| Geant4 Cross Reference |
1 //
2 // MIT License
3 // Copyright (c) 2020 Jonathan R. Madsen
4 // Permission is hereby granted, free of charge, to any person obtaining a copy
5 // of this software and associated documentation files (the "Software"), to deal
6 // in the Software without restriction, including without limitation the rights
7 // to use, copy, modify, merge, publish, distribute, sublicense, and
8 // copies of the Software, and to permit persons to whom the Software is
9 // furnished to do so, subject to the following conditions:
10 // The above copyright notice and this permission notice shall be included in
11 // all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED
12 // "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT
13 // LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR
14 // PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
15 // HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
16 // ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
17 // WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
18 //
19 // ---------------------------------------------------------------
20 // Tasking class header file
21 //
22 // Class Description:
23 //
24 // This file creates a class for an efficient thread-pool that
25 // accepts work in the form of tasks.
26 //
27 // ---------------------------------------------------------------
28 // Author: Jonathan Madsen (Feb 13th 2018)
29 // ---------------------------------------------------------------
30
31 #pragma once
32
33 #include "PTL/AutoLock.hh"
34 #ifndef G4GMAKE
35 #include ""
36 #endif
37 #include "PTL/ThreadData.hh"
38 #include "PTL/Threading.hh"
39 #include "PTL/Types.hh"
40 #include "PTL/VTask.hh"
41 #include "PTL/VUserTaskQueue.hh"
42
43 #if defined(PTL_USE_TBB)
44 # if !defined(TBB_SUPPRESS_DEPRECATED_MESSAGES)
45 # define TBB_SUPPRESS_DEPRECATED_MESSAGES 1
46 # endif
47 # if !defined(TBB_PREVIEW_GLOBAL_CONTROL)
48 # define TBB_PREVIEW_GLOBAL_CONTROL 1
49 # endif
50 # include <tbb/global_control.h>
51 # include <tbb/task_arena.h>
52 # include <tbb/task_group.h>
53 #endif
54
55 #include <algorithm>
56 #include <atomic>
57 #include <chrono>
58 #include <cstdint>
59 #include <cstdlib>
60 #include <deque>
61 #include <functional>
62 #include <iostream>
63 #include <map>
64 #include <memory>
65 #include <mutex> // IWYU pragma: keep
66 #include <set>
67 #include <thread>
68 #include <type_traits> // IWYU pragma: keep
69 #include <unordered_map>
70 #include <utility>
71 #include <vector>
72
73 namespace PTL
74 {
75 namespace thread_pool
76 {
77 namespace state
78 {
79 static const short STARTED = 0;
80 static const short PARTIAL = 1;
81 static const short STOPPED = 2;
82 static const short NONINIT = 3;
83
84 } // namespace state
85 } // namespace thread_pool
86
87 class ThreadPool
88 {
89 public:
90 template <typename KeyT, typename MappedT, typename HashT = KeyT>
91 using uomap = std::unordered_map<KeyT, MappedT, std::hash<HashT>>;
92
93 // pod-types
94 using size_type = size_t;
95 using task_count_type = std::shared_ptr<std::atomic_uintmax_t>;
96 using atomic_int_type = std::shared_ptr<std::atomic_uintmax_t>;
97 using pool_state_type = std::shared_ptr<std::atomic_short>;
98 using atomic_bool_type = std::shared_ptr<std::atomic_bool>;
99 // objects
100 using task_type = VTask;
101 using lock_t = std::shared_ptr<Mutex>;
102 using condition_t = std::shared_ptr<Condition>;
103 using task_pointer = std::shared_ptr<task_type>;
104 using task_queue_t = VUserTaskQueue;
105 // containers
106 using thread_list_t = std::deque<ThreadId>;
107 using bool_list_t = std::vector<bool>;
108 using thread_id_map_t = std::map<ThreadId, uintmax_t>;
109 using thread_index_map_t = std::map<uintmax_t, ThreadId>;
110 using thread_vec_t = std::vector<Thread>;
111 using thread_data_t = std::vector<std::shared_ptr<ThreadData>>;
112 // functions
113 using initialize_func_t = std::function<void()>;
114 using finalize_func_t = std::function<void()>;
115 using affinity_func_t = std::function<intmax_t(intmax_t)>;
116
117 static affinity_func_t& affinity_functor()
118 {
119 static affinity_func_t _v = [](intmax_t) {
120 static std::atomic<intmax_t> assigned;
121 intmax_t _assign = assigned++;
122 return _assign % Thread::hardware_concurrency();
123 };
124 return _v;
125 }
126
127 static initialize_func_t& initialization_functor()
128 {
129 static initialize_func_t _v = []() {};
130 return _v;
131 }
132
133 static finalize_func_t& finalization_functor()
134 {
135 static finalize_func_t _v = []() {};
136 return _v;
137 }
138
139 struct Config
140 {
141 bool init = true;
142 bool use_tbb = false;
143 bool use_affinity = false;
144 int verbose = 0;
145 int priority = 0;
146 size_type pool_size = f_default_pool_size();
147 VUserTaskQueue* task_queue = nullptr;
148 affinity_func_t set_affinity = affinity_functor();
149 initialize_func_t initializer = initialization_functor();
150 finalize_func_t finalizer = finalization_functor();
151 };
152
153 public:
154 // Constructor and Destructors
155 explicit ThreadPool(const Config&);
156 ~ThreadPool();
157 ThreadPool(const ThreadPool&) = delete;
158 ThreadPool(ThreadPool&&) = default;
159 ThreadPool& operator=(const ThreadPool&) = delete;
160 ThreadPool& operator=(ThreadPool&&) = default;
161
162 public:
163 // Public functions
164 size_type initialize_threadpool(size_type); // start the threads
165 size_type destroy_threadpool(); // destroy the threads
166 size_type stop_thread();
167
168 template <typename FuncT>
169 void execute_on_all_threads(FuncT&& _func);
170
171 template <typename FuncT>
172 void execute_on_specific_threads(const std::set<std::thread::id>& _tid,
173 FuncT&& _func);
174
175 task_queue_t* get_queue() const { return m_task_queue; }
176 task_queue_t*& get_valid_queue(task_queue_t*&) const;
177
178 bool is_tbb_threadpool() const { return m_tbb_tp; }
179
180 public:
181 /// set the default pool size
182 static void set_default_size(size_type _v) { f_default_pool_size() = _v; }
183
184 /// get the default pool size
185 static size_type get_default_size() { return f_default_pool_size(); }
186
187 public:
188 // add tasks for threads to process
189 size_type add_task(task_pointer&& task, int bin = -1);
190 // size_type add_thread_task(ThreadId id, task_pointer&& task);
191 // add a generic container with iterator
192 template <typename ListT>
193 size_type add_tasks(ListT&);
194
195 Thread* get_thread(size_type _n) const;
196 Thread* get_thread(std::thread::id id) const;
197
198 // only relevant when compiled with PTL_USE_TBB
199 static tbb_global_control_t*& tbb_global_control();
200
201 void set_initialization(initialize_func_t f) { m_init_func = std::move(f); }
202 void set_finalization(finalize_func_t f) { m_fini_func = std::move(f); }
203
204 void reset_initialization()
205 {
206 m_init_func = []() {};
207 }
208 void reset_finalization()
209 {
210 m_fini_func = []() {};
211 }
212
213 public:
214 // get the pool state
215 const pool_state_type& state() const { return m_pool_state; }
216 // see how many main task threads there are
217 size_type size() const { return m_pool_size; }
218 // set the thread pool size
219 void resize(size_type _n);
220 // affinity assigns threads to cores, assignment at constructor
221 bool using_affinity() const { return m_use_affinity; }
222 bool is_alive() { return m_alive_flag->load(); }
223 void notify();
224 void notify_all();
225 void notify(size_type);
226 bool is_initialized() const;
227 int get_active_threads_count() const { return (int)m_thread_awake->load(); }
228
229 void set_affinity(affinity_func_t f) { m_affinity_func = std::move(f); }
230 void set_affinity(intmax_t i, Thread&) const;
231 void set_priority(int _prio, Thread&) const;
232
233 void set_verbose(int n) { m_verbose = n; }
234 int get_verbose() const { return m_verbose; }
235 bool is_main() const { return ThisThread::get_id() == m_main_tid; }
236
237 tbb_task_arena_t* get_task_arena();
238
239 public:
240 // read FORCE_NUM_THREADS environment variable
241 static const thread_id_map_t& get_thread_ids();
242 static uintmax_t get_thread_id(ThreadId);
243 static uintmax_t get_this_thread_id();
244 static uintmax_t add_thread_id(ThreadId = ThisThread::get_id());
245
246 private:
247 void execute_thread(VUserTaskQueue*); // function thread sits in
248 int insert(task_pointer&&, int = -1);
249 int run_on_this(task_pointer&&);
250
251 private:
252 // called in THREAD INIT
253 static void start_thread(ThreadPool*, thread_data_t*, intmax_t = -1);
254
255 void record_entry();
256 void record_exit();
257
258 private:
259 // Private variables
260 // random
261 bool m_use_affinity = false;
262 bool m_tbb_tp = false;
263 bool m_delete_task_queue = false;
264 int m_verbose = 0;
265 int m_priority = 0;
266 size_type m_pool_size = 0;
267 ThreadId m_main_tid = ThisThread::get_id();
268 atomic_bool_type m_alive_flag = std::make_shared<std::atomic_bool>(false);
269 pool_state_type m_pool_state = std::make_shared<std::atomic_short>(0);
270 atomic_int_type m_thread_awake = std::make_shared<std::atomic_uintmax_t>(0);
271 atomic_int_type m_thread_active = std::make_shared<std::atomic_uintmax_t>(0);
272
273 // locks
274 lock_t m_task_lock = std::make_shared<Mutex>();
275 // conditions
276 condition_t m_task_cond = std::make_shared<Condition>();
277
278 // containers
279 bool_list_t m_is_joined = {}; // join list
280 bool_list_t m_is_stopped = {}; // lets thread know to stop
281 thread_list_t m_main_threads = {}; // storage for active threads
282 thread_list_t m_stop_threads = {}; // storage for stopped threads
283 thread_vec_t m_threads = {};
284 thread_data_t m_thread_data = {};
285
286 // task queue
287 task_queue_t* m_task_queue = nullptr;
288 tbb_task_arena_t* m_tbb_task_arena = nullptr;
289 tbb_task_group_t* m_tbb_task_group = nullptr;
290
291 // functions
292 initialize_func_t m_init_func = initialization_functor();
293 finalize_func_t m_fini_func = finalization_functor();
294 affinity_func_t m_affinity_func = affinity_functor();
295
296 private:
297 static size_type& f_default_pool_size();
298 static thread_id_map_t& f_thread_ids();
299 };
300
301 //--------------------------------------------------------------------------------------//
302 inline void
303 ThreadPool::notify()
304 {
305 // wake up one thread that is waiting for a task to be available
306 if(m_thread_awake->load() < m_pool_size)
307 {
308 AutoLock l(*m_task_lock);
309 m_task_cond->notify_one();
310 }
311 }
312 //--------------------------------------------------------------------------------------//
313 inline void
314 ThreadPool::notify_all()
315 {
316 // wake all threads
317 AutoLock l(*m_task_lock);
318 m_task_cond->notify_all();
319 }
320 //--------------------------------------------------------------------------------------//
321 inline void
322 ThreadPool::notify(size_type ntasks)
323 {
324 if(ntasks == 0)
325 return;
326
327 // wake up as many threads that tasks just added
328 if(m_thread_awake->load() < m_pool_size)
329 {
330 AutoLock l(*m_task_lock);
331 if(ntasks < this->size())
332 {
333 for(size_type i = 0; i < ntasks; ++i)
334 m_task_cond->notify_one();
335 }
336 else
337 {
338 m_task_cond->notify_all();
339 }
340 }
341 }
342 //--------------------------------------------------------------------------------------//
343 // local function for getting the tbb task scheduler
344 inline tbb_global_control_t*&
345 ThreadPool::tbb_global_control()
346 {
347 static thread_local tbb_global_control_t* _instance = nullptr;
348 return _instance;
349 }
350 //--------------------------------------------------------------------------------------//
351 // task arena
352 inline tbb_task_arena_t*
353 ThreadPool::get_task_arena()
354 {
355 #if defined(PTL_USE_TBB)
356 // create a task arena
357 if(!m_tbb_task_arena)
358 {
359 auto _sz = (tbb_global_control())
360 ? tbb_global_control()->active_value(
361 tbb::global_control::max_allowed_parallelism)
362 : size();
363 m_tbb_task_arena = new tbb_task_arena_t(::tbb::task_arena::attach{});
364 m_tbb_task_arena->initialize(_sz, 1);
365 }
366 #else
367 if(!m_tbb_task_arena)
368 m_tbb_task_arena = new tbb_task_arena_t{};
369 #endif
370 return m_tbb_task_arena;
371 }
372 //--------------------------------------------------------------------------------------//
373 inline void
374 ThreadPool::resize(size_type _n)
375 {
376 initialize_threadpool(_n);
377 if(m_task_queue)
378 m_task_queue->resize(static_cast<intmax_t>(_n));
379 }
380 //--------------------------------------------------------------------------------------//
381 inline int
382 ThreadPool::run_on_this(task_pointer&& _task)
383 {
384 auto&& _func = [_task]() { (*_task)(); };
385
386 if(m_tbb_tp && m_tbb_task_group)
387 {
388 auto* _arena = get_task_arena();
389 _arena->execute([this, _func]() { this->m_tbb_task_group->run(_func); });
390 }
391 else
392 {
393 _func();
394 }
395 // return the number of tasks added to task-list
396 return 0;
397 }
398 //--------------------------------------------------------------------------------------//
399 inline int
400 ThreadPool::insert(task_pointer&& task, int bin)
401 {
402 static thread_local ThreadData* _data = ThreadData::GetInstance();
403
404 // pass the task to the queue
405 auto ibin = get_valid_queue(m_task_queue)->InsertTask(std::move(task), _data, bin);
406 notify();
407 return (int)ibin;
408 }
409 //--------------------------------------------------------------------------------------//
410 inline ThreadPool::size_type
411 ThreadPool::add_task(task_pointer&& task, int bin)
412 {
413 // if not native (i.e. TBB) or we haven't built thread-pool, just execute
414 if(m_tbb_tp || !task->is_native_task() || !m_alive_flag->load())
415 return static_cast<size_type>(run_on_this(std::move(task)));
416
417 return static_cast<size_type>(insert(std::move(task), bin));
418 }
419 //--------------------------------------------------------------------------------------//
420 template <typename ListT>
421 inline ThreadPool::size_type
422 ThreadPool::add_tasks(ListT& c)
423 {
424 if(!m_alive_flag) // if we haven't built thread-pool, just execute
425 {
426 for(auto& itr : c)
427 run(itr);
428 c.clear();
429 return 0;
430 }
431
432 // TODO: put a limit on how many tasks can be added at most
433 auto c_size = c.size();
434 for(auto& itr : c)
435 {
436 if(!itr->is_native_task())
437 --c_size;
438 else
439 {
440 //++(m_task_queue);
441 get_valid_queue(m_task_queue)->InsertTask(itr);
442 }
443 }
444 c.clear();
445
446 // notify sleeping threads
447 notify(c_size);
448
449 return c_size;
450 }
451 //--------------------------------------------------------------------------------------//
452 template <typename FuncT>
453 inline void
454 ThreadPool::execute_on_all_threads(FuncT&& _func)
455 {
456 if(m_tbb_tp && m_tbb_task_group)
457 {
458 #if defined(PTL_USE_TBB)
459 // TBB lazily activates threads to process tasks and the main thread
460 // participates in processing the tasks so getting a specific
461 // function to execute only on the worker threads requires some trickery
462 //
463 std::set<std::thread::id> _first{};
464 Mutex _mutex{};
465 // init function which executes function and returns 1 only once
466 auto _init = [&]() {
467 int _once = 0;
468 _mutex.lock();
469 if(_first.find(std::this_thread::get_id()) == _first.end())
470 {
471 // we need to reset this thread-local static for multiple invocations
472 // of the same template instantiation
473 _once = 1;
474 _first.insert(std::this_thread::get_id());
475 }
476 _mutex.unlock();
477 if(_once != 0)
478 {
479 _func();
480 return 1;
481 }
482 return 0;
483 };
484 // this will collect the number of threads which have
485 // executed the _init function above
486 std::atomic<size_t> _total_init{ 0 };
487 // max parallelism by TBB
488 size_t _maxp = tbb_global_control()->active_value(
489 tbb::global_control::max_allowed_parallelism);
490 // create a task arean
491 auto* _arena = get_task_arena();
492 // size of the thread-pool
493 size_t _sz = size();
494 // number of cores
495 size_t _ncore = GetNumberOfCores();
496 // maximum depth for recursion
497 size_t _dmax = std::max<size_t>(_ncore, 8);
498 // how many threads we need to initialize
499 size_t _num = std::min(_maxp, std::min(_sz, _ncore));
500 // this is the task passed to the task-group
501 std::function<void()> _init_task;
502 _init_task = [&]() {
503 add_thread_id();
504 static thread_local size_type _depth = 0;
505 int _ret = 0;
506 // don't let the main thread execute the function
507 if(!is_main())
508 {
509 // execute the function
510 _ret = _init();
511 // add the result
512 _total_init += _ret;
513 }
514 // if the function did not return anything, recursively execute
515 // two more tasks
516 ++_depth;
517 if(_ret == 0 && _depth < _dmax && _total_init.load() < _num)
518 {
519 tbb::task_group tg{};
520 tg.run([&]() { _init_task(); });
521 tg.run([&]() { _init_task(); });
522 ThisThread::sleep_for(std::chrono::milliseconds{ 1 });
523 tg.wait();
524 }
525 --_depth;
526 };
527
528 // TBB won't oversubscribe so we need to limit by ncores - 1
529 size_t nitr = 0;
530 auto _fname = __FUNCTION__;
531 auto _write_info = [&]() {
532 std::cout << "[" << _fname << "]> Total initialized: " << _total_init
533 << ", expected: " << _num << ", max-parallel: " << _maxp
534 << ", size: " << _sz << ", ncore: " << _ncore << std::endl;
535 };
536 while(_total_init < _num)
537 {
538 auto _n = 2 * _num;
539 while(--_n > 0)
540 {
541 _arena->execute(
542 [&]() { m_tbb_task_group->run([&]() { _init_task(); }); });
543 }
544 _arena->execute([&]() { m_tbb_task_group->wait(); });
545 // don't loop infinitely but use a strict condition
546 if(nitr++ > 2 * (_num + 1) && (_total_init - 1) == _num)
547 {
548 _write_info();
549 break;
550 }
551 // at this point we need to exit
552 if(nitr > 4 * (_ncore + 1))
553 {
554 _write_info();
555 break;
556 }
557 }
558 if(get_verbose() > 3)
559 _write_info();
560 #endif
561 }
562 else if(get_queue())
563 {
564 get_queue()->ExecuteOnAllThreads(this, std::forward<FuncT>(_func));
565 }
566 }
567
568 //--------------------------------------------------------------------------------------//
569
570 template <typename FuncT>
571 inline void
572 ThreadPool::execute_on_specific_threads(const std::set<std::thread::id>& _tids,
573 FuncT&& _func)
574 {
575 if(m_tbb_tp && m_tbb_task_group)
576 {
577 #if defined(PTL_USE_TBB)
578 // TBB lazily activates threads to process tasks and the main thread
579 // participates in processing the tasks so getting a specific
580 // function to execute only on the worker threads requires some trickery
581 //
582 std::set<std::thread::id> _first{};
583 Mutex _mutex{};
584 // init function which executes function and returns 1 only once
585 auto _exec = [&]() {
586 int _once = 0;
587 _mutex.lock();
588 if(_first.find(std::this_thread::get_id()) == _first.end())
589 {
590 // we need to reset this thread-local static for multiple invocations
591 // of the same template instantiation
592 _once = 1;
593 _first.insert(std::this_thread::get_id());
594 }
595 _mutex.unlock();
596 if(_once != 0)
597 {
598 _func();
599 return 1;
600 }
601 return 0;
602 };
603 // this will collect the number of threads which have
604 // executed the _exec function above
605 std::atomic<size_t> _total_exec{ 0 };
606 // number of cores
607 size_t _ncore = GetNumberOfCores();
608 // maximum depth for recursion
609 size_t _dmax = std::max<size_t>(_ncore, 8);
610 // how many threads we need to initialize
611 size_t _num = _tids.size();
612 // create a task arena
613 auto* _arena = get_task_arena();
614 // this is the task passed to the task-group
615 std::function<void()> _exec_task;
616 _exec_task = [&]() {
617 add_thread_id();
618 static thread_local size_type _depth = 0;
619 int _ret = 0;
620 auto _this_tid = std::this_thread::get_id();
621 // don't let the main thread execute the function
622 if(_tids.count(_this_tid) > 0)
623 {
624 // execute the function
625 _ret = _exec();
626 // add the result
627 _total_exec += _ret;
628 }
629 // if the function did not return anything, recursively execute
630 // two more tasks
631 ++_depth;
632 if(_ret == 0 && _depth < _dmax && _total_exec.load() < _num)
633 {
634 tbb::task_group tg{};
635 tg.run([&]() { _exec_task(); });
636 tg.run([&]() { _exec_task(); });
637 ThisThread::sleep_for(std::chrono::milliseconds{ 1 });
638 tg.wait();
639 }
640 --_depth;
641 };
642
643 // TBB won't oversubscribe so we need to limit by ncores - 1
644 size_t nitr = 0;
645 auto _fname = __FUNCTION__;
646 auto _write_info = [&]() {
647 std::cout << "[" << _fname << "]> Total executed: " << _total_exec
648 << ", expected: " << _num << ", size: " << size() << std::endl;
649 };
650 while(_total_exec < _num)
651 {
652 auto _n = 2 * _num;
653 while(--_n > 0)
654 {
655 _arena->execute(
656 [&]() { m_tbb_task_group->run([&]() { _exec_task(); }); });
657 }
658 _arena->execute([&]() { m_tbb_task_group->wait(); });
659 // don't loop infinitely but use a strict condition
660 if(nitr++ > 2 * (_num + 1) && (_total_exec - 1) == _num)
661 {
662 _write_info();
663 break;
664 }
665 // at this point we need to exit
666 if(nitr > 8 * (_num + 1))
667 {
668 _write_info();
669 break;
670 }
671 }
672 if(get_verbose() > 3)
673 _write_info();
674 #endif
675 }
676 else if(get_queue())
677 {
678 get_queue()->ExecuteOnSpecificThreads(_tids, this, std::forward<FuncT>(_func));
679 }
680 }
681
682 //======================================================================================//
683
684 } // namespace PTL
685