Intel(R) Threading Building Blocks Doxygen Documentation  version 4.2.3
tbb::internal::custom_scheduler< SchedulerTraits > Class Template Reference

A scheduler with a customized evaluation loop. More...

#include <custom_scheduler.h>

Inheritance diagram for tbb::internal::custom_scheduler< SchedulerTraits >:
Collaboration diagram for tbb::internal::custom_scheduler< SchedulerTraits >:

Public Member Functions

taskreceive_or_steal_task (__TBB_ISOLATION_ARG(__TBB_atomic reference_count &completion_ref_count, isolation_tag isolation)) __TBB_override
 Try getting a task from the mailbox or stealing from another scheduler. More...
 

Static Public Member Functions

static generic_schedulerallocate_scheduler (market &m)
 

Private Types

typedef custom_scheduler< SchedulerTraits > scheduler_type
 

Private Member Functions

 custom_scheduler (market &m)
 
void local_wait_for_all (task &parent, task *child) __TBB_override
 Scheduler loop that dispatches tasks. More...
 
void wait_for_all (task &parent, task *child) __TBB_override
 Entry point from client code to the scheduler loop that dispatches tasks. More...
 
void tally_completion_of_predecessor (task &s, __TBB_ISOLATION_ARG(task *&bypass_slot, isolation_tag isolation))
 Decrements ref_count of a predecessor. More...
 
- Private Member Functions inherited from tbb::internal::generic_scheduler
bool is_task_pool_published () const
 
bool is_local_task_pool_quiescent () const
 
bool is_quiescent_local_task_pool_empty () const
 
bool is_quiescent_local_task_pool_reset () const
 
void attach_mailbox (affinity_id id)
 
void init_stack_info ()
 Sets up the data necessary for the stealing limiting heuristics. More...
 
bool can_steal ()
 Returns true if stealing is allowed. More...
 
void publish_task_pool ()
 Used by workers to enter the task pool. More...
 
void leave_task_pool ()
 Leave the task pool. More...
 
void reset_task_pool_and_leave ()
 Resets head and tail indices to 0, and leaves task pool. More...
 
task ** lock_task_pool (arena_slot *victim_arena_slot) const
 Locks victim's task pool, and returns pointer to it. The pointer can be NULL. More...
 
void unlock_task_pool (arena_slot *victim_arena_slot, task **victim_task_pool) const
 Unlocks victim's task pool. More...
 
void acquire_task_pool () const
 Locks the local task pool. More...
 
void release_task_pool () const
 Unlocks the local task pool. More...
 
taskprepare_for_spawning (task *t)
 Checks if t is affinitized to another thread, and if so, bundles it as proxy. More...
 
void commit_spawned_tasks (size_t new_tail)
 Makes newly spawned tasks visible to thieves. More...
 
void commit_relocated_tasks (size_t new_tail)
 Makes relocated tasks visible to thieves and releases the local task pool. More...
 
taskget_task (__TBB_ISOLATION_EXPR(isolation_tag isolation))
 Get a task from the local pool. More...
 
taskget_task (size_t T)
 Get a task from the local pool at specified location T. More...
 
taskget_mailbox_task (__TBB_ISOLATION_EXPR(isolation_tag isolation))
 Attempt to get a task from the mailbox. More...
 
tasksteal_task (__TBB_ISOLATION_EXPR(isolation_tag isolation))
 Attempts to steal a task from a randomly chosen thread/scheduler. More...
 
tasksteal_task_from (__TBB_ISOLATION_ARG(arena_slot &victim_arena_slot, isolation_tag isolation))
 Steal task from another scheduler's ready pool. More...
 
size_t prepare_task_pool (size_t n)
 Makes sure that the task pool can accommodate at least n more elements. More...
 
bool cleanup_master (bool blocking_terminate)
 Perform necessary cleanup when a master thread stops using TBB. More...
 
void assert_task_pool_valid () const
 
void attach_arena (arena *, size_t index, bool is_master)
 
void nested_arena_entry (arena *, size_t)
 
void nested_arena_exit ()
 
void wait_until_empty ()
 
void spawn (task &first, task *&next) __TBB_override
 For internal use only. More...
 
void spawn_root_and_wait (task &first, task *&next) __TBB_override
 For internal use only. More...
 
void enqueue (task &, void *reserved) __TBB_override
 For internal use only. More...
 
void local_spawn (task *first, task *&next)
 
void local_spawn_root_and_wait (task *first, task *&next)
 
void free_scheduler ()
 Destroy and deallocate this scheduler object. More...
 
taskallocate_task (size_t number_of_bytes, __TBB_CONTEXT_ARG(task *parent, task_group_context *context))
 Allocate task object, either from the heap or a free list. More...
 
template<free_task_hint h>
void free_task (task &t)
 Put task on free list. More...
 
void deallocate_task (task &t)
 Return task object to the memory allocator. More...
 
bool is_worker () const
 True if running on a worker thread, false otherwise. More...
 
bool outermost_level () const
 True if the scheduler is on the outermost dispatch level. More...
 
bool master_outermost_level () const
 True if the scheduler is on the outermost dispatch level in a master thread. More...
 
bool worker_outermost_level () const
 True if the scheduler is on the outermost dispatch level in a worker thread. More...
 
unsigned max_threads_in_arena ()
 Returns the concurrency limit of the current arena. More...
 
void free_nonlocal_small_task (task &t)
 Free a small task t that that was allocated by a different scheduler. More...
 
 generic_scheduler (market &)
 
- Private Member Functions inherited from tbb::internal::scheduler
virtual ~scheduler ()=0
 Pure virtual destructor;. More...
 

Additional Inherited Members

- Static Private Member Functions inherited from tbb::internal::generic_scheduler
static bool is_version_3_task (task &t)
 
static bool is_proxy (const task &t)
 True if t is a task_proxy. More...
 
static generic_schedulercreate_master (arena *a)
 Initialize a scheduler for a master thread. More...
 
static generic_schedulercreate_worker (market &m, size_t index)
 Initialize a scheduler for a worker thread. More...
 
static void cleanup_worker (void *arg, bool worker)
 Perform necessary cleanup when a worker thread finishes. More...
 
static taskplugged_return_list ()
 Special value used to mark my_return_list as not taking any more entries. More...
 
- Private Attributes inherited from tbb::internal::generic_scheduler
uintptr_t my_stealing_threshold
 Position in the call stack specifying its maximal filling when stealing is still allowed. More...
 
marketmy_market
 The market I am in. More...
 
FastRandom my_random
 Random number generator used for picking a random victim from which to steal. More...
 
taskmy_free_list
 Free list of small tasks that can be reused. More...
 
taskmy_dummy_task
 Fake root task created by slave threads. More...
 
long my_ref_count
 Reference count for scheduler. More...
 
bool my_auto_initialized
 True if *this was created by automatic TBB initialization. More...
 
__TBB_atomic intptr_t my_small_task_count
 Number of small tasks that have been allocated by this scheduler. More...
 
taskmy_return_list
 List of small tasks that have been returned to this scheduler by other schedulers. More...
 
- Private Attributes inherited from tbb::internal::intrusive_list_node
intrusive_list_nodemy_prev_node
 
intrusive_list_nodemy_next_node
 
- Private Attributes inherited from tbb::internal::scheduler_state
size_t my_arena_index
 Index of the arena slot the scheduler occupies now, or occupied last time. More...
 
arena_slotmy_arena_slot
 Pointer to the slot in the arena we own at the moment. More...
 
arenamy_arena
 The arena that I own (if master) or am servicing at the moment (if worker) More...
 
taskmy_innermost_running_task
 Innermost task whose task::execute() is running. A dummy task on the outermost level. More...
 
mail_inbox my_inbox
 
affinity_id my_affinity_id
 The mailbox id assigned to this scheduler. More...
 
scheduler_properties my_properties
 
- Static Private Attributes inherited from tbb::internal::generic_scheduler
static const size_t quick_task_size = 256-task_prefix_reservation_size
 If sizeof(task) is <=quick_task_size, it is handled on a free list instead of malloc'd. More...
 
static const size_t null_arena_index = ~size_t(0)
 
static const size_t min_task_pool_size = 64
 

Detailed Description

template<typename SchedulerTraits>
class tbb::internal::custom_scheduler< SchedulerTraits >

A scheduler with a customized evaluation loop.

The customization can use SchedulerTraits to make decisions without needing a run-time check.

Definition at line 52 of file custom_scheduler.h.

Member Typedef Documentation

◆ scheduler_type

template<typename SchedulerTraits >
typedef custom_scheduler<SchedulerTraits> tbb::internal::custom_scheduler< SchedulerTraits >::scheduler_type
private

Definition at line 53 of file custom_scheduler.h.

Constructor & Destructor Documentation

◆ custom_scheduler()

template<typename SchedulerTraits >
tbb::internal::custom_scheduler< SchedulerTraits >::custom_scheduler ( market m)
inlineprivate

Definition at line 55 of file custom_scheduler.h.

Member Function Documentation

◆ allocate_scheduler()

template<typename SchedulerTraits >
static generic_scheduler* tbb::internal::custom_scheduler< SchedulerTraits >::allocate_scheduler ( market m)
inlinestatic

Definition at line 119 of file custom_scheduler.h.

119  {
120  void* p = NFS_Allocate(1, sizeof(scheduler_type), NULL);
121  std::memset(p, 0, sizeof(scheduler_type));
122  scheduler_type* s = new( p ) scheduler_type( m );
123  s->assert_task_pool_valid();
124  ITT_SYNC_CREATE(s, SyncType_Scheduler, SyncObj_TaskPoolSpinning);
125  return s;
126  }
#define ITT_SYNC_CREATE(obj, type, name)
Definition: itt_notify.h:119
void const char const char int ITT_FORMAT __itt_group_sync s
custom_scheduler< SchedulerTraits > scheduler_type
void const char const char int ITT_FORMAT __itt_group_sync p
void *__TBB_EXPORTED_FUNC NFS_Allocate(size_t n_element, size_t element_size, void *hint)
Allocate memory on cache/sector line boundary.

References ITT_SYNC_CREATE, tbb::internal::NFS_Allocate(), p, and s.

Here is the call graph for this function:

◆ local_wait_for_all()

template<typename SchedulerTraits >
void tbb::internal::custom_scheduler< SchedulerTraits >::local_wait_for_all ( task parent,
task child 
)
privatevirtual

Scheduler loop that dispatches tasks.

If child is non-NULL, it is dispatched first. Then, until "parent" has a reference count of 1, other task are dispatched or stolen.

Implements tbb::internal::generic_scheduler.

Definition at line 356 of file custom_scheduler.h.

356  {
357  __TBB_ASSERT( governor::is_set(this), NULL );
358  __TBB_ASSERT( parent.ref_count() >= (child && child->parent() == &parent ? 2 : 1), "ref_count is too small" );
361  // Using parent's refcount in sync_prepare (in the stealing loop below) is
362  // a workaround for TP. We need to name it here to display correctly in Ampl.
363  if( SchedulerTraits::itt_possible )
364  ITT_SYNC_CREATE(&parent.prefix().ref_count, SyncType_Scheduler, SyncObj_TaskStealingLoop);
365 #if __TBB_TASK_GROUP_CONTEXT
366  __TBB_ASSERT( parent.prefix().context, "parent task does not have context" );
367 #endif /* __TBB_TASK_GROUP_CONTEXT */
368  task* t = child;
369  // Constant all_local_work_done is an unreachable refcount value that prevents
370  // early quitting the dispatch loop. It is defined to be in the middle of the range
371  // of negative values representable by the reference_count type.
372  static const reference_count
373  // For normal dispatch loops
374  parents_work_done = 1,
375  // For termination dispatch loops in masters
376  all_local_work_done = (reference_count)3 << (sizeof(reference_count) * 8 - 2);
377  reference_count quit_point;
378 #if __TBB_TASK_PRIORITY
379  __TBB_ASSERT( (uintptr_t)*my_ref_top_priority < (uintptr_t)num_priority_levels, NULL );
380  volatile intptr_t *old_ref_top_priority = my_ref_top_priority;
381  // When entering nested parallelism level market level counter
382  // must be replaced with the one local to this arena.
383  volatile uintptr_t *old_ref_reload_epoch = my_ref_reload_epoch;
384 #endif /* __TBB_TASK_PRIORITY */
385  task* old_innermost_running_task = my_innermost_running_task;
386  scheduler_properties old_properties = my_properties;
387  // Remove outermost property to indicate nested level.
388  __TBB_ASSERT( my_properties.outermost || my_innermost_running_task!=my_dummy_task, "The outermost property should be set out of a dispatch loop" );
390 #if __TBB_TASK_ISOLATION
391  isolation_tag isolation = my_innermost_running_task->prefix().isolation;
392 #endif /* __TBB_TASK_ISOLATION */
393  if( master_outermost_level() ) {
394  // We are in the outermost task dispatch loop of a master thread or a worker which mimics master
395  quit_point = &parent == my_dummy_task ? all_local_work_done : parents_work_done;
396  } else {
397  quit_point = parents_work_done;
398 #if __TBB_TASK_PRIORITY
399  if ( &parent != my_dummy_task ) {
400  // We are in a nested dispatch loop.
401  // Market or arena priority must not prevent child tasks from being
402  // executed so that dynamic priority changes did not cause deadlock.
403  my_ref_top_priority = &parent.prefix().context->my_priority;
404  my_ref_reload_epoch = &my_arena->my_reload_epoch;
405  if(my_ref_reload_epoch != old_ref_reload_epoch)
406  my_local_reload_epoch = *my_ref_reload_epoch-1;
407  }
408 #endif /* __TBB_TASK_PRIORITY */
409  }
410 
411  context_guard_helper</*report_tasks=*/SchedulerTraits::itt_possible> context_guard;
412  if ( t ) {
413  context_guard.set_ctx( __TBB_CONTEXT_ARG1(t->prefix().context) );
414 #if __TBB_TASK_ISOLATION
415  if ( isolation != no_isolation ) {
416  __TBB_ASSERT( t->prefix().isolation == no_isolation, NULL );
417  // Propagate the isolation to the task executed without spawn.
418  t->prefix().isolation = isolation;
419  }
420 #endif /* __TBB_TASK_ISOLATION */
421  }
422 #if TBB_USE_EXCEPTIONS
423  // Infinite safeguard EH loop
424  for (;;) {
425  try {
426 #endif /* TBB_USE_EXCEPTIONS */
427  // Outer loop receives tasks from global environment (via mailbox, FIFO queue(s),
428  // and by stealing from other threads' task pools).
429  // All exit points from the dispatch loop are located in its immediate scope.
430  for(;;) {
431  // Middle loop retrieves tasks from the local task pool.
432  for(;;) {
433  // Inner loop evaluates tasks coming from nesting loops and those returned
434  // by just executed tasks (bypassing spawn or enqueue calls).
435  while(t) {
436  __TBB_ASSERT( my_inbox.is_idle_state(false), NULL );
437  __TBB_ASSERT(!is_proxy(*t),"unexpected proxy");
438  __TBB_ASSERT( t->prefix().owner, NULL );
439 #if __TBB_TASK_ISOLATION
440  __TBB_ASSERT( isolation == no_isolation || isolation == t->prefix().isolation,
441  "A task from another isolated region is going to be executed" );
442 #endif /* __TBB_TASK_ISOLATION */
444 #if __TBB_TASK_GROUP_CONTEXT && TBB_USE_ASSERT
445  assert_context_valid(t->prefix().context);
446  if ( !t->prefix().context->my_cancellation_requested )
447 #endif
448  // TODO: make the assert stronger by prohibiting allocated state.
449  __TBB_ASSERT( 1L<<t->state() & (1L<<task::allocated|1L<<task::ready|1L<<task::reexecute), NULL );
451 #if __TBB_PREVIEW_CRITICAL_TASKS
452  // TODO: check performance and optimize if needed for added conditions on the
453  // hot-path.
454  if( !internal::is_critical(*t) ) {
455  if( task* critical_task = get_critical_task( __TBB_ISOLATION_EXPR(isolation) ) ) {
456  __TBB_ASSERT( internal::is_critical(*critical_task),
457  "Received task must be critical one" );
458  ITT_NOTIFY(sync_acquired, &my_arena->my_critical_task_stream);
459  t->prefix().state = task::allocated;
460  my_innermost_running_task = t; // required during spawn to propagate isolation
461  local_spawn(t, t->prefix().next);
462  t = critical_task;
463  } else {
464 #endif /* __TBB_PREVIEW_CRITICAL_TASKS */
465 #if __TBB_TASK_PRIORITY
466  intptr_t p = priority(*t);
467  if ( p != *my_ref_top_priority
468  && (t->prefix().extra_state & es_task_enqueued) == 0 ) {
469  assert_priority_valid(p);
470  if ( p != my_arena->my_top_priority ) {
471  my_market->update_arena_priority( *my_arena, p );
472  }
473  if ( p < effective_reference_priority() ) {
474  if ( !my_offloaded_tasks ) {
475  my_offloaded_task_list_tail_link = &t->prefix().next_offloaded;
476  // Erase possible reference to the owner scheduler
477  // (next_offloaded is a union member)
478  *my_offloaded_task_list_tail_link = NULL;
479  }
480  offload_task( *t, p );
481  if ( is_task_pool_published() ) {
482  t = winnow_task_pool( __TBB_ISOLATION_EXPR( isolation ) );
483  if ( t )
484  continue;
485  } else {
486  // Mark arena as full to unlock arena priority level adjustment
487  // by arena::is_out_of_work(), and ensure worker's presence.
489  }
490  goto stealing_ground;
491  }
492  }
493 #endif /* __TBB_TASK_PRIORITY */
494 #if __TBB_PREVIEW_CRITICAL_TASKS
495  }
496  } // if is not critical
497 #endif
498  task* t_next = NULL;
500  t->prefix().owner = this;
501  t->prefix().state = task::executing;
502 #if __TBB_TASK_GROUP_CONTEXT
503  if ( !t->prefix().context->my_cancellation_requested )
504 #endif
505  {
506  GATHER_STATISTIC( ++my_counters.tasks_executed );
507  GATHER_STATISTIC( my_counters.avg_arena_concurrency += my_arena->num_workers_active() );
508  GATHER_STATISTIC( my_counters.avg_assigned_workers += my_arena->my_num_workers_allotted );
509 #if __TBB_TASK_PRIORITY
510  GATHER_STATISTIC( my_counters.avg_arena_prio += p );
511  GATHER_STATISTIC( my_counters.avg_market_prio += my_market->my_global_top_priority );
512 #endif /* __TBB_TASK_PRIORITY */
513  ITT_STACK(SchedulerTraits::itt_possible, callee_enter, t->prefix().context->itt_caller);
514 #if __TBB_PREVIEW_CRITICAL_TASKS
515  internal::critical_task_count_guard tc_guard(my_properties, *t);
516 #endif
517  t_next = t->execute();
518  ITT_STACK(SchedulerTraits::itt_possible, callee_leave, t->prefix().context->itt_caller);
519  if (t_next) {
520  __TBB_ASSERT( t_next->state()==task::allocated,
521  "if task::execute() returns task, it must be marked as allocated" );
522  reset_extra_state(t_next);
523  __TBB_ISOLATION_EXPR( t_next->prefix().isolation = t->prefix().isolation );
524 #if TBB_USE_ASSERT
525  affinity_id next_affinity=t_next->prefix().affinity;
526  if (next_affinity != 0 && next_affinity != my_affinity_id)
527  GATHER_STATISTIC( ++my_counters.affinity_ignored );
528 #endif
529  } // if there is bypassed task
530  }
532  switch( t->state() ) {
533  case task::executing: {
534  task* s = t->parent();
536  __TBB_ASSERT( t->prefix().ref_count==0, "Task still has children after it has been executed" );
537  t->~task();
538  if( s )
539  tally_completion_of_predecessor( *s, __TBB_ISOLATION_ARG( t_next, t->prefix().isolation ) );
540  free_task<no_hint>( *t );
543  break;
544  }
545 
546  case task::recycle: // set by recycle_as_safe_continuation()
547  t->prefix().state = task::allocated;
548 #if __TBB_RECYCLE_TO_ENQUEUE
550  case task::to_enqueue: // set by recycle_to_enqueue()
551 #endif
552  __TBB_ASSERT( t_next != t, "a task returned from method execute() can not be recycled in another way" );
554  // for safe continuation, need atomically decrement ref_count;
555  tally_completion_of_predecessor(*t, __TBB_ISOLATION_ARG( t_next, t->prefix().isolation ) );
557  break;
558 
559  case task::reexecute: // set by recycle_to_reexecute()
560  __TBB_ASSERT( t_next, "reexecution requires that method execute() return another task" );
561  __TBB_ASSERT( t_next != t, "a task returned from method execute() can not be recycled in another way" );
562  t->prefix().state = task::allocated;
564  local_spawn( t, t->prefix().next );
566  break;
567  case task::allocated:
569  break;
570 #if TBB_USE_ASSERT
571  case task::ready:
572  __TBB_ASSERT( false, "task is in READY state upon return from method execute()" );
573  break;
574  default:
575  __TBB_ASSERT( false, "illegal state" );
576 #else
577  default: // just to shut up some compilation warnings
578  break;
579 #endif /* TBB_USE_ASSERT */
580  }
581  GATHER_STATISTIC( t_next ? ++my_counters.spawns_bypassed : 0 );
582  t = t_next;
583  } // end of scheduler bypass loop
584 
586  if ( parent.prefix().ref_count == quit_point ) {
587  __TBB_ASSERT( quit_point != all_local_work_done, NULL );
588  __TBB_control_consistency_helper(); // on ref_count
589  ITT_NOTIFY(sync_acquired, &parent.prefix().ref_count);
590  goto done;
591  }
592  if ( is_task_pool_published() ) {
593  t = get_task( __TBB_ISOLATION_EXPR( isolation ) );
594  } else {
596  break;
597  }
599 
600  if ( !t ) break;
601 
602  context_guard.set_ctx( __TBB_CONTEXT_ARG1(t->prefix().context) );
603  }; // end of local task pool retrieval loop
604 
605 #if __TBB_TASK_PRIORITY
606 stealing_ground:
607 #endif /* __TBB_TASK_PRIORITY */
608 #if __TBB_HOARD_NONLOCAL_TASKS
609  // before stealing, previously stolen task objects are returned
610  for (; my_nonlocal_free_list; my_nonlocal_free_list = t ) {
611  t = my_nonlocal_free_list->prefix().next;
612  free_nonlocal_small_task( *my_nonlocal_free_list );
613  }
614 #endif
615  if ( quit_point == all_local_work_done ) {
618  my_innermost_running_task = old_innermost_running_task;
619  my_properties = old_properties;
620 #if __TBB_TASK_PRIORITY
621  my_ref_top_priority = old_ref_top_priority;
622  if(my_ref_reload_epoch != old_ref_reload_epoch)
623  my_local_reload_epoch = *old_ref_reload_epoch-1;
624  my_ref_reload_epoch = old_ref_reload_epoch;
625 #endif /* __TBB_TASK_PRIORITY */
626  return;
627  }
628 
629  t = receive_or_steal_task( __TBB_ISOLATION_ARG( parent.prefix().ref_count, isolation ) );
630  if ( !t )
631  goto done;
632  // The user can capture another the FPU settings to the context so the
633  // cached data in the helper can be out-of-date and we cannot do fast
634  // check.
635  context_guard.set_ctx( __TBB_CONTEXT_ARG1(t->prefix().context) );
636  } // end of infinite stealing loop
637 #if TBB_USE_EXCEPTIONS
638  __TBB_ASSERT( false, "Must never get here" );
639  } // end of try-block
640  TbbCatchAll( t->prefix().context );
641  // Complete post-processing ...
642  if( t->state() == task::recycle
643 #if __TBB_RECYCLE_TO_ENQUEUE
644  // TODO: the enqueue semantics gets lost below, consider reimplementing
645  || t->state() == task::to_enqueue
646 #endif
647  ) {
648  // ... for recycled tasks to atomically decrement ref_count
649  t->prefix().state = task::allocated;
650  if( SchedulerTraits::itt_possible )
651  ITT_NOTIFY(sync_releasing, &t->prefix().ref_count);
652  if( __TBB_FetchAndDecrementWrelease(&t->prefix().ref_count)==1 ) {
653  if( SchedulerTraits::itt_possible )
654  ITT_NOTIFY(sync_acquired, &t->prefix().ref_count);
655  }else{
656  t = NULL;
657  }
658  }
659  } // end of infinite EH loop
660  __TBB_ASSERT( false, "Must never get here too" );
661 #endif /* TBB_USE_EXCEPTIONS */
662 done:
663  my_innermost_running_task = old_innermost_running_task;
664  my_properties = old_properties;
665 #if __TBB_TASK_PRIORITY
666  my_ref_top_priority = old_ref_top_priority;
667  if(my_ref_reload_epoch != old_ref_reload_epoch)
668  my_local_reload_epoch = *old_ref_reload_epoch-1;
669  my_ref_reload_epoch = old_ref_reload_epoch;
670 #endif /* __TBB_TASK_PRIORITY */
671  if ( !ConcurrentWaitsEnabled(parent) ) {
672  if ( parent.prefix().ref_count != parents_work_done ) {
673  // This is a worker that was revoked by the market.
675  "Worker thread exits nested dispatch loop prematurely" );
676  return;
677  }
678  parent.prefix().ref_count = 0;
679  }
680 #if TBB_USE_ASSERT
681  parent.prefix().extra_state &= ~es_ref_count_active;
682 #endif /* TBB_USE_ASSERT */
683 #if __TBB_TASK_GROUP_CONTEXT
684  __TBB_ASSERT(parent.prefix().context && default_context(), NULL);
685  task_group_context* parent_ctx = parent.prefix().context;
686  if ( parent_ctx->my_cancellation_requested ) {
687  task_group_context::exception_container_type *pe = parent_ctx->my_exception;
688  if ( master_outermost_level() && parent_ctx == default_context() ) {
689  // We are in the outermost task dispatch loop of a master thread, and
690  // the whole task tree has been collapsed. So we may clear cancellation data.
691  parent_ctx->my_cancellation_requested = 0;
692  // TODO: Add assertion that master's dummy task context does not have children
693  parent_ctx->my_state &= ~(uintptr_t)task_group_context::may_have_children;
694  }
695  if ( pe ) {
696  // On Windows, FPU control settings changed in the helper destructor are not visible
697  // outside a catch block. So restore the default settings manually before rethrowing
698  // the exception.
699  context_guard.restore_default();
700  TbbRethrowException( pe );
701  }
702  }
703  __TBB_ASSERT(!is_worker() || !CancellationInfoPresent(*my_dummy_task),
704  "Worker's dummy task context modified");
705  __TBB_ASSERT(!master_outermost_level() || !CancellationInfoPresent(*my_dummy_task),
706  "Unexpected exception or cancellation data in the master's dummy task");
707 #endif /* __TBB_TASK_GROUP_CONTEXT */
709 }
static const intptr_t num_priority_levels
task * my_dummy_task
Fake root task created by slave threads.
Definition: scheduler.h:169
internal::tbb_exception_ptr exception_container_type
Definition: task.h:341
bool is_idle_state(bool value) const
Indicate whether thread that reads this mailbox is idle.
Definition: mailbox.h:218
#define __TBB_ISOLATION_EXPR(isolation)
intptr_t reference_count
A reference count.
Definition: task.h:117
void local_spawn(task *first, task *&next)
Definition: scheduler.cpp:614
#define ITT_SYNC_CREATE(obj, type, name)
Definition: itt_notify.h:119
void const char const char int ITT_FORMAT __itt_group_sync s
#define __TBB_FetchAndDecrementWrelease(P)
Definition: tbb_machine.h:311
#define GATHER_STATISTIC(x)
unsigned num_workers_active()
The number of workers active in the arena.
Definition: arena.h:233
void poison_pointer(T *__TBB_atomic &)
Definition: tbb_stddef.h:305
scheduler_properties my_properties
Definition: scheduler.h:91
affinity_id my_affinity_id
The mailbox id assigned to this scheduler.
Definition: scheduler.h:89
void free_nonlocal_small_task(task &t)
Free a small task t that that was allocated by a different scheduler.
Definition: scheduler.cpp:375
arena * my_arena
The arena that I own (if master) or am servicing at the moment (if worker)
Definition: scheduler.h:74
task object is freshly allocated or recycled.
Definition: task.h:617
void tally_completion_of_predecessor(task &s, __TBB_ISOLATION_ARG(task *&bypass_slot, isolation_tag isolation))
Decrements ref_count of a predecessor.
internal::task_prefix & prefix(internal::version_tag *=NULL) const
Get reference to corresponding task_prefix.
Definition: task.h:946
unsigned short affinity_id
An id as used for specifying affinity.
Definition: task.h:120
#define ITT_STACK(precond, name, obj)
Definition: itt_notify.h:122
task to be rescheduled.
Definition: task.h:613
static bool is_set(generic_scheduler *s)
Used to check validity of the local scheduler TLS contents.
Definition: governor.cpp:120
void const char const char int ITT_FORMAT __itt_group_sync x void const char ITT_FORMAT __itt_group_sync s void ITT_FORMAT __itt_group_sync p void ITT_FORMAT p void ITT_FORMAT p no args __itt_suppress_mode_t unsigned int void size_t ITT_FORMAT d void ITT_FORMAT p void ITT_FORMAT p __itt_model_site __itt_model_site_instance ITT_FORMAT p __itt_model_task __itt_model_task_instance ITT_FORMAT p void ITT_FORMAT p void ITT_FORMAT p void size_t ITT_FORMAT d void ITT_FORMAT p const wchar_t ITT_FORMAT s const char ITT_FORMAT s const char ITT_FORMAT s const char ITT_FORMAT s no args void ITT_FORMAT p size_t ITT_FORMAT d no args const wchar_t const wchar_t ITT_FORMAT s __itt_heap_function void size_t int ITT_FORMAT d __itt_heap_function void ITT_FORMAT p __itt_heap_function void void size_t int ITT_FORMAT d no args no args unsigned int ITT_FORMAT u const __itt_domain __itt_id ITT_FORMAT lu const __itt_domain __itt_id __itt_id parent
#define __TBB_fallthrough
Definition: tbb_stddef.h:250
unsigned my_num_workers_allotted
The number of workers that have been marked out by the resource manager to service the arena.
Definition: arena.h:51
#define __TBB_ASSERT(predicate, comment)
No-op version of __TBB_ASSERT.
Definition: tbb_stddef.h:165
bool is_quiescent_local_task_pool_reset() const
Definition: scheduler.h:562
void assert_task_valid(const task *)
void advertise_new_work()
If necessary, raise a flag that there is new job in arena.
Definition: arena.h:389
bool ConcurrentWaitsEnabled(task &t)
task * get_task(__TBB_ISOLATION_EXPR(isolation_tag isolation))
Get a task from the local pool.
Definition: scheduler.cpp:974
#define __TBB_control_consistency_helper()
Definition: gcc_generic.h:60
bool outermost
Indicates that a scheduler is on outermost level.
Definition: scheduler.h:53
void const char const char int ITT_FORMAT __itt_group_sync x void const char ITT_FORMAT __itt_group_sync s void ITT_FORMAT __itt_group_sync p void ITT_FORMAT p sync_releasing
task * receive_or_steal_task(__TBB_ISOLATION_ARG(__TBB_atomic reference_count &completion_ref_count, isolation_tag isolation)) __TBB_override
Try getting a task from the mailbox or stealing from another scheduler.
#define __TBB_CONTEXT_ARG1(context)
#define ITT_NOTIFY(name, obj)
Definition: itt_notify.h:116
void const char const char int ITT_FORMAT __itt_group_sync x void const char ITT_FORMAT __itt_group_sync s void ITT_FORMAT __itt_group_sync p void ITT_FORMAT p void ITT_FORMAT p no args __itt_suppress_mode_t unsigned int void size_t ITT_FORMAT d void ITT_FORMAT p void ITT_FORMAT p __itt_model_site __itt_model_site_instance ITT_FORMAT p __itt_model_task * task
bool is_critical(task &t)
Definition: task.h:958
market * my_market
The market I am in.
Definition: scheduler.h:155
Set if ref_count might be changed by another thread. Used for debugging.
intptr_t isolation_tag
A tag for task isolation.
Definition: task.h:124
bool worker_outermost_level() const
True if the scheduler is on the outermost dispatch level in a worker thread.
Definition: scheduler.h:575
#define __TBB_ISOLATION_ARG(arg1, isolation)
const isolation_tag no_isolation
Definition: task.h:125
task to be recycled as continuation
Definition: task.h:621
static bool is_proxy(const task &t)
True if t is a task_proxy.
Definition: scheduler.h:269
bool master_outermost_level() const
True if the scheduler is on the outermost dispatch level in a master thread.
Definition: scheduler.h:571
task is in ready pool, or is going to be put there, or was just taken off.
Definition: task.h:615
void const char const char int ITT_FORMAT __itt_group_sync p
bool is_worker() const
True if running on a worker thread, false otherwise.
Definition: scheduler.h:591
void reset_extra_state(task *t)
task is running, and will be destroyed after method execute() completes.
Definition: task.h:611
task * my_innermost_running_task
Innermost task whose task::execute() is running. A dummy task on the outermost level.
Definition: scheduler.h:77

References __TBB_ASSERT, __TBB_CONTEXT_ARG1, __TBB_control_consistency_helper, __TBB_fallthrough, __TBB_FetchAndDecrementWrelease, __TBB_ISOLATION_ARG, __TBB_ISOLATION_EXPR, tbb::task::allocated, tbb::internal::assert_task_valid(), tbb::internal::ConcurrentWaitsEnabled(), tbb::internal::es_ref_count_active, tbb::internal::es_task_enqueued, tbb::task::execute(), tbb::task::executing, GATHER_STATISTIC, tbb::internal::is_critical(), tbb::internal::governor::is_set(), ITT_NOTIFY, ITT_STACK, ITT_SYNC_CREATE, tbb::task_group_context::may_have_children, tbb::internal::no_isolation, tbb::internal::num_priority_levels, p, parent, tbb::task::parent(), tbb::internal::poison_pointer(), tbb::task::prefix(), tbb::task::ready, tbb::task::recycle, tbb::task::reexecute, tbb::internal::reset_extra_state(), s, tbb::internal::context_guard_helper< T >::set_ctx(), tbb::task::state(), sync_releasing, and tbb::internal::arena::wakeup.

Here is the call graph for this function:

◆ receive_or_steal_task()

template<typename SchedulerTraits >
task * tbb::internal::custom_scheduler< SchedulerTraits >::receive_or_steal_task ( __TBB_ISOLATION_ARG(__TBB_atomic reference_count &completion_ref_count, isolation_tag isolation)  )
virtual

Try getting a task from the mailbox or stealing from another scheduler.

Returns the stolen task or NULL if all attempts fail.

Implements tbb::internal::generic_scheduler.

Definition at line 138 of file custom_scheduler.h.

138  {
139  task* t = NULL;
140  bool outermost_worker_level = worker_outermost_level();
141  bool outermost_dispatch_level = outermost_worker_level || master_outermost_level();
142  bool can_steal_here = can_steal();
143  my_inbox.set_is_idle( true );
144 #if __TBB_HOARD_NONLOCAL_TASKS
145  __TBB_ASSERT(!my_nonlocal_free_list, NULL);
146 #endif
147 #if __TBB_TASK_PRIORITY
148  if ( outermost_dispatch_level ) {
149  if ( intptr_t skipped_priority = my_arena->my_skipped_fifo_priority ) {
150  // This thread can dequeue FIFO tasks, and some priority levels of
151  // FIFO tasks have been bypassed (to prevent deadlock caused by
152  // dynamic priority changes in nested task group hierarchy).
153  if ( my_arena->my_skipped_fifo_priority.compare_and_swap(0, skipped_priority) == skipped_priority
154  && skipped_priority > my_arena->my_top_priority )
155  {
156  my_market->update_arena_priority( *my_arena, skipped_priority );
157  }
158  }
159  }
160 #endif /* !__TBB_TASK_PRIORITY */
161  // TODO: Try to find a place to reset my_limit (under market's lock)
162  // The number of slots potentially used in the arena. Updated once in a while, as my_limit changes rarely.
163  size_t n = my_arena->my_limit-1;
164  int yield_count = 0;
165  // The state "failure_count==-1" is used only when itt_possible is true,
166  // and denotes that a sync_prepare has not yet been issued.
167  for( int failure_count = -static_cast<int>(SchedulerTraits::itt_possible);; ++failure_count) {
168  __TBB_ASSERT( my_arena->my_limit > 0, NULL );
169  __TBB_ASSERT( my_arena_index <= n, NULL );
170  if( completion_ref_count==1 ) {
171  if( SchedulerTraits::itt_possible ) {
172  if( failure_count!=-1 ) {
173  ITT_NOTIFY(sync_prepare, &completion_ref_count);
174  // Notify Intel(R) Thread Profiler that thread has stopped spinning.
175  ITT_NOTIFY(sync_acquired, this);
176  }
177  ITT_NOTIFY(sync_acquired, &completion_ref_count);
178  }
179  __TBB_ASSERT( !t, NULL );
180  // A worker thread in its outermost dispatch loop (i.e. its execution stack is empty) should
181  // exit it either when there is no more work in the current arena, or when revoked by the market.
182  __TBB_ASSERT( !outermost_worker_level, NULL );
183  __TBB_control_consistency_helper(); // on ref_count
184  break; // exit stealing loop and return;
185  }
186  // Check if the resource manager requires our arena to relinquish some threads
187  if ( outermost_worker_level && (my_arena->my_num_workers_allotted < my_arena->num_workers_active()
188 #if __TBB_ENQUEUE_ENFORCED_CONCURRENCY
189  || my_arena->recall_by_mandatory_request()
190 #endif
191  ) ) {
192  if( SchedulerTraits::itt_possible && failure_count != -1 )
193  ITT_NOTIFY(sync_cancel, this);
194  return NULL;
195  }
196 #if __TBB_TASK_PRIORITY
197  const int p = int(my_arena->my_top_priority);
198 #else /* !__TBB_TASK_PRIORITY */
199  static const int p = 0;
200 #endif
201  // Check if there are tasks mailed to this thread via task-to-thread affinity mechanism.
203  if ( n && !my_inbox.empty() ) {
204  t = get_mailbox_task( __TBB_ISOLATION_EXPR( isolation ) );
205 #if __TBB_TASK_ISOLATION
206  // There is a race with a thread adding a new task (possibly with suitable isolation)
207  // to our mailbox, so the below conditions might result in a false positive.
208  // Then set_is_idle(false) allows that task to be stolen; it's OK.
209  if ( isolation != no_isolation && !t && !my_inbox.empty()
210  && my_inbox.is_idle_state( true ) ) {
211  // We have proxy tasks in our mailbox but the isolation blocks their execution.
212  // So publish the proxy tasks in mailbox to be available for stealing from owner's task pool.
213  my_inbox.set_is_idle( false );
214  }
215 #endif /* __TBB_TASK_ISOLATION */
216  }
217  if ( t ) {
218  GATHER_STATISTIC( ++my_counters.mails_received );
219  }
220  // Check if there are tasks in starvation-resistant stream.
221  // Only allowed at the outermost dispatch level without isolation.
222  else if (__TBB_ISOLATION_EXPR(isolation == no_isolation &&) outermost_dispatch_level &&
224 #if __TBB_PREVIEW_CRITICAL_TASKS && __TBB_CPF_BUILD
225  t = my_arena->my_task_stream.pop( p, subsequent_lane_selector(my_arena_slot->hint_for_pop) )
226 #else
228 #endif
229  ) ) {
230  ITT_NOTIFY(sync_acquired, &my_arena->my_task_stream);
231  // just proceed with the obtained task
232  }
233 #if __TBB_TASK_PRIORITY
234  // Check if any earlier offloaded non-top priority tasks become returned to the top level
235  else if ( my_offloaded_tasks && (t = reload_tasks( __TBB_ISOLATION_EXPR( isolation ) )) ) {
236  __TBB_ASSERT( !is_proxy(*t), "The proxy task cannot be offloaded" );
237  // just proceed with the obtained task
238  }
239 #endif /* __TBB_TASK_PRIORITY */
240  else if ( can_steal_here && n && (t = steal_task( __TBB_ISOLATION_EXPR(isolation) )) ) {
241  // just proceed with the obtained task
242  }
243 #if __TBB_PREVIEW_CRITICAL_TASKS
244  else if( (t = get_critical_task( __TBB_ISOLATION_EXPR(isolation) )) ) {
245  __TBB_ASSERT( internal::is_critical(*t), "Received task must be critical one" );
246  ITT_NOTIFY(sync_acquired, &my_arena->my_critical_task_stream);
247  // just proceed with the obtained task
248  }
249 #endif // __TBB_PREVIEW_CRITICAL_TASKS
250  else
251  goto fail;
252  // A task was successfully obtained somewhere
253  __TBB_ASSERT(t,NULL);
254 #if __TBB_ARENA_OBSERVER
255  my_arena->my_observers.notify_entry_observers( my_last_local_observer, is_worker() );
256 #endif
257 #if __TBB_SCHEDULER_OBSERVER
258  the_global_observer_list.notify_entry_observers( my_last_global_observer, is_worker() );
259 #endif /* __TBB_SCHEDULER_OBSERVER */
260  if ( SchedulerTraits::itt_possible && failure_count != -1 ) {
261  // FIXME - might be victim, or might be selected from a mailbox
262  // Notify Intel(R) Thread Profiler that thread has stopped spinning.
263  ITT_NOTIFY(sync_acquired, this);
264  }
265  break; // exit stealing loop and return
266 fail:
267  GATHER_STATISTIC( ++my_counters.steals_failed );
268  if( SchedulerTraits::itt_possible && failure_count==-1 ) {
269  // The first attempt to steal work failed, so notify Intel(R) Thread Profiler that
270  // the thread has started spinning. Ideally, we would do this notification
271  // *before* the first failed attempt to steal, but at that point we do not
272  // know that the steal will fail.
273  ITT_NOTIFY(sync_prepare, this);
274  failure_count = 0;
275  }
276  // Pause, even if we are going to yield, because the yield might return immediately.
277  prolonged_pause();
278  const int failure_threshold = 2*int(n+1);
279  if( failure_count>=failure_threshold ) {
280 #if __TBB_YIELD2P
281  failure_count = 0;
282 #else
283  failure_count = failure_threshold;
284 #endif
285  __TBB_Yield();
286 #if __TBB_TASK_PRIORITY
287  // Check if there are tasks abandoned by other workers
288  if ( my_arena->my_orphaned_tasks ) {
289  // Epoch must be advanced before seizing the list pointer
290  ++my_arena->my_abandonment_epoch;
291  task* orphans = (task*)__TBB_FetchAndStoreW( &my_arena->my_orphaned_tasks, 0 );
292  if ( orphans ) {
293  task** link = NULL;
294  // Get local counter out of the way (we've just brought in external tasks)
295  my_local_reload_epoch--;
296  t = reload_tasks( orphans, link, __TBB_ISOLATION_ARG( effective_reference_priority(), isolation ) );
297  if ( orphans ) {
298  *link = my_offloaded_tasks;
299  if ( !my_offloaded_tasks )
300  my_offloaded_task_list_tail_link = link;
301  my_offloaded_tasks = orphans;
302  }
303  __TBB_ASSERT( !my_offloaded_tasks == !my_offloaded_task_list_tail_link, NULL );
304  if ( t ) {
305  if( SchedulerTraits::itt_possible )
306  ITT_NOTIFY(sync_cancel, this);
307  __TBB_ASSERT( !is_proxy(*t), "The proxy task cannot be offloaded" );
308  break; // exit stealing loop and return
309  }
310  }
311  }
312 #endif /* __TBB_TASK_PRIORITY */
313  const int yield_threshold = 100;
314  if( yield_count++ >= yield_threshold ) {
315  // When a worker thread has nothing to do, return it to RML.
316  // For purposes of affinity support, the thread is considered idle while in RML.
317 #if __TBB_TASK_PRIORITY
318  if( outermost_worker_level || my_arena->my_top_priority > my_arena->my_bottom_priority ) {
319  if ( my_arena->is_out_of_work() && outermost_worker_level ) {
320 #else /* !__TBB_TASK_PRIORITY */
321  if ( outermost_worker_level && my_arena->is_out_of_work() ) {
322 #endif /* !__TBB_TASK_PRIORITY */
323  if( SchedulerTraits::itt_possible )
324  ITT_NOTIFY(sync_cancel, this);
325  return NULL;
326  }
327 #if __TBB_TASK_PRIORITY
328  }
329  if ( my_offloaded_tasks ) {
330  // Safeguard against any sloppiness in managing reload epoch
331  // counter (e.g. on the hot path because of performance reasons).
332  my_local_reload_epoch--;
333  // Break the deadlock caused by a higher priority dispatch loop
334  // stealing and offloading a lower priority task. Priority check
335  // at the stealing moment cannot completely preclude such cases
336  // because priorities can changes dynamically.
337  if ( !outermost_worker_level && *my_ref_top_priority > my_arena->my_top_priority ) {
338  GATHER_STATISTIC( ++my_counters.prio_ref_fixups );
339  my_ref_top_priority = &my_arena->my_top_priority;
340  // it's expected that only outermost workers can use global reload epoch
341  __TBB_ASSERT(my_ref_reload_epoch == &my_arena->my_reload_epoch, NULL);
342  }
343  }
344 #endif /* __TBB_TASK_PRIORITY */
345  } // end of arena snapshot branch
346  // If several attempts did not find work, re-read the arena limit.
347  n = my_arena->my_limit-1;
348  } // end of yielding branch
349  } // end of nonlocal task retrieval loop
350  if ( my_inbox.is_idle_state( true ) )
351  my_inbox.set_is_idle( false );
352  return t;
353 }
size_t my_arena_index
Index of the arena slot the scheduler occupies now, or occupied last time.
Definition: scheduler.h:68
bool is_idle_state(bool value) const
Indicate whether thread that reads this mailbox is idle.
Definition: mailbox.h:218
atomic< unsigned > my_limit
The maximal number of currently busy slots.
Definition: arena.h:65
#define __TBB_ISOLATION_EXPR(isolation)
bool can_steal()
Returns true if stealing is allowed.
Definition: scheduler.h:191
#define GATHER_STATISTIC(x)
unsigned num_workers_active()
The number of workers active in the arena.
Definition: arena.h:233
void const char const char int ITT_FORMAT __itt_group_sync x void const char ITT_FORMAT __itt_group_sync s void ITT_FORMAT __itt_group_sync p sync_cancel
affinity_id my_affinity_id
The mailbox id assigned to this scheduler.
Definition: scheduler.h:89
arena * my_arena
The arena that I own (if master) or am servicing at the moment (if worker)
Definition: scheduler.h:74
bool empty(int level)
Checks existence of a task.
Definition: task_stream.h:138
unsigned hint_for_pop
Hint provided for operations with the container of starvation-resistant tasks.
void set_is_idle(bool value)
Indicate whether thread that reads this mailbox is idle.
Definition: mailbox.h:211
task * get_mailbox_task(__TBB_ISOLATION_EXPR(isolation_tag isolation))
Attempt to get a task from the mailbox.
Definition: scheduler.cpp:1196
bool empty()
Return true if mailbox is empty.
Definition: mailbox.h:206
unsigned my_num_workers_allotted
The number of workers that have been marked out by the resource manager to service the arena.
Definition: arena.h:51
#define __TBB_ASSERT(predicate, comment)
No-op version of __TBB_ASSERT.
Definition: tbb_stddef.h:165
#define __TBB_control_consistency_helper()
Definition: gcc_generic.h:60
#define __TBB_Yield()
Definition: ibm_aix51.h:44
#define ITT_NOTIFY(name, obj)
Definition: itt_notify.h:116
void const char const char int ITT_FORMAT __itt_group_sync x void const char ITT_FORMAT __itt_group_sync s void ITT_FORMAT __itt_group_sync p void ITT_FORMAT p void ITT_FORMAT p no args __itt_suppress_mode_t unsigned int void size_t ITT_FORMAT d void ITT_FORMAT p void ITT_FORMAT p __itt_model_site __itt_model_site_instance ITT_FORMAT p __itt_model_task * task
bool is_critical(task &t)
Definition: task.h:958
task_stream< num_priority_levels > my_task_stream
Task pool for the tasks scheduled via task::enqueue() method.
Definition: arena.h:76
market * my_market
The market I am in.
Definition: scheduler.h:155
arena_slot * my_arena_slot
Pointer to the slot in the arena we own at the moment.
Definition: scheduler.h:71
bool worker_outermost_level() const
True if the scheduler is on the outermost dispatch level in a worker thread.
Definition: scheduler.h:575
#define __TBB_ISOLATION_ARG(arg1, isolation)
const isolation_tag no_isolation
Definition: task.h:125
static bool is_proxy(const task &t)
True if t is a task_proxy.
Definition: scheduler.h:269
task * pop(int level, unsigned &last_used_lane)
Try finding and popping a task.
Definition: task_stream.h:116
bool master_outermost_level() const
True if the scheduler is on the outermost dispatch level in a master thread.
Definition: scheduler.h:571
void const char const char int ITT_FORMAT __itt_group_sync x void const char ITT_FORMAT __itt_group_sync s void ITT_FORMAT __itt_group_sync p void ITT_FORMAT p void ITT_FORMAT p no args __itt_suppress_mode_t unsigned int void size_t ITT_FORMAT d void ITT_FORMAT p void ITT_FORMAT p __itt_model_site __itt_model_site_instance ITT_FORMAT p __itt_model_task __itt_model_task_instance ITT_FORMAT p void ITT_FORMAT p void ITT_FORMAT p void size_t ITT_FORMAT d void ITT_FORMAT p const wchar_t ITT_FORMAT s const char ITT_FORMAT s const char ITT_FORMAT s const char ITT_FORMAT s no args void ITT_FORMAT p size_t ITT_FORMAT d no args const wchar_t const wchar_t ITT_FORMAT s __itt_heap_function void size_t int ITT_FORMAT d __itt_heap_function void ITT_FORMAT p __itt_heap_function void void size_t int ITT_FORMAT d no args no args unsigned int ITT_FORMAT u const __itt_domain __itt_id ITT_FORMAT lu const __itt_domain __itt_id __itt_id __itt_string_handle ITT_FORMAT p const __itt_domain __itt_id ITT_FORMAT p const __itt_domain __itt_id __itt_timestamp __itt_timestamp ITT_FORMAT lu const __itt_domain __itt_id __itt_id __itt_string_handle ITT_FORMAT p const __itt_domain ITT_FORMAT p const __itt_domain __itt_string_handle unsigned long long ITT_FORMAT lu const __itt_domain __itt_id __itt_string_handle __itt_metadata_type size_t void ITT_FORMAT p const __itt_domain __itt_id __itt_string_handle const wchar_t size_t ITT_FORMAT lu const __itt_domain __itt_id __itt_relation __itt_id ITT_FORMAT p const wchar_t int ITT_FORMAT __itt_group_mark d int
bool is_out_of_work()
Check if there is job anywhere in arena.
Definition: arena.cpp:403
void const char const char int ITT_FORMAT __itt_group_sync p
bool is_worker() const
True if running on a worker thread, false otherwise.
Definition: scheduler.h:591
task * steal_task(__TBB_ISOLATION_EXPR(isolation_tag isolation))
Attempts to steal a task from a randomly chosen thread/scheduler.
Definition: scheduler.cpp:1071

References __TBB_ASSERT, __TBB_control_consistency_helper, __TBB_ISOLATION_ARG, __TBB_ISOLATION_EXPR, __TBB_Yield, GATHER_STATISTIC, int, tbb::internal::is_critical(), ITT_NOTIFY, tbb::internal::no_isolation, p, tbb::internal::prolonged_pause(), and sync_cancel.

Here is the call graph for this function:

◆ tally_completion_of_predecessor()

template<typename SchedulerTraits >
void tbb::internal::custom_scheduler< SchedulerTraits >::tally_completion_of_predecessor ( task s,
__TBB_ISOLATION_ARG(task *&bypass_slot, isolation_tag isolation)   
)
inlineprivate

Decrements ref_count of a predecessor.

If it achieves 0, the predecessor is scheduled for execution. When changing, remember that this is a hot path function.

Definition at line 72 of file custom_scheduler.h.

72  {
73  task_prefix& p = s.prefix();
74  if( SchedulerTraits::itt_possible )
75  ITT_NOTIFY(sync_releasing, &p.ref_count);
76  if( SchedulerTraits::has_slow_atomic && p.ref_count==1 )
77  p.ref_count=0;
78  else if( __TBB_FetchAndDecrementWrelease(&p.ref_count) > 1 ) {// more references exist
79  // '__TBB_cl_evict(&p)' degraded performance of parallel_preorder example
80  return;
81  }
82 
83  // Ordering on p.ref_count (superfluous if SchedulerTraits::has_slow_atomic)
85  __TBB_ASSERT(p.ref_count==0, "completion of task caused predecessor's reference count to underflow");
86  if( SchedulerTraits::itt_possible )
87  ITT_NOTIFY(sync_acquired, &p.ref_count);
88 #if TBB_USE_ASSERT
89  p.extra_state &= ~es_ref_count_active;
90 #endif /* TBB_USE_ASSERT */
91 #if __TBB_TASK_ISOLATION
92  if ( isolation != no_isolation ) {
93  // The parent is allowed not to have isolation (even if a child has isolation) because it has never spawned.
94  __TBB_ASSERT(p.isolation == no_isolation || p.isolation == isolation, NULL);
95  p.isolation = isolation;
96  }
97 #endif /* __TBB_TASK_ISOLATION */
98 
99 #if __TBB_RECYCLE_TO_ENQUEUE
100  if (p.state==task::to_enqueue) {
101  // related to __TBB_TASK_ARENA TODO: try keep priority of the task
102  // e.g. rework task_prefix to remember priority of received task and use here
104  } else
105 #endif /*__TBB_RECYCLE_TO_ENQUEUE*/
106  if( bypass_slot==NULL )
107  bypass_slot = &s;
108 #if __TBB_PREVIEW_CRITICAL_TASKS
109  else if( internal::is_critical( s ) ) {
110  local_spawn( bypass_slot, bypass_slot->prefix().next );
111  bypass_slot = &s;
112  }
113 #endif /* __TBB_PREVIEW_CRITICAL_TASKS */
114  else
115  local_spawn( &s, s.prefix().next );
116  }
void local_spawn(task *first, task *&next)
Definition: scheduler.cpp:614
void const char const char int ITT_FORMAT __itt_group_sync s
#define __TBB_FetchAndDecrementWrelease(P)
Definition: tbb_machine.h:311
arena * my_arena
The arena that I own (if master) or am servicing at the moment (if worker)
Definition: scheduler.h:74
FastRandom my_random
Random number generator used for picking a random victim from which to steal.
Definition: scheduler.h:158
#define __TBB_ASSERT(predicate, comment)
No-op version of __TBB_ASSERT.
Definition: tbb_stddef.h:165
#define __TBB_control_consistency_helper()
Definition: gcc_generic.h:60
void const char const char int ITT_FORMAT __itt_group_sync x void const char ITT_FORMAT __itt_group_sync s void ITT_FORMAT __itt_group_sync p void ITT_FORMAT p sync_releasing
#define ITT_NOTIFY(name, obj)
Definition: itt_notify.h:116
bool is_critical(task &t)
Definition: task.h:958
Set if ref_count might be changed by another thread. Used for debugging.
const isolation_tag no_isolation
Definition: task.h:125
void const char const char int ITT_FORMAT __itt_group_sync p
void enqueue_task(task &, intptr_t, FastRandom &)
enqueue a task into starvation-resistance queue
Definition: arena.cpp:554

References __TBB_ASSERT, __TBB_control_consistency_helper, __TBB_FetchAndDecrementWrelease, tbb::internal::arena::enqueue_task(), tbb::internal::es_ref_count_active, tbb::internal::is_critical(), ITT_NOTIFY, tbb::internal::generic_scheduler::local_spawn(), tbb::internal::scheduler_state::my_arena, tbb::internal::generic_scheduler::my_random, tbb::internal::no_isolation, p, tbb::task::prefix(), s, and sync_releasing.

Here is the call graph for this function:

◆ wait_for_all()

template<typename SchedulerTraits >
void tbb::internal::custom_scheduler< SchedulerTraits >::wait_for_all ( task parent,
task child 
)
inlineprivatevirtual

Entry point from client code to the scheduler loop that dispatches tasks.

The method is virtual, but the *this object is used only for sake of dispatching on the correct vtable, not necessarily the correct *this object. The correct *this object is looked up in TLS.

Implements tbb::internal::scheduler.

Definition at line 65 of file custom_scheduler.h.

65  {
66  static_cast<custom_scheduler*>(governor::local_scheduler())->scheduler_type::local_wait_for_all( parent, child );
67  }
static generic_scheduler * local_scheduler()
Obtain the thread-local instance of the TBB scheduler.
Definition: governor.h:122
void const char const char int ITT_FORMAT __itt_group_sync x void const char ITT_FORMAT __itt_group_sync s void ITT_FORMAT __itt_group_sync p void ITT_FORMAT p void ITT_FORMAT p no args __itt_suppress_mode_t unsigned int void size_t ITT_FORMAT d void ITT_FORMAT p void ITT_FORMAT p __itt_model_site __itt_model_site_instance ITT_FORMAT p __itt_model_task __itt_model_task_instance ITT_FORMAT p void ITT_FORMAT p void ITT_FORMAT p void size_t ITT_FORMAT d void ITT_FORMAT p const wchar_t ITT_FORMAT s const char ITT_FORMAT s const char ITT_FORMAT s const char ITT_FORMAT s no args void ITT_FORMAT p size_t ITT_FORMAT d no args const wchar_t const wchar_t ITT_FORMAT s __itt_heap_function void size_t int ITT_FORMAT d __itt_heap_function void ITT_FORMAT p __itt_heap_function void void size_t int ITT_FORMAT d no args no args unsigned int ITT_FORMAT u const __itt_domain __itt_id ITT_FORMAT lu const __itt_domain __itt_id __itt_id parent

References tbb::internal::governor::local_scheduler(), and parent.

Here is the call graph for this function:

The documentation for this class was generated from the following file:

Copyright © 2005-2019 Intel Corporation. All Rights Reserved.

Intel, Pentium, Intel Xeon, Itanium, Intel XScale and VTune are registered trademarks or trademarks of Intel Corporation or its subsidiaries in the United States and other countries.

* Other names and brands may be claimed as the property of others.