LLVM OpenMP* Runtime Library
kmp_taskdeps.cpp
1/*
2 * kmp_taskdeps.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//#define KMP_SUPPORT_GRAPH_OUTPUT 1
14
15#include "kmp.h"
16#include "kmp_io.h"
17#include "kmp_wait_release.h"
18#include "kmp_taskdeps.h"
19#if OMPT_SUPPORT
20#include "ompt-specific.h"
21#endif
22
23// TODO: Improve memory allocation? keep a list of pre-allocated structures?
24// allocate in blocks? re-use list finished list entries?
25// TODO: don't use atomic ref counters for stack-allocated nodes.
26// TODO: find an alternate to atomic refs for heap-allocated nodes?
27// TODO: Finish graph output support
28// TODO: kmp_lock_t seems a tad to big (and heavy weight) for this. Check other
29// runtime locks
30// TODO: Any ITT support needed?
31
32#ifdef KMP_SUPPORT_GRAPH_OUTPUT
33static std::atomic<kmp_int32> kmp_node_id_seed = ATOMIC_VAR_INIT(0);
34#endif
35
36static void __kmp_init_node(kmp_depnode_t *node) {
37 node->dn.successors = NULL;
38 node->dn.task = NULL; // will point to the right task
39 // once dependences have been processed
40 for (int i = 0; i < MAX_MTX_DEPS; ++i)
41 node->dn.mtx_locks[i] = NULL;
42 node->dn.mtx_num_locks = 0;
43 __kmp_init_lock(&node->dn.lock);
44 KMP_ATOMIC_ST_RLX(&node->dn.nrefs, 1); // init creates the first reference
45#ifdef KMP_SUPPORT_GRAPH_OUTPUT
46 node->dn.id = KMP_ATOMIC_INC(&kmp_node_id_seed);
47#endif
48}
49
50static inline kmp_depnode_t *__kmp_node_ref(kmp_depnode_t *node) {
51 KMP_ATOMIC_INC(&node->dn.nrefs);
52 return node;
53}
54
55enum { KMP_DEPHASH_OTHER_SIZE = 97, KMP_DEPHASH_MASTER_SIZE = 997 };
56
57size_t sizes[] = {997, 2003, 4001, 8191, 16001, 32003, 64007, 131071, 270029};
58const size_t MAX_GEN = 8;
59
60static inline size_t __kmp_dephash_hash(kmp_intptr_t addr, size_t hsize) {
61 // TODO alternate to try: set = (((Addr64)(addrUsefulBits * 9.618)) %
62 // m_num_sets );
63 return ((addr >> 6) ^ (addr >> 2)) % hsize;
64}
65
66static kmp_dephash_t *__kmp_dephash_extend(kmp_info_t *thread,
67 kmp_dephash_t *current_dephash) {
68 kmp_dephash_t *h;
69
70 size_t gen = current_dephash->generation + 1;
71 if (gen >= MAX_GEN)
72 return current_dephash;
73 size_t new_size = sizes[gen];
74
75 size_t size_to_allocate =
76 new_size * sizeof(kmp_dephash_entry_t *) + sizeof(kmp_dephash_t);
77
78#if USE_FAST_MEMORY
79 h = (kmp_dephash_t *)__kmp_fast_allocate(thread, size_to_allocate);
80#else
81 h = (kmp_dephash_t *)__kmp_thread_malloc(thread, size_to_allocate);
82#endif
83
84 h->size = new_size;
85 h->nelements = current_dephash->nelements;
86 h->buckets = (kmp_dephash_entry **)(h + 1);
87 h->generation = gen;
88 h->nconflicts = 0;
89
90 // make sure buckets are properly initialized
91 for (size_t i = 0; i < new_size; i++) {
92 h->buckets[i] = NULL;
93 }
94
95 // insert existing elements in the new table
96 for (size_t i = 0; i < current_dephash->size; i++) {
97 kmp_dephash_entry_t *next, *entry;
98 for (entry = current_dephash->buckets[i]; entry; entry = next) {
99 next = entry->next_in_bucket;
100 // Compute the new hash using the new size, and insert the entry in
101 // the new bucket.
102 size_t new_bucket = __kmp_dephash_hash(entry->addr, h->size);
103 entry->next_in_bucket = h->buckets[new_bucket];
104 if (entry->next_in_bucket) {
105 h->nconflicts++;
106 }
107 h->buckets[new_bucket] = entry;
108 }
109 }
110
111 // Free old hash table
112#if USE_FAST_MEMORY
113 __kmp_fast_free(thread, current_dephash);
114#else
115 __kmp_thread_free(thread, current_dephash);
116#endif
117
118 return h;
119}
120
121static kmp_dephash_t *__kmp_dephash_create(kmp_info_t *thread,
122 kmp_taskdata_t *current_task) {
123 kmp_dephash_t *h;
124
125 size_t h_size;
126
127 if (current_task->td_flags.tasktype == TASK_IMPLICIT)
128 h_size = KMP_DEPHASH_MASTER_SIZE;
129 else
130 h_size = KMP_DEPHASH_OTHER_SIZE;
131
132 size_t size = h_size * sizeof(kmp_dephash_entry_t *) + sizeof(kmp_dephash_t);
133
134#if USE_FAST_MEMORY
135 h = (kmp_dephash_t *)__kmp_fast_allocate(thread, size);
136#else
137 h = (kmp_dephash_t *)__kmp_thread_malloc(thread, size);
138#endif
139 h->size = h_size;
140
141 h->generation = 0;
142 h->nelements = 0;
143 h->nconflicts = 0;
144 h->buckets = (kmp_dephash_entry **)(h + 1);
145
146 for (size_t i = 0; i < h_size; i++)
147 h->buckets[i] = 0;
148
149 return h;
150}
151
152static kmp_dephash_entry *__kmp_dephash_find(kmp_info_t *thread,
153 kmp_dephash_t **hash,
154 kmp_intptr_t addr) {
155 kmp_dephash_t *h = *hash;
156 if (h->nelements != 0 && h->nconflicts / h->size >= 1) {
157 *hash = __kmp_dephash_extend(thread, h);
158 h = *hash;
159 }
160 size_t bucket = __kmp_dephash_hash(addr, h->size);
161
162 kmp_dephash_entry_t *entry;
163 for (entry = h->buckets[bucket]; entry; entry = entry->next_in_bucket)
164 if (entry->addr == addr)
165 break;
166
167 if (entry == NULL) {
168// create entry. This is only done by one thread so no locking required
169#if USE_FAST_MEMORY
170 entry = (kmp_dephash_entry_t *)__kmp_fast_allocate(
171 thread, sizeof(kmp_dephash_entry_t));
172#else
173 entry = (kmp_dephash_entry_t *)__kmp_thread_malloc(
174 thread, sizeof(kmp_dephash_entry_t));
175#endif
176 entry->addr = addr;
177 entry->last_out = NULL;
178 entry->last_set = NULL;
179 entry->prev_set = NULL;
180 entry->last_flag = 0;
181 entry->mtx_lock = NULL;
182 entry->next_in_bucket = h->buckets[bucket];
183 h->buckets[bucket] = entry;
184 h->nelements++;
185 if (entry->next_in_bucket)
186 h->nconflicts++;
187 }
188 return entry;
189}
190
191static kmp_depnode_list_t *__kmp_add_node(kmp_info_t *thread,
192 kmp_depnode_list_t *list,
193 kmp_depnode_t *node) {
194 kmp_depnode_list_t *new_head;
195
196#if USE_FAST_MEMORY
197 new_head = (kmp_depnode_list_t *)__kmp_fast_allocate(
198 thread, sizeof(kmp_depnode_list_t));
199#else
200 new_head = (kmp_depnode_list_t *)__kmp_thread_malloc(
201 thread, sizeof(kmp_depnode_list_t));
202#endif
203
204 new_head->node = __kmp_node_ref(node);
205 new_head->next = list;
206
207 return new_head;
208}
209
210static inline void __kmp_track_dependence(kmp_int32 gtid, kmp_depnode_t *source,
211 kmp_depnode_t *sink,
212 kmp_task_t *sink_task) {
213#ifdef KMP_SUPPORT_GRAPH_OUTPUT
214 kmp_taskdata_t *task_source = KMP_TASK_TO_TASKDATA(source->dn.task);
215 // do not use sink->dn.task as that is only filled after the dependences
216 // are already processed!
217 kmp_taskdata_t *task_sink = KMP_TASK_TO_TASKDATA(sink_task);
218
219 __kmp_printf("%d(%s) -> %d(%s)\n", source->dn.id,
220 task_source->td_ident->psource, sink->dn.id,
221 task_sink->td_ident->psource);
222#endif
223#if OMPT_SUPPORT && OMPT_OPTIONAL
224 /* OMPT tracks dependences between task (a=source, b=sink) in which
225 task a blocks the execution of b through the ompt_new_dependence_callback
226 */
227 if (ompt_enabled.ompt_callback_task_dependence) {
228 kmp_taskdata_t *task_source = KMP_TASK_TO_TASKDATA(source->dn.task);
229 ompt_data_t *sink_data;
230 if (sink_task)
231 sink_data = &(KMP_TASK_TO_TASKDATA(sink_task)->ompt_task_info.task_data);
232 else
233 sink_data = &__kmp_threads[gtid]->th.ompt_thread_info.task_data;
234
235 ompt_callbacks.ompt_callback(ompt_callback_task_dependence)(
236 &(task_source->ompt_task_info.task_data), sink_data);
237 }
238#endif /* OMPT_SUPPORT && OMPT_OPTIONAL */
239}
240
241static inline kmp_int32
242__kmp_depnode_link_successor(kmp_int32 gtid, kmp_info_t *thread,
243 kmp_task_t *task, kmp_depnode_t *node,
244 kmp_depnode_list_t *plist) {
245 if (!plist)
246 return 0;
247 kmp_int32 npredecessors = 0;
248 // link node as successor of list elements
249 for (kmp_depnode_list_t *p = plist; p; p = p->next) {
250 kmp_depnode_t *dep = p->node;
251 if (dep->dn.task) {
252 KMP_ACQUIRE_DEPNODE(gtid, dep);
253 if (dep->dn.task) {
254 __kmp_track_dependence(gtid, dep, node, task);
255 dep->dn.successors = __kmp_add_node(thread, dep->dn.successors, node);
256 KA_TRACE(40, ("__kmp_process_deps: T#%d adding dependence from %p to "
257 "%p\n",
258 gtid, KMP_TASK_TO_TASKDATA(dep->dn.task),
259 KMP_TASK_TO_TASKDATA(task)));
260 npredecessors++;
261 }
262 KMP_RELEASE_DEPNODE(gtid, dep);
263 }
264 }
265 return npredecessors;
266}
267
268static inline kmp_int32 __kmp_depnode_link_successor(kmp_int32 gtid,
269 kmp_info_t *thread,
270 kmp_task_t *task,
271 kmp_depnode_t *source,
272 kmp_depnode_t *sink) {
273 if (!sink)
274 return 0;
275 kmp_int32 npredecessors = 0;
276 if (sink->dn.task) {
277 // synchronously add source to sink' list of successors
278 KMP_ACQUIRE_DEPNODE(gtid, sink);
279 if (sink->dn.task) {
280 __kmp_track_dependence(gtid, sink, source, task);
281 sink->dn.successors = __kmp_add_node(thread, sink->dn.successors, source);
282 KA_TRACE(40, ("__kmp_process_deps: T#%d adding dependence from %p to "
283 "%p\n",
284 gtid, KMP_TASK_TO_TASKDATA(sink->dn.task),
285 KMP_TASK_TO_TASKDATA(task)));
286 npredecessors++;
287 }
288 KMP_RELEASE_DEPNODE(gtid, sink);
289 }
290 return npredecessors;
291}
292
293template <bool filter>
294static inline kmp_int32
295__kmp_process_deps(kmp_int32 gtid, kmp_depnode_t *node, kmp_dephash_t **hash,
296 bool dep_barrier, kmp_int32 ndeps,
297 kmp_depend_info_t *dep_list, kmp_task_t *task) {
298 KA_TRACE(30, ("__kmp_process_deps<%d>: T#%d processing %d dependences : "
299 "dep_barrier = %d\n",
300 filter, gtid, ndeps, dep_barrier));
301
302 kmp_info_t *thread = __kmp_threads[gtid];
303 kmp_int32 npredecessors = 0;
304 for (kmp_int32 i = 0; i < ndeps; i++) {
305 const kmp_depend_info_t *dep = &dep_list[i];
306
307 if (filter && dep->base_addr == 0)
308 continue; // skip filtered entries
309
310 kmp_dephash_entry_t *info =
311 __kmp_dephash_find(thread, hash, dep->base_addr);
312 kmp_depnode_t *last_out = info->last_out;
313 kmp_depnode_list_t *last_set = info->last_set;
314 kmp_depnode_list_t *prev_set = info->prev_set;
315
316 if (dep->flags.out) { // out or inout --> clean lists if any
317 if (last_set) {
318 npredecessors +=
319 __kmp_depnode_link_successor(gtid, thread, task, node, last_set);
320 __kmp_depnode_list_free(thread, last_set);
321 __kmp_depnode_list_free(thread, prev_set);
322 info->last_set = NULL;
323 info->prev_set = NULL;
324 info->last_flag = 0; // no sets in this dephash entry
325 } else {
326 npredecessors +=
327 __kmp_depnode_link_successor(gtid, thread, task, node, last_out);
328 }
329 __kmp_node_deref(thread, last_out);
330 if (!dep_barrier) {
331 info->last_out = __kmp_node_ref(node);
332 } else {
333 // if this is a sync point in the serial sequence, then the previous
334 // outputs are guaranteed to be completed after the execution of this
335 // task so the previous output nodes can be cleared.
336 info->last_out = NULL;
337 }
338 } else { // either IN or MTX or SET
339 if (info->last_flag == 0 || info->last_flag == dep->flag) {
340 // last_set either didn't exist or of same dep kind
341 // link node as successor of the last_out if any
342 npredecessors +=
343 __kmp_depnode_link_successor(gtid, thread, task, node, last_out);
344 // link node as successor of all nodes in the prev_set if any
345 npredecessors +=
346 __kmp_depnode_link_successor(gtid, thread, task, node, prev_set);
347 if (dep_barrier) {
348 // clean last_out and prev_set if any; don't touch last_set
349 __kmp_node_deref(thread, last_out);
350 info->last_out = NULL;
351 __kmp_depnode_list_free(thread, prev_set);
352 info->prev_set = NULL;
353 }
354 } else { // last_set is of different dep kind, make it prev_set
355 // link node as successor of all nodes in the last_set
356 npredecessors +=
357 __kmp_depnode_link_successor(gtid, thread, task, node, last_set);
358 // clean last_out if any
359 __kmp_node_deref(thread, last_out);
360 info->last_out = NULL;
361 // clean prev_set if any
362 __kmp_depnode_list_free(thread, prev_set);
363 if (!dep_barrier) {
364 // move last_set to prev_set, new last_set will be allocated
365 info->prev_set = last_set;
366 } else {
367 info->prev_set = NULL;
368 info->last_flag = 0;
369 }
370 info->last_set = NULL;
371 }
372 // for dep_barrier last_flag value should remain:
373 // 0 if last_set is empty, unchanged otherwise
374 if (!dep_barrier) {
375 info->last_flag = dep->flag; // store dep kind of the last_set
376 info->last_set = __kmp_add_node(thread, info->last_set, node);
377 }
378 // check if we are processing MTX dependency
379 if (dep->flag == KMP_DEP_MTX) {
380 if (info->mtx_lock == NULL) {
381 info->mtx_lock = (kmp_lock_t *)__kmp_allocate(sizeof(kmp_lock_t));
382 __kmp_init_lock(info->mtx_lock);
383 }
384 KMP_DEBUG_ASSERT(node->dn.mtx_num_locks < MAX_MTX_DEPS);
385 kmp_int32 m;
386 // Save lock in node's array
387 for (m = 0; m < MAX_MTX_DEPS; ++m) {
388 // sort pointers in decreasing order to avoid potential livelock
389 if (node->dn.mtx_locks[m] < info->mtx_lock) {
390 KMP_DEBUG_ASSERT(!node->dn.mtx_locks[node->dn.mtx_num_locks]);
391 for (int n = node->dn.mtx_num_locks; n > m; --n) {
392 // shift right all lesser non-NULL pointers
393 KMP_DEBUG_ASSERT(node->dn.mtx_locks[n - 1] != NULL);
394 node->dn.mtx_locks[n] = node->dn.mtx_locks[n - 1];
395 }
396 node->dn.mtx_locks[m] = info->mtx_lock;
397 break;
398 }
399 }
400 KMP_DEBUG_ASSERT(m < MAX_MTX_DEPS); // must break from loop
401 node->dn.mtx_num_locks++;
402 }
403 }
404 }
405 KA_TRACE(30, ("__kmp_process_deps<%d>: T#%d found %d predecessors\n", filter,
406 gtid, npredecessors));
407 return npredecessors;
408}
409
410#define NO_DEP_BARRIER (false)
411#define DEP_BARRIER (true)
412
413// returns true if the task has any outstanding dependence
414static bool __kmp_check_deps(kmp_int32 gtid, kmp_depnode_t *node,
415 kmp_task_t *task, kmp_dephash_t **hash,
416 bool dep_barrier, kmp_int32 ndeps,
417 kmp_depend_info_t *dep_list,
418 kmp_int32 ndeps_noalias,
419 kmp_depend_info_t *noalias_dep_list) {
420 int i, n_mtxs = 0;
421#if KMP_DEBUG
422 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
423#endif
424 KA_TRACE(20, ("__kmp_check_deps: T#%d checking dependences for task %p : %d "
425 "possibly aliased dependences, %d non-aliased dependences : "
426 "dep_barrier=%d .\n",
427 gtid, taskdata, ndeps, ndeps_noalias, dep_barrier));
428
429 // Filter deps in dep_list
430 // TODO: Different algorithm for large dep_list ( > 10 ? )
431 for (i = 0; i < ndeps; i++) {
432 if (dep_list[i].base_addr != 0) {
433 KMP_DEBUG_ASSERT(
434 dep_list[i].flag == KMP_DEP_IN || dep_list[i].flag == KMP_DEP_OUT ||
435 dep_list[i].flag == KMP_DEP_INOUT ||
436 dep_list[i].flag == KMP_DEP_MTX || dep_list[i].flag == KMP_DEP_SET);
437 for (int j = i + 1; j < ndeps; j++) {
438 if (dep_list[i].base_addr == dep_list[j].base_addr) {
439 if (dep_list[i].flag != dep_list[j].flag) {
440 // two different dependences on same address work identical to OUT
441 dep_list[i].flag = KMP_DEP_OUT;
442 }
443 dep_list[j].base_addr = 0; // Mark j element as void
444 }
445 }
446 if (dep_list[i].flag == KMP_DEP_MTX) {
447 // limit number of mtx deps to MAX_MTX_DEPS per node
448 if (n_mtxs < MAX_MTX_DEPS && task != NULL) {
449 ++n_mtxs;
450 } else {
451 dep_list[i].flag = KMP_DEP_OUT; // downgrade mutexinoutset to inout
452 }
453 }
454 }
455 }
456
457 // doesn't need to be atomic as no other thread is going to be accessing this
458 // node just yet.
459 // npredecessors is set -1 to ensure that none of the releasing tasks queues
460 // this task before we have finished processing all the dependences
461 node->dn.npredecessors = -1;
462
463 // used to pack all npredecessors additions into a single atomic operation at
464 // the end
465 int npredecessors;
466
467 npredecessors = __kmp_process_deps<true>(gtid, node, hash, dep_barrier, ndeps,
468 dep_list, task);
469 npredecessors += __kmp_process_deps<false>(
470 gtid, node, hash, dep_barrier, ndeps_noalias, noalias_dep_list, task);
471
472 node->dn.task = task;
473 KMP_MB();
474
475 // Account for our initial fake value
476 npredecessors++;
477
478 // Update predecessors and obtain current value to check if there are still
479 // any outstanding dependences (some tasks may have finished while we
480 // processed the dependences)
481 npredecessors =
482 node->dn.npredecessors.fetch_add(npredecessors) + npredecessors;
483
484 KA_TRACE(20, ("__kmp_check_deps: T#%d found %d predecessors for task %p \n",
485 gtid, npredecessors, taskdata));
486
487 // beyond this point the task could be queued (and executed) by a releasing
488 // task...
489 return npredecessors > 0 ? true : false;
490}
491
508kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid,
509 kmp_task_t *new_task, kmp_int32 ndeps,
510 kmp_depend_info_t *dep_list,
511 kmp_int32 ndeps_noalias,
512 kmp_depend_info_t *noalias_dep_list) {
513
514 kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
515 KA_TRACE(10, ("__kmpc_omp_task_with_deps(enter): T#%d loc=%p task=%p\n", gtid,
516 loc_ref, new_taskdata));
517 __kmp_assert_valid_gtid(gtid);
518 kmp_info_t *thread = __kmp_threads[gtid];
519 kmp_taskdata_t *current_task = thread->th.th_current_task;
520
521#if OMPT_SUPPORT
522 if (ompt_enabled.enabled) {
523 if (!current_task->ompt_task_info.frame.enter_frame.ptr)
524 current_task->ompt_task_info.frame.enter_frame.ptr =
525 OMPT_GET_FRAME_ADDRESS(0);
526 if (ompt_enabled.ompt_callback_task_create) {
527 ompt_callbacks.ompt_callback(ompt_callback_task_create)(
528 &(current_task->ompt_task_info.task_data),
529 &(current_task->ompt_task_info.frame),
530 &(new_taskdata->ompt_task_info.task_data),
531 ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 1,
532 OMPT_LOAD_OR_GET_RETURN_ADDRESS(gtid));
533 }
534
535 new_taskdata->ompt_task_info.frame.enter_frame.ptr =
536 OMPT_GET_FRAME_ADDRESS(0);
537 }
538
539#if OMPT_OPTIONAL
540 /* OMPT grab all dependences if requested by the tool */
541 if (ndeps + ndeps_noalias > 0 && ompt_enabled.ompt_callback_dependences) {
542 kmp_int32 i;
543
544 int ompt_ndeps = ndeps + ndeps_noalias;
545 ompt_dependence_t *ompt_deps = (ompt_dependence_t *)KMP_OMPT_DEPS_ALLOC(
546 thread, (ndeps + ndeps_noalias) * sizeof(ompt_dependence_t));
547
548 KMP_ASSERT(ompt_deps != NULL);
549
550 for (i = 0; i < ndeps; i++) {
551 ompt_deps[i].variable.ptr = (void *)dep_list[i].base_addr;
552 if (dep_list[i].flags.in && dep_list[i].flags.out)
553 ompt_deps[i].dependence_type = ompt_dependence_type_inout;
554 else if (dep_list[i].flags.out)
555 ompt_deps[i].dependence_type = ompt_dependence_type_out;
556 else if (dep_list[i].flags.in)
557 ompt_deps[i].dependence_type = ompt_dependence_type_in;
558 else if (dep_list[i].flags.mtx)
559 ompt_deps[i].dependence_type = ompt_dependence_type_mutexinoutset;
560 else if (dep_list[i].flags.set)
561 ompt_deps[i].dependence_type = ompt_dependence_type_inoutset;
562 }
563 for (i = 0; i < ndeps_noalias; i++) {
564 ompt_deps[ndeps + i].variable.ptr = (void *)noalias_dep_list[i].base_addr;
565 if (noalias_dep_list[i].flags.in && noalias_dep_list[i].flags.out)
566 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inout;
567 else if (noalias_dep_list[i].flags.out)
568 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_out;
569 else if (noalias_dep_list[i].flags.in)
570 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_in;
571 else if (noalias_dep_list[i].flags.mtx)
572 ompt_deps[ndeps + i].dependence_type =
573 ompt_dependence_type_mutexinoutset;
574 else if (noalias_dep_list[i].flags.set)
575 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inoutset;
576 }
577 ompt_callbacks.ompt_callback(ompt_callback_dependences)(
578 &(new_taskdata->ompt_task_info.task_data), ompt_deps, ompt_ndeps);
579 /* We can now free the allocated memory for the dependences */
580 /* For OMPD we might want to delay the free until end of this function */
581 KMP_OMPT_DEPS_FREE(thread, ompt_deps);
582 }
583#endif /* OMPT_OPTIONAL */
584#endif /* OMPT_SUPPORT */
585
586 bool serial = current_task->td_flags.team_serial ||
587 current_task->td_flags.tasking_ser ||
588 current_task->td_flags.final;
589 kmp_task_team_t *task_team = thread->th.th_task_team;
590 serial = serial &&
591 !(task_team && (task_team->tt.tt_found_proxy_tasks ||
592 task_team->tt.tt_hidden_helper_task_encountered));
593
594 if (!serial && (ndeps > 0 || ndeps_noalias > 0)) {
595 /* if no dependences have been tracked yet, create the dependence hash */
596 if (current_task->td_dephash == NULL)
597 current_task->td_dephash = __kmp_dephash_create(thread, current_task);
598
599#if USE_FAST_MEMORY
600 kmp_depnode_t *node =
601 (kmp_depnode_t *)__kmp_fast_allocate(thread, sizeof(kmp_depnode_t));
602#else
603 kmp_depnode_t *node =
604 (kmp_depnode_t *)__kmp_thread_malloc(thread, sizeof(kmp_depnode_t));
605#endif
606
607 __kmp_init_node(node);
608 new_taskdata->td_depnode = node;
609
610 if (__kmp_check_deps(gtid, node, new_task, &current_task->td_dephash,
611 NO_DEP_BARRIER, ndeps, dep_list, ndeps_noalias,
612 noalias_dep_list)) {
613 KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d task had blocking "
614 "dependences: "
615 "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n",
616 gtid, loc_ref, new_taskdata));
617#if OMPT_SUPPORT
618 if (ompt_enabled.enabled) {
619 current_task->ompt_task_info.frame.enter_frame = ompt_data_none;
620 }
621#endif
622 return TASK_CURRENT_NOT_QUEUED;
623 }
624 } else {
625 KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d ignored dependences "
626 "for task (serialized) loc=%p task=%p\n",
627 gtid, loc_ref, new_taskdata));
628 }
629
630 KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d task had no blocking "
631 "dependences : "
632 "loc=%p task=%p, transferring to __kmp_omp_task\n",
633 gtid, loc_ref, new_taskdata));
634
635 kmp_int32 ret = __kmp_omp_task(gtid, new_task, true);
636#if OMPT_SUPPORT
637 if (ompt_enabled.enabled) {
638 current_task->ompt_task_info.frame.enter_frame = ompt_data_none;
639 }
640#endif
641 return ret;
642}
643
644#if OMPT_SUPPORT
645void __ompt_taskwait_dep_finish(kmp_taskdata_t *current_task,
646 ompt_data_t *taskwait_task_data) {
647 if (ompt_enabled.ompt_callback_task_schedule) {
648 ompt_callbacks.ompt_callback(ompt_callback_task_schedule)(
649 taskwait_task_data, ompt_taskwait_complete, NULL);
650 }
651 current_task->ompt_task_info.frame.enter_frame.ptr = NULL;
652 *taskwait_task_data = ompt_data_none;
653}
654#endif /* OMPT_SUPPORT */
655
667void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps,
668 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
669 kmp_depend_info_t *noalias_dep_list) {
670 KA_TRACE(10, ("__kmpc_omp_wait_deps(enter): T#%d loc=%p\n", gtid, loc_ref));
671
672 if (ndeps == 0 && ndeps_noalias == 0) {
673 KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d has no dependences to "
674 "wait upon : loc=%p\n",
675 gtid, loc_ref));
676 return;
677 }
678 __kmp_assert_valid_gtid(gtid);
679 kmp_info_t *thread = __kmp_threads[gtid];
680 kmp_taskdata_t *current_task = thread->th.th_current_task;
681
682#if OMPT_SUPPORT
683 // this function represents a taskwait construct with depend clause
684 // We signal 4 events:
685 // - creation of the taskwait task
686 // - dependences of the taskwait task
687 // - schedule and finish of the taskwait task
688 ompt_data_t *taskwait_task_data = &thread->th.ompt_thread_info.task_data;
689 KMP_ASSERT(taskwait_task_data->ptr == NULL);
690 if (ompt_enabled.enabled) {
691 if (!current_task->ompt_task_info.frame.enter_frame.ptr)
692 current_task->ompt_task_info.frame.enter_frame.ptr =
693 OMPT_GET_FRAME_ADDRESS(0);
694 if (ompt_enabled.ompt_callback_task_create) {
695 ompt_callbacks.ompt_callback(ompt_callback_task_create)(
696 &(current_task->ompt_task_info.task_data),
697 &(current_task->ompt_task_info.frame), taskwait_task_data,
698 ompt_task_taskwait | ompt_task_undeferred | ompt_task_mergeable, 1,
699 OMPT_LOAD_OR_GET_RETURN_ADDRESS(gtid));
700 }
701 }
702
703#if OMPT_OPTIONAL
704 /* OMPT grab all dependences if requested by the tool */
705 if (ndeps + ndeps_noalias > 0 && ompt_enabled.ompt_callback_dependences) {
706 kmp_int32 i;
707
708 int ompt_ndeps = ndeps + ndeps_noalias;
709 ompt_dependence_t *ompt_deps = (ompt_dependence_t *)KMP_OMPT_DEPS_ALLOC(
710 thread, (ndeps + ndeps_noalias) * sizeof(ompt_dependence_t));
711
712 KMP_ASSERT(ompt_deps != NULL);
713
714 for (i = 0; i < ndeps; i++) {
715 ompt_deps[i].variable.ptr = (void *)dep_list[i].base_addr;
716 if (dep_list[i].flags.in && dep_list[i].flags.out)
717 ompt_deps[i].dependence_type = ompt_dependence_type_inout;
718 else if (dep_list[i].flags.out)
719 ompt_deps[i].dependence_type = ompt_dependence_type_out;
720 else if (dep_list[i].flags.in)
721 ompt_deps[i].dependence_type = ompt_dependence_type_in;
722 else if (dep_list[i].flags.mtx)
723 ompt_deps[ndeps + i].dependence_type =
724 ompt_dependence_type_mutexinoutset;
725 else if (dep_list[i].flags.set)
726 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inoutset;
727 }
728 for (i = 0; i < ndeps_noalias; i++) {
729 ompt_deps[ndeps + i].variable.ptr = (void *)noalias_dep_list[i].base_addr;
730 if (noalias_dep_list[i].flags.in && noalias_dep_list[i].flags.out)
731 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inout;
732 else if (noalias_dep_list[i].flags.out)
733 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_out;
734 else if (noalias_dep_list[i].flags.in)
735 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_in;
736 else if (noalias_dep_list[i].flags.mtx)
737 ompt_deps[ndeps + i].dependence_type =
738 ompt_dependence_type_mutexinoutset;
739 else if (noalias_dep_list[i].flags.set)
740 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inoutset;
741 }
742 ompt_callbacks.ompt_callback(ompt_callback_dependences)(
743 taskwait_task_data, ompt_deps, ompt_ndeps);
744 /* We can now free the allocated memory for the dependences */
745 /* For OMPD we might want to delay the free until end of this function */
746 KMP_OMPT_DEPS_FREE(thread, ompt_deps);
747 ompt_deps = NULL;
748 }
749#endif /* OMPT_OPTIONAL */
750#endif /* OMPT_SUPPORT */
751
752 // We can return immediately as:
753 // - dependences are not computed in serial teams (except with proxy tasks)
754 // - if the dephash is not yet created it means we have nothing to wait for
755 bool ignore = current_task->td_flags.team_serial ||
756 current_task->td_flags.tasking_ser ||
757 current_task->td_flags.final;
758 ignore = ignore && thread->th.th_task_team != NULL &&
759 thread->th.th_task_team->tt.tt_found_proxy_tasks == FALSE;
760 ignore = ignore || current_task->td_dephash == NULL;
761
762 if (ignore) {
763 KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d has no blocking "
764 "dependences : loc=%p\n",
765 gtid, loc_ref));
766#if OMPT_SUPPORT
767 __ompt_taskwait_dep_finish(current_task, taskwait_task_data);
768#endif /* OMPT_SUPPORT */
769 return;
770 }
771
772 kmp_depnode_t node = {0};
773 __kmp_init_node(&node);
774
775 if (!__kmp_check_deps(gtid, &node, NULL, &current_task->td_dephash,
776 DEP_BARRIER, ndeps, dep_list, ndeps_noalias,
777 noalias_dep_list)) {
778 KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d has no blocking "
779 "dependences : loc=%p\n",
780 gtid, loc_ref));
781#if OMPT_SUPPORT
782 __ompt_taskwait_dep_finish(current_task, taskwait_task_data);
783#endif /* OMPT_SUPPORT */
784 return;
785 }
786
787 int thread_finished = FALSE;
788 kmp_flag_32<false, false> flag(
789 (std::atomic<kmp_uint32> *)&node.dn.npredecessors, 0U);
790 while (node.dn.npredecessors > 0) {
791 flag.execute_tasks(thread, gtid, FALSE,
792 &thread_finished USE_ITT_BUILD_ARG(NULL),
793 __kmp_task_stealing_constraint);
794 }
795
796#if OMPT_SUPPORT
797 __ompt_taskwait_dep_finish(current_task, taskwait_task_data);
798#endif /* OMPT_SUPPORT */
799 KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d finished waiting : loc=%p\n",
800 gtid, loc_ref));
801}
kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
Definition: kmp.h:233