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 = 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(lck: &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#if USE_ITT_BUILD && USE_ITT_NOTIFY
49 __itt_sync_create(node, "OMP task dep node", NULL, 0);
50#endif
51}
52
53static inline kmp_depnode_t *__kmp_node_ref(kmp_depnode_t *node) {
54 KMP_ATOMIC_INC(&node->dn.nrefs);
55 return node;
56}
57
58enum { KMP_DEPHASH_OTHER_SIZE = 97, KMP_DEPHASH_MASTER_SIZE = 997 };
59
60size_t sizes[] = {997, 2003, 4001, 8191, 16001, 32003, 64007, 131071, 270029};
61const size_t MAX_GEN = 8;
62
63static inline size_t __kmp_dephash_hash(kmp_intptr_t addr, size_t hsize) {
64 // TODO alternate to try: set = (((Addr64)(addrUsefulBits * 9.618)) %
65 // m_num_sets );
66 return ((addr >> 6) ^ (addr >> 2)) % hsize;
67}
68
69static kmp_dephash_t *__kmp_dephash_extend(kmp_info_t *thread,
70 kmp_dephash_t *current_dephash) {
71 kmp_dephash_t *h;
72
73 size_t gen = current_dephash->generation + 1;
74 if (gen >= MAX_GEN)
75 return current_dephash;
76 size_t new_size = sizes[gen];
77
78 size_t size_to_allocate =
79 new_size * sizeof(kmp_dephash_entry_t *) + sizeof(kmp_dephash_t);
80
81#if USE_FAST_MEMORY
82 h = (kmp_dephash_t *)__kmp_fast_allocate(thread, size_to_allocate);
83#else
84 h = (kmp_dephash_t *)__kmp_thread_malloc(thread, size_to_allocate);
85#endif
86
87 h->size = new_size;
88 h->nelements = current_dephash->nelements;
89 h->buckets = (kmp_dephash_entry **)(h + 1);
90 h->generation = gen;
91 h->nconflicts = 0;
92 h->last_all = current_dephash->last_all;
93
94 // make sure buckets are properly initialized
95 for (size_t i = 0; i < new_size; i++) {
96 h->buckets[i] = NULL;
97 }
98
99 // insert existing elements in the new table
100 for (size_t i = 0; i < current_dephash->size; i++) {
101 kmp_dephash_entry_t *next, *entry;
102 for (entry = current_dephash->buckets[i]; entry; entry = next) {
103 next = entry->next_in_bucket;
104 // Compute the new hash using the new size, and insert the entry in
105 // the new bucket.
106 size_t new_bucket = __kmp_dephash_hash(addr: entry->addr, hsize: h->size);
107 entry->next_in_bucket = h->buckets[new_bucket];
108 if (entry->next_in_bucket) {
109 h->nconflicts++;
110 }
111 h->buckets[new_bucket] = entry;
112 }
113 }
114
115 // Free old hash table
116#if USE_FAST_MEMORY
117 __kmp_fast_free(thread, current_dephash);
118#else
119 __kmp_thread_free(thread, current_dephash);
120#endif
121
122 return h;
123}
124
125static kmp_dephash_t *__kmp_dephash_create(kmp_info_t *thread,
126 kmp_taskdata_t *current_task) {
127 kmp_dephash_t *h;
128
129 size_t h_size;
130
131 if (current_task->td_flags.tasktype == TASK_IMPLICIT)
132 h_size = KMP_DEPHASH_MASTER_SIZE;
133 else
134 h_size = KMP_DEPHASH_OTHER_SIZE;
135
136 size_t size = h_size * sizeof(kmp_dephash_entry_t *) + sizeof(kmp_dephash_t);
137
138#if USE_FAST_MEMORY
139 h = (kmp_dephash_t *)__kmp_fast_allocate(thread, size);
140#else
141 h = (kmp_dephash_t *)__kmp_thread_malloc(thread, size);
142#endif
143 h->size = h_size;
144
145 h->generation = 0;
146 h->nelements = 0;
147 h->nconflicts = 0;
148 h->buckets = (kmp_dephash_entry **)(h + 1);
149 h->last_all = NULL;
150
151 for (size_t i = 0; i < h_size; i++)
152 h->buckets[i] = 0;
153
154 return h;
155}
156
157static kmp_dephash_entry *__kmp_dephash_find(kmp_info_t *thread,
158 kmp_dephash_t **hash,
159 kmp_intptr_t addr) {
160 kmp_dephash_t *h = *hash;
161 if (h->nelements != 0 && h->nconflicts / h->size >= 1) {
162 *hash = __kmp_dephash_extend(thread, current_dephash: h);
163 h = *hash;
164 }
165 size_t bucket = __kmp_dephash_hash(addr, hsize: h->size);
166
167 kmp_dephash_entry_t *entry;
168 for (entry = h->buckets[bucket]; entry; entry = entry->next_in_bucket)
169 if (entry->addr == addr)
170 break;
171
172 if (entry == NULL) {
173// create entry. This is only done by one thread so no locking required
174#if USE_FAST_MEMORY
175 entry = (kmp_dephash_entry_t *)__kmp_fast_allocate(
176 thread, sizeof(kmp_dephash_entry_t));
177#else
178 entry = (kmp_dephash_entry_t *)__kmp_thread_malloc(
179 thread, sizeof(kmp_dephash_entry_t));
180#endif
181 entry->addr = addr;
182 if (!h->last_all) // no predecessor task with omp_all_memory dependence
183 entry->last_out = NULL;
184 else // else link the omp_all_memory depnode to the new entry
185 entry->last_out = __kmp_node_ref(node: h->last_all);
186 entry->last_set = NULL;
187 entry->prev_set = NULL;
188 entry->last_flag = 0;
189 entry->mtx_lock = NULL;
190 entry->next_in_bucket = h->buckets[bucket];
191 h->buckets[bucket] = entry;
192 h->nelements++;
193 if (entry->next_in_bucket)
194 h->nconflicts++;
195 }
196 return entry;
197}
198
199static kmp_depnode_list_t *__kmp_add_node(kmp_info_t *thread,
200 kmp_depnode_list_t *list,
201 kmp_depnode_t *node) {
202 kmp_depnode_list_t *new_head;
203
204#if USE_FAST_MEMORY
205 new_head = (kmp_depnode_list_t *)__kmp_fast_allocate(
206 thread, sizeof(kmp_depnode_list_t));
207#else
208 new_head = (kmp_depnode_list_t *)__kmp_thread_malloc(
209 thread, sizeof(kmp_depnode_list_t));
210#endif
211
212 new_head->node = __kmp_node_ref(node);
213 new_head->next = list;
214
215 return new_head;
216}
217
218static inline void __kmp_track_dependence(kmp_int32 gtid, kmp_depnode_t *source,
219 kmp_depnode_t *sink,
220 kmp_task_t *sink_task) {
221#if OMPX_TASKGRAPH
222 kmp_taskdata_t *task_source = KMP_TASK_TO_TASKDATA(source->dn.task);
223 kmp_taskdata_t *task_sink = KMP_TASK_TO_TASKDATA(sink_task);
224 if (source->dn.task && sink_task) {
225 // Not supporting dependency between two tasks that one is within the TDG
226 // and the other is not
227 KMP_ASSERT(task_source->is_taskgraph == task_sink->is_taskgraph);
228 }
229 if (task_sink->is_taskgraph &&
230 __kmp_tdg_is_recording(task_sink->tdg->tdg_status)) {
231 kmp_node_info_t *source_info =
232 &task_sink->tdg->record_map[task_source->td_task_id];
233 bool exists = false;
234 for (int i = 0; i < source_info->nsuccessors; i++) {
235 if (source_info->successors[i] == task_sink->td_task_id) {
236 exists = true;
237 break;
238 }
239 }
240 if (!exists) {
241 if (source_info->nsuccessors >= source_info->successors_size) {
242 source_info->successors_size = 2 * source_info->successors_size;
243 kmp_int32 *old_succ_ids = source_info->successors;
244 kmp_int32 *new_succ_ids = (kmp_int32 *)__kmp_allocate(
245 source_info->successors_size * sizeof(kmp_int32));
246 source_info->successors = new_succ_ids;
247 __kmp_free(old_succ_ids);
248 }
249
250 source_info->successors[source_info->nsuccessors] = task_sink->td_task_id;
251 source_info->nsuccessors++;
252
253 kmp_node_info_t *sink_info =
254 &(task_sink->tdg->record_map[task_sink->td_task_id]);
255 sink_info->npredecessors++;
256 }
257 }
258#endif
259#ifdef KMP_SUPPORT_GRAPH_OUTPUT
260 kmp_taskdata_t *task_source = KMP_TASK_TO_TASKDATA(source->dn.task);
261 // do not use sink->dn.task as that is only filled after the dependences
262 // are already processed!
263 kmp_taskdata_t *task_sink = KMP_TASK_TO_TASKDATA(sink_task);
264
265 __kmp_printf("%d(%s) -> %d(%s)\n", source->dn.id,
266 task_source->td_ident->psource, sink->dn.id,
267 task_sink->td_ident->psource);
268#endif
269#if OMPT_SUPPORT && OMPT_OPTIONAL
270 /* OMPT tracks dependences between task (a=source, b=sink) in which
271 task a blocks the execution of b through the ompt_new_dependence_callback
272 */
273 if (ompt_enabled.ompt_callback_task_dependence) {
274 kmp_taskdata_t *task_source = KMP_TASK_TO_TASKDATA(source->dn.task);
275 ompt_data_t *sink_data;
276 if (sink_task)
277 sink_data = &(KMP_TASK_TO_TASKDATA(sink_task)->ompt_task_info.task_data);
278 else
279 sink_data = &__kmp_threads[gtid]->th.ompt_thread_info.task_data;
280
281 ompt_callbacks.ompt_callback(ompt_callback_task_dependence)(
282 &(task_source->ompt_task_info.task_data), sink_data);
283 }
284#endif /* OMPT_SUPPORT && OMPT_OPTIONAL */
285}
286
287kmp_base_depnode_t *__kmpc_task_get_depnode(kmp_task_t *task) {
288 kmp_taskdata_t *td = KMP_TASK_TO_TASKDATA(task);
289 return td->td_depnode ? &(td->td_depnode->dn) : NULL;
290}
291
292kmp_depnode_list_t *__kmpc_task_get_successors(kmp_task_t *task) {
293 kmp_taskdata_t *td = KMP_TASK_TO_TASKDATA(task);
294 return td->td_depnode->dn.successors;
295}
296
297static inline kmp_int32
298__kmp_depnode_link_successor(kmp_int32 gtid, kmp_info_t *thread,
299 kmp_task_t *task, kmp_depnode_t *node,
300 kmp_depnode_list_t *plist) {
301 if (!plist)
302 return 0;
303 kmp_int32 npredecessors = 0;
304 // link node as successor of list elements
305 for (kmp_depnode_list_t *p = plist; p; p = p->next) {
306 kmp_depnode_t *dep = p->node;
307#if OMPX_TASKGRAPH
308 kmp_tdg_status tdg_status = KMP_TDG_NONE;
309 if (task) {
310 kmp_taskdata_t *td = KMP_TASK_TO_TASKDATA(task);
311 if (td->is_taskgraph)
312 tdg_status = KMP_TASK_TO_TASKDATA(task)->tdg->tdg_status;
313 if (__kmp_tdg_is_recording(tdg_status))
314 __kmp_track_dependence(gtid, dep, node, task);
315 }
316#endif
317 if (dep->dn.task) {
318 KMP_ACQUIRE_DEPNODE(gtid, dep);
319 if (dep->dn.task) {
320 if (!dep->dn.successors || dep->dn.successors->node != node) {
321#if OMPX_TASKGRAPH
322 if (!(__kmp_tdg_is_recording(tdg_status)) && task)
323#endif
324 __kmp_track_dependence(gtid, source: dep, sink: node, sink_task: task);
325 dep->dn.successors = __kmp_add_node(thread, list: dep->dn.successors, node);
326 KA_TRACE(40, ("__kmp_process_deps: T#%d adding dependence from %p to "
327 "%p\n",
328 gtid, KMP_TASK_TO_TASKDATA(dep->dn.task),
329 KMP_TASK_TO_TASKDATA(task)));
330 npredecessors++;
331 }
332 }
333 KMP_RELEASE_DEPNODE(gtid, dep);
334 }
335 }
336 return npredecessors;
337}
338
339// Add the edge 'sink' -> 'source' in the task dependency graph
340static inline kmp_int32 __kmp_depnode_link_successor(kmp_int32 gtid,
341 kmp_info_t *thread,
342 kmp_task_t *task,
343 kmp_depnode_t *source,
344 kmp_depnode_t *sink) {
345 if (!sink)
346 return 0;
347 kmp_int32 npredecessors = 0;
348#if OMPX_TASKGRAPH
349 kmp_tdg_status tdg_status = KMP_TDG_NONE;
350 kmp_taskdata_t *td = KMP_TASK_TO_TASKDATA(task);
351 if (task) {
352 if (td->is_taskgraph)
353 tdg_status = KMP_TASK_TO_TASKDATA(task)->tdg->tdg_status;
354 if (__kmp_tdg_is_recording(tdg_status) && sink->dn.task)
355 __kmp_track_dependence(gtid, sink, source, task);
356 }
357#endif
358 if (sink->dn.task) {
359 // synchronously add source to sink' list of successors
360 KMP_ACQUIRE_DEPNODE(gtid, sink);
361 if (sink->dn.task) {
362 if (!sink->dn.successors || sink->dn.successors->node != source) {
363#if OMPX_TASKGRAPH
364 if (!(__kmp_tdg_is_recording(tdg_status)) && task)
365#endif
366 __kmp_track_dependence(gtid, source: sink, sink: source, sink_task: task);
367 sink->dn.successors = __kmp_add_node(thread, list: sink->dn.successors, node: source);
368 KA_TRACE(40, ("__kmp_process_deps: T#%d adding dependence from %p to "
369 "%p\n",
370 gtid, KMP_TASK_TO_TASKDATA(sink->dn.task),
371 KMP_TASK_TO_TASKDATA(task)));
372#if OMPX_TASKGRAPH
373 if (__kmp_tdg_is_recording(tdg_status)) {
374 kmp_taskdata_t *tdd = KMP_TASK_TO_TASKDATA(sink->dn.task);
375 if (tdd->is_taskgraph) {
376 if (tdd->td_flags.onced)
377 // decrement npredecessors if sink->dn.task belongs to a taskgraph
378 // and
379 // 1) the task is reset to its initial state (by kmp_free_task) or
380 // 2) the task is complete but not yet reset
381 npredecessors--;
382 }
383 }
384#endif
385 npredecessors++;
386 }
387 }
388 KMP_RELEASE_DEPNODE(gtid, sink);
389 }
390 return npredecessors;
391}
392
393static inline kmp_int32
394__kmp_process_dep_all(kmp_int32 gtid, kmp_depnode_t *node, kmp_dephash_t *h,
395 bool dep_barrier, kmp_task_t *task) {
396 KA_TRACE(30, ("__kmp_process_dep_all: T#%d processing dep_all, "
397 "dep_barrier = %d\n",
398 gtid, dep_barrier));
399 kmp_info_t *thread = __kmp_threads[gtid];
400 kmp_int32 npredecessors = 0;
401
402 // process previous omp_all_memory node if any
403 npredecessors +=
404 __kmp_depnode_link_successor(gtid, thread, task, source: node, sink: h->last_all);
405 __kmp_node_deref(thread, node: h->last_all);
406 if (!dep_barrier) {
407 h->last_all = __kmp_node_ref(node);
408 } else {
409 // if this is a sync point in the serial sequence, then the previous
410 // outputs are guaranteed to be completed after the execution of this
411 // task so the previous output nodes can be cleared.
412 h->last_all = NULL;
413 }
414
415 // process all regular dependences
416 for (size_t i = 0; i < h->size; i++) {
417 kmp_dephash_entry_t *info = h->buckets[i];
418 if (!info) // skip empty slots in dephash
419 continue;
420 for (; info; info = info->next_in_bucket) {
421 // for each entry the omp_all_memory works as OUT dependence
422 kmp_depnode_t *last_out = info->last_out;
423 kmp_depnode_list_t *last_set = info->last_set;
424 kmp_depnode_list_t *prev_set = info->prev_set;
425 if (last_set) {
426 npredecessors +=
427 __kmp_depnode_link_successor(gtid, thread, task, node, plist: last_set);
428 __kmp_depnode_list_free(thread, list: last_set);
429 __kmp_depnode_list_free(thread, list: prev_set);
430 info->last_set = NULL;
431 info->prev_set = NULL;
432 info->last_flag = 0; // no sets in this dephash entry
433 } else {
434 npredecessors +=
435 __kmp_depnode_link_successor(gtid, thread, task, source: node, sink: last_out);
436 }
437 __kmp_node_deref(thread, node: last_out);
438 if (!dep_barrier) {
439 info->last_out = __kmp_node_ref(node);
440 } else {
441 info->last_out = NULL;
442 }
443 }
444 }
445 KA_TRACE(30, ("__kmp_process_dep_all: T#%d found %d predecessors\n", gtid,
446 npredecessors));
447 return npredecessors;
448}
449
450template <bool filter>
451static inline kmp_int32
452__kmp_process_deps(kmp_int32 gtid, kmp_depnode_t *node, kmp_dephash_t **hash,
453 bool dep_barrier, kmp_int32 ndeps,
454 kmp_depend_info_t *dep_list, kmp_task_t *task) {
455 KA_TRACE(30, ("__kmp_process_deps<%d>: T#%d processing %d dependences : "
456 "dep_barrier = %d\n",
457 filter, gtid, ndeps, dep_barrier));
458
459 kmp_info_t *thread = __kmp_threads[gtid];
460 kmp_int32 npredecessors = 0;
461 for (kmp_int32 i = 0; i < ndeps; i++) {
462 const kmp_depend_info_t *dep = &dep_list[i];
463
464 if (filter && dep->base_addr == 0)
465 continue; // skip filtered entries
466
467 kmp_dephash_entry_t *info =
468 __kmp_dephash_find(thread, hash, addr: dep->base_addr);
469 kmp_depnode_t *last_out = info->last_out;
470 kmp_depnode_list_t *last_set = info->last_set;
471 kmp_depnode_list_t *prev_set = info->prev_set;
472
473 if (dep->flags.out) { // out or inout --> clean lists if any
474 if (last_set) {
475 npredecessors +=
476 __kmp_depnode_link_successor(gtid, thread, task, node, plist: last_set);
477 __kmp_depnode_list_free(thread, list: last_set);
478 __kmp_depnode_list_free(thread, list: prev_set);
479 info->last_set = NULL;
480 info->prev_set = NULL;
481 info->last_flag = 0; // no sets in this dephash entry
482 } else {
483 npredecessors +=
484 __kmp_depnode_link_successor(gtid, thread, task, source: node, sink: last_out);
485 }
486 __kmp_node_deref(thread, node: last_out);
487 if (!dep_barrier) {
488 info->last_out = __kmp_node_ref(node);
489 } else {
490 // if this is a sync point in the serial sequence, then the previous
491 // outputs are guaranteed to be completed after the execution of this
492 // task so the previous output nodes can be cleared.
493 info->last_out = NULL;
494 }
495 } else { // either IN or MTX or SET
496 if (info->last_flag == 0 || info->last_flag == dep->flag) {
497 // last_set either didn't exist or of same dep kind
498 // link node as successor of the last_out if any
499 npredecessors +=
500 __kmp_depnode_link_successor(gtid, thread, task, source: node, sink: last_out);
501 // link node as successor of all nodes in the prev_set if any
502 npredecessors +=
503 __kmp_depnode_link_successor(gtid, thread, task, node, plist: prev_set);
504 if (dep_barrier) {
505 // clean last_out and prev_set if any; don't touch last_set
506 __kmp_node_deref(thread, node: last_out);
507 info->last_out = NULL;
508 __kmp_depnode_list_free(thread, list: prev_set);
509 info->prev_set = NULL;
510 }
511 } else { // last_set is of different dep kind, make it prev_set
512 // link node as successor of all nodes in the last_set
513 npredecessors +=
514 __kmp_depnode_link_successor(gtid, thread, task, node, plist: last_set);
515 // clean last_out if any
516 __kmp_node_deref(thread, node: last_out);
517 info->last_out = NULL;
518 // clean prev_set if any
519 __kmp_depnode_list_free(thread, list: prev_set);
520 if (!dep_barrier) {
521 // move last_set to prev_set, new last_set will be allocated
522 info->prev_set = last_set;
523 } else {
524 info->prev_set = NULL;
525 info->last_flag = 0;
526 }
527 info->last_set = NULL;
528 }
529 // for dep_barrier last_flag value should remain:
530 // 0 if last_set is empty, unchanged otherwise
531 if (!dep_barrier) {
532 info->last_flag = dep->flag; // store dep kind of the last_set
533 info->last_set = __kmp_add_node(thread, list: info->last_set, node);
534 }
535 // check if we are processing MTX dependency
536 if (dep->flag == KMP_DEP_MTX) {
537 if (info->mtx_lock == NULL) {
538 info->mtx_lock = (kmp_lock_t *)__kmp_allocate(sizeof(kmp_lock_t));
539 __kmp_init_lock(lck: info->mtx_lock);
540 }
541 KMP_DEBUG_ASSERT(node->dn.mtx_num_locks < MAX_MTX_DEPS);
542 kmp_int32 m;
543 // Save lock in node's array
544 for (m = 0; m < MAX_MTX_DEPS; ++m) {
545 // sort pointers in decreasing order to avoid potential livelock
546 if (node->dn.mtx_locks[m] < info->mtx_lock) {
547 KMP_DEBUG_ASSERT(!node->dn.mtx_locks[node->dn.mtx_num_locks]);
548 for (int n = node->dn.mtx_num_locks; n > m; --n) {
549 // shift right all lesser non-NULL pointers
550 KMP_DEBUG_ASSERT(node->dn.mtx_locks[n - 1] != NULL);
551 node->dn.mtx_locks[n] = node->dn.mtx_locks[n - 1];
552 }
553 node->dn.mtx_locks[m] = info->mtx_lock;
554 break;
555 }
556 }
557 KMP_DEBUG_ASSERT(m < MAX_MTX_DEPS); // must break from loop
558 node->dn.mtx_num_locks++;
559 }
560 }
561 }
562 KA_TRACE(30, ("__kmp_process_deps<%d>: T#%d found %d predecessors\n", filter,
563 gtid, npredecessors));
564 return npredecessors;
565}
566
567#define NO_DEP_BARRIER (false)
568#define DEP_BARRIER (true)
569
570// returns true if the task has any outstanding dependence
571static bool __kmp_check_deps(kmp_int32 gtid, kmp_depnode_t *node,
572 kmp_task_t *task, kmp_dephash_t **hash,
573 bool dep_barrier, kmp_int32 ndeps,
574 kmp_depend_info_t *dep_list,
575 kmp_int32 ndeps_noalias,
576 kmp_depend_info_t *noalias_dep_list) {
577 int i, n_mtxs = 0, dep_all = 0;
578#if KMP_DEBUG
579 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
580#endif
581 KA_TRACE(20, ("__kmp_check_deps: T#%d checking dependences for task %p : %d "
582 "possibly aliased dependences, %d non-aliased dependences : "
583 "dep_barrier=%d .\n",
584 gtid, taskdata, ndeps, ndeps_noalias, dep_barrier));
585
586 // Filter deps in dep_list
587 // TODO: Different algorithm for large dep_list ( > 10 ? )
588 for (i = 0; i < ndeps; i++) {
589 if (dep_list[i].base_addr != 0 &&
590 dep_list[i].base_addr != (kmp_intptr_t)KMP_SIZE_T_MAX) {
591 KMP_DEBUG_ASSERT(
592 dep_list[i].flag == KMP_DEP_IN || dep_list[i].flag == KMP_DEP_OUT ||
593 dep_list[i].flag == KMP_DEP_INOUT ||
594 dep_list[i].flag == KMP_DEP_MTX || dep_list[i].flag == KMP_DEP_SET);
595 for (int j = i + 1; j < ndeps; j++) {
596 if (dep_list[i].base_addr == dep_list[j].base_addr) {
597 if (dep_list[i].flag != dep_list[j].flag) {
598 // two different dependences on same address work identical to OUT
599 dep_list[i].flag = KMP_DEP_OUT;
600 }
601 dep_list[j].base_addr = 0; // Mark j element as void
602 }
603 }
604 if (dep_list[i].flag == KMP_DEP_MTX) {
605 // limit number of mtx deps to MAX_MTX_DEPS per node
606 if (n_mtxs < MAX_MTX_DEPS && task != NULL) {
607 ++n_mtxs;
608 } else {
609 dep_list[i].flag = KMP_DEP_OUT; // downgrade mutexinoutset to inout
610 }
611 }
612 } else if (dep_list[i].flag == KMP_DEP_ALL ||
613 dep_list[i].base_addr == (kmp_intptr_t)KMP_SIZE_T_MAX) {
614 // omp_all_memory dependence can be marked by compiler by either
615 // (addr=0 && flag=0x80) (flag KMP_DEP_ALL), or (addr=-1).
616 // omp_all_memory overrides all other dependences if any
617 dep_all = 1;
618 break;
619 }
620 }
621
622 // doesn't need to be atomic as no other thread is going to be accessing this
623 // node just yet.
624 // npredecessors is set -1 to ensure that none of the releasing tasks queues
625 // this task before we have finished processing all the dependences
626 node->dn.npredecessors = -1;
627
628 // used to pack all npredecessors additions into a single atomic operation at
629 // the end
630 int npredecessors;
631
632 if (!dep_all) { // regular dependences
633 npredecessors = __kmp_process_deps<true>(gtid, node, hash, dep_barrier,
634 ndeps, dep_list, task);
635 npredecessors += __kmp_process_deps<false>(
636 gtid, node, hash, dep_barrier, ndeps: ndeps_noalias, dep_list: noalias_dep_list, task);
637 } else { // omp_all_memory dependence
638 npredecessors = __kmp_process_dep_all(gtid, node, h: *hash, dep_barrier, task);
639 }
640
641 node->dn.task = task;
642 KMP_MB();
643
644 // Account for our initial fake value
645 npredecessors++;
646
647 // Update predecessors and obtain current value to check if there are still
648 // any outstanding dependences (some tasks may have finished while we
649 // processed the dependences)
650 npredecessors =
651 node->dn.npredecessors.fetch_add(i: npredecessors) + npredecessors;
652
653 KA_TRACE(20, ("__kmp_check_deps: T#%d found %d predecessors for task %p \n",
654 gtid, npredecessors, taskdata));
655
656 // beyond this point the task could be queued (and executed) by a releasing
657 // task...
658 return npredecessors > 0 ? true : false;
659}
660
661/*!
662@ingroup TASKING
663@param loc_ref location of the original task directive
664@param gtid Global Thread ID of encountering thread
665@param new_task task thunk allocated by __kmp_omp_task_alloc() for the ''new
666task''
667@param ndeps Number of depend items with possible aliasing
668@param dep_list List of depend items with possible aliasing
669@param ndeps_noalias Number of depend items with no aliasing
670@param noalias_dep_list List of depend items with no aliasing
671
672@return Returns either TASK_CURRENT_NOT_QUEUED if the current task was not
673suspended and queued, or TASK_CURRENT_QUEUED if it was suspended and queued
674
675Schedule a non-thread-switchable task with dependences for execution
676*/
677kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid,
678 kmp_task_t *new_task, kmp_int32 ndeps,
679 kmp_depend_info_t *dep_list,
680 kmp_int32 ndeps_noalias,
681 kmp_depend_info_t *noalias_dep_list) {
682
683 kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
684 KA_TRACE(10, ("__kmpc_omp_task_with_deps(enter): T#%d loc=%p task=%p\n", gtid,
685 loc_ref, new_taskdata));
686 __kmp_assert_valid_gtid(gtid);
687 kmp_info_t *thread = __kmp_threads[gtid];
688 kmp_taskdata_t *current_task = thread->th.th_current_task;
689
690#if OMPX_TASKGRAPH
691 // record TDG with deps
692 if (new_taskdata->is_taskgraph &&
693 __kmp_tdg_is_recording(new_taskdata->tdg->tdg_status)) {
694 kmp_tdg_info_t *tdg = new_taskdata->tdg;
695 // extend record_map if needed
696 if (new_taskdata->td_task_id >= tdg->map_size) {
697 __kmp_acquire_bootstrap_lock(&tdg->graph_lock);
698 if (new_taskdata->td_task_id >= tdg->map_size) {
699 kmp_uint old_size = tdg->map_size;
700 kmp_uint new_size = old_size * 2;
701 kmp_node_info_t *old_record = tdg->record_map;
702 kmp_node_info_t *new_record = (kmp_node_info_t *)__kmp_allocate(
703 new_size * sizeof(kmp_node_info_t));
704 KMP_MEMCPY(new_record, tdg->record_map,
705 old_size * sizeof(kmp_node_info_t));
706 tdg->record_map = new_record;
707
708 __kmp_free(old_record);
709
710 for (kmp_int i = old_size; i < new_size; i++) {
711 kmp_int32 *successorsList = (kmp_int32 *)__kmp_allocate(
712 __kmp_successors_size * sizeof(kmp_int32));
713 new_record[i].task = nullptr;
714 new_record[i].successors = successorsList;
715 new_record[i].nsuccessors = 0;
716 new_record[i].npredecessors = 0;
717 new_record[i].successors_size = __kmp_successors_size;
718 KMP_ATOMIC_ST_REL(&new_record[i].npredecessors_counter, 0);
719 }
720 // update the size at the end, so that we avoid other
721 // threads use old_record while map_size is already updated
722 tdg->map_size = new_size;
723 }
724 __kmp_release_bootstrap_lock(&tdg->graph_lock);
725 }
726 tdg->record_map[new_taskdata->td_task_id].task = new_task;
727 tdg->record_map[new_taskdata->td_task_id].parent_task =
728 new_taskdata->td_parent;
729 KMP_ATOMIC_INC(&tdg->num_tasks);
730 }
731#endif
732#if OMPT_SUPPORT
733 if (ompt_enabled.enabled) {
734 if (!current_task->ompt_task_info.frame.enter_frame.ptr)
735 current_task->ompt_task_info.frame.enter_frame.ptr =
736 OMPT_GET_FRAME_ADDRESS(0);
737 if (ompt_enabled.ompt_callback_task_create) {
738 ompt_callbacks.ompt_callback(ompt_callback_task_create)(
739 &(current_task->ompt_task_info.task_data),
740 &(current_task->ompt_task_info.frame),
741 &(new_taskdata->ompt_task_info.task_data),
742 ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 1,
743 OMPT_LOAD_OR_GET_RETURN_ADDRESS(gtid));
744 }
745
746 new_taskdata->ompt_task_info.frame.enter_frame.ptr =
747 OMPT_GET_FRAME_ADDRESS(0);
748 }
749
750#if OMPT_OPTIONAL
751 /* OMPT grab all dependences if requested by the tool */
752 if (ndeps + ndeps_noalias > 0 && ompt_enabled.ompt_callback_dependences) {
753 kmp_int32 i;
754
755 int ompt_ndeps = ndeps + ndeps_noalias;
756 ompt_dependence_t *ompt_deps = (ompt_dependence_t *)KMP_OMPT_DEPS_ALLOC(
757 thread, (ndeps + ndeps_noalias) * sizeof(ompt_dependence_t));
758
759 KMP_ASSERT(ompt_deps != NULL);
760
761 for (i = 0; i < ndeps; i++) {
762 ompt_deps[i].variable.ptr = (void *)dep_list[i].base_addr;
763 if (dep_list[i].base_addr == KMP_SIZE_T_MAX)
764 ompt_deps[i].dependence_type = ompt_dependence_type_out_all_memory;
765 else if (dep_list[i].flags.in && dep_list[i].flags.out)
766 ompt_deps[i].dependence_type = ompt_dependence_type_inout;
767 else if (dep_list[i].flags.out)
768 ompt_deps[i].dependence_type = ompt_dependence_type_out;
769 else if (dep_list[i].flags.in)
770 ompt_deps[i].dependence_type = ompt_dependence_type_in;
771 else if (dep_list[i].flags.mtx)
772 ompt_deps[i].dependence_type = ompt_dependence_type_mutexinoutset;
773 else if (dep_list[i].flags.set)
774 ompt_deps[i].dependence_type = ompt_dependence_type_inoutset;
775 else if (dep_list[i].flags.all)
776 ompt_deps[i].dependence_type = ompt_dependence_type_out_all_memory;
777 }
778 for (i = 0; i < ndeps_noalias; i++) {
779 ompt_deps[ndeps + i].variable.ptr = (void *)noalias_dep_list[i].base_addr;
780 if (noalias_dep_list[i].base_addr == KMP_SIZE_T_MAX)
781 ompt_deps[ndeps + i].dependence_type =
782 ompt_dependence_type_out_all_memory;
783 else if (noalias_dep_list[i].flags.in && noalias_dep_list[i].flags.out)
784 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inout;
785 else if (noalias_dep_list[i].flags.out)
786 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_out;
787 else if (noalias_dep_list[i].flags.in)
788 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_in;
789 else if (noalias_dep_list[i].flags.mtx)
790 ompt_deps[ndeps + i].dependence_type =
791 ompt_dependence_type_mutexinoutset;
792 else if (noalias_dep_list[i].flags.set)
793 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inoutset;
794 else if (noalias_dep_list[i].flags.all)
795 ompt_deps[ndeps + i].dependence_type =
796 ompt_dependence_type_out_all_memory;
797 }
798 ompt_callbacks.ompt_callback(ompt_callback_dependences)(
799 &(new_taskdata->ompt_task_info.task_data), ompt_deps, ompt_ndeps);
800 /* We can now free the allocated memory for the dependences */
801 /* For OMPD we might want to delay the free until end of this function */
802 KMP_OMPT_DEPS_FREE(thread, ompt_deps);
803 }
804#endif /* OMPT_OPTIONAL */
805#endif /* OMPT_SUPPORT */
806
807 bool serial = current_task->td_flags.team_serial ||
808 current_task->td_flags.tasking_ser ||
809 current_task->td_flags.final;
810 kmp_task_team_t *task_team = thread->th.th_task_team;
811 serial = serial &&
812 !(task_team && (task_team->tt.tt_found_proxy_tasks ||
813 task_team->tt.tt_hidden_helper_task_encountered));
814
815 if (!serial && (ndeps > 0 || ndeps_noalias > 0)) {
816 /* if no dependences have been tracked yet, create the dependence hash */
817 if (current_task->td_dephash == NULL)
818 current_task->td_dephash = __kmp_dephash_create(thread, current_task);
819
820#if USE_FAST_MEMORY
821 kmp_depnode_t *node =
822 (kmp_depnode_t *)__kmp_fast_allocate(thread, sizeof(kmp_depnode_t));
823#else
824 kmp_depnode_t *node =
825 (kmp_depnode_t *)__kmp_thread_malloc(thread, sizeof(kmp_depnode_t));
826#endif
827
828 __kmp_init_node(node);
829 new_taskdata->td_depnode = node;
830
831 if (__kmp_check_deps(gtid, node, task: new_task, hash: &current_task->td_dephash,
832 NO_DEP_BARRIER, ndeps, dep_list, ndeps_noalias,
833 noalias_dep_list)) {
834 KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d task had blocking "
835 "dependences: "
836 "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n",
837 gtid, loc_ref, new_taskdata));
838#if OMPT_SUPPORT
839 if (ompt_enabled.enabled) {
840 current_task->ompt_task_info.frame.enter_frame = ompt_data_none;
841 }
842#endif
843 return TASK_CURRENT_NOT_QUEUED;
844 }
845 } else {
846 KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d ignored dependences "
847 "for task (serialized) loc=%p task=%p\n",
848 gtid, loc_ref, new_taskdata));
849 }
850
851 KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d task had no blocking "
852 "dependences : "
853 "loc=%p task=%p, transferring to __kmp_omp_task\n",
854 gtid, loc_ref, new_taskdata));
855
856 kmp_int32 ret = __kmp_omp_task(gtid, new_task, serialize_immediate: true);
857#if OMPT_SUPPORT
858 if (ompt_enabled.enabled) {
859 current_task->ompt_task_info.frame.enter_frame = ompt_data_none;
860 }
861#endif
862 return ret;
863}
864
865#if OMPT_SUPPORT
866void __ompt_taskwait_dep_finish(kmp_taskdata_t *current_task,
867 ompt_data_t *taskwait_task_data) {
868 if (ompt_enabled.ompt_callback_task_schedule) {
869 ompt_callbacks.ompt_callback(ompt_callback_task_schedule)(
870 taskwait_task_data, ompt_taskwait_complete, NULL);
871 }
872 current_task->ompt_task_info.frame.enter_frame.ptr = NULL;
873 *taskwait_task_data = ompt_data_none;
874}
875#endif /* OMPT_SUPPORT */
876
877/*!
878@ingroup TASKING
879@param loc_ref location of the original task directive
880@param gtid Global Thread ID of encountering thread
881@param ndeps Number of depend items with possible aliasing
882@param dep_list List of depend items with possible aliasing
883@param ndeps_noalias Number of depend items with no aliasing
884@param noalias_dep_list List of depend items with no aliasing
885
886Blocks the current task until all specifies dependences have been fulfilled.
887*/
888void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps,
889 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
890 kmp_depend_info_t *noalias_dep_list) {
891 __kmpc_omp_taskwait_deps_51(loc_ref, gtid, ndeps, dep_list, ndeps_noalias,
892 noalias_dep_list, has_no_wait: false);
893}
894
895/* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause.
896 Placeholder for taskwait with nowait clause.
897 Earlier code of __kmpc_omp_wait_deps() is now
898 in this function.
899*/
900void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid,
901 kmp_int32 ndeps, kmp_depend_info_t *dep_list,
902 kmp_int32 ndeps_noalias,
903 kmp_depend_info_t *noalias_dep_list,
904 kmp_int32 has_no_wait) {
905 KA_TRACE(10, ("__kmpc_omp_taskwait_deps(enter): T#%d loc=%p nowait#%d\n",
906 gtid, loc_ref, has_no_wait));
907 if (ndeps == 0 && ndeps_noalias == 0) {
908 KA_TRACE(10, ("__kmpc_omp_taskwait_deps(exit): T#%d has no dependences to "
909 "wait upon : loc=%p\n",
910 gtid, loc_ref));
911 return;
912 }
913 __kmp_assert_valid_gtid(gtid);
914 kmp_info_t *thread = __kmp_threads[gtid];
915 kmp_taskdata_t *current_task = thread->th.th_current_task;
916
917#if OMPT_SUPPORT
918 // this function represents a taskwait construct with depend clause
919 // We signal 4 events:
920 // - creation of the taskwait task
921 // - dependences of the taskwait task
922 // - schedule and finish of the taskwait task
923 ompt_data_t *taskwait_task_data = &thread->th.ompt_thread_info.task_data;
924 KMP_ASSERT(taskwait_task_data->ptr == NULL);
925 if (ompt_enabled.enabled) {
926 if (!current_task->ompt_task_info.frame.enter_frame.ptr)
927 current_task->ompt_task_info.frame.enter_frame.ptr =
928 OMPT_GET_FRAME_ADDRESS(0);
929 if (ompt_enabled.ompt_callback_task_create) {
930 ompt_callbacks.ompt_callback(ompt_callback_task_create)(
931 &(current_task->ompt_task_info.task_data),
932 &(current_task->ompt_task_info.frame), taskwait_task_data,
933 ompt_task_taskwait | ompt_task_undeferred | ompt_task_mergeable, 1,
934 OMPT_LOAD_OR_GET_RETURN_ADDRESS(gtid));
935 }
936 }
937
938#if OMPT_OPTIONAL
939 /* OMPT grab all dependences if requested by the tool */
940 if (ndeps + ndeps_noalias > 0 && ompt_enabled.ompt_callback_dependences) {
941 kmp_int32 i;
942
943 int ompt_ndeps = ndeps + ndeps_noalias;
944 ompt_dependence_t *ompt_deps = (ompt_dependence_t *)KMP_OMPT_DEPS_ALLOC(
945 thread, (ndeps + ndeps_noalias) * sizeof(ompt_dependence_t));
946
947 KMP_ASSERT(ompt_deps != NULL);
948
949 for (i = 0; i < ndeps; i++) {
950 ompt_deps[i].variable.ptr = (void *)dep_list[i].base_addr;
951 if (dep_list[i].flags.in && dep_list[i].flags.out)
952 ompt_deps[i].dependence_type = ompt_dependence_type_inout;
953 else if (dep_list[i].flags.out)
954 ompt_deps[i].dependence_type = ompt_dependence_type_out;
955 else if (dep_list[i].flags.in)
956 ompt_deps[i].dependence_type = ompt_dependence_type_in;
957 else if (dep_list[i].flags.mtx)
958 ompt_deps[ndeps + i].dependence_type =
959 ompt_dependence_type_mutexinoutset;
960 else if (dep_list[i].flags.set)
961 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inoutset;
962 }
963 for (i = 0; i < ndeps_noalias; i++) {
964 ompt_deps[ndeps + i].variable.ptr = (void *)noalias_dep_list[i].base_addr;
965 if (noalias_dep_list[i].flags.in && noalias_dep_list[i].flags.out)
966 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inout;
967 else if (noalias_dep_list[i].flags.out)
968 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_out;
969 else if (noalias_dep_list[i].flags.in)
970 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_in;
971 else if (noalias_dep_list[i].flags.mtx)
972 ompt_deps[ndeps + i].dependence_type =
973 ompt_dependence_type_mutexinoutset;
974 else if (noalias_dep_list[i].flags.set)
975 ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inoutset;
976 }
977 ompt_callbacks.ompt_callback(ompt_callback_dependences)(
978 taskwait_task_data, ompt_deps, ompt_ndeps);
979 /* We can now free the allocated memory for the dependences */
980 /* For OMPD we might want to delay the free until end of this function */
981 KMP_OMPT_DEPS_FREE(thread, ompt_deps);
982 ompt_deps = NULL;
983 }
984#endif /* OMPT_OPTIONAL */
985#endif /* OMPT_SUPPORT */
986
987 // We can return immediately as:
988 // - dependences are not computed in serial teams (except with proxy tasks)
989 // - if the dephash is not yet created it means we have nothing to wait for
990 bool ignore = current_task->td_flags.team_serial ||
991 current_task->td_flags.tasking_ser ||
992 current_task->td_flags.final;
993 ignore =
994 ignore && thread->th.th_task_team != NULL &&
995 thread->th.th_task_team->tt.tt_found_proxy_tasks == FALSE &&
996 thread->th.th_task_team->tt.tt_hidden_helper_task_encountered == FALSE;
997 ignore = ignore || current_task->td_dephash == NULL;
998
999 if (ignore) {
1000 KA_TRACE(10, ("__kmpc_omp_taskwait_deps(exit): T#%d has no blocking "
1001 "dependences : loc=%p\n",
1002 gtid, loc_ref));
1003#if OMPT_SUPPORT
1004 __ompt_taskwait_dep_finish(current_task, taskwait_task_data);
1005#endif /* OMPT_SUPPORT */
1006 return;
1007 }
1008
1009 kmp_depnode_t node = {.dn_align: 0};
1010 __kmp_init_node(node: &node);
1011
1012 if (!__kmp_check_deps(gtid, node: &node, NULL, hash: &current_task->td_dephash,
1013 DEP_BARRIER, ndeps, dep_list, ndeps_noalias,
1014 noalias_dep_list)) {
1015 KA_TRACE(10, ("__kmpc_omp_taskwait_deps(exit): T#%d has no blocking "
1016 "dependences : loc=%p\n",
1017 gtid, loc_ref));
1018#if OMPT_SUPPORT
1019 __ompt_taskwait_dep_finish(current_task, taskwait_task_data);
1020#endif /* OMPT_SUPPORT */
1021 return;
1022 }
1023
1024 int thread_finished = FALSE;
1025 kmp_flag_32<false, false> flag(
1026 (std::atomic<kmp_uint32> *)&node.dn.npredecessors, 0U);
1027 while (node.dn.npredecessors > 0) {
1028 flag.execute_tasks(this_thr: thread, gtid, FALSE,
1029 thread_finished: &thread_finished USE_ITT_BUILD_ARG(NULL),
1030 is_constrained: __kmp_task_stealing_constraint);
1031 }
1032
1033 // Wait until the last __kmp_release_deps is finished before we free the
1034 // current stack frame holding the "node" variable; once its nrefs count
1035 // reaches 1, we're sure nobody else can try to reference it again.
1036 while (node.dn.nrefs > 1)
1037 KMP_YIELD(TRUE);
1038
1039#if OMPT_SUPPORT
1040 __ompt_taskwait_dep_finish(current_task, taskwait_task_data);
1041#endif /* OMPT_SUPPORT */
1042 KA_TRACE(10, ("__kmpc_omp_taskwait_deps(exit): T#%d finished waiting : loc=%p\
1043 \n",
1044 gtid, loc_ref));
1045}
1046

source code of openmp/runtime/src/kmp_taskdeps.cpp