: T#%d returning %u\n", gtid,
1383 (unsigned)lb));
1384 return (unsigned)lb;
1385 }
1386
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_NEXT)1387 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_NEXT)(void) {
1388 int status;
1389 kmp_int lb, ub, stride;
1390 int gtid = __kmp_get_gtid();
1391 MKLOC(loc, "GOMP_sections_next");
1392 KA_TRACE(20, ("GOMP_sections_next: T#%d\n", gtid));
1393
1394 #if OMPT_SUPPORT
1395 OMPT_STORE_RETURN_ADDRESS(gtid);
1396 #endif
1397
1398 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride);
1399 if (status) {
1400 KMP_DEBUG_ASSERT(stride == 1);
1401 KMP_DEBUG_ASSERT(lb > 0);
1402 KMP_ASSERT(lb == ub);
1403 } else {
1404 lb = 0;
1405 }
1406
1407 KA_TRACE(
1408 20, ("GOMP_sections_next exit: T#%d returning %u\n", gtid, (unsigned)lb));
1409 return (unsigned)lb;
1410 }
1411
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START)1412 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START)(
1413 void (*task)(void *), void *data, unsigned num_threads, unsigned count) {
1414 int gtid = __kmp_entry_gtid();
1415
1416 #if OMPT_SUPPORT
1417 ompt_frame_t *parent_frame;
1418
1419 if (ompt_enabled.enabled) {
1420 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL);
1421 parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1422 }
1423 OMPT_STORE_RETURN_ADDRESS(gtid);
1424 #endif
1425
1426 MKLOC(loc, "GOMP_parallel_sections_start");
1427 KA_TRACE(20, ("GOMP_parallel_sections_start: T#%d\n", gtid));
1428
1429 __kmp_GOMP_fork_call(&loc, gtid, num_threads, 0u, task,
1430 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9,
1431 task, data, num_threads, &loc, kmp_nm_dynamic_chunked,
1432 (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1);
1433
1434 #if OMPT_SUPPORT
1435 if (ompt_enabled.enabled) {
1436 parent_frame->enter_frame = ompt_data_none;
1437 }
1438 #endif
1439
1440 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE);
1441
1442 KA_TRACE(20, ("GOMP_parallel_sections_start exit: T#%d\n", gtid));
1443 }
1444
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END)1445 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END)(void) {
1446 int gtid = __kmp_get_gtid();
1447 KA_TRACE(20, ("GOMP_sections_end: T#%d\n", gtid))
1448
1449 #if OMPT_SUPPORT
1450 ompt_frame_t *ompt_frame;
1451 if (ompt_enabled.enabled) {
1452 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
1453 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1454 }
1455 OMPT_STORE_RETURN_ADDRESS(gtid);
1456 #endif
1457 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
1458 #if OMPT_SUPPORT
1459 if (ompt_enabled.enabled) {
1460 ompt_frame->enter_frame = ompt_data_none;
1461 }
1462 #endif
1463
1464 KA_TRACE(20, ("GOMP_sections_end exit: T#%d\n", gtid))
1465 }
1466
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT)1467 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT)(void) {
1468 KA_TRACE(20, ("GOMP_sections_end_nowait: T#%d\n", __kmp_get_gtid()))
1469 }
1470
1471 // libgomp has an empty function for GOMP_taskyield as of 2013-10-10
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKYIELD)1472 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKYIELD)(void) {
1473 KA_TRACE(20, ("GOMP_taskyield: T#%d\n", __kmp_get_gtid()))
1474 return;
1475 }
1476
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL)1477 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL)(void (*task)(void *),
1478 void *data,
1479 unsigned num_threads,
1480 unsigned int flags) {
1481 int gtid = __kmp_entry_gtid();
1482 MKLOC(loc, "GOMP_parallel");
1483 KA_TRACE(20, ("GOMP_parallel: T#%d\n", gtid));
1484
1485 #if OMPT_SUPPORT
1486 ompt_task_info_t *parent_task_info, *task_info;
1487 if (ompt_enabled.enabled) {
1488 parent_task_info = __ompt_get_task_info_object(0);
1489 parent_task_info->frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1490 }
1491 OMPT_STORE_RETURN_ADDRESS(gtid);
1492 #endif
1493 __kmp_GOMP_fork_call(&loc, gtid, num_threads, flags, task,
1494 (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task,
1495 data);
1496 #if OMPT_SUPPORT
1497 if (ompt_enabled.enabled) {
1498 task_info = __ompt_get_task_info_object(0);
1499 task_info->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1500 }
1501 #endif
1502 task(data);
1503 {
1504 #if OMPT_SUPPORT
1505 OMPT_STORE_RETURN_ADDRESS(gtid);
1506 #endif
1507 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)();
1508 }
1509 #if OMPT_SUPPORT
1510 if (ompt_enabled.enabled) {
1511 task_info->frame.exit_frame = ompt_data_none;
1512 parent_task_info->frame.enter_frame = ompt_data_none;
1513 }
1514 #endif
1515 }
1516
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS)1517 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS)(void (*task)(void *),
1518 void *data,
1519 unsigned num_threads,
1520 unsigned count,
1521 unsigned flags) {
1522 int gtid = __kmp_entry_gtid();
1523 MKLOC(loc, "GOMP_parallel_sections");
1524 KA_TRACE(20, ("GOMP_parallel_sections: T#%d\n", gtid));
1525
1526 #if OMPT_SUPPORT
1527 ompt_frame_t *task_frame;
1528 kmp_info_t *thr;
1529 if (ompt_enabled.enabled) {
1530 thr = __kmp_threads[gtid];
1531 task_frame = &(thr->th.th_current_task->ompt_task_info.frame);
1532 task_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1533 }
1534 OMPT_STORE_RETURN_ADDRESS(gtid);
1535 #endif
1536
1537 __kmp_GOMP_fork_call(&loc, gtid, num_threads, flags, task,
1538 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9,
1539 task, data, num_threads, &loc, kmp_nm_dynamic_chunked,
1540 (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1);
1541
1542 {
1543 #if OMPT_SUPPORT
1544 OMPT_STORE_RETURN_ADDRESS(gtid);
1545 #endif
1546
1547 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE);
1548 }
1549
1550 #if OMPT_SUPPORT
1551 ompt_frame_t *child_frame;
1552 if (ompt_enabled.enabled) {
1553 child_frame = &(thr->th.th_current_task->ompt_task_info.frame);
1554 child_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1555 }
1556 #endif
1557
1558 task(data);
1559
1560 #if OMPT_SUPPORT
1561 if (ompt_enabled.enabled) {
1562 child_frame->exit_frame = ompt_data_none;
1563 }
1564 #endif
1565
1566 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)();
1567 KA_TRACE(20, ("GOMP_parallel_sections exit: T#%d\n", gtid));
1568
1569 #if OMPT_SUPPORT
1570 if (ompt_enabled.enabled) {
1571 task_frame->enter_frame = ompt_data_none;
1572 }
1573 #endif
1574 }
1575
1576 #define PARALLEL_LOOP(func, schedule, ompt_pre, ompt_post) \
1577 void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \
1578 long ub, long str, long chunk_sz, unsigned flags) { \
1579 int gtid = __kmp_entry_gtid(); \
1580 MKLOC(loc, KMP_STR(func)); \
1581 KA_TRACE( \
1582 20, \
1583 (KMP_STR( \
1584 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \
1585 gtid, lb, ub, str, chunk_sz)); \
1586 \
1587 ompt_pre(); \
1588 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
1589 __kmp_GOMP_fork_call(&loc, gtid, num_threads, flags, task, \
1590 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \
1591 9, task, data, num_threads, &loc, (schedule), lb, \
1592 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \
1593 \
1594 { \
1595 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
1596 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
1597 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
1598 (schedule) != kmp_sch_static); \
1599 } \
1600 task(data); \
1601 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); \
1602 ompt_post(); \
1603 \
1604 KA_TRACE(20, (KMP_STR(func) " exit: T#%d\n", gtid)); \
1605 }
1606
PARALLEL_LOOP(KMP_EXPAND_NAME (KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC),kmp_sch_static,OMPT_LOOP_PRE,OMPT_LOOP_POST)1607 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC),
1608 kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1609 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC),
1610 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1611 PARALLEL_LOOP(
1612 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_GUIDED),
1613 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1614 PARALLEL_LOOP(
1615 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_DYNAMIC),
1616 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1617 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED),
1618 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1619 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME),
1620 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1621 PARALLEL_LOOP(
1622 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_MAYBE_NONMONOTONIC_RUNTIME),
1623 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1624 PARALLEL_LOOP(
1625 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_RUNTIME),
1626 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1627
1628 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_START)(void) {
1629 int gtid = __kmp_entry_gtid();
1630 MKLOC(loc, "GOMP_taskgroup_start");
1631 KA_TRACE(20, ("GOMP_taskgroup_start: T#%d\n", gtid));
1632
1633 #if OMPT_SUPPORT
1634 OMPT_STORE_RETURN_ADDRESS(gtid);
1635 #endif
1636
1637 __kmpc_taskgroup(&loc, gtid);
1638
1639 return;
1640 }
1641
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_END)1642 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_END)(void) {
1643 int gtid = __kmp_get_gtid();
1644 MKLOC(loc, "GOMP_taskgroup_end");
1645 KA_TRACE(20, ("GOMP_taskgroup_end: T#%d\n", gtid));
1646
1647 #if OMPT_SUPPORT
1648 OMPT_STORE_RETURN_ADDRESS(gtid);
1649 #endif
1650
1651 __kmpc_end_taskgroup(&loc, gtid);
1652
1653 return;
1654 }
1655
__kmp_gomp_to_omp_cancellation_kind(int gomp_kind)1656 static kmp_int32 __kmp_gomp_to_omp_cancellation_kind(int gomp_kind) {
1657 kmp_int32 cncl_kind = 0;
1658 switch (gomp_kind) {
1659 case 1:
1660 cncl_kind = cancel_parallel;
1661 break;
1662 case 2:
1663 cncl_kind = cancel_loop;
1664 break;
1665 case 4:
1666 cncl_kind = cancel_sections;
1667 break;
1668 case 8:
1669 cncl_kind = cancel_taskgroup;
1670 break;
1671 }
1672 return cncl_kind;
1673 }
1674
1675 // Return true if cancellation should take place, false otherwise
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCELLATION_POINT)1676 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCELLATION_POINT)(int which) {
1677 int gtid = __kmp_get_gtid();
1678 MKLOC(loc, "GOMP_cancellation_point");
1679 KA_TRACE(20, ("GOMP_cancellation_point: T#%d which:%d\n", gtid, which));
1680 kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which);
1681 return __kmpc_cancellationpoint(&loc, gtid, cncl_kind);
1682 }
1683
1684 // Return true if cancellation should take place, false otherwise
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCEL)1685 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCEL)(int which, bool do_cancel) {
1686 int gtid = __kmp_get_gtid();
1687 MKLOC(loc, "GOMP_cancel");
1688 KA_TRACE(20, ("GOMP_cancel: T#%d which:%d do_cancel:%d\n", gtid, which,
1689 (int)do_cancel));
1690 kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which);
1691
1692 if (do_cancel == FALSE) {
1693 return __kmpc_cancellationpoint(&loc, gtid, cncl_kind);
1694 } else {
1695 return __kmpc_cancel(&loc, gtid, cncl_kind);
1696 }
1697 }
1698
1699 // Return true if cancellation should take place, false otherwise
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER_CANCEL)1700 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER_CANCEL)(void) {
1701 int gtid = __kmp_get_gtid();
1702 KA_TRACE(20, ("GOMP_barrier_cancel: T#%d\n", gtid));
1703 return __kmp_barrier_gomp_cancel(gtid);
1704 }
1705
1706 // Return true if cancellation should take place, false otherwise
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL)1707 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL)(void) {
1708 int gtid = __kmp_get_gtid();
1709 KA_TRACE(20, ("GOMP_sections_end_cancel: T#%d\n", gtid));
1710 return __kmp_barrier_gomp_cancel(gtid);
1711 }
1712
1713 // Return true if cancellation should take place, false otherwise
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_CANCEL)1714 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_CANCEL)(void) {
1715 int gtid = __kmp_get_gtid();
1716 KA_TRACE(20, ("GOMP_loop_end_cancel: T#%d\n", gtid));
1717 return __kmp_barrier_gomp_cancel(gtid);
1718 }
1719
1720 // All target functions are empty as of 2014-05-29
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET)1721 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET)(int device, void (*fn)(void *),
1722 const void *openmp_target,
1723 size_t mapnum, void **hostaddrs,
1724 size_t *sizes,
1725 unsigned char *kinds) {
1726 return;
1727 }
1728
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_DATA)1729 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_DATA)(
1730 int device, const void *openmp_target, size_t mapnum, void **hostaddrs,
1731 size_t *sizes, unsigned char *kinds) {
1732 return;
1733 }
1734
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_END_DATA)1735 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_END_DATA)(void) { return; }
1736
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_UPDATE)1737 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_UPDATE)(
1738 int device, const void *openmp_target, size_t mapnum, void **hostaddrs,
1739 size_t *sizes, unsigned char *kinds) {
1740 return;
1741 }
1742
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS)1743 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS)(unsigned int num_teams,
1744 unsigned int thread_limit) {
1745 return;
1746 }
1747
1748 // Task duplication function which copies src to dest (both are
1749 // preallocated task structures)
__kmp_gomp_task_dup(kmp_task_t * dest,kmp_task_t * src,kmp_int32 last_private)1750 static void __kmp_gomp_task_dup(kmp_task_t *dest, kmp_task_t *src,
1751 kmp_int32 last_private) {
1752 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(src);
1753 if (taskdata->td_copy_func) {
1754 (taskdata->td_copy_func)(dest->shareds, src->shareds);
1755 }
1756 }
1757
1758 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_REGISTER)(
1759 uintptr_t *);
1760
1761 #ifdef __cplusplus
1762 } // extern "C"
1763 #endif
1764
1765 template <typename T>
__GOMP_taskloop(void (* func)(void *),void * data,void (* copy_func)(void *,void *),long arg_size,long arg_align,unsigned gomp_flags,unsigned long num_tasks,int priority,T start,T end,T step)1766 void __GOMP_taskloop(void (*func)(void *), void *data,
1767 void (*copy_func)(void *, void *), long arg_size,
1768 long arg_align, unsigned gomp_flags,
1769 unsigned long num_tasks, int priority, T start, T end,
1770 T step) {
1771 typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32);
1772 MKLOC(loc, "GOMP_taskloop");
1773 int sched;
1774 T *loop_bounds;
1775 int gtid = __kmp_entry_gtid();
1776 kmp_int32 flags = 0;
1777 int if_val = gomp_flags & (1u << 10);
1778 int nogroup = gomp_flags & (1u << 11);
1779 int up = gomp_flags & (1u << 8);
1780 int reductions = gomp_flags & (1u << 12);
1781 p_task_dup_t task_dup = NULL;
1782 kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags;
1783 #ifdef KMP_DEBUG
1784 {
1785 char *buff;
1786 buff = __kmp_str_format(
1787 "GOMP_taskloop: T#%%d: func:%%p data:%%p copy_func:%%p "
1788 "arg_size:%%ld arg_align:%%ld gomp_flags:0x%%x num_tasks:%%lu "
1789 "priority:%%d start:%%%s end:%%%s step:%%%s\n",
1790 traits_t<T>::spec, traits_t<T>::spec, traits_t<T>::spec);
1791 KA_TRACE(20, (buff, gtid, func, data, copy_func, arg_size, arg_align,
1792 gomp_flags, num_tasks, priority, start, end, step));
1793 __kmp_str_free(&buff);
1794 }
1795 #endif
1796 KMP_ASSERT((size_t)arg_size >= 2 * sizeof(T));
1797 KMP_ASSERT(arg_align > 0);
1798 // The low-order bit is the "untied" flag
1799 if (!(gomp_flags & 1)) {
1800 input_flags->tiedness = TASK_TIED;
1801 }
1802 // The second low-order bit is the "final" flag
1803 if (gomp_flags & 2) {
1804 input_flags->final = 1;
1805 }
1806 // Negative step flag
1807 if (!up) {
1808 // If step is flagged as negative, but isn't properly sign extended
1809 // Then manually sign extend it. Could be a short, int, char embedded
1810 // in a long. So cannot assume any cast.
1811 if (step > 0) {
1812 for (int i = sizeof(T) * CHAR_BIT - 1; i >= 0L; --i) {
1813 // break at the first 1 bit
1814 if (step & ((T)1 << i))
1815 break;
1816 step |= ((T)1 << i);
1817 }
1818 }
1819 }
1820 input_flags->native = 1;
1821 // Figure out if none/grainsize/num_tasks clause specified
1822 if (num_tasks > 0) {
1823 if (gomp_flags & (1u << 9))
1824 sched = 1; // grainsize specified
1825 else
1826 sched = 2; // num_tasks specified
1827 // neither grainsize nor num_tasks specified
1828 } else {
1829 sched = 0;
1830 }
1831
1832 // __kmp_task_alloc() sets up all other flags
1833 kmp_task_t *task =
1834 __kmp_task_alloc(&loc, gtid, input_flags, sizeof(kmp_task_t),
1835 arg_size + arg_align - 1, (kmp_routine_entry_t)func);
1836 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
1837 taskdata->td_copy_func = copy_func;
1838 taskdata->td_size_loop_bounds = sizeof(T);
1839
1840 // re-align shareds if needed and setup firstprivate copy constructors
1841 // through the task_dup mechanism
1842 task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) /
1843 arg_align * arg_align);
1844 if (copy_func) {
1845 task_dup = __kmp_gomp_task_dup;
1846 }
1847 KMP_MEMCPY(task->shareds, data, arg_size);
1848
1849 loop_bounds = (T *)task->shareds;
1850 loop_bounds[0] = start;
1851 loop_bounds[1] = end + (up ? -1 : 1);
1852
1853 if (!nogroup) {
1854 #if OMPT_SUPPORT && OMPT_OPTIONAL
1855 OMPT_STORE_RETURN_ADDRESS(gtid);
1856 #endif
1857 __kmpc_taskgroup(&loc, gtid);
1858 if (reductions) {
1859 // The data pointer points to lb, ub, then reduction data
1860 struct data_t {
1861 T a, b;
1862 uintptr_t *d;
1863 };
1864 uintptr_t *d = ((data_t *)data)->d;
1865 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_REGISTER)(d);
1866 }
1867 }
1868 __kmpc_taskloop(&loc, gtid, task, if_val, (kmp_uint64 *)&(loop_bounds[0]),
1869 (kmp_uint64 *)&(loop_bounds[1]), (kmp_int64)step, 1, sched,
1870 (kmp_uint64)num_tasks, (void *)task_dup);
1871 if (!nogroup) {
1872 #if OMPT_SUPPORT && OMPT_OPTIONAL
1873 OMPT_STORE_RETURN_ADDRESS(gtid);
1874 #endif
1875 __kmpc_end_taskgroup(&loc, gtid);
1876 }
1877 }
1878
1879 // 4 byte version of GOMP_doacross_post
1880 // This verison needs to create a temporary array which converts 4 byte
1881 // integers into 8 byte integers
1882 template <typename T, bool need_conversion = (sizeof(long) == 4)>
1883 void __kmp_GOMP_doacross_post(T *count);
1884
__kmp_GOMP_doacross_post(long * count)1885 template <> void __kmp_GOMP_doacross_post<long, true>(long *count) {
1886 int gtid = __kmp_entry_gtid();
1887 kmp_info_t *th = __kmp_threads[gtid];
1888 MKLOC(loc, "GOMP_doacross_post");
1889 kmp_int64 num_dims = th->th.th_dispatch->th_doacross_info[0];
1890 kmp_int64 *vec = (kmp_int64 *)__kmp_thread_malloc(
1891 th, (size_t)(sizeof(kmp_int64) * num_dims));
1892 for (kmp_int64 i = 0; i < num_dims; ++i) {
1893 vec[i] = (kmp_int64)count[i];
1894 }
1895 __kmpc_doacross_post(&loc, gtid, vec);
1896 __kmp_thread_free(th, vec);
1897 }
1898
1899 // 8 byte versions of GOMP_doacross_post
1900 // This version can just pass in the count array directly instead of creating
1901 // a temporary array
__kmp_GOMP_doacross_post(long * count)1902 template <> void __kmp_GOMP_doacross_post<long, false>(long *count) {
1903 int gtid = __kmp_entry_gtid();
1904 MKLOC(loc, "GOMP_doacross_post");
1905 __kmpc_doacross_post(&loc, gtid, RCAST(kmp_int64 *, count));
1906 }
1907
__kmp_GOMP_doacross_wait(T first,va_list args)1908 template <typename T> void __kmp_GOMP_doacross_wait(T first, va_list args) {
1909 int gtid = __kmp_entry_gtid();
1910 kmp_info_t *th = __kmp_threads[gtid];
1911 MKLOC(loc, "GOMP_doacross_wait");
1912 kmp_int64 num_dims = th->th.th_dispatch->th_doacross_info[0];
1913 kmp_int64 *vec = (kmp_int64 *)__kmp_thread_malloc(
1914 th, (size_t)(sizeof(kmp_int64) * num_dims));
1915 vec[0] = (kmp_int64)first;
1916 for (kmp_int64 i = 1; i < num_dims; ++i) {
1917 T item = va_arg(args, T);
1918 vec[i] = (kmp_int64)item;
1919 }
1920 __kmpc_doacross_wait(&loc, gtid, vec);
1921 __kmp_thread_free(th, vec);
1922 return;
1923 }
1924
1925 #ifdef __cplusplus
1926 extern "C" {
1927 #endif // __cplusplus
1928
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP)1929 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP)(
1930 void (*func)(void *), void *data, void (*copy_func)(void *, void *),
1931 long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks,
1932 int priority, long start, long end, long step) {
1933 __GOMP_taskloop<long>(func, data, copy_func, arg_size, arg_align, gomp_flags,
1934 num_tasks, priority, start, end, step);
1935 }
1936
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP_ULL)1937 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP_ULL)(
1938 void (*func)(void *), void *data, void (*copy_func)(void *, void *),
1939 long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks,
1940 int priority, unsigned long long start, unsigned long long end,
1941 unsigned long long step) {
1942 __GOMP_taskloop<unsigned long long>(func, data, copy_func, arg_size,
1943 arg_align, gomp_flags, num_tasks,
1944 priority, start, end, step);
1945 }
1946
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_POST)1947 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_POST)(long *count) {
1948 __kmp_GOMP_doacross_post(count);
1949 }
1950
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_WAIT)1951 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_WAIT)(long first, ...) {
1952 va_list args;
1953 va_start(args, first);
1954 __kmp_GOMP_doacross_wait<long>(first, args);
1955 va_end(args);
1956 }
1957
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_POST)1958 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_POST)(
1959 unsigned long long *count) {
1960 int gtid = __kmp_entry_gtid();
1961 MKLOC(loc, "GOMP_doacross_ull_post");
1962 __kmpc_doacross_post(&loc, gtid, RCAST(kmp_int64 *, count));
1963 }
1964
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT)1965 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT)(
1966 unsigned long long first, ...) {
1967 va_list args;
1968 va_start(args, first);
1969 __kmp_GOMP_doacross_wait<unsigned long long>(first, args);
1970 va_end(args);
1971 }
1972
1973 // fn: the function each primary thread of new team will call
1974 // data: argument to fn
1975 // num_teams, thread_limit: max bounds on respective ICV
1976 // flags: unused
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS_REG)1977 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS_REG)(void (*fn)(void *),
1978 void *data,
1979 unsigned num_teams,
1980 unsigned thread_limit,
1981 unsigned flags) {
1982 MKLOC(loc, "GOMP_teams_reg");
1983 int gtid = __kmp_entry_gtid();
1984 KA_TRACE(20, ("GOMP_teams_reg: T#%d num_teams=%u thread_limit=%u flag=%u\n",
1985 gtid, num_teams, thread_limit, flags));
1986 __kmpc_push_num_teams(&loc, gtid, num_teams, thread_limit);
1987 __kmpc_fork_teams(&loc, 2, (microtask_t)__kmp_GOMP_microtask_wrapper, fn,
1988 data);
1989 KA_TRACE(20, ("GOMP_teams_reg exit: T#%d\n", gtid));
1990 }
1991
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT_DEPEND)1992 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT_DEPEND)(void **depend) {
1993 MKLOC(loc, "GOMP_taskwait_depend");
1994 int gtid = __kmp_entry_gtid();
1995 KA_TRACE(20, ("GOMP_taskwait_depend: T#%d\n", gtid));
1996 kmp_gomp_depends_info_t gomp_depends(depend);
1997 kmp_int32 ndeps = gomp_depends.get_num_deps();
1998 SimpleVLA<kmp_depend_info_t> dep_list(ndeps);
1999 for (kmp_int32 i = 0; i < ndeps; i++)
2000 dep_list[i] = gomp_depends.get_kmp_depend(i);
2001 #if OMPT_SUPPORT
2002 OMPT_STORE_RETURN_ADDRESS(gtid);
2003 #endif
2004 __kmpc_omp_wait_deps(&loc, gtid, ndeps, dep_list, 0, NULL);
2005 KA_TRACE(20, ("GOMP_taskwait_depend exit: T#%d\n", gtid));
2006 }
2007
2008 static inline void
__kmp_GOMP_taskgroup_reduction_register(uintptr_t * data,kmp_taskgroup_t * tg,int nthreads,uintptr_t * allocated=nullptr)2009 __kmp_GOMP_taskgroup_reduction_register(uintptr_t *data, kmp_taskgroup_t *tg,
2010 int nthreads,
2011 uintptr_t *allocated = nullptr) {
2012 KMP_ASSERT(data);
2013 KMP_ASSERT(nthreads > 0);
2014 // Have private copy pointers point to previously allocated
2015 // reduction data or allocate new data here
2016 if (allocated) {
2017 data[2] = allocated[2];
2018 data[6] = allocated[6];
2019 } else {
2020 data[2] = (uintptr_t)__kmp_allocate(nthreads * data[1]);
2021 data[6] = data[2] + (nthreads * data[1]);
2022 }
2023 if (tg)
2024 tg->gomp_data = data;
2025 }
2026
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_REGISTER)2027 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_REGISTER)(
2028 uintptr_t *data) {
2029 int gtid = __kmp_entry_gtid();
2030 KA_TRACE(20, ("GOMP_taskgroup_reduction_register: T#%d\n", gtid));
2031 kmp_info_t *thread = __kmp_threads[gtid];
2032 kmp_taskgroup_t *tg = thread->th.th_current_task->td_taskgroup;
2033 int nthreads = thread->th.th_team_nproc;
2034 __kmp_GOMP_taskgroup_reduction_register(data, tg, nthreads);
2035 }
2036
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_UNREGISTER)2037 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_UNREGISTER)(
2038 uintptr_t *data) {
2039 KA_TRACE(20,
2040 ("GOMP_taskgroup_reduction_unregister: T#%d\n", __kmp_get_gtid()));
2041 KMP_ASSERT(data && data[2]);
2042 __kmp_free((void *)data[2]);
2043 }
2044
2045 // Search through reduction data and set ptrs[] elements
2046 // to proper privatized copy address
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASK_REDUCTION_REMAP)2047 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASK_REDUCTION_REMAP)(size_t cnt,
2048 size_t cntorig,
2049 void **ptrs) {
2050 int gtid = __kmp_entry_gtid();
2051 KA_TRACE(20, ("GOMP_task_reduction_remap: T#%d\n", gtid));
2052 kmp_info_t *thread = __kmp_threads[gtid];
2053 kmp_int32 tid = __kmp_get_tid();
2054 for (size_t i = 0; i < cnt; ++i) {
2055 uintptr_t address = (uintptr_t)ptrs[i];
2056 void *propagated_address = NULL;
2057 void *mapped_address = NULL;
2058 // Check taskgroups reduce data
2059 kmp_taskgroup_t *tg = thread->th.th_current_task->td_taskgroup;
2060 while (tg) {
2061 uintptr_t *gomp_data = tg->gomp_data;
2062 if (!gomp_data) {
2063 tg = tg->parent;
2064 continue;
2065 }
2066 // Check the shared addresses list
2067 size_t num_vars = (size_t)gomp_data[0];
2068 uintptr_t per_thread_size = gomp_data[1];
2069 uintptr_t reduce_data = gomp_data[2];
2070 uintptr_t end_reduce_data = gomp_data[6];
2071 for (size_t j = 0; j < num_vars; ++j) {
2072 uintptr_t *entry = gomp_data + 7 + 3 * j;
2073 if (entry[0] == address) {
2074 uintptr_t offset = entry[1];
2075 mapped_address =
2076 (void *)(reduce_data + tid * per_thread_size + offset);
2077 if (i < cntorig)
2078 propagated_address = (void *)entry[0];
2079 break;
2080 }
2081 }
2082 if (mapped_address)
2083 break;
2084 // Check if address is within privatized copies range
2085 if (!mapped_address && address >= reduce_data &&
2086 address < end_reduce_data) {
2087 uintptr_t offset = (address - reduce_data) % per_thread_size;
2088 mapped_address = (void *)(reduce_data + tid * per_thread_size + offset);
2089 if (i < cntorig) {
2090 for (size_t j = 0; j < num_vars; ++j) {
2091 uintptr_t *entry = gomp_data + 7 + 3 * j;
2092 if (entry[1] == offset) {
2093 propagated_address = (void *)entry[0];
2094 break;
2095 }
2096 }
2097 }
2098 }
2099 if (mapped_address)
2100 break;
2101 tg = tg->parent;
2102 }
2103 KMP_ASSERT(mapped_address);
2104 ptrs[i] = mapped_address;
2105 if (i < cntorig) {
2106 KMP_ASSERT(propagated_address);
2107 ptrs[cnt + i] = propagated_address;
2108 }
2109 }
2110 }
2111
__kmp_GOMP_init_reductions(int gtid,uintptr_t * data,int is_ws)2112 static void __kmp_GOMP_init_reductions(int gtid, uintptr_t *data, int is_ws) {
2113 kmp_info_t *thr = __kmp_threads[gtid];
2114 kmp_team_t *team = thr->th.th_team;
2115 // First start a taskgroup
2116 __kmpc_taskgroup(NULL, gtid);
2117 // Then setup reduction data
2118 void *reduce_data = KMP_ATOMIC_LD_RLX(&team->t.t_tg_reduce_data[is_ws]);
2119 if (reduce_data == NULL &&
2120 __kmp_atomic_compare_store(&team->t.t_tg_reduce_data[is_ws], reduce_data,
2121 (void *)1)) {
2122 // Single thread enters this block to initialize common reduction data
2123 KMP_DEBUG_ASSERT(reduce_data == NULL);
2124 __kmp_GOMP_taskgroup_reduction_register(data, NULL, thr->th.th_team_nproc);
2125 KMP_ATOMIC_ST_REL(&team->t.t_tg_fini_counter[is_ws], 0);
2126 KMP_ATOMIC_ST_REL(&team->t.t_tg_reduce_data[is_ws], (void *)data);
2127 } else {
2128 // Wait for task reduction initialization
2129 while ((reduce_data = KMP_ATOMIC_LD_ACQ(
2130 &team->t.t_tg_reduce_data[is_ws])) == (void *)1) {
2131 KMP_CPU_PAUSE();
2132 }
2133 KMP_DEBUG_ASSERT(reduce_data > (void *)1); // should be valid pointer here
2134 }
2135 // For worksharing constructs, each thread has its own reduction structure.
2136 // Have each reduction structure point to same privatized copies of vars.
2137 // For parallel, each thread points to same reduction structure and privatized
2138 // copies of vars
2139 if (is_ws) {
2140 __kmp_GOMP_taskgroup_reduction_register(
2141 data, NULL, thr->th.th_team_nproc,
2142 (uintptr_t *)KMP_ATOMIC_LD_ACQ(&team->t.t_tg_reduce_data[is_ws]));
2143 }
2144 kmp_taskgroup_t *tg = thr->th.th_current_task->td_taskgroup;
2145 tg->gomp_data = data;
2146 }
2147
2148 static unsigned
__kmp_GOMP_par_reductions_microtask_wrapper(int * gtid,int * npr,void (* task)(void *),void * data)2149 __kmp_GOMP_par_reductions_microtask_wrapper(int *gtid, int *npr,
2150 void (*task)(void *), void *data) {
2151 kmp_info_t *thr = __kmp_threads[*gtid];
2152 kmp_team_t *team = thr->th.th_team;
2153 uintptr_t *reduce_data = *(uintptr_t **)data;
2154 __kmp_GOMP_init_reductions(*gtid, reduce_data, 0);
2155
2156 #if OMPT_SUPPORT
2157 ompt_frame_t *ompt_frame;
2158 ompt_state_t enclosing_state;
2159
2160 if (ompt_enabled.enabled) {
2161 // save enclosing task state; set current state for task
2162 enclosing_state = thr->th.ompt_thread_info.state;
2163 thr->th.ompt_thread_info.state = ompt_state_work_parallel;
2164
2165 // set task frame
2166 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
2167 ompt_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
2168 }
2169 #endif
2170
2171 task(data);
2172
2173 #if OMPT_SUPPORT
2174 if (ompt_enabled.enabled) {
2175 // clear task frame
2176 ompt_frame->exit_frame = ompt_data_none;
2177
2178 // restore enclosing state
2179 thr->th.ompt_thread_info.state = enclosing_state;
2180 }
2181 #endif
2182 __kmpc_end_taskgroup(NULL, *gtid);
2183 // if last thread out, then reset the team's reduce data
2184 // the GOMP_taskgroup_reduction_unregister() function will deallocate
2185 // private copies after reduction calculations take place.
2186 int count = KMP_ATOMIC_INC(&team->t.t_tg_fini_counter[0]);
2187 if (count == thr->th.th_team_nproc - 1) {
2188 KMP_ATOMIC_ST_REL(&team->t.t_tg_reduce_data[0], NULL);
2189 KMP_ATOMIC_ST_REL(&team->t.t_tg_fini_counter[0], 0);
2190 }
2191 return (unsigned)thr->th.th_team_nproc;
2192 }
2193
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_REDUCTIONS)2194 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_REDUCTIONS)(
2195 void (*task)(void *), void *data, unsigned num_threads,
2196 unsigned int flags) {
2197 MKLOC(loc, "GOMP_parallel_reductions");
2198 int gtid = __kmp_entry_gtid();
2199 KA_TRACE(20, ("GOMP_parallel_reductions: T#%d\n", gtid));
2200 __kmp_GOMP_fork_call(&loc, gtid, num_threads, flags, task,
2201 (microtask_t)__kmp_GOMP_par_reductions_microtask_wrapper,
2202 2, task, data);
2203 unsigned retval =
2204 __kmp_GOMP_par_reductions_microtask_wrapper(>id, NULL, task, data);
2205 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)();
2206 KA_TRACE(20, ("GOMP_parallel_reductions exit: T#%d\n", gtid));
2207 return retval;
2208 }
2209
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_START)2210 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_START)(
2211 long start, long end, long incr, long sched, long chunk_size, long *istart,
2212 long *iend, uintptr_t *reductions, void **mem) {
2213 int status = 0;
2214 int gtid = __kmp_entry_gtid();
2215 KA_TRACE(20, ("GOMP_loop_start: T#%d, reductions: %p\n", gtid, reductions));
2216 if (reductions)
2217 __kmp_GOMP_init_reductions(gtid, reductions, 1);
2218 if (mem)
2219 KMP_FATAL(GompFeatureNotSupported, "scan");
2220 if (istart == NULL)
2221 return true;
2222 const long MONOTONIC_FLAG = (long)(kmp_sched_monotonic);
2223 long monotonic = sched & MONOTONIC_FLAG;
2224 sched &= ~MONOTONIC_FLAG;
2225 if (sched == 0) {
2226 if (monotonic)
2227 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_START)(
2228 start, end, incr, istart, iend);
2229 else
2230 status = KMP_EXPAND_NAME(
2231 KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_START)(
2232 start, end, incr, istart, iend);
2233 } else if (sched == 1) {
2234 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_START)(
2235 start, end, incr, chunk_size, istart, iend);
2236 } else if (sched == 2) {
2237 if (monotonic)
2238 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START)(
2239 start, end, incr, chunk_size, istart, iend);
2240 else
2241 status =
2242 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_START)(
2243 start, end, incr, chunk_size, istart, iend);
2244 } else if (sched == 3) {
2245 if (monotonic)
2246 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_START)(
2247 start, end, incr, chunk_size, istart, iend);
2248 else
2249 status =
2250 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_START)(
2251 start, end, incr, chunk_size, istart, iend);
2252 } else if (sched == 4) {
2253 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_START)(
2254 start, end, incr, istart, iend);
2255 } else {
2256 KMP_ASSERT(0);
2257 }
2258 return status;
2259 }
2260
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_START)2261 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_START)(
2262 bool up, unsigned long long start, unsigned long long end,
2263 unsigned long long incr, long sched, unsigned long long chunk_size,
2264 unsigned long long *istart, unsigned long long *iend, uintptr_t *reductions,
2265 void **mem) {
2266 int status = 0;
2267 int gtid = __kmp_entry_gtid();
2268 KA_TRACE(20,
2269 ("GOMP_loop_ull_start: T#%d, reductions: %p\n", gtid, reductions));
2270 if (reductions)
2271 __kmp_GOMP_init_reductions(gtid, reductions, 1);
2272 if (mem)
2273 KMP_FATAL(GompFeatureNotSupported, "scan");
2274 if (istart == NULL)
2275 return true;
2276 const long MONOTONIC_FLAG = (long)(kmp_sched_monotonic);
2277 long monotonic = sched & MONOTONIC_FLAG;
2278 sched &= ~MONOTONIC_FLAG;
2279 if (sched == 0) {
2280 if (monotonic)
2281 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START)(
2282 up, start, end, incr, istart, iend);
2283 else
2284 status = KMP_EXPAND_NAME(
2285 KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_START)(
2286 up, start, end, incr, istart, iend);
2287 } else if (sched == 1) {
2288 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START)(
2289 up, start, end, incr, chunk_size, istart, iend);
2290 } else if (sched == 2) {
2291 if (monotonic)
2292 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START)(
2293 up, start, end, incr, chunk_size, istart, iend);
2294 else
2295 status = KMP_EXPAND_NAME(
2296 KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_START)(
2297 up, start, end, incr, chunk_size, istart, iend);
2298 } else if (sched == 3) {
2299 if (monotonic)
2300 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START)(
2301 up, start, end, incr, chunk_size, istart, iend);
2302 else
2303 status =
2304 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_START)(
2305 up, start, end, incr, chunk_size, istart, iend);
2306 } else if (sched == 4) {
2307 status =
2308 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_START)(
2309 up, start, end, incr, istart, iend);
2310 } else {
2311 KMP_ASSERT(0);
2312 }
2313 return status;
2314 }
2315
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_START)2316 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_START)(
2317 unsigned ncounts, long *counts, long sched, long chunk_size, long *istart,
2318 long *iend, uintptr_t *reductions, void **mem) {
2319 int status = 0;
2320 int gtid = __kmp_entry_gtid();
2321 KA_TRACE(20, ("GOMP_loop_doacross_start: T#%d, reductions: %p\n", gtid,
2322 reductions));
2323 if (reductions)
2324 __kmp_GOMP_init_reductions(gtid, reductions, 1);
2325 if (mem)
2326 KMP_FATAL(GompFeatureNotSupported, "scan");
2327 if (istart == NULL)
2328 return true;
2329 // Ignore any monotonic flag
2330 const long MONOTONIC_FLAG = (long)(kmp_sched_monotonic);
2331 sched &= ~MONOTONIC_FLAG;
2332 if (sched == 0) {
2333 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START)(
2334 ncounts, counts, istart, iend);
2335 } else if (sched == 1) {
2336 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START)(
2337 ncounts, counts, chunk_size, istart, iend);
2338 } else if (sched == 2) {
2339 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START)(
2340 ncounts, counts, chunk_size, istart, iend);
2341 } else if (sched == 3) {
2342 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START)(
2343 ncounts, counts, chunk_size, istart, iend);
2344 } else {
2345 KMP_ASSERT(0);
2346 }
2347 return status;
2348 }
2349
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_START)2350 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_START)(
2351 unsigned ncounts, unsigned long long *counts, long sched,
2352 unsigned long long chunk_size, unsigned long long *istart,
2353 unsigned long long *iend, uintptr_t *reductions, void **mem) {
2354 int status = 0;
2355 int gtid = __kmp_entry_gtid();
2356 KA_TRACE(20, ("GOMP_loop_ull_doacross_start: T#%d, reductions: %p\n", gtid,
2357 reductions));
2358 if (reductions)
2359 __kmp_GOMP_init_reductions(gtid, reductions, 1);
2360 if (mem)
2361 KMP_FATAL(GompFeatureNotSupported, "scan");
2362 if (istart == NULL)
2363 return true;
2364 // Ignore any monotonic flag
2365 const long MONOTONIC_FLAG = (long)(kmp_sched_monotonic);
2366 sched &= ~MONOTONIC_FLAG;
2367 if (sched == 0) {
2368 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START)(
2369 ncounts, counts, istart, iend);
2370 } else if (sched == 1) {
2371 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START)(
2372 ncounts, counts, chunk_size, istart, iend);
2373 } else if (sched == 2) {
2374 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START)(
2375 ncounts, counts, chunk_size, istart, iend);
2376 } else if (sched == 3) {
2377 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START)(
2378 ncounts, counts, chunk_size, istart, iend);
2379 } else {
2380 KMP_ASSERT(0);
2381 }
2382 return status;
2383 }
2384
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_START)2385 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_START)(
2386 long start, long end, long incr, long sched, long chunk_size, long *istart,
2387 long *iend, uintptr_t *reductions, void **mem) {
2388 int status = 0;
2389 int gtid = __kmp_entry_gtid();
2390 KA_TRACE(20, ("GOMP_loop_ordered_start: T#%d, reductions: %p\n", gtid,
2391 reductions));
2392 if (reductions)
2393 __kmp_GOMP_init_reductions(gtid, reductions, 1);
2394 if (mem)
2395 KMP_FATAL(GompFeatureNotSupported, "scan");
2396 if (istart == NULL)
2397 return true;
2398 // Ignore any monotonic flag
2399 const long MONOTONIC_FLAG = (long)(kmp_sched_monotonic);
2400 sched &= ~MONOTONIC_FLAG;
2401 if (sched == 0) {
2402 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START)(
2403 start, end, incr, istart, iend);
2404 } else if (sched == 1) {
2405 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START)(
2406 start, end, incr, chunk_size, istart, iend);
2407 } else if (sched == 2) {
2408 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START)(
2409 start, end, incr, chunk_size, istart, iend);
2410 } else if (sched == 3) {
2411 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START)(
2412 start, end, incr, chunk_size, istart, iend);
2413 } else {
2414 KMP_ASSERT(0);
2415 }
2416 return status;
2417 }
2418
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_START)2419 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_START)(
2420 bool up, unsigned long long start, unsigned long long end,
2421 unsigned long long incr, long sched, unsigned long long chunk_size,
2422 unsigned long long *istart, unsigned long long *iend, uintptr_t *reductions,
2423 void **mem) {
2424 int status = 0;
2425 int gtid = __kmp_entry_gtid();
2426 KA_TRACE(20, ("GOMP_loop_ull_ordered_start: T#%d, reductions: %p\n", gtid,
2427 reductions));
2428 if (reductions)
2429 __kmp_GOMP_init_reductions(gtid, reductions, 1);
2430 if (mem)
2431 KMP_FATAL(GompFeatureNotSupported, "scan");
2432 if (istart == NULL)
2433 return true;
2434 // Ignore any monotonic flag
2435 const long MONOTONIC_FLAG = (long)(kmp_sched_monotonic);
2436 sched &= ~MONOTONIC_FLAG;
2437 if (sched == 0) {
2438 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START)(
2439 up, start, end, incr, istart, iend);
2440 } else if (sched == 1) {
2441 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START)(
2442 up, start, end, incr, chunk_size, istart, iend);
2443 } else if (sched == 2) {
2444 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START)(
2445 up, start, end, incr, chunk_size, istart, iend);
2446 } else if (sched == 3) {
2447 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START)(
2448 up, start, end, incr, chunk_size, istart, iend);
2449 } else {
2450 KMP_ASSERT(0);
2451 }
2452 return status;
2453 }
2454
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS2_START)2455 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS2_START)(
2456 unsigned count, uintptr_t *reductions, void **mem) {
2457 int gtid = __kmp_entry_gtid();
2458 KA_TRACE(20,
2459 ("GOMP_sections2_start: T#%d, reductions: %p\n", gtid, reductions));
2460 if (reductions)
2461 __kmp_GOMP_init_reductions(gtid, reductions, 1);
2462 if (mem)
2463 KMP_FATAL(GompFeatureNotSupported, "scan");
2464 return KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_START)(count);
2465 }
2466
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_WORKSHARE_TASK_REDUCTION_UNREGISTER)2467 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_WORKSHARE_TASK_REDUCTION_UNREGISTER)(
2468 bool cancelled) {
2469 int gtid = __kmp_get_gtid();
2470 MKLOC(loc, "GOMP_workshare_task_reduction_unregister");
2471 KA_TRACE(20, ("GOMP_workshare_task_reduction_unregister: T#%d\n", gtid));
2472 kmp_info_t *thr = __kmp_threads[gtid];
2473 kmp_team_t *team = thr->th.th_team;
2474 __kmpc_end_taskgroup(NULL, gtid);
2475 // If last thread out of workshare, then reset the team's reduce data
2476 // the GOMP_taskgroup_reduction_unregister() function will deallocate
2477 // private copies after reduction calculations take place.
2478 int count = KMP_ATOMIC_INC(&team->t.t_tg_fini_counter[1]);
2479 if (count == thr->th.th_team_nproc - 1) {
2480 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_UNREGISTER)
2481 ((uintptr_t *)KMP_ATOMIC_LD_RLX(&team->t.t_tg_reduce_data[1]));
2482 KMP_ATOMIC_ST_REL(&team->t.t_tg_reduce_data[1], NULL);
2483 KMP_ATOMIC_ST_REL(&team->t.t_tg_fini_counter[1], 0);
2484 }
2485 if (!cancelled) {
2486 __kmpc_barrier(&loc, gtid);
2487 }
2488 }
2489
2490 // allocator construct
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ALLOC)2491 void *KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ALLOC)(size_t alignment, size_t size,
2492 uintptr_t allocator) {
2493 int gtid = __kmp_entry_gtid();
2494 KA_TRACE(20, ("GOMP_alloc: T#%d\n", gtid));
2495 #if OMPT_SUPPORT && OMPT_OPTIONAL
2496 OMPT_STORE_RETURN_ADDRESS(gtid);
2497 #endif
2498 return __kmp_alloc(gtid, alignment, size, (omp_allocator_handle_t)allocator);
2499 }
2500
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_FREE)2501 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_FREE)(void *ptr, uintptr_t allocator) {
2502 int gtid = __kmp_entry_gtid();
2503 KA_TRACE(20, ("GOMP_free: T#%d\n", gtid));
2504 #if OMPT_SUPPORT && OMPT_OPTIONAL
2505 OMPT_STORE_RETURN_ADDRESS(gtid);
2506 #endif
2507 return ___kmpc_free(gtid, ptr, (omp_allocator_handle_t)allocator);
2508 }
2509
2510 /* The following sections of code create aliases for the GOMP_* functions, then
2511 create versioned symbols using the assembler directive .symver. This is only
2512 pertinent for ELF .so library. The KMP_VERSION_SYMBOL macro is defined in
2513 kmp_os.h */
2514
2515 #ifdef KMP_USE_VERSION_SYMBOLS
2516 // GOMP_1.0 versioned symbols
2517 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_END, 10, "GOMP_1.0");
2518 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_START, 10, "GOMP_1.0");
2519 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER, 10, "GOMP_1.0");
2520 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_END, 10, "GOMP_1.0");
2521 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_END, 10, "GOMP_1.0");
2522 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_START, 10, "GOMP_1.0");
2523 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_START, 10, "GOMP_1.0");
2524 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT, 10, "GOMP_1.0");
2525 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START, 10, "GOMP_1.0");
2526 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END, 10, "GOMP_1.0");
2527 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_NOWAIT, 10, "GOMP_1.0");
2528 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT, 10, "GOMP_1.0");
2529 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_START, 10, "GOMP_1.0");
2530 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT, 10, "GOMP_1.0");
2531 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START, 10,
2532 "GOMP_1.0");
2533 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT, 10, "GOMP_1.0");
2534 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START, 10, "GOMP_1.0");
2535 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT, 10, "GOMP_1.0");
2536 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START, 10,
2537 "GOMP_1.0");
2538 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT, 10, "GOMP_1.0");
2539 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START, 10, "GOMP_1.0");
2540 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT, 10, "GOMP_1.0");
2541 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_START, 10, "GOMP_1.0");
2542 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT, 10, "GOMP_1.0");
2543 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_START, 10, "GOMP_1.0");
2544 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_END, 10, "GOMP_1.0");
2545 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_START, 10, "GOMP_1.0");
2546 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_END, 10, "GOMP_1.0");
2547 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START, 10,
2548 "GOMP_1.0");
2549 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START, 10,
2550 "GOMP_1.0");
2551 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START, 10,
2552 "GOMP_1.0");
2553 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START, 10,
2554 "GOMP_1.0");
2555 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START, 10, "GOMP_1.0");
2556 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_START, 10, "GOMP_1.0");
2557 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END, 10, "GOMP_1.0");
2558 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT, 10, "GOMP_1.0");
2559 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_NEXT, 10, "GOMP_1.0");
2560 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_START, 10, "GOMP_1.0");
2561 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_END, 10, "GOMP_1.0");
2562 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_START, 10, "GOMP_1.0");
2563 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_START, 10, "GOMP_1.0");
2564
2565 // GOMP_2.0 versioned symbols
2566 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASK, 20, "GOMP_2.0");
2567 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKWAIT, 20, "GOMP_2.0");
2568 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT, 20, "GOMP_2.0");
2569 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START, 20, "GOMP_2.0");
2570 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT, 20, "GOMP_2.0");
2571 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START, 20, "GOMP_2.0");
2572 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT, 20,
2573 "GOMP_2.0");
2574 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START, 20,
2575 "GOMP_2.0");
2576 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT, 20,
2577 "GOMP_2.0");
2578 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START, 20,
2579 "GOMP_2.0");
2580 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT, 20,
2581 "GOMP_2.0");
2582 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START, 20,
2583 "GOMP_2.0");
2584 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT, 20,
2585 "GOMP_2.0");
2586 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START, 20,
2587 "GOMP_2.0");
2588 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT, 20, "GOMP_2.0");
2589 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START, 20, "GOMP_2.0");
2590 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT, 20, "GOMP_2.0");
2591 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START, 20, "GOMP_2.0");
2592
2593 // GOMP_3.0 versioned symbols
2594 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKYIELD, 30, "GOMP_3.0");
2595
2596 // GOMP_4.0 versioned symbols
2597 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL, 40, "GOMP_4.0");
2598 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS, 40, "GOMP_4.0");
2599 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC, 40, "GOMP_4.0");
2600 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED, 40, "GOMP_4.0");
2601 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME, 40, "GOMP_4.0");
2602 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC, 40, "GOMP_4.0");
2603 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_START, 40, "GOMP_4.0");
2604 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_END, 40, "GOMP_4.0");
2605 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER_CANCEL, 40, "GOMP_4.0");
2606 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCEL, 40, "GOMP_4.0");
2607 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCELLATION_POINT, 40, "GOMP_4.0");
2608 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_CANCEL, 40, "GOMP_4.0");
2609 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL, 40, "GOMP_4.0");
2610 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET, 40, "GOMP_4.0");
2611 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_DATA, 40, "GOMP_4.0");
2612 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_END_DATA, 40, "GOMP_4.0");
2613 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_UPDATE, 40, "GOMP_4.0");
2614 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TEAMS, 40, "GOMP_4.0");
2615
2616 // GOMP_4.5 versioned symbols
2617 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP, 45, "GOMP_4.5");
2618 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP_ULL, 45, "GOMP_4.5");
2619 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_POST, 45, "GOMP_4.5");
2620 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_WAIT, 45, "GOMP_4.5");
2621 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START, 45,
2622 "GOMP_4.5");
2623 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START, 45,
2624 "GOMP_4.5");
2625 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START, 45,
2626 "GOMP_4.5");
2627 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START, 45,
2628 "GOMP_4.5");
2629 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_ULL_POST, 45, "GOMP_4.5");
2630 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT, 45, "GOMP_4.5");
2631 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START, 45,
2632 "GOMP_4.5");
2633 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START, 45,
2634 "GOMP_4.5");
2635 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START, 45,
2636 "GOMP_4.5");
2637 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START, 45,
2638 "GOMP_4.5");
2639 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_START, 45,
2640 "GOMP_4.5");
2641 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_NEXT, 45,
2642 "GOMP_4.5");
2643 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_START, 45,
2644 "GOMP_4.5");
2645 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_NEXT, 45,
2646 "GOMP_4.5");
2647 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_START, 45,
2648 "GOMP_4.5");
2649 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_NEXT, 45,
2650 "GOMP_4.5");
2651 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_START, 45,
2652 "GOMP_4.5");
2653 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_NEXT, 45,
2654 "GOMP_4.5");
2655 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_DYNAMIC, 45,
2656 "GOMP_4.5");
2657 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_GUIDED, 45,
2658 "GOMP_4.5");
2659
2660 // GOMP_5.0 versioned symbols
2661 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_NEXT, 50,
2662 "GOMP_5.0");
2663 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_START, 50,
2664 "GOMP_5.0");
2665 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_NEXT, 50,
2666 "GOMP_5.0");
2667 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_START, 50,
2668 "GOMP_5.0");
2669 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_NEXT,
2670 50, "GOMP_5.0");
2671 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_START,
2672 50, "GOMP_5.0");
2673 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_NEXT, 50,
2674 "GOMP_5.0");
2675 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_START, 50,
2676 "GOMP_5.0");
2677 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_RUNTIME, 50,
2678 "GOMP_5.0");
2679 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_MAYBE_NONMONOTONIC_RUNTIME,
2680 50, "GOMP_5.0");
2681 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TEAMS_REG, 50, "GOMP_5.0");
2682 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKWAIT_DEPEND, 50, "GOMP_5.0");
2683 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_REGISTER, 50,
2684 "GOMP_5.0");
2685 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_UNREGISTER, 50,
2686 "GOMP_5.0");
2687 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASK_REDUCTION_REMAP, 50, "GOMP_5.0");
2688 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_REDUCTIONS, 50, "GOMP_5.0");
2689 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_START, 50, "GOMP_5.0");
2690 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_START, 50, "GOMP_5.0");
2691 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_START, 50, "GOMP_5.0");
2692 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_START, 50, "GOMP_5.0");
2693 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_START, 50, "GOMP_5.0");
2694 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_START, 50, "GOMP_5.0");
2695 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS2_START, 50, "GOMP_5.0");
2696 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_WORKSHARE_TASK_REDUCTION_UNREGISTER, 50,
2697 "GOMP_5.0");
2698
2699 // GOMP_5.0.1 versioned symbols
2700 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ALLOC, 501, "GOMP_5.0.1");
2701 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_FREE, 501, "GOMP_5.0.1");
2702 #endif // KMP_USE_VERSION_SYMBOLS
2703
2704 #ifdef __cplusplus
2705 } // extern "C"
2706 #endif // __cplusplus
2707