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
16 #if OMPT_SUPPORT
17 #include "ompt-specific.h"
18 #endif
19
20 enum {
21 KMP_GOMP_TASK_UNTIED_FLAG = 1,
22 KMP_GOMP_TASK_FINAL_FLAG = 2,
23 KMP_GOMP_TASK_DEPENDS_FLAG = 8
24 };
25
26 // This class helps convert gomp dependency info into
27 // kmp_depend_info_t structures
28 class kmp_gomp_depends_info_t {
29 void **depend;
30 kmp_int32 num_deps;
31 size_t num_out, num_mutexinout, num_in;
32 size_t offset;
33
34 public:
kmp_gomp_depends_info_t(void ** depend)35 kmp_gomp_depends_info_t(void **depend) : depend(depend) {
36 size_t ndeps = (kmp_intptr_t)depend[0];
37 size_t num_doable;
38 // GOMP taskdep structure:
39 // if depend[0] != 0:
40 // depend = [ ndeps | nout | &out | ... | &out | &in | ... | &in ]
41 //
42 // if depend[0] == 0:
43 // depend = [ 0 | ndeps | nout | nmtx | nin | &out | ... | &out | &mtx |
44 // ... | &mtx | &in | ... | &in | &depobj | ... | &depobj ]
45 if (ndeps) {
46 num_out = (kmp_intptr_t)depend[1];
47 num_in = ndeps - num_out;
48 num_mutexinout = 0;
49 num_doable = ndeps;
50 offset = 2;
51 } else {
52 ndeps = (kmp_intptr_t)depend[1];
53 num_out = (kmp_intptr_t)depend[2];
54 num_mutexinout = (kmp_intptr_t)depend[3];
55 num_in = (kmp_intptr_t)depend[4];
56 num_doable = num_out + num_mutexinout + num_in;
57 offset = 5;
58 }
59 // TODO: Support gomp depobj
60 if (ndeps != num_doable) {
61 KMP_FATAL(GompFeatureNotSupported, "depobj");
62 }
63 num_deps = static_cast<kmp_int32>(ndeps);
64 }
get_num_deps() const65 kmp_int32 get_num_deps() const { return num_deps; }
get_kmp_depend(size_t index) const66 kmp_depend_info_t get_kmp_depend(size_t index) const {
67 kmp_depend_info_t retval;
68 memset(&retval, '\0', sizeof(retval));
69 KMP_ASSERT(index < (size_t)num_deps);
70 retval.base_addr = (kmp_intptr_t)depend[offset + index];
71 retval.len = 0;
72 // Because inout and out are logically equivalent,
73 // use inout and in dependency flags. GOMP does not provide a
74 // way to distinguish if user specified out vs. inout.
75 if (index < num_out) {
76 retval.flags.in = 1;
77 retval.flags.out = 1;
78 } else if (index >= num_out && index < (num_out + num_mutexinout)) {
79 retval.flags.mtx = 1;
80 } else {
81 retval.flags.in = 1;
82 }
83 return retval;
84 }
85 };
86
87 #ifdef __cplusplus
88 extern "C" {
89 #endif // __cplusplus
90
91 #define MKLOC(loc, routine) \
92 static ident_t loc = {0, KMP_IDENT_KMPC, 0, 0, ";unknown;unknown;0;0;;"};
93
94 #include "kmp_ftn_os.h"
95
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER)96 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER)(void) {
97 int gtid = __kmp_entry_gtid();
98 MKLOC(loc, "GOMP_barrier");
99 KA_TRACE(20, ("GOMP_barrier: T#%d\n", gtid));
100 #if OMPT_SUPPORT && OMPT_OPTIONAL
101 ompt_frame_t *ompt_frame;
102 if (ompt_enabled.enabled) {
103 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
104 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
105 }
106 OMPT_STORE_RETURN_ADDRESS(gtid);
107 #endif
108 __kmpc_barrier(&loc, gtid);
109 #if OMPT_SUPPORT && OMPT_OPTIONAL
110 if (ompt_enabled.enabled) {
111 ompt_frame->enter_frame = ompt_data_none;
112 }
113 #endif
114 }
115
116 // Mutual exclusion
117
118 // The symbol that icc/ifort generates for unnamed for unnamed critical sections
119 // - .gomp_critical_user_ - is defined using .comm in any objects reference it.
120 // We can't reference it directly here in C code, as the symbol contains a ".".
121 //
122 // The RTL contains an assembly language definition of .gomp_critical_user_
123 // with another symbol __kmp_unnamed_critical_addr initialized with it's
124 // address.
125 extern kmp_critical_name *__kmp_unnamed_critical_addr;
126
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_START)127 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_START)(void) {
128 int gtid = __kmp_entry_gtid();
129 MKLOC(loc, "GOMP_critical_start");
130 KA_TRACE(20, ("GOMP_critical_start: T#%d\n", gtid));
131 #if OMPT_SUPPORT && OMPT_OPTIONAL
132 OMPT_STORE_RETURN_ADDRESS(gtid);
133 #endif
134 __kmpc_critical(&loc, gtid, __kmp_unnamed_critical_addr);
135 }
136
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_END)137 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_END)(void) {
138 int gtid = __kmp_get_gtid();
139 MKLOC(loc, "GOMP_critical_end");
140 KA_TRACE(20, ("GOMP_critical_end: T#%d\n", gtid));
141 #if OMPT_SUPPORT && OMPT_OPTIONAL
142 OMPT_STORE_RETURN_ADDRESS(gtid);
143 #endif
144 __kmpc_end_critical(&loc, gtid, __kmp_unnamed_critical_addr);
145 }
146
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_START)147 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_START)(void **pptr) {
148 int gtid = __kmp_entry_gtid();
149 MKLOC(loc, "GOMP_critical_name_start");
150 KA_TRACE(20, ("GOMP_critical_name_start: T#%d\n", gtid));
151 __kmpc_critical(&loc, gtid, (kmp_critical_name *)pptr);
152 }
153
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_END)154 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_END)(void **pptr) {
155 int gtid = __kmp_get_gtid();
156 MKLOC(loc, "GOMP_critical_name_end");
157 KA_TRACE(20, ("GOMP_critical_name_end: T#%d\n", gtid));
158 __kmpc_end_critical(&loc, gtid, (kmp_critical_name *)pptr);
159 }
160
161 // The Gnu codegen tries to use locked operations to perform atomic updates
162 // inline. If it can't, then it calls GOMP_atomic_start() before performing
163 // the update and GOMP_atomic_end() afterward, regardless of the data type.
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_START)164 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_START)(void) {
165 int gtid = __kmp_entry_gtid();
166 KA_TRACE(20, ("GOMP_atomic_start: T#%d\n", gtid));
167
168 #if OMPT_SUPPORT
169 __ompt_thread_assign_wait_id(0);
170 #endif
171
172 __kmp_acquire_atomic_lock(&__kmp_atomic_lock, gtid);
173 }
174
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_END)175 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_END)(void) {
176 int gtid = __kmp_get_gtid();
177 KA_TRACE(20, ("GOMP_atomic_end: T#%d\n", gtid));
178 __kmp_release_atomic_lock(&__kmp_atomic_lock, gtid);
179 }
180
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_START)181 int KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_START)(void) {
182 int gtid = __kmp_entry_gtid();
183 MKLOC(loc, "GOMP_single_start");
184 KA_TRACE(20, ("GOMP_single_start: T#%d\n", gtid));
185
186 if (!TCR_4(__kmp_init_parallel))
187 __kmp_parallel_initialize();
188 __kmp_resume_if_soft_paused();
189
190 // 3rd parameter == FALSE prevents kmp_enter_single from pushing a
191 // workshare when USE_CHECKS is defined. We need to avoid the push,
192 // as there is no corresponding GOMP_single_end() call.
193 kmp_int32 rc = __kmp_enter_single(gtid, &loc, FALSE);
194
195 #if OMPT_SUPPORT && OMPT_OPTIONAL
196 kmp_info_t *this_thr = __kmp_threads[gtid];
197 kmp_team_t *team = this_thr->th.th_team;
198 int tid = __kmp_tid_from_gtid(gtid);
199
200 if (ompt_enabled.enabled) {
201 if (rc) {
202 if (ompt_enabled.ompt_callback_work) {
203 ompt_callbacks.ompt_callback(ompt_callback_work)(
204 ompt_work_single_executor, ompt_scope_begin,
205 &(team->t.ompt_team_info.parallel_data),
206 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
207 1, OMPT_GET_RETURN_ADDRESS(0));
208 }
209 } else {
210 if (ompt_enabled.ompt_callback_work) {
211 ompt_callbacks.ompt_callback(ompt_callback_work)(
212 ompt_work_single_other, ompt_scope_begin,
213 &(team->t.ompt_team_info.parallel_data),
214 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
215 1, OMPT_GET_RETURN_ADDRESS(0));
216 ompt_callbacks.ompt_callback(ompt_callback_work)(
217 ompt_work_single_other, ompt_scope_end,
218 &(team->t.ompt_team_info.parallel_data),
219 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
220 1, OMPT_GET_RETURN_ADDRESS(0));
221 }
222 }
223 }
224 #endif
225
226 return rc;
227 }
228
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_START)229 void *KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_START)(void) {
230 void *retval;
231 int gtid = __kmp_entry_gtid();
232 MKLOC(loc, "GOMP_single_copy_start");
233 KA_TRACE(20, ("GOMP_single_copy_start: T#%d\n", gtid));
234
235 if (!TCR_4(__kmp_init_parallel))
236 __kmp_parallel_initialize();
237 __kmp_resume_if_soft_paused();
238
239 // If this is the first thread to enter, return NULL. The generated code will
240 // then call GOMP_single_copy_end() for this thread only, with the
241 // copyprivate data pointer as an argument.
242 if (__kmp_enter_single(gtid, &loc, FALSE))
243 return NULL;
244
245 // Wait for the first thread to set the copyprivate data pointer,
246 // and for all other threads to reach this point.
247
248 #if OMPT_SUPPORT && OMPT_OPTIONAL
249 ompt_frame_t *ompt_frame;
250 if (ompt_enabled.enabled) {
251 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
252 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
253 }
254 OMPT_STORE_RETURN_ADDRESS(gtid);
255 #endif
256 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
257
258 // Retrieve the value of the copyprivate data point, and wait for all
259 // threads to do likewise, then return.
260 retval = __kmp_team_from_gtid(gtid)->t.t_copypriv_data;
261 {
262 #if OMPT_SUPPORT && OMPT_OPTIONAL
263 OMPT_STORE_RETURN_ADDRESS(gtid);
264 #endif
265 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
266 }
267 #if OMPT_SUPPORT && OMPT_OPTIONAL
268 if (ompt_enabled.enabled) {
269 ompt_frame->enter_frame = ompt_data_none;
270 }
271 #endif
272 return retval;
273 }
274
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_END)275 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_END)(void *data) {
276 int gtid = __kmp_get_gtid();
277 KA_TRACE(20, ("GOMP_single_copy_end: T#%d\n", gtid));
278
279 // Set the copyprivate data pointer fo the team, then hit the barrier so that
280 // the other threads will continue on and read it. Hit another barrier before
281 // continuing, so that the know that the copyprivate data pointer has been
282 // propagated to all threads before trying to reuse the t_copypriv_data field.
283 __kmp_team_from_gtid(gtid)->t.t_copypriv_data = data;
284 #if OMPT_SUPPORT && OMPT_OPTIONAL
285 ompt_frame_t *ompt_frame;
286 if (ompt_enabled.enabled) {
287 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
288 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
289 }
290 OMPT_STORE_RETURN_ADDRESS(gtid);
291 #endif
292 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
293 {
294 #if OMPT_SUPPORT && OMPT_OPTIONAL
295 OMPT_STORE_RETURN_ADDRESS(gtid);
296 #endif
297 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
298 }
299 #if OMPT_SUPPORT && OMPT_OPTIONAL
300 if (ompt_enabled.enabled) {
301 ompt_frame->enter_frame = ompt_data_none;
302 }
303 #endif
304 }
305
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_START)306 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_START)(void) {
307 int gtid = __kmp_entry_gtid();
308 MKLOC(loc, "GOMP_ordered_start");
309 KA_TRACE(20, ("GOMP_ordered_start: T#%d\n", gtid));
310 #if OMPT_SUPPORT && OMPT_OPTIONAL
311 OMPT_STORE_RETURN_ADDRESS(gtid);
312 #endif
313 __kmpc_ordered(&loc, gtid);
314 }
315
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_END)316 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_END)(void) {
317 int gtid = __kmp_get_gtid();
318 MKLOC(loc, "GOMP_ordered_end");
319 KA_TRACE(20, ("GOMP_ordered_start: T#%d\n", gtid));
320 #if OMPT_SUPPORT && OMPT_OPTIONAL
321 OMPT_STORE_RETURN_ADDRESS(gtid);
322 #endif
323 __kmpc_end_ordered(&loc, gtid);
324 }
325
326 // Dispatch macro defs
327 //
328 // They come in two flavors: 64-bit unsigned, and either 32-bit signed
329 // (IA-32 architecture) or 64-bit signed (Intel(R) 64).
330
331 #if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_MIPS
332 #define KMP_DISPATCH_INIT __kmp_aux_dispatch_init_4
333 #define KMP_DISPATCH_FINI_CHUNK __kmp_aux_dispatch_fini_chunk_4
334 #define KMP_DISPATCH_NEXT __kmpc_dispatch_next_4
335 #else
336 #define KMP_DISPATCH_INIT __kmp_aux_dispatch_init_8
337 #define KMP_DISPATCH_FINI_CHUNK __kmp_aux_dispatch_fini_chunk_8
338 #define KMP_DISPATCH_NEXT __kmpc_dispatch_next_8
339 #endif /* KMP_ARCH_X86 */
340
341 #define KMP_DISPATCH_INIT_ULL __kmp_aux_dispatch_init_8u
342 #define KMP_DISPATCH_FINI_CHUNK_ULL __kmp_aux_dispatch_fini_chunk_8u
343 #define KMP_DISPATCH_NEXT_ULL __kmpc_dispatch_next_8u
344
345 // The parallel construct
346
347 #ifndef KMP_DEBUG
348 static
349 #endif /* KMP_DEBUG */
350 void
__kmp_GOMP_microtask_wrapper(int * gtid,int * npr,void (* task)(void *),void * data)351 __kmp_GOMP_microtask_wrapper(int *gtid, int *npr, void (*task)(void *),
352 void *data) {
353 #if OMPT_SUPPORT
354 kmp_info_t *thr;
355 ompt_frame_t *ompt_frame;
356 ompt_state_t enclosing_state;
357
358 if (ompt_enabled.enabled) {
359 // get pointer to thread data structure
360 thr = __kmp_threads[*gtid];
361
362 // save enclosing task state; set current state for task
363 enclosing_state = thr->th.ompt_thread_info.state;
364 thr->th.ompt_thread_info.state = ompt_state_work_parallel;
365
366 // set task frame
367 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
368 ompt_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
369 }
370 #endif
371
372 task(data);
373
374 #if OMPT_SUPPORT
375 if (ompt_enabled.enabled) {
376 // clear task frame
377 ompt_frame->exit_frame = ompt_data_none;
378
379 // restore enclosing state
380 thr->th.ompt_thread_info.state = enclosing_state;
381 }
382 #endif
383 }
384
385 #ifndef KMP_DEBUG
386 static
387 #endif /* KMP_DEBUG */
388 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)389 __kmp_GOMP_parallel_microtask_wrapper(int *gtid, int *npr,
390 void (*task)(void *), void *data,
391 unsigned num_threads, ident_t *loc,
392 enum sched_type schedule, long start,
393 long end, long incr,
394 long chunk_size) {
395 // Initialize the loop worksharing construct.
396
397 KMP_DISPATCH_INIT(loc, *gtid, schedule, start, end, incr, chunk_size,
398 schedule != kmp_sch_static);
399
400 #if OMPT_SUPPORT
401 kmp_info_t *thr;
402 ompt_frame_t *ompt_frame;
403 ompt_state_t enclosing_state;
404
405 if (ompt_enabled.enabled) {
406 thr = __kmp_threads[*gtid];
407 // save enclosing task state; set current state for task
408 enclosing_state = thr->th.ompt_thread_info.state;
409 thr->th.ompt_thread_info.state = ompt_state_work_parallel;
410
411 // set task frame
412 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
413 ompt_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
414 }
415 #endif
416
417 // Now invoke the microtask.
418 task(data);
419
420 #if OMPT_SUPPORT
421 if (ompt_enabled.enabled) {
422 // clear task frame
423 ompt_frame->exit_frame = ompt_data_none;
424
425 // reset enclosing state
426 thr->th.ompt_thread_info.state = enclosing_state;
427 }
428 #endif
429 }
430
__kmp_GOMP_fork_call(ident_t * loc,int gtid,unsigned num_threads,unsigned flags,void (* unwrapped_task)(void *),microtask_t wrapper,int argc,...)431 static void __kmp_GOMP_fork_call(ident_t *loc, int gtid, unsigned num_threads,
432 unsigned flags, void (*unwrapped_task)(void *),
433 microtask_t wrapper, int argc, ...) {
434 int rc;
435 kmp_info_t *thr = __kmp_threads[gtid];
436 kmp_team_t *team = thr->th.th_team;
437 int tid = __kmp_tid_from_gtid(gtid);
438
439 va_list ap;
440 va_start(ap, argc);
441
442 if (num_threads != 0)
443 __kmp_push_num_threads(loc, gtid, num_threads);
444 if (flags != 0)
445 __kmp_push_proc_bind(loc, gtid, (kmp_proc_bind_t)flags);
446 rc = __kmp_fork_call(loc, gtid, fork_context_gnu, argc, wrapper,
447 __kmp_invoke_task_func, kmp_va_addr_of(ap));
448
449 va_end(ap);
450
451 if (rc) {
452 __kmp_run_before_invoked_task(gtid, tid, thr, team);
453 }
454
455 #if OMPT_SUPPORT
456 int ompt_team_size;
457 if (ompt_enabled.enabled) {
458 ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
459 ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
460
461 // implicit task callback
462 if (ompt_enabled.ompt_callback_implicit_task) {
463 ompt_team_size = __kmp_team_from_gtid(gtid)->t.t_nproc;
464 ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
465 ompt_scope_begin, &(team_info->parallel_data),
466 &(task_info->task_data), ompt_team_size, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
467 task_info->thread_num = __kmp_tid_from_gtid(gtid);
468 }
469 thr->th.ompt_thread_info.state = ompt_state_work_parallel;
470 }
471 #endif
472 }
473
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_START)474 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_START)(void (*task)(void *),
475 void *data,
476 unsigned num_threads) {
477 int gtid = __kmp_entry_gtid();
478
479 #if OMPT_SUPPORT
480 ompt_frame_t *parent_frame, *frame;
481
482 if (ompt_enabled.enabled) {
483 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL);
484 parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
485 }
486 OMPT_STORE_RETURN_ADDRESS(gtid);
487 #endif
488
489 MKLOC(loc, "GOMP_parallel_start");
490 KA_TRACE(20, ("GOMP_parallel_start: T#%d\n", gtid));
491 __kmp_GOMP_fork_call(&loc, gtid, num_threads, 0u, task,
492 (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task,
493 data);
494 #if OMPT_SUPPORT
495 if (ompt_enabled.enabled) {
496 __ompt_get_task_info_internal(0, NULL, NULL, &frame, NULL, NULL);
497 frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
498 }
499 #endif
500 }
501
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)502 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(void) {
503 int gtid = __kmp_get_gtid();
504 kmp_info_t *thr;
505
506 thr = __kmp_threads[gtid];
507
508 MKLOC(loc, "GOMP_parallel_end");
509 KA_TRACE(20, ("GOMP_parallel_end: T#%d\n", gtid));
510
511 if (!thr->th.th_team->t.t_serialized) {
512 __kmp_run_after_invoked_task(gtid, __kmp_tid_from_gtid(gtid), thr,
513 thr->th.th_team);
514 }
515 #if OMPT_SUPPORT
516 if (ompt_enabled.enabled) {
517 // Implicit task is finished here, in the barrier we might schedule
518 // deferred tasks,
519 // these don't see the implicit task on the stack
520 OMPT_CUR_TASK_INFO(thr)->frame.exit_frame = ompt_data_none;
521 }
522 #endif
523
524 __kmp_join_call(&loc, gtid
525 #if OMPT_SUPPORT
526 ,
527 fork_context_gnu
528 #endif
529 );
530 }
531
532 // Loop worksharing constructs
533
534 // The Gnu codegen passes in an exclusive upper bound for the overall range,
535 // but the libguide dispatch code expects an inclusive upper bound, hence the
536 // "end - incr" 5th argument to KMP_DISPATCH_INIT (and the " ub - str" 11th
537 // argument to __kmp_GOMP_fork_call).
538 //
539 // Conversely, KMP_DISPATCH_NEXT returns and inclusive upper bound in *p_ub,
540 // but the Gnu codegen expects an exclusive upper bound, so the adjustment
541 // "*p_ub += stride" compensates for the discrepancy.
542 //
543 // Correction: the gnu codegen always adjusts the upper bound by +-1, not the
544 // stride value. We adjust the dispatch parameters accordingly (by +-1), but
545 // we still adjust p_ub by the actual stride value.
546 //
547 // The "runtime" versions do not take a chunk_sz parameter.
548 //
549 // The profile lib cannot support construct checking of unordered loops that
550 // are predetermined by the compiler to be statically scheduled, as the gcc
551 // codegen will not always emit calls to GOMP_loop_static_next() to get the
552 // next iteration. Instead, it emits inline code to call omp_get_thread_num()
553 // num and calculate the iteration space using the result. It doesn't do this
554 // with ordered static loop, so they can be checked.
555
556 #if OMPT_SUPPORT
557 #define IF_OMPT_SUPPORT(code) code
558 #else
559 #define IF_OMPT_SUPPORT(code)
560 #endif
561
562 #define LOOP_START(func, schedule) \
563 int func(long lb, long ub, long str, long chunk_sz, long *p_lb, \
564 long *p_ub) { \
565 int status; \
566 long stride; \
567 int gtid = __kmp_entry_gtid(); \
568 MKLOC(loc, KMP_STR(func)); \
569 KA_TRACE( \
570 20, \
571 (KMP_STR( \
572 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \
573 gtid, lb, ub, str, chunk_sz)); \
574 \
575 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
576 { \
577 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
578 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
579 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
580 (schedule) != kmp_sch_static); \
581 } \
582 { \
583 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
584 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
585 (kmp_int *)p_ub, (kmp_int *)&stride); \
586 } \
587 if (status) { \
588 KMP_DEBUG_ASSERT(stride == str); \
589 *p_ub += (str > 0) ? 1 : -1; \
590 } \
591 } else { \
592 status = 0; \
593 } \
594 \
595 KA_TRACE( \
596 20, \
597 (KMP_STR( \
598 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \
599 gtid, *p_lb, *p_ub, status)); \
600 return status; \
601 }
602
603 #define LOOP_RUNTIME_START(func, schedule) \
604 int func(long lb, long ub, long str, long *p_lb, long *p_ub) { \
605 int status; \
606 long stride; \
607 long chunk_sz = 0; \
608 int gtid = __kmp_entry_gtid(); \
609 MKLOC(loc, KMP_STR(func)); \
610 KA_TRACE( \
611 20, \
612 (KMP_STR(func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \
613 gtid, lb, ub, str, chunk_sz)); \
614 \
615 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
616 { \
617 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
618 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
619 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
620 TRUE); \
621 } \
622 { \
623 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
624 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
625 (kmp_int *)p_ub, (kmp_int *)&stride); \
626 } \
627 if (status) { \
628 KMP_DEBUG_ASSERT(stride == str); \
629 *p_ub += (str > 0) ? 1 : -1; \
630 } \
631 } else { \
632 status = 0; \
633 } \
634 \
635 KA_TRACE( \
636 20, \
637 (KMP_STR( \
638 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \
639 gtid, *p_lb, *p_ub, status)); \
640 return status; \
641 }
642
643 #define KMP_DOACROSS_FINI(status, gtid) \
644 if (!status && __kmp_threads[gtid]->th.th_dispatch->th_doacross_flags) { \
645 __kmpc_doacross_fini(NULL, gtid); \
646 }
647
648 #define LOOP_NEXT(func, fini_code) \
649 int func(long *p_lb, long *p_ub) { \
650 int status; \
651 long stride; \
652 int gtid = __kmp_get_gtid(); \
653 MKLOC(loc, KMP_STR(func)); \
654 KA_TRACE(20, (KMP_STR(func) ": T#%d\n", gtid)); \
655 \
656 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
657 fini_code status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
658 (kmp_int *)p_ub, (kmp_int *)&stride); \
659 if (status) { \
660 *p_ub += (stride > 0) ? 1 : -1; \
661 } \
662 KMP_DOACROSS_FINI(status, gtid) \
663 \
664 KA_TRACE( \
665 20, \
666 (KMP_STR(func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, stride 0x%lx, " \
667 "returning %d\n", \
668 gtid, *p_lb, *p_ub, stride, status)); \
669 return status; \
670 }
671
LOOP_START(KMP_EXPAND_NAME (KMP_API_NAME_GOMP_LOOP_STATIC_START),kmp_sch_static)672 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_START), kmp_sch_static)
673 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT), {})
674 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START),
675 kmp_sch_dynamic_chunked)
676 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_START),
677 kmp_sch_dynamic_chunked)
678 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT), {})
679 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_NEXT), {})
680 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_START),
681 kmp_sch_guided_chunked)
682 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_START),
683 kmp_sch_guided_chunked)
684 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT), {})
685 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_NEXT), {})
686 LOOP_RUNTIME_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_START),
687 kmp_sch_runtime)
688 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT), {})
689 LOOP_RUNTIME_START(
690 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_START),
691 kmp_sch_runtime)
692 LOOP_RUNTIME_START(
693 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_START),
694 kmp_sch_runtime)
695 LOOP_NEXT(
696 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_NEXT), {})
697 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_NEXT), {})
698
699 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START),
700 kmp_ord_static)
701 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT),
702 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); })
703 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START),
704 kmp_ord_dynamic_chunked)
705 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT),
706 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); })
707 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START),
708 kmp_ord_guided_chunked)
709 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT),
710 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); })
711 LOOP_RUNTIME_START(
712 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START),
713 kmp_ord_runtime)
714 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT),
715 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); })
716
717 #define LOOP_DOACROSS_START(func, schedule) \
718 bool func(unsigned ncounts, long *counts, long chunk_sz, long *p_lb, \
719 long *p_ub) { \
720 int status; \
721 long stride, lb, ub, str; \
722 int gtid = __kmp_entry_gtid(); \
723 struct kmp_dim *dims = \
724 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \
725 MKLOC(loc, KMP_STR(func)); \
726 for (unsigned i = 0; i < ncounts; ++i) { \
727 dims[i].lo = 0; \
728 dims[i].up = counts[i] - 1; \
729 dims[i].st = 1; \
730 } \
731 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \
732 lb = 0; \
733 ub = counts[0]; \
734 str = 1; \
735 KA_TRACE(20, (KMP_STR(func) ": T#%d, ncounts %u, lb 0x%lx, ub 0x%lx, str " \
736 "0x%lx, chunk_sz " \
737 "0x%lx\n", \
738 gtid, ncounts, lb, ub, str, chunk_sz)); \
739 \
740 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
741 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
742 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
743 (schedule) != kmp_sch_static); \
744 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
745 (kmp_int *)p_ub, (kmp_int *)&stride); \
746 if (status) { \
747 KMP_DEBUG_ASSERT(stride == str); \
748 *p_ub += (str > 0) ? 1 : -1; \
749 } \
750 } else { \
751 status = 0; \
752 } \
753 KMP_DOACROSS_FINI(status, gtid); \
754 \
755 KA_TRACE( \
756 20, \
757 (KMP_STR( \
758 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \
759 gtid, *p_lb, *p_ub, status)); \
760 __kmp_free(dims); \
761 return status; \
762 }
763
764 #define LOOP_DOACROSS_RUNTIME_START(func, schedule) \
765 int func(unsigned ncounts, long *counts, long *p_lb, long *p_ub) { \
766 int status; \
767 long stride, lb, ub, str; \
768 long chunk_sz = 0; \
769 int gtid = __kmp_entry_gtid(); \
770 struct kmp_dim *dims = \
771 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \
772 MKLOC(loc, KMP_STR(func)); \
773 for (unsigned i = 0; i < ncounts; ++i) { \
774 dims[i].lo = 0; \
775 dims[i].up = counts[i] - 1; \
776 dims[i].st = 1; \
777 } \
778 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \
779 lb = 0; \
780 ub = counts[0]; \
781 str = 1; \
782 KA_TRACE( \
783 20, \
784 (KMP_STR(func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \
785 gtid, lb, ub, str, chunk_sz)); \
786 \
787 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
788 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
789 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, TRUE); \
790 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
791 (kmp_int *)p_ub, (kmp_int *)&stride); \
792 if (status) { \
793 KMP_DEBUG_ASSERT(stride == str); \
794 *p_ub += (str > 0) ? 1 : -1; \
795 } \
796 } else { \
797 status = 0; \
798 } \
799 KMP_DOACROSS_FINI(status, gtid); \
800 \
801 KA_TRACE( \
802 20, \
803 (KMP_STR( \
804 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \
805 gtid, *p_lb, *p_ub, status)); \
806 __kmp_free(dims); \
807 return status; \
808 }
809
810 LOOP_DOACROSS_START(
811 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START),
812 kmp_sch_static)
813 LOOP_DOACROSS_START(
814 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START),
815 kmp_sch_dynamic_chunked)
816 LOOP_DOACROSS_START(
817 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START),
818 kmp_sch_guided_chunked)
819 LOOP_DOACROSS_RUNTIME_START(
820 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START),
821 kmp_sch_runtime)
822
823 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END)(void) {
824 int gtid = __kmp_get_gtid();
825 KA_TRACE(20, ("GOMP_loop_end: T#%d\n", gtid))
826
827 #if OMPT_SUPPORT && OMPT_OPTIONAL
828 ompt_frame_t *ompt_frame;
829 if (ompt_enabled.enabled) {
830 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
831 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
832 OMPT_STORE_RETURN_ADDRESS(gtid);
833 }
834 #endif
835 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
836 #if OMPT_SUPPORT && OMPT_OPTIONAL
837 if (ompt_enabled.enabled) {
838 ompt_frame->enter_frame = ompt_data_none;
839 }
840 #endif
841
842 KA_TRACE(20, ("GOMP_loop_end exit: T#%d\n", gtid))
843 }
844
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_NOWAIT)845 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_NOWAIT)(void) {
846 KA_TRACE(20, ("GOMP_loop_end_nowait: T#%d\n", __kmp_get_gtid()))
847 }
848
849 // Unsigned long long loop worksharing constructs
850 //
851 // These are new with gcc 4.4
852
853 #define LOOP_START_ULL(func, schedule) \
854 int func(int up, unsigned long long lb, unsigned long long ub, \
855 unsigned long long str, unsigned long long chunk_sz, \
856 unsigned long long *p_lb, unsigned long long *p_ub) { \
857 int status; \
858 long long str2 = up ? ((long long)str) : -((long long)str); \
859 long long stride; \
860 int gtid = __kmp_entry_gtid(); \
861 MKLOC(loc, KMP_STR(func)); \
862 \
863 KA_TRACE(20, (KMP_STR(func) ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str " \
864 "0x%llx, chunk_sz 0x%llx\n", \
865 gtid, up, lb, ub, str, chunk_sz)); \
866 \
867 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
868 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \
869 (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \
870 (schedule) != kmp_sch_static); \
871 status = \
872 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
873 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
874 if (status) { \
875 KMP_DEBUG_ASSERT(stride == str2); \
876 *p_ub += (str > 0) ? 1 : -1; \
877 } \
878 } else { \
879 status = 0; \
880 } \
881 \
882 KA_TRACE( \
883 20, \
884 (KMP_STR( \
885 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \
886 gtid, *p_lb, *p_ub, status)); \
887 return status; \
888 }
889
890 #define LOOP_RUNTIME_START_ULL(func, schedule) \
891 int func(int up, unsigned long long lb, unsigned long long ub, \
892 unsigned long long str, unsigned long long *p_lb, \
893 unsigned long long *p_ub) { \
894 int status; \
895 long long str2 = up ? ((long long)str) : -((long long)str); \
896 unsigned long long stride; \
897 unsigned long long chunk_sz = 0; \
898 int gtid = __kmp_entry_gtid(); \
899 MKLOC(loc, KMP_STR(func)); \
900 \
901 KA_TRACE(20, (KMP_STR(func) ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str " \
902 "0x%llx, chunk_sz 0x%llx\n", \
903 gtid, up, lb, ub, str, chunk_sz)); \
904 \
905 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
906 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \
907 (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \
908 TRUE); \
909 status = \
910 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
911 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
912 if (status) { \
913 KMP_DEBUG_ASSERT((long long)stride == str2); \
914 *p_ub += (str > 0) ? 1 : -1; \
915 } \
916 } else { \
917 status = 0; \
918 } \
919 \
920 KA_TRACE( \
921 20, \
922 (KMP_STR( \
923 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \
924 gtid, *p_lb, *p_ub, status)); \
925 return status; \
926 }
927
928 #define LOOP_NEXT_ULL(func, fini_code) \
929 int func(unsigned long long *p_lb, unsigned long long *p_ub) { \
930 int status; \
931 long long stride; \
932 int gtid = __kmp_get_gtid(); \
933 MKLOC(loc, KMP_STR(func)); \
934 KA_TRACE(20, (KMP_STR(func) ": T#%d\n", gtid)); \
935 \
936 fini_code status = \
937 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
938 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
939 if (status) { \
940 *p_ub += (stride > 0) ? 1 : -1; \
941 } \
942 \
943 KA_TRACE( \
944 20, \
945 (KMP_STR( \
946 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, stride 0x%llx, " \
947 "returning %d\n", \
948 gtid, *p_lb, *p_ub, stride, status)); \
949 return status; \
950 }
951
LOOP_START_ULL(KMP_EXPAND_NAME (KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START),kmp_sch_static)952 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START),
953 kmp_sch_static)
954 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT), {})
955 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START),
956 kmp_sch_dynamic_chunked)
957 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT), {})
958 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START),
959 kmp_sch_guided_chunked)
960 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT), {})
961 LOOP_START_ULL(
962 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_START),
963 kmp_sch_dynamic_chunked)
964 LOOP_NEXT_ULL(
965 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_NEXT), {})
966 LOOP_START_ULL(
967 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_START),
968 kmp_sch_guided_chunked)
969 LOOP_NEXT_ULL(
970 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_NEXT), {})
971 LOOP_RUNTIME_START_ULL(
972 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START), kmp_sch_runtime)
973 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT), {})
974 LOOP_RUNTIME_START_ULL(
975 KMP_EXPAND_NAME(
976 KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_START),
977 kmp_sch_runtime)
978 LOOP_RUNTIME_START_ULL(
979 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_START),
980 kmp_sch_runtime)
981 LOOP_NEXT_ULL(
982 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_NEXT),
983 {})
984 LOOP_NEXT_ULL(
985 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_NEXT), {})
986
987 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START),
988 kmp_ord_static)
989 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT),
990 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); })
991 LOOP_START_ULL(
992 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START),
993 kmp_ord_dynamic_chunked)
994 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT),
995 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); })
996 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START),
997 kmp_ord_guided_chunked)
998 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT),
999 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); })
1000 LOOP_RUNTIME_START_ULL(
1001 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START),
1002 kmp_ord_runtime)
1003 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT),
1004 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); })
1005
1006 #define LOOP_DOACROSS_START_ULL(func, schedule) \
1007 int func(unsigned ncounts, unsigned long long *counts, \
1008 unsigned long long chunk_sz, unsigned long long *p_lb, \
1009 unsigned long long *p_ub) { \
1010 int status; \
1011 long long stride, str, lb, ub; \
1012 int gtid = __kmp_entry_gtid(); \
1013 struct kmp_dim *dims = \
1014 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \
1015 MKLOC(loc, KMP_STR(func)); \
1016 for (unsigned i = 0; i < ncounts; ++i) { \
1017 dims[i].lo = 0; \
1018 dims[i].up = counts[i] - 1; \
1019 dims[i].st = 1; \
1020 } \
1021 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \
1022 lb = 0; \
1023 ub = counts[0]; \
1024 str = 1; \
1025 \
1026 KA_TRACE(20, (KMP_STR(func) ": T#%d, lb 0x%llx, ub 0x%llx, str " \
1027 "0x%llx, chunk_sz 0x%llx\n", \
1028 gtid, lb, ub, str, chunk_sz)); \
1029 \
1030 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
1031 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \
1032 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
1033 (schedule) != kmp_sch_static); \
1034 status = \
1035 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
1036 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
1037 if (status) { \
1038 KMP_DEBUG_ASSERT(stride == str); \
1039 *p_ub += (str > 0) ? 1 : -1; \
1040 } \
1041 } else { \
1042 status = 0; \
1043 } \
1044 KMP_DOACROSS_FINI(status, gtid); \
1045 \
1046 KA_TRACE( \
1047 20, \
1048 (KMP_STR( \
1049 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \
1050 gtid, *p_lb, *p_ub, status)); \
1051 __kmp_free(dims); \
1052 return status; \
1053 }
1054
1055 #define LOOP_DOACROSS_RUNTIME_START_ULL(func, schedule) \
1056 int func(unsigned ncounts, unsigned long long *counts, \
1057 unsigned long long *p_lb, unsigned long long *p_ub) { \
1058 int status; \
1059 unsigned long long stride, str, lb, ub; \
1060 unsigned long long chunk_sz = 0; \
1061 int gtid = __kmp_entry_gtid(); \
1062 struct kmp_dim *dims = \
1063 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \
1064 MKLOC(loc, KMP_STR(func)); \
1065 for (unsigned i = 0; i < ncounts; ++i) { \
1066 dims[i].lo = 0; \
1067 dims[i].up = counts[i] - 1; \
1068 dims[i].st = 1; \
1069 } \
1070 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \
1071 lb = 0; \
1072 ub = counts[0]; \
1073 str = 1; \
1074 KA_TRACE(20, (KMP_STR(func) ": T#%d, lb 0x%llx, ub 0x%llx, str " \
1075 "0x%llx, chunk_sz 0x%llx\n", \
1076 gtid, lb, ub, str, chunk_sz)); \
1077 \
1078 if ((str > 0) ? (lb < ub) : (lb > ub)) { \
1079 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \
1080 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
1081 TRUE); \
1082 status = \
1083 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
1084 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
1085 if (status) { \
1086 KMP_DEBUG_ASSERT(stride == str); \
1087 *p_ub += (str > 0) ? 1 : -1; \
1088 } \
1089 } else { \
1090 status = 0; \
1091 } \
1092 KMP_DOACROSS_FINI(status, gtid); \
1093 \
1094 KA_TRACE( \
1095 20, \
1096 (KMP_STR( \
1097 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \
1098 gtid, *p_lb, *p_ub, status)); \
1099 __kmp_free(dims); \
1100 return status; \
1101 }
1102
1103 LOOP_DOACROSS_START_ULL(
1104 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START),
1105 kmp_sch_static)
1106 LOOP_DOACROSS_START_ULL(
1107 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START),
1108 kmp_sch_dynamic_chunked)
1109 LOOP_DOACROSS_START_ULL(
1110 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START),
1111 kmp_sch_guided_chunked)
1112 LOOP_DOACROSS_RUNTIME_START_ULL(
1113 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START),
1114 kmp_sch_runtime)
1115
1116 // Combined parallel / loop worksharing constructs
1117 //
1118 // There are no ull versions (yet).
1119
1120 #define PARALLEL_LOOP_START(func, schedule, ompt_pre, ompt_post) \
1121 void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \
1122 long ub, long str, long chunk_sz) { \
1123 int gtid = __kmp_entry_gtid(); \
1124 MKLOC(loc, KMP_STR(func)); \
1125 KA_TRACE( \
1126 20, \
1127 (KMP_STR( \
1128 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \
1129 gtid, lb, ub, str, chunk_sz)); \
1130 \
1131 ompt_pre(); \
1132 \
1133 __kmp_GOMP_fork_call(&loc, gtid, num_threads, 0u, task, \
1134 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \
1135 9, task, data, num_threads, &loc, (schedule), lb, \
1136 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \
1137 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid)); \
1138 \
1139 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
1140 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
1141 (schedule) != kmp_sch_static); \
1142 \
1143 ompt_post(); \
1144 \
1145 KA_TRACE(20, (KMP_STR(func) " exit: T#%d\n", gtid)); \
1146 }
1147
1148 #if OMPT_SUPPORT && OMPT_OPTIONAL
1149
1150 #define OMPT_LOOP_PRE() \
1151 ompt_frame_t *parent_frame; \
1152 if (ompt_enabled.enabled) { \
1153 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); \
1154 parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); \
1155 OMPT_STORE_RETURN_ADDRESS(gtid); \
1156 }
1157
1158 #define OMPT_LOOP_POST() \
1159 if (ompt_enabled.enabled) { \
1160 parent_frame->enter_frame = ompt_data_none; \
1161 }
1162
1163 #else
1164
1165 #define OMPT_LOOP_PRE()
1166
1167 #define OMPT_LOOP_POST()
1168
1169 #endif
1170
1171 PARALLEL_LOOP_START(
1172 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START),
1173 kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1174 PARALLEL_LOOP_START(
1175 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START),
1176 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1177 PARALLEL_LOOP_START(
1178 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START),
1179 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1180 PARALLEL_LOOP_START(
1181 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START),
1182 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1183
1184 // Tasking constructs
1185
1186 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASK)(void (*func)(void *), void *data,
1187 void (*copy_func)(void *, void *),
1188 long arg_size, long arg_align,
1189 bool if_cond, unsigned gomp_flags,
1190 void **depend) {
1191 MKLOC(loc, "GOMP_task");
1192 int gtid = __kmp_entry_gtid();
1193 kmp_int32 flags = 0;
1194 kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags;
1195
1196 KA_TRACE(20, ("GOMP_task: T#%d\n", gtid));
1197
1198 // The low-order bit is the "untied" flag
1199 if (!(gomp_flags & KMP_GOMP_TASK_UNTIED_FLAG)) {
1200 input_flags->tiedness = 1;
1201 }
1202 // The second low-order bit is the "final" flag
1203 if (gomp_flags & KMP_GOMP_TASK_FINAL_FLAG) {
1204 input_flags->final = 1;
1205 }
1206 input_flags->native = 1;
1207 // __kmp_task_alloc() sets up all other flags
1208
1209 if (!if_cond) {
1210 arg_size = 0;
1211 }
1212
1213 kmp_task_t *task = __kmp_task_alloc(
1214 &loc, gtid, input_flags, sizeof(kmp_task_t),
1215 arg_size ? arg_size + arg_align - 1 : 0, (kmp_routine_entry_t)func);
1216
1217 if (arg_size > 0) {
1218 if (arg_align > 0) {
1219 task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) /
1220 arg_align * arg_align);
1221 }
1222 // else error??
1223
1224 if (copy_func) {
1225 (*copy_func)(task->shareds, data);
1226 } else {
1227 KMP_MEMCPY(task->shareds, data, arg_size);
1228 }
1229 }
1230
1231 #if OMPT_SUPPORT
1232 kmp_taskdata_t *current_task;
1233 if (ompt_enabled.enabled) {
1234 current_task = __kmp_threads[gtid]->th.th_current_task;
1235 current_task->ompt_task_info.frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1236 }
1237 OMPT_STORE_RETURN_ADDRESS(gtid);
1238 #endif
1239
1240 if (if_cond) {
1241 if (gomp_flags & KMP_GOMP_TASK_DEPENDS_FLAG) {
1242 KMP_ASSERT(depend);
1243 kmp_gomp_depends_info_t gomp_depends(depend);
1244 kmp_int32 ndeps = gomp_depends.get_num_deps();
1245 kmp_depend_info_t dep_list[ndeps];
1246 for (kmp_int32 i = 0; i < ndeps; i++)
1247 dep_list[i] = gomp_depends.get_kmp_depend(i);
1248 __kmpc_omp_task_with_deps(&loc, gtid, task, ndeps, dep_list, 0, NULL);
1249 } else {
1250 __kmpc_omp_task(&loc, gtid, task);
1251 }
1252 } else {
1253 #if OMPT_SUPPORT
1254 ompt_thread_info_t oldInfo;
1255 kmp_info_t *thread;
1256 kmp_taskdata_t *taskdata;
1257 if (ompt_enabled.enabled) {
1258 // Store the threads states and restore them after the task
1259 thread = __kmp_threads[gtid];
1260 taskdata = KMP_TASK_TO_TASKDATA(task);
1261 oldInfo = thread->th.ompt_thread_info;
1262 thread->th.ompt_thread_info.wait_id = 0;
1263 thread->th.ompt_thread_info.state = ompt_state_work_parallel;
1264 taskdata->ompt_task_info.frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1265 }
1266 OMPT_STORE_RETURN_ADDRESS(gtid);
1267 #endif
1268 if (gomp_flags & KMP_GOMP_TASK_DEPENDS_FLAG) {
1269 KMP_ASSERT(depend);
1270 kmp_gomp_depends_info_t gomp_depends(depend);
1271 kmp_int32 ndeps = gomp_depends.get_num_deps();
1272 kmp_depend_info_t dep_list[ndeps];
1273 for (kmp_int32 i = 0; i < ndeps; i++)
1274 dep_list[i] = gomp_depends.get_kmp_depend(i);
1275 __kmpc_omp_wait_deps(&loc, gtid, ndeps, dep_list, 0, NULL);
1276 }
1277
1278 __kmpc_omp_task_begin_if0(&loc, gtid, task);
1279 func(data);
1280 __kmpc_omp_task_complete_if0(&loc, gtid, task);
1281
1282 #if OMPT_SUPPORT
1283 if (ompt_enabled.enabled) {
1284 thread->th.ompt_thread_info = oldInfo;
1285 taskdata->ompt_task_info.frame.exit_frame = ompt_data_none;
1286 }
1287 #endif
1288 }
1289 #if OMPT_SUPPORT
1290 if (ompt_enabled.enabled) {
1291 current_task->ompt_task_info.frame.enter_frame = ompt_data_none;
1292 }
1293 #endif
1294
1295 KA_TRACE(20, ("GOMP_task exit: T#%d\n", gtid));
1296 }
1297
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT)1298 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT)(void) {
1299 MKLOC(loc, "GOMP_taskwait");
1300 int gtid = __kmp_entry_gtid();
1301
1302 #if OMPT_SUPPORT
1303 OMPT_STORE_RETURN_ADDRESS(gtid);
1304 #endif
1305
1306 KA_TRACE(20, ("GOMP_taskwait: T#%d\n", gtid));
1307
1308 __kmpc_omp_taskwait(&loc, gtid);
1309
1310 KA_TRACE(20, ("GOMP_taskwait exit: T#%d\n", gtid));
1311 }
1312
1313 // Sections worksharing constructs
1314 //
1315 // For the sections construct, we initialize a dynamically scheduled loop
1316 // worksharing construct with lb 1 and stride 1, and use the iteration #'s
1317 // that its returns as sections ids.
1318 //
1319 // There are no special entry points for ordered sections, so we always use
1320 // the dynamically scheduled workshare, even if the sections aren't ordered.
1321
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_START)1322 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_START)(unsigned count) {
1323 int status;
1324 kmp_int lb, ub, stride;
1325 int gtid = __kmp_entry_gtid();
1326 MKLOC(loc, "GOMP_sections_start");
1327 KA_TRACE(20, ("GOMP_sections_start: T#%d\n", gtid));
1328
1329 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE);
1330
1331 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride);
1332 if (status) {
1333 KMP_DEBUG_ASSERT(stride == 1);
1334 KMP_DEBUG_ASSERT(lb > 0);
1335 KMP_ASSERT(lb == ub);
1336 } else {
1337 lb = 0;
1338 }
1339
1340 KA_TRACE(20, ("GOMP_sections_start exit: T#%d returning %u\n", gtid,
1341 (unsigned)lb));
1342 return (unsigned)lb;
1343 }
1344
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_NEXT)1345 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_NEXT)(void) {
1346 int status;
1347 kmp_int lb, ub, stride;
1348 int gtid = __kmp_get_gtid();
1349 MKLOC(loc, "GOMP_sections_next");
1350 KA_TRACE(20, ("GOMP_sections_next: T#%d\n", gtid));
1351
1352 #if OMPT_SUPPORT
1353 OMPT_STORE_RETURN_ADDRESS(gtid);
1354 #endif
1355
1356 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride);
1357 if (status) {
1358 KMP_DEBUG_ASSERT(stride == 1);
1359 KMP_DEBUG_ASSERT(lb > 0);
1360 KMP_ASSERT(lb == ub);
1361 } else {
1362 lb = 0;
1363 }
1364
1365 KA_TRACE(
1366 20, ("GOMP_sections_next exit: T#%d returning %u\n", gtid, (unsigned)lb));
1367 return (unsigned)lb;
1368 }
1369
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START)1370 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START)(
1371 void (*task)(void *), void *data, unsigned num_threads, unsigned count) {
1372 int gtid = __kmp_entry_gtid();
1373
1374 #if OMPT_SUPPORT
1375 ompt_frame_t *parent_frame;
1376
1377 if (ompt_enabled.enabled) {
1378 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL);
1379 parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1380 }
1381 OMPT_STORE_RETURN_ADDRESS(gtid);
1382 #endif
1383
1384 MKLOC(loc, "GOMP_parallel_sections_start");
1385 KA_TRACE(20, ("GOMP_parallel_sections_start: T#%d\n", gtid));
1386
1387 __kmp_GOMP_fork_call(&loc, gtid, num_threads, 0u, task,
1388 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9,
1389 task, data, num_threads, &loc, kmp_nm_dynamic_chunked,
1390 (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1);
1391
1392 #if OMPT_SUPPORT
1393 if (ompt_enabled.enabled) {
1394 parent_frame->enter_frame = ompt_data_none;
1395 }
1396 #endif
1397
1398 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE);
1399
1400 KA_TRACE(20, ("GOMP_parallel_sections_start exit: T#%d\n", gtid));
1401 }
1402
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END)1403 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END)(void) {
1404 int gtid = __kmp_get_gtid();
1405 KA_TRACE(20, ("GOMP_sections_end: T#%d\n", gtid))
1406
1407 #if OMPT_SUPPORT
1408 ompt_frame_t *ompt_frame;
1409 if (ompt_enabled.enabled) {
1410 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
1411 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1412 }
1413 OMPT_STORE_RETURN_ADDRESS(gtid);
1414 #endif
1415 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
1416 #if OMPT_SUPPORT
1417 if (ompt_enabled.enabled) {
1418 ompt_frame->enter_frame = ompt_data_none;
1419 }
1420 #endif
1421
1422 KA_TRACE(20, ("GOMP_sections_end exit: T#%d\n", gtid))
1423 }
1424
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT)1425 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT)(void) {
1426 KA_TRACE(20, ("GOMP_sections_end_nowait: T#%d\n", __kmp_get_gtid()))
1427 }
1428
1429 // libgomp has an empty function for GOMP_taskyield as of 2013-10-10
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKYIELD)1430 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKYIELD)(void) {
1431 KA_TRACE(20, ("GOMP_taskyield: T#%d\n", __kmp_get_gtid()))
1432 return;
1433 }
1434
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL)1435 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL)(void (*task)(void *),
1436 void *data,
1437 unsigned num_threads,
1438 unsigned int flags) {
1439 int gtid = __kmp_entry_gtid();
1440 MKLOC(loc, "GOMP_parallel");
1441 KA_TRACE(20, ("GOMP_parallel: T#%d\n", gtid));
1442
1443 #if OMPT_SUPPORT
1444 ompt_task_info_t *parent_task_info, *task_info;
1445 if (ompt_enabled.enabled) {
1446 parent_task_info = __ompt_get_task_info_object(0);
1447 parent_task_info->frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1448 }
1449 OMPT_STORE_RETURN_ADDRESS(gtid);
1450 #endif
1451 __kmp_GOMP_fork_call(&loc, gtid, num_threads, flags, task,
1452 (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task,
1453 data);
1454 #if OMPT_SUPPORT
1455 if (ompt_enabled.enabled) {
1456 task_info = __ompt_get_task_info_object(0);
1457 task_info->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1458 }
1459 #endif
1460 task(data);
1461 {
1462 #if OMPT_SUPPORT
1463 OMPT_STORE_RETURN_ADDRESS(gtid);
1464 #endif
1465 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)();
1466 }
1467 #if OMPT_SUPPORT
1468 if (ompt_enabled.enabled) {
1469 task_info->frame.exit_frame = ompt_data_none;
1470 parent_task_info->frame.enter_frame = ompt_data_none;
1471 }
1472 #endif
1473 }
1474
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS)1475 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS)(void (*task)(void *),
1476 void *data,
1477 unsigned num_threads,
1478 unsigned count,
1479 unsigned flags) {
1480 int gtid = __kmp_entry_gtid();
1481 MKLOC(loc, "GOMP_parallel_sections");
1482 KA_TRACE(20, ("GOMP_parallel_sections: T#%d\n", gtid));
1483
1484 #if OMPT_SUPPORT
1485 OMPT_STORE_RETURN_ADDRESS(gtid);
1486 #endif
1487
1488 __kmp_GOMP_fork_call(&loc, gtid, num_threads, flags, task,
1489 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9,
1490 task, data, num_threads, &loc, kmp_nm_dynamic_chunked,
1491 (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1);
1492
1493 {
1494 #if OMPT_SUPPORT
1495 OMPT_STORE_RETURN_ADDRESS(gtid);
1496 #endif
1497
1498 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE);
1499 }
1500 task(data);
1501 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)();
1502 KA_TRACE(20, ("GOMP_parallel_sections exit: T#%d\n", gtid));
1503 }
1504
1505 #define PARALLEL_LOOP(func, schedule, ompt_pre, ompt_post) \
1506 void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \
1507 long ub, long str, long chunk_sz, unsigned flags) { \
1508 int gtid = __kmp_entry_gtid(); \
1509 MKLOC(loc, KMP_STR(func)); \
1510 KA_TRACE( \
1511 20, \
1512 (KMP_STR( \
1513 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \
1514 gtid, lb, ub, str, chunk_sz)); \
1515 \
1516 ompt_pre(); \
1517 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
1518 __kmp_GOMP_fork_call(&loc, gtid, num_threads, flags, task, \
1519 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \
1520 9, task, data, num_threads, &loc, (schedule), lb, \
1521 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \
1522 \
1523 { \
1524 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
1525 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
1526 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
1527 (schedule) != kmp_sch_static); \
1528 } \
1529 task(data); \
1530 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); \
1531 ompt_post(); \
1532 \
1533 KA_TRACE(20, (KMP_STR(func) " exit: T#%d\n", gtid)); \
1534 }
1535
PARALLEL_LOOP(KMP_EXPAND_NAME (KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC),kmp_sch_static,OMPT_LOOP_PRE,OMPT_LOOP_POST)1536 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC),
1537 kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1538 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC),
1539 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1540 PARALLEL_LOOP(
1541 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_GUIDED),
1542 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1543 PARALLEL_LOOP(
1544 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_DYNAMIC),
1545 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1546 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED),
1547 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1548 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME),
1549 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1550 PARALLEL_LOOP(
1551 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_MAYBE_NONMONOTONIC_RUNTIME),
1552 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1553 PARALLEL_LOOP(
1554 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_RUNTIME),
1555 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1556
1557 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_START)(void) {
1558 int gtid = __kmp_entry_gtid();
1559 MKLOC(loc, "GOMP_taskgroup_start");
1560 KA_TRACE(20, ("GOMP_taskgroup_start: T#%d\n", gtid));
1561
1562 #if OMPT_SUPPORT
1563 OMPT_STORE_RETURN_ADDRESS(gtid);
1564 #endif
1565
1566 __kmpc_taskgroup(&loc, gtid);
1567
1568 return;
1569 }
1570
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_END)1571 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_END)(void) {
1572 int gtid = __kmp_get_gtid();
1573 MKLOC(loc, "GOMP_taskgroup_end");
1574 KA_TRACE(20, ("GOMP_taskgroup_end: T#%d\n", gtid));
1575
1576 #if OMPT_SUPPORT
1577 OMPT_STORE_RETURN_ADDRESS(gtid);
1578 #endif
1579
1580 __kmpc_end_taskgroup(&loc, gtid);
1581
1582 return;
1583 }
1584
__kmp_gomp_to_omp_cancellation_kind(int gomp_kind)1585 static kmp_int32 __kmp_gomp_to_omp_cancellation_kind(int gomp_kind) {
1586 kmp_int32 cncl_kind = 0;
1587 switch (gomp_kind) {
1588 case 1:
1589 cncl_kind = cancel_parallel;
1590 break;
1591 case 2:
1592 cncl_kind = cancel_loop;
1593 break;
1594 case 4:
1595 cncl_kind = cancel_sections;
1596 break;
1597 case 8:
1598 cncl_kind = cancel_taskgroup;
1599 break;
1600 }
1601 return cncl_kind;
1602 }
1603
1604 // Return true if cancellation should take place, false otherwise
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCELLATION_POINT)1605 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCELLATION_POINT)(int which) {
1606 int gtid = __kmp_get_gtid();
1607 MKLOC(loc, "GOMP_cancellation_point");
1608 KA_TRACE(20, ("GOMP_cancellation_point: T#%d which:%d\n", gtid, which));
1609 kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which);
1610 return __kmpc_cancellationpoint(&loc, gtid, cncl_kind);
1611 }
1612
1613 // Return true if cancellation should take place, false otherwise
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCEL)1614 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCEL)(int which, bool do_cancel) {
1615 int gtid = __kmp_get_gtid();
1616 MKLOC(loc, "GOMP_cancel");
1617 KA_TRACE(20, ("GOMP_cancel: T#%d which:%d do_cancel:%d\n", gtid, which,
1618 (int)do_cancel));
1619 kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which);
1620
1621 if (do_cancel == FALSE) {
1622 return __kmpc_cancellationpoint(&loc, gtid, cncl_kind);
1623 } else {
1624 return __kmpc_cancel(&loc, gtid, cncl_kind);
1625 }
1626 }
1627
1628 // Return true if cancellation should take place, false otherwise
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER_CANCEL)1629 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER_CANCEL)(void) {
1630 int gtid = __kmp_get_gtid();
1631 KA_TRACE(20, ("GOMP_barrier_cancel: T#%d\n", gtid));
1632 return __kmp_barrier_gomp_cancel(gtid);
1633 }
1634
1635 // Return true if cancellation should take place, false otherwise
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL)1636 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL)(void) {
1637 int gtid = __kmp_get_gtid();
1638 KA_TRACE(20, ("GOMP_sections_end_cancel: T#%d\n", gtid));
1639 return __kmp_barrier_gomp_cancel(gtid);
1640 }
1641
1642 // Return true if cancellation should take place, false otherwise
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_CANCEL)1643 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_CANCEL)(void) {
1644 int gtid = __kmp_get_gtid();
1645 KA_TRACE(20, ("GOMP_loop_end_cancel: T#%d\n", gtid));
1646 return __kmp_barrier_gomp_cancel(gtid);
1647 }
1648
1649 // All target functions are empty as of 2014-05-29
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET)1650 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET)(int device, void (*fn)(void *),
1651 const void *openmp_target,
1652 size_t mapnum, void **hostaddrs,
1653 size_t *sizes,
1654 unsigned char *kinds) {
1655 return;
1656 }
1657
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_DATA)1658 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_DATA)(
1659 int device, const void *openmp_target, size_t mapnum, void **hostaddrs,
1660 size_t *sizes, unsigned char *kinds) {
1661 return;
1662 }
1663
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_END_DATA)1664 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_END_DATA)(void) { return; }
1665
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_UPDATE)1666 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_UPDATE)(
1667 int device, const void *openmp_target, size_t mapnum, void **hostaddrs,
1668 size_t *sizes, unsigned char *kinds) {
1669 return;
1670 }
1671
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS)1672 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS)(unsigned int num_teams,
1673 unsigned int thread_limit) {
1674 return;
1675 }
1676
1677 // Task duplication function which copies src to dest (both are
1678 // preallocated task structures)
__kmp_gomp_task_dup(kmp_task_t * dest,kmp_task_t * src,kmp_int32 last_private)1679 static void __kmp_gomp_task_dup(kmp_task_t *dest, kmp_task_t *src,
1680 kmp_int32 last_private) {
1681 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(src);
1682 if (taskdata->td_copy_func) {
1683 (taskdata->td_copy_func)(dest->shareds, src->shareds);
1684 }
1685 }
1686
1687 #ifdef __cplusplus
1688 } // extern "C"
1689 #endif
1690
1691 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)1692 void __GOMP_taskloop(void (*func)(void *), void *data,
1693 void (*copy_func)(void *, void *), long arg_size,
1694 long arg_align, unsigned gomp_flags,
1695 unsigned long num_tasks, int priority, T start, T end,
1696 T step) {
1697 typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32);
1698 MKLOC(loc, "GOMP_taskloop");
1699 int sched;
1700 T *loop_bounds;
1701 int gtid = __kmp_entry_gtid();
1702 kmp_int32 flags = 0;
1703 int if_val = gomp_flags & (1u << 10);
1704 int nogroup = gomp_flags & (1u << 11);
1705 int up = gomp_flags & (1u << 8);
1706 p_task_dup_t task_dup = NULL;
1707 kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags;
1708 #ifdef KMP_DEBUG
1709 {
1710 char *buff;
1711 buff = __kmp_str_format(
1712 "GOMP_taskloop: T#%%d: func:%%p data:%%p copy_func:%%p "
1713 "arg_size:%%ld arg_align:%%ld gomp_flags:0x%%x num_tasks:%%lu "
1714 "priority:%%d start:%%%s end:%%%s step:%%%s\n",
1715 traits_t<T>::spec, traits_t<T>::spec, traits_t<T>::spec);
1716 KA_TRACE(20, (buff, gtid, func, data, copy_func, arg_size, arg_align,
1717 gomp_flags, num_tasks, priority, start, end, step));
1718 __kmp_str_free(&buff);
1719 }
1720 #endif
1721 KMP_ASSERT((size_t)arg_size >= 2 * sizeof(T));
1722 KMP_ASSERT(arg_align > 0);
1723 // The low-order bit is the "untied" flag
1724 if (!(gomp_flags & 1)) {
1725 input_flags->tiedness = 1;
1726 }
1727 // The second low-order bit is the "final" flag
1728 if (gomp_flags & 2) {
1729 input_flags->final = 1;
1730 }
1731 // Negative step flag
1732 if (!up) {
1733 // If step is flagged as negative, but isn't properly sign extended
1734 // Then manually sign extend it. Could be a short, int, char embedded
1735 // in a long. So cannot assume any cast.
1736 if (step > 0) {
1737 for (int i = sizeof(T) * CHAR_BIT - 1; i >= 0L; --i) {
1738 // break at the first 1 bit
1739 if (step & ((T)1 << i))
1740 break;
1741 step |= ((T)1 << i);
1742 }
1743 }
1744 }
1745 input_flags->native = 1;
1746 // Figure out if none/grainsize/num_tasks clause specified
1747 if (num_tasks > 0) {
1748 if (gomp_flags & (1u << 9))
1749 sched = 1; // grainsize specified
1750 else
1751 sched = 2; // num_tasks specified
1752 // neither grainsize nor num_tasks specified
1753 } else {
1754 sched = 0;
1755 }
1756
1757 // __kmp_task_alloc() sets up all other flags
1758 kmp_task_t *task =
1759 __kmp_task_alloc(&loc, gtid, input_flags, sizeof(kmp_task_t),
1760 arg_size + arg_align - 1, (kmp_routine_entry_t)func);
1761 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
1762 taskdata->td_copy_func = copy_func;
1763 taskdata->td_size_loop_bounds = sizeof(T);
1764
1765 // re-align shareds if needed and setup firstprivate copy constructors
1766 // through the task_dup mechanism
1767 task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) /
1768 arg_align * arg_align);
1769 if (copy_func) {
1770 task_dup = __kmp_gomp_task_dup;
1771 }
1772 KMP_MEMCPY(task->shareds, data, arg_size);
1773
1774 loop_bounds = (T *)task->shareds;
1775 loop_bounds[0] = start;
1776 loop_bounds[1] = end + (up ? -1 : 1);
1777 __kmpc_taskloop(&loc, gtid, task, if_val, (kmp_uint64 *)&(loop_bounds[0]),
1778 (kmp_uint64 *)&(loop_bounds[1]), (kmp_int64)step, nogroup,
1779 sched, (kmp_uint64)num_tasks, (void *)task_dup);
1780 }
1781
1782 // 4 byte version of GOMP_doacross_post
1783 // This verison needs to create a temporary array which converts 4 byte
1784 // integers into 8 byte integers
1785 template <typename T, bool need_conversion = (sizeof(long) == 4)>
1786 void __kmp_GOMP_doacross_post(T *count);
1787
__kmp_GOMP_doacross_post(long * count)1788 template <> void __kmp_GOMP_doacross_post<long, true>(long *count) {
1789 int gtid = __kmp_entry_gtid();
1790 kmp_info_t *th = __kmp_threads[gtid];
1791 MKLOC(loc, "GOMP_doacross_post");
1792 kmp_int64 num_dims = th->th.th_dispatch->th_doacross_info[0];
1793 kmp_int64 *vec =
1794 (kmp_int64 *)__kmp_thread_malloc(th, sizeof(kmp_int64) * num_dims);
1795 for (kmp_int64 i = 0; i < num_dims; ++i) {
1796 vec[i] = (kmp_int64)count[i];
1797 }
1798 __kmpc_doacross_post(&loc, gtid, vec);
1799 __kmp_thread_free(th, vec);
1800 }
1801
1802 // 8 byte versions of GOMP_doacross_post
1803 // This version can just pass in the count array directly instead of creating
1804 // a temporary array
__kmp_GOMP_doacross_post(long * count)1805 template <> void __kmp_GOMP_doacross_post<long, false>(long *count) {
1806 int gtid = __kmp_entry_gtid();
1807 MKLOC(loc, "GOMP_doacross_post");
1808 __kmpc_doacross_post(&loc, gtid, RCAST(kmp_int64 *, count));
1809 }
1810
__kmp_GOMP_doacross_wait(T first,va_list args)1811 template <typename T> void __kmp_GOMP_doacross_wait(T first, va_list args) {
1812 int gtid = __kmp_entry_gtid();
1813 kmp_info_t *th = __kmp_threads[gtid];
1814 MKLOC(loc, "GOMP_doacross_wait");
1815 kmp_int64 num_dims = th->th.th_dispatch->th_doacross_info[0];
1816 kmp_int64 *vec =
1817 (kmp_int64 *)__kmp_thread_malloc(th, sizeof(kmp_int64) * num_dims);
1818 vec[0] = (kmp_int64)first;
1819 for (kmp_int64 i = 1; i < num_dims; ++i) {
1820 T item = va_arg(args, T);
1821 vec[i] = (kmp_int64)item;
1822 }
1823 __kmpc_doacross_wait(&loc, gtid, vec);
1824 __kmp_thread_free(th, vec);
1825 return;
1826 }
1827
1828 #ifdef __cplusplus
1829 extern "C" {
1830 #endif // __cplusplus
1831
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP)1832 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP)(
1833 void (*func)(void *), void *data, void (*copy_func)(void *, void *),
1834 long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks,
1835 int priority, long start, long end, long step) {
1836 __GOMP_taskloop<long>(func, data, copy_func, arg_size, arg_align, gomp_flags,
1837 num_tasks, priority, start, end, step);
1838 }
1839
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP_ULL)1840 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP_ULL)(
1841 void (*func)(void *), void *data, void (*copy_func)(void *, void *),
1842 long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks,
1843 int priority, unsigned long long start, unsigned long long end,
1844 unsigned long long step) {
1845 __GOMP_taskloop<unsigned long long>(func, data, copy_func, arg_size,
1846 arg_align, gomp_flags, num_tasks,
1847 priority, start, end, step);
1848 }
1849
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_POST)1850 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_POST)(long *count) {
1851 __kmp_GOMP_doacross_post(count);
1852 }
1853
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_WAIT)1854 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_WAIT)(long first, ...) {
1855 va_list args;
1856 va_start(args, first);
1857 __kmp_GOMP_doacross_wait<long>(first, args);
1858 va_end(args);
1859 }
1860
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_POST)1861 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_POST)(
1862 unsigned long long *count) {
1863 int gtid = __kmp_entry_gtid();
1864 MKLOC(loc, "GOMP_doacross_ull_post");
1865 __kmpc_doacross_post(&loc, gtid, RCAST(kmp_int64 *, count));
1866 }
1867
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT)1868 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT)(
1869 unsigned long long first, ...) {
1870 va_list args;
1871 va_start(args, first);
1872 __kmp_GOMP_doacross_wait<unsigned long long>(first, args);
1873 va_end(args);
1874 }
1875
1876 // fn: the function each master thread of new team will call
1877 // data: argument to fn
1878 // num_teams, thread_limit: max bounds on respective ICV
1879 // flags: unused
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS_REG)1880 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS_REG)(void (*fn)(void *),
1881 void *data,
1882 unsigned num_teams,
1883 unsigned thread_limit,
1884 unsigned flags) {
1885 MKLOC(loc, "GOMP_teams_reg");
1886 int gtid = __kmp_entry_gtid();
1887 KA_TRACE(20, ("GOMP_teams_reg: T#%d num_teams=%u thread_limit=%u flag=%u\n",
1888 gtid, num_teams, thread_limit, flags));
1889 __kmpc_push_num_teams(&loc, gtid, num_teams, thread_limit);
1890 __kmpc_fork_teams(&loc, 2, (microtask_t)__kmp_GOMP_microtask_wrapper, fn,
1891 data);
1892 KA_TRACE(20, ("GOMP_teams_reg exit: T#%d\n", gtid));
1893 }
1894
KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT_DEPEND)1895 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT_DEPEND)(void **depend) {
1896 MKLOC(loc, "GOMP_taskwait_depend");
1897 int gtid = __kmp_entry_gtid();
1898 KA_TRACE(20, ("GOMP_taskwait_depend: T#%d\n", gtid));
1899 kmp_gomp_depends_info_t gomp_depends(depend);
1900 kmp_int32 ndeps = gomp_depends.get_num_deps();
1901 kmp_depend_info_t dep_list[ndeps];
1902 for (kmp_int32 i = 0; i < ndeps; i++)
1903 dep_list[i] = gomp_depends.get_kmp_depend(i);
1904 #if OMPT_SUPPORT
1905 OMPT_STORE_RETURN_ADDRESS(gtid);
1906 #endif
1907 __kmpc_omp_wait_deps(&loc, gtid, ndeps, dep_list, 0, NULL);
1908 KA_TRACE(20, ("GOMP_taskwait_depend exit: T#%d\n", gtid));
1909 }
1910
1911 /* The following sections of code create aliases for the GOMP_* functions, then
1912 create versioned symbols using the assembler directive .symver. This is only
1913 pertinent for ELF .so library. The KMP_VERSION_SYMBOL macro is defined in
1914 kmp_os.h */
1915
1916 #ifdef KMP_USE_VERSION_SYMBOLS
1917 // GOMP_1.0 versioned symbols
1918 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_END, 10, "GOMP_1.0");
1919 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_START, 10, "GOMP_1.0");
1920 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER, 10, "GOMP_1.0");
1921 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_END, 10, "GOMP_1.0");
1922 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_END, 10, "GOMP_1.0");
1923 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_START, 10, "GOMP_1.0");
1924 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_START, 10, "GOMP_1.0");
1925 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT, 10, "GOMP_1.0");
1926 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START, 10, "GOMP_1.0");
1927 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END, 10, "GOMP_1.0");
1928 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_NOWAIT, 10, "GOMP_1.0");
1929 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT, 10, "GOMP_1.0");
1930 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_START, 10, "GOMP_1.0");
1931 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT, 10, "GOMP_1.0");
1932 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START, 10,
1933 "GOMP_1.0");
1934 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT, 10, "GOMP_1.0");
1935 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START, 10, "GOMP_1.0");
1936 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT, 10, "GOMP_1.0");
1937 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START, 10,
1938 "GOMP_1.0");
1939 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT, 10, "GOMP_1.0");
1940 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START, 10, "GOMP_1.0");
1941 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT, 10, "GOMP_1.0");
1942 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_START, 10, "GOMP_1.0");
1943 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT, 10, "GOMP_1.0");
1944 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_START, 10, "GOMP_1.0");
1945 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_END, 10, "GOMP_1.0");
1946 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_START, 10, "GOMP_1.0");
1947 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_END, 10, "GOMP_1.0");
1948 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START, 10,
1949 "GOMP_1.0");
1950 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START, 10,
1951 "GOMP_1.0");
1952 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START, 10,
1953 "GOMP_1.0");
1954 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START, 10,
1955 "GOMP_1.0");
1956 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START, 10, "GOMP_1.0");
1957 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_START, 10, "GOMP_1.0");
1958 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END, 10, "GOMP_1.0");
1959 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT, 10, "GOMP_1.0");
1960 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_NEXT, 10, "GOMP_1.0");
1961 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_START, 10, "GOMP_1.0");
1962 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_END, 10, "GOMP_1.0");
1963 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_START, 10, "GOMP_1.0");
1964 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_START, 10, "GOMP_1.0");
1965
1966 // GOMP_2.0 versioned symbols
1967 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASK, 20, "GOMP_2.0");
1968 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKWAIT, 20, "GOMP_2.0");
1969 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT, 20, "GOMP_2.0");
1970 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START, 20, "GOMP_2.0");
1971 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT, 20, "GOMP_2.0");
1972 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START, 20, "GOMP_2.0");
1973 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT, 20,
1974 "GOMP_2.0");
1975 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START, 20,
1976 "GOMP_2.0");
1977 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT, 20,
1978 "GOMP_2.0");
1979 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START, 20,
1980 "GOMP_2.0");
1981 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT, 20,
1982 "GOMP_2.0");
1983 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START, 20,
1984 "GOMP_2.0");
1985 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT, 20,
1986 "GOMP_2.0");
1987 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START, 20,
1988 "GOMP_2.0");
1989 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT, 20, "GOMP_2.0");
1990 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START, 20, "GOMP_2.0");
1991 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT, 20, "GOMP_2.0");
1992 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START, 20, "GOMP_2.0");
1993
1994 // GOMP_3.0 versioned symbols
1995 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKYIELD, 30, "GOMP_3.0");
1996
1997 // GOMP_4.0 versioned symbols
1998 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL, 40, "GOMP_4.0");
1999 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS, 40, "GOMP_4.0");
2000 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC, 40, "GOMP_4.0");
2001 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED, 40, "GOMP_4.0");
2002 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME, 40, "GOMP_4.0");
2003 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC, 40, "GOMP_4.0");
2004 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_START, 40, "GOMP_4.0");
2005 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_END, 40, "GOMP_4.0");
2006 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER_CANCEL, 40, "GOMP_4.0");
2007 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCEL, 40, "GOMP_4.0");
2008 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCELLATION_POINT, 40, "GOMP_4.0");
2009 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_CANCEL, 40, "GOMP_4.0");
2010 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL, 40, "GOMP_4.0");
2011 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET, 40, "GOMP_4.0");
2012 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_DATA, 40, "GOMP_4.0");
2013 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_END_DATA, 40, "GOMP_4.0");
2014 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_UPDATE, 40, "GOMP_4.0");
2015 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TEAMS, 40, "GOMP_4.0");
2016
2017 // GOMP_4.5 versioned symbols
2018 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP, 45, "GOMP_4.5");
2019 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP_ULL, 45, "GOMP_4.5");
2020 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_POST, 45, "GOMP_4.5");
2021 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_WAIT, 45, "GOMP_4.5");
2022 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START, 45,
2023 "GOMP_4.5");
2024 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START, 45,
2025 "GOMP_4.5");
2026 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START, 45,
2027 "GOMP_4.5");
2028 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START, 45,
2029 "GOMP_4.5");
2030 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_ULL_POST, 45, "GOMP_4.5");
2031 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT, 45, "GOMP_4.5");
2032 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START, 45,
2033 "GOMP_4.5");
2034 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START, 45,
2035 "GOMP_4.5");
2036 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START, 45,
2037 "GOMP_4.5");
2038 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START, 45,
2039 "GOMP_4.5");
2040 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_START, 45,
2041 "GOMP_4.5");
2042 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_NEXT, 45,
2043 "GOMP_4.5");
2044 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_START, 45,
2045 "GOMP_4.5");
2046 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_NEXT, 45,
2047 "GOMP_4.5");
2048 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_START, 45,
2049 "GOMP_4.5");
2050 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_NEXT, 45,
2051 "GOMP_4.5");
2052 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_START, 45,
2053 "GOMP_4.5");
2054 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_NEXT, 45,
2055 "GOMP_4.5");
2056 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_DYNAMIC, 45,
2057 "GOMP_4.5");
2058 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_GUIDED, 45,
2059 "GOMP_4.5");
2060
2061 // GOMP_5.0 versioned symbols
2062 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_NEXT, 50,
2063 "GOMP_5.0");
2064 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_START, 50,
2065 "GOMP_5.0");
2066 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_NEXT, 50,
2067 "GOMP_5.0");
2068 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_START, 50,
2069 "GOMP_5.0");
2070 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_NEXT,
2071 50, "GOMP_5.0");
2072 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_START,
2073 50, "GOMP_5.0");
2074 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_NEXT, 50,
2075 "GOMP_5.0");
2076 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_START, 50,
2077 "GOMP_5.0");
2078 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_RUNTIME, 50,
2079 "GOMP_5.0");
2080 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_MAYBE_NONMONOTONIC_RUNTIME,
2081 50, "GOMP_5.0");
2082 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TEAMS_REG, 50, "GOMP_5.0");
2083 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKWAIT_DEPEND, 50, "GOMP_5.0");
2084
2085 #endif // KMP_USE_VERSION_SYMBOLS
2086
2087 #ifdef __cplusplus
2088 } // extern "C"
2089 #endif // __cplusplus
2090