1 /*
2  * kmp_taskdeps.cpp
3  * $Revision: 42539 $
4  * $Date: 2013-07-17 11:20:01 -0500 (Wed, 17 Jul 2013) $
5  */
6 
7 
8 //===----------------------------------------------------------------------===//
9 //
10 //                     The LLVM Compiler Infrastructure
11 //
12 // This file is dual licensed under the MIT and the University of Illinois Open
13 // Source Licenses. See LICENSE.txt for details.
14 //
15 //===----------------------------------------------------------------------===//
16 
17 
18 //#define KMP_SUPPORT_GRAPH_OUTPUT 1
19 
20 #include "kmp.h"
21 #include "kmp_io.h"
22 
23 #if OMP_40_ENABLED
24 
25 //TODO: Improve memory allocation? keep a list of pre-allocated structures? allocate in blocks? re-use list finished list entries?
26 //TODO: don't use atomic ref counters for stack-allocated nodes.
27 //TODO: find an alternate to atomic refs for heap-allocated nodes?
28 //TODO: Finish graph output support
29 //TODO: kmp_lock_t seems a tad to big (and heavy weight) for this. Check other runtime locks
30 //TODO: Any ITT support needed?
31 
32 #ifdef KMP_SUPPORT_GRAPH_OUTPUT
33 static kmp_int32 kmp_node_id_seed = 0;
34 #endif
35 
36 static void
37 __kmp_init_node ( kmp_depnode_t *node )
38 {
39     node->dn.task = NULL; // set to null initially, it will point to the right task once dependences have been processed
40     node->dn.successors = NULL;
41     __kmp_init_lock(&node->dn.lock);
42     node->dn.nrefs = 1; // init creates the first reference to the node
43 #ifdef KMP_SUPPORT_GRAPH_OUTPUT
44     node->dn.id = KMP_TEST_THEN_INC32(&kmp_node_id_seed);
45 #endif
46 }
47 
48 static inline kmp_depnode_t *
49 __kmp_node_ref ( kmp_depnode_t *node )
50 {
51     KMP_TEST_THEN_INC32(&node->dn.nrefs);
52     return node;
53 }
54 
55 static inline void
56 __kmp_node_deref ( kmp_info_t *thread, kmp_depnode_t *node )
57 {
58     if (!node) return;
59 
60     kmp_int32 n = KMP_TEST_THEN_DEC32(&node->dn.nrefs) - 1;
61     if ( n == 0 ) {
62         KMP_ASSERT(node->dn.nrefs == 0);
63 #if USE_FAST_MEMORY
64         __kmp_fast_free(thread,node);
65 #else
66         __kmp_thread_free(thread,node);
67 #endif
68     }
69 }
70 
71 #define KMP_ACQUIRE_DEPNODE(gtid,n) __kmp_acquire_lock(&(n)->dn.lock,(gtid))
72 #define KMP_RELEASE_DEPNODE(gtid,n) __kmp_release_lock(&(n)->dn.lock,(gtid))
73 
74 static void
75 __kmp_depnode_list_free ( kmp_info_t *thread, kmp_depnode_list *list );
76 
77 static const kmp_int32 kmp_dephash_log2 = 6;
78 static const kmp_int32 kmp_dephash_size = (1 << kmp_dephash_log2);
79 
80 static inline kmp_int32
81 __kmp_dephash_hash ( kmp_intptr_t addr )
82 {
83     //TODO alternate to try: set = (((Addr64)(addrUsefulBits * 9.618)) % m_num_sets );
84     return ((addr >> kmp_dephash_log2) ^ addr) % kmp_dephash_size;
85 }
86 
87 static kmp_dephash_t *
88 __kmp_dephash_create ( kmp_info_t *thread )
89 {
90     kmp_dephash_t *h;
91 
92     kmp_int32 size = kmp_dephash_size * sizeof(kmp_dephash_entry_t) + sizeof(kmp_dephash_t);
93 
94 #if USE_FAST_MEMORY
95     h = (kmp_dephash_t *) __kmp_fast_allocate( thread, size );
96 #else
97     h = (kmp_dephash_t *) __kmp_thread_malloc( thread, size );
98 #endif
99 
100 #ifdef KMP_DEBUG
101     h->nelements = 0;
102 #endif
103     h->buckets = (kmp_dephash_entry **)(h+1);
104 
105     for ( kmp_int32 i = 0; i < kmp_dephash_size; i++ )
106         h->buckets[i] = 0;
107 
108     return h;
109 }
110 
111 static void
112 __kmp_dephash_free ( kmp_info_t *thread, kmp_dephash_t *h )
113 {
114     for ( kmp_int32 i=0; i < kmp_dephash_size; i++ ) {
115         if ( h->buckets[i] ) {
116             kmp_dephash_entry_t *next;
117             for ( kmp_dephash_entry_t *entry = h->buckets[i]; entry; entry = next ) {
118                 next = entry->next_in_bucket;
119                 __kmp_depnode_list_free(thread,entry->last_ins);
120                 __kmp_node_deref(thread,entry->last_out);
121 #if USE_FAST_MEMORY
122                 __kmp_fast_free(thread,entry);
123 #else
124                 __kmp_thread_free(thread,entry);
125 #endif
126             }
127         }
128     }
129 #if USE_FAST_MEMORY
130     __kmp_fast_free(thread,h);
131 #else
132     __kmp_thread_free(thread,h);
133 #endif
134 }
135 
136 static kmp_dephash_entry *
137 __kmp_dephash_find ( kmp_info_t *thread, kmp_dephash_t *h, kmp_intptr_t addr )
138 {
139     kmp_int32 bucket = __kmp_dephash_hash(addr);
140 
141     kmp_dephash_entry_t *entry;
142     for ( entry = h->buckets[bucket]; entry; entry = entry->next_in_bucket )
143         if ( entry->addr == addr ) break;
144 
145     if ( entry == NULL ) {
146         // create entry. This is only done by one thread so no locking required
147 #if USE_FAST_MEMORY
148         entry = (kmp_dephash_entry_t *) __kmp_fast_allocate( thread, sizeof(kmp_dephash_entry_t) );
149 #else
150         entry = (kmp_dephash_entry_t *) __kmp_thread_malloc( thread, sizeof(kmp_dephash_entry_t) );
151 #endif
152         entry->addr = addr;
153         entry->last_out = NULL;
154         entry->last_ins = NULL;
155         entry->next_in_bucket = h->buckets[bucket];
156         h->buckets[bucket] = entry;
157 #ifdef KMP_DEBUG
158         h->nelements++;
159         if ( entry->next_in_bucket ) h->nconflicts++;
160 #endif
161     }
162     return entry;
163 }
164 
165 static kmp_depnode_list_t *
166 __kmp_add_node ( kmp_info_t *thread, kmp_depnode_list_t *list, kmp_depnode_t *node )
167 {
168     kmp_depnode_list_t *new_head;
169 
170 #if USE_FAST_MEMORY
171     new_head = (kmp_depnode_list_t *) __kmp_fast_allocate(thread,sizeof(kmp_depnode_list_t));
172 #else
173     new_head = (kmp_depnode_list_t *) __kmp_thread_malloc(thread,sizeof(kmp_depnode_list_t));
174 #endif
175 
176     new_head->node = __kmp_node_ref(node);
177     new_head->next = list;
178 
179     return new_head;
180 }
181 
182 static void
183 __kmp_depnode_list_free ( kmp_info_t *thread, kmp_depnode_list *list )
184 {
185     kmp_depnode_list *next;
186 
187     for ( ; list ; list = next ) {
188         next = list->next;
189 
190         __kmp_node_deref(thread,list->node);
191 #if USE_FAST_MEMORY
192         __kmp_fast_free(thread,list);
193 #else
194         __kmp_thread_free(thread,list);
195 #endif
196     }
197 }
198 
199 static inline void
200 __kmp_track_dependence ( kmp_depnode_t *source, kmp_depnode_t *sink )
201 {
202 #ifdef KMP_SUPPORT_GRAPH_OUTPUT
203     kmp_taskdata_t * task_source = KMP_TASK_TO_TASKDATA(source->dn.task);
204     kmp_taskdata_t * task_sink = KMP_TASK_TO_TASKDATA(sink->dn.task);    // this can be NULL when if(0) ...
205 
206     __kmp_printf("%d(%s) -> %d(%s)\n", source->dn.id, task_source->td_ident->psource, sink->dn.id, task_sink->td_ident->psource);
207 #endif
208 }
209 
210 template< bool filter >
211 static inline kmp_int32
212 __kmp_process_deps ( kmp_int32 gtid, kmp_depnode_t *node, kmp_dephash_t *hash,
213                      bool dep_barrier,kmp_int32 ndeps, kmp_depend_info_t *dep_list)
214 {
215     kmp_info_t *thread = __kmp_threads[ gtid ];
216     kmp_int32 npredecessors=0;
217     for ( kmp_int32 i = 0; i < ndeps ; i++ ) {
218         const kmp_depend_info_t * dep = &dep_list[i];
219 
220         KMP_DEBUG_ASSERT(dep->flags.in);
221 
222         if ( filter && dep->base_addr == 0 ) continue; // skip filtered entries
223 
224         kmp_dephash_entry_t *info = __kmp_dephash_find(thread,hash,dep->base_addr);
225         kmp_depnode_t *last_out = info->last_out;
226 
227         if ( dep->flags.out && info->last_ins ) {
228             for ( kmp_depnode_list_t * p = info->last_ins; p; p = p->next ) {
229                 kmp_depnode_t * indep = p->node;
230                 if ( indep->dn.task ) {
231                     KMP_ACQUIRE_DEPNODE(gtid,indep);
232                     if ( indep->dn.task ) {
233                         __kmp_track_dependence(indep,node);
234                         indep->dn.successors = __kmp_add_node(thread, indep->dn.successors, node);
235                         npredecessors++;
236                     }
237                     KMP_RELEASE_DEPNODE(gtid,indep);
238                 }
239             }
240 
241             __kmp_depnode_list_free(thread,info->last_ins);
242             info->last_ins = NULL;
243 
244         } else if ( last_out && last_out->dn.task ) {
245             KMP_ACQUIRE_DEPNODE(gtid,last_out);
246             if ( last_out->dn.task ) {
247                 __kmp_track_dependence(last_out,node);
248                 last_out->dn.successors = __kmp_add_node(thread, last_out->dn.successors, node);
249                 npredecessors++;
250             }
251             KMP_RELEASE_DEPNODE(gtid,last_out);
252         }
253 
254         if ( dep_barrier ) {
255             // if this is a sync point in the serial sequence and previous outputs are guaranteed to be completed after
256             // the execution of this task so the previous output nodes can be cleared.
257             __kmp_node_deref(thread,last_out);
258             info->last_out = NULL;
259         } else {
260             if ( dep->flags.out ) {
261                 __kmp_node_deref(thread,last_out);
262                 info->last_out = __kmp_node_ref(node);
263             } else
264                 info->last_ins = __kmp_add_node(thread, info->last_ins, node);
265         }
266 
267     }
268     return npredecessors;
269 }
270 
271 #define NO_DEP_BARRIER (false)
272 #define DEP_BARRIER (true)
273 
274 // returns true if the task has any outstanding dependence
275 static bool
276 __kmp_check_deps ( kmp_int32 gtid, kmp_depnode_t *node, kmp_task_t *task, kmp_dephash_t *hash, bool dep_barrier,
277                    kmp_int32 ndeps, kmp_depend_info_t *dep_list,
278                    kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list )
279 {
280     int i;
281 
282     // Filter deps in dep_list
283     // TODO: Different algorithm for large dep_list ( > 10 ? )
284     for ( i = 0; i < ndeps; i ++ ) {
285         if ( dep_list[i].base_addr != 0 )
286             for ( int j = i+1; j < ndeps; j++ )
287                 if ( dep_list[i].base_addr == dep_list[j].base_addr ) {
288                     dep_list[i].flags.in |= dep_list[j].flags.in;
289                     dep_list[i].flags.out |= dep_list[j].flags.out;
290                     dep_list[j].base_addr = 0; // Mark j element as void
291                 }
292     }
293 
294     // doesn't need to be atomic as no other thread is going to be accessing this node just yet
295     // npredecessors is set 1 to ensure that none of the releasing tasks queues this task before we have finished processing all the dependencies
296     node->dn.npredecessors = 1;
297 
298     // used to pack all npredecessors additions into a single atomic operation at the end
299     int npredecessors;
300 
301     npredecessors = __kmp_process_deps<true>(gtid, node, hash, dep_barrier, ndeps, dep_list);
302     npredecessors += __kmp_process_deps<false>(gtid, node, hash, dep_barrier, ndeps_noalias, noalias_dep_list);
303 
304     KMP_TEST_THEN_ADD32(&node->dn.npredecessors, npredecessors);
305 
306     // Remove the fake predecessor and find out if there's any outstanding dependence (some tasks may have finished while we processed the dependences)
307     node->dn.task = task;
308     KMP_MB();
309     npredecessors = KMP_TEST_THEN_DEC32(&node->dn.npredecessors) - 1;
310 
311     // beyond this point the task could be queued (and executed) by a releasing task...
312     return npredecessors > 0 ? true : false;
313 }
314 
315 void
316 __kmp_release_deps ( kmp_int32 gtid, kmp_taskdata_t *task )
317 {
318     kmp_info_t *thread = __kmp_threads[ gtid ];
319     kmp_depnode_t *node = task->td_depnode;
320 
321     if ( task->td_dephash )
322         __kmp_dephash_free(thread,task->td_dephash);
323 
324     if ( !node ) return;
325 
326     KMP_ACQUIRE_DEPNODE(gtid,node);
327     node->dn.task = NULL; // mark this task as finished, so no new dependencies are generated
328     KMP_RELEASE_DEPNODE(gtid,node);
329 
330     kmp_depnode_list_t *next;
331     for ( kmp_depnode_list_t *p = node->dn.successors; p; p = next ) {
332         kmp_depnode_t *successor = p->node;
333         kmp_int32 npredecessors = KMP_TEST_THEN_DEC32(&successor->dn.npredecessors) - 1;
334 
335         // successor task can be NULL for wait_depends or because deps are still being processed
336         if ( npredecessors == 0 ) {
337             KMP_MB();
338             if ( successor->dn.task )
339             // loc_ref was already stored in successor's task_data
340                 __kmpc_omp_task(NULL,gtid,successor->dn.task);
341         }
342 
343         next = p->next;
344         __kmp_node_deref(thread,p->node);
345 #if USE_FAST_MEMORY
346         __kmp_fast_free(thread,p);
347 #else
348         __kmp_thread_free(thread,p);
349 #endif
350     }
351 
352     __kmp_node_deref(thread,node);
353 }
354 
355 /*!
356 @ingroup TASKING
357 @param loc_ref location of the original task directive
358 @param gtid Global Thread ID of encountering thread
359 @param new_task task thunk allocated by __kmp_omp_task_alloc() for the ''new task''
360 @param ndeps Number of depend items with possible aliasing
361 @param dep_list List of depend items with possible aliasing
362 @param ndeps_noalias Number of depend items with no aliasing
363 @param noalias_dep_list List of depend items with no aliasing
364 
365 @return Returns either TASK_CURRENT_NOT_QUEUED if the current task was not suspendend and queued, or TASK_CURRENT_QUEUED if it was suspended and queued
366 
367 Schedule a non-thread-switchable task with dependences for execution
368 */
369 kmp_int32
370 __kmpc_omp_task_with_deps( ident_t *loc_ref, kmp_int32 gtid, kmp_task_t * new_task,
371                                  kmp_int32 ndeps, kmp_depend_info_t *dep_list,
372 				 kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list )
373 {
374     kmp_info_t *thread = __kmp_threads[ gtid ];
375     kmp_taskdata_t * current_task = thread->th.th_current_task;
376 
377     bool serial = current_task->td_flags.team_serial || current_task->td_flags.tasking_ser || current_task->td_flags.final;
378 
379     if ( !serial && ( ndeps > 0 || ndeps_noalias > 0 )) {
380         /* if no dependencies have been tracked yet, create the dependence hash */
381         if ( current_task->td_dephash == NULL )
382             current_task->td_dephash = __kmp_dephash_create(thread);
383 
384 #if USE_FAST_MEMORY
385         kmp_depnode_t *node = (kmp_depnode_t *) __kmp_fast_allocate(thread,sizeof(kmp_depnode_t));
386 #else
387         kmp_depnode_t *node = (kmp_depnode_t *) __kmp_thread_malloc(thread,sizeof(kmp_depnode_t));
388 #endif
389 
390         __kmp_init_node(node);
391         KMP_TASK_TO_TASKDATA(new_task)->td_depnode = node;
392 
393         if ( __kmp_check_deps( gtid, node, new_task, current_task->td_dephash, NO_DEP_BARRIER,
394                                ndeps, dep_list, ndeps_noalias,noalias_dep_list ) )
395             return TASK_CURRENT_NOT_QUEUED;
396     }
397 
398     return __kmpc_omp_task(loc_ref,gtid,new_task);
399 }
400 
401 /*!
402 @ingroup TASKING
403 @param loc_ref location of the original task directive
404 @param gtid Global Thread ID of encountering thread
405 @param ndeps Number of depend items with possible aliasing
406 @param dep_list List of depend items with possible aliasing
407 @param ndeps_noalias Number of depend items with no aliasing
408 @param noalias_dep_list List of depend items with no aliasing
409 
410 Blocks the current task until all specifies dependencies have been fulfilled.
411 */
412 void
413 __kmpc_omp_wait_deps ( ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list,
414                        kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list )
415 {
416     if ( ndeps == 0 && ndeps_noalias == 0 ) return;
417 
418     kmp_info_t *thread = __kmp_threads[ gtid ];
419     kmp_taskdata_t * current_task = thread->th.th_current_task;
420 
421     // dependences are not computed in serial teams
422     if ( current_task->td_flags.team_serial || current_task->td_flags.tasking_ser || current_task->td_flags.final)
423         return;
424 
425     // if the dephash is not yet created it means we have nothing to wait for
426     if ( current_task->td_dephash == NULL ) return;
427 
428     kmp_depnode_t node;
429     __kmp_init_node(&node);
430 
431     if (!__kmp_check_deps( gtid, &node, NULL, current_task->td_dephash, DEP_BARRIER,
432                            ndeps, dep_list, ndeps_noalias, noalias_dep_list ))
433         return;
434 
435     int thread_finished = FALSE;
436     while ( node.dn.npredecessors > 0 ) {
437         __kmp_execute_tasks( thread, gtid, (volatile kmp_uint32 *)&(node.dn.npredecessors),
438                              0, FALSE, &thread_finished,
439 #if USE_ITT_BUILD
440                              NULL,
441 #endif
442                              __kmp_task_stealing_constraint );
443     }
444 
445 }
446 
447 #endif /* OMP_40_ENABLED */
448 
449