1 | /* |
2 | Copyright (c) 2005-2019 Intel Corporation |
3 | |
4 | Licensed under the Apache License, Version 2.0 (the "License"); |
5 | you may not use this file except in compliance with the License. |
6 | You may obtain a copy of the License at |
7 | |
8 | http://www.apache.org/licenses/LICENSE-2.0 |
9 | |
10 | Unless required by applicable law or agreed to in writing, software |
11 | distributed under the License is distributed on an "AS IS" BASIS, |
12 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
13 | See the License for the specific language governing permissions and |
14 | limitations under the License. |
15 | */ |
16 | |
17 | #ifndef __TBB_parallel_invoke_H |
18 | #define __TBB_parallel_invoke_H |
19 | |
20 | #include "task.h" |
21 | #include "tbb_profiling.h" |
22 | |
23 | #if __TBB_VARIADIC_PARALLEL_INVOKE |
24 | #include <utility> // std::forward |
25 | #endif |
26 | |
27 | namespace tbb { |
28 | |
29 | #if !__TBB_TASK_GROUP_CONTEXT |
30 | /** Dummy to avoid cluttering the bulk of the header with enormous amount of ifdefs. **/ |
31 | struct task_group_context { |
32 | task_group_context(tbb::internal::string_index){} |
33 | }; |
34 | #endif /* __TBB_TASK_GROUP_CONTEXT */ |
35 | |
36 | //! @cond INTERNAL |
37 | namespace internal { |
38 | // Simple task object, executing user method |
39 | template<typename function> |
40 | class function_invoker : public task{ |
41 | public: |
42 | function_invoker(const function& _function) : my_function(_function) {} |
43 | private: |
44 | const function &my_function; |
45 | task* execute() __TBB_override |
46 | { |
47 | my_function(); |
48 | return NULL; |
49 | } |
50 | }; |
51 | |
52 | // The class spawns two or three child tasks |
53 | template <size_t N, typename function1, typename function2, typename function3> |
54 | class spawner : public task { |
55 | private: |
56 | const function1& my_func1; |
57 | const function2& my_func2; |
58 | const function3& my_func3; |
59 | bool is_recycled; |
60 | |
61 | task* execute () __TBB_override { |
62 | if(is_recycled){ |
63 | return NULL; |
64 | }else{ |
65 | __TBB_ASSERT(N==2 || N==3, "Number of arguments passed to spawner is wrong" ); |
66 | set_ref_count(N); |
67 | recycle_as_safe_continuation(); |
68 | internal::function_invoker<function2>* invoker2 = new (allocate_child()) internal::function_invoker<function2>(my_func2); |
69 | __TBB_ASSERT(invoker2, "Child task allocation failed" ); |
70 | spawn(*invoker2); |
71 | size_t n = N; // To prevent compiler warnings |
72 | if (n>2) { |
73 | internal::function_invoker<function3>* invoker3 = new (allocate_child()) internal::function_invoker<function3>(my_func3); |
74 | __TBB_ASSERT(invoker3, "Child task allocation failed" ); |
75 | spawn(*invoker3); |
76 | } |
77 | my_func1(); |
78 | is_recycled = true; |
79 | return NULL; |
80 | } |
81 | } // execute |
82 | |
83 | public: |
84 | spawner(const function1& _func1, const function2& _func2, const function3& _func3) : my_func1(_func1), my_func2(_func2), my_func3(_func3), is_recycled(false) {} |
85 | }; |
86 | |
87 | // Creates and spawns child tasks |
88 | class parallel_invoke_helper : public empty_task { |
89 | public: |
90 | // Dummy functor class |
91 | class parallel_invoke_noop { |
92 | public: |
93 | void operator() () const {} |
94 | }; |
95 | // Creates a helper object with user-defined number of children expected |
96 | parallel_invoke_helper(int number_of_children) |
97 | { |
98 | set_ref_count(number_of_children + 1); |
99 | } |
100 | |
101 | #if __TBB_VARIADIC_PARALLEL_INVOKE |
102 | void add_children() {} |
103 | void add_children(tbb::task_group_context&) {} |
104 | |
105 | template <typename function> |
106 | void add_children(function&& _func) |
107 | { |
108 | internal::function_invoker<function>* invoker = new (allocate_child()) internal::function_invoker<function>(std::forward<function>(_func)); |
109 | __TBB_ASSERT(invoker, "Child task allocation failed" ); |
110 | spawn(*invoker); |
111 | } |
112 | |
113 | template<typename function> |
114 | void add_children(function&& _func, tbb::task_group_context&) |
115 | { |
116 | add_children(std::forward<function>(_func)); |
117 | } |
118 | |
119 | // Adds child(ren) task(s) and spawns them |
120 | template <typename function1, typename function2, typename... function> |
121 | void add_children(function1&& _func1, function2&& _func2, function&&... _func) |
122 | { |
123 | // The third argument is dummy, it is ignored actually. |
124 | parallel_invoke_noop noop; |
125 | typedef internal::spawner<2, function1, function2, parallel_invoke_noop> spawner_type; |
126 | spawner_type & sub_root = *new(allocate_child()) spawner_type(std::forward<function1>(_func1), std::forward<function2>(_func2), noop); |
127 | spawn(sub_root); |
128 | add_children(std::forward<function>(_func)...); |
129 | } |
130 | #else |
131 | // Adds child task and spawns it |
132 | template <typename function> |
133 | void add_children (const function &_func) |
134 | { |
135 | internal::function_invoker<function>* invoker = new (allocate_child()) internal::function_invoker<function>(_func); |
136 | __TBB_ASSERT(invoker, "Child task allocation failed" ); |
137 | spawn(*invoker); |
138 | } |
139 | |
140 | // Adds a task with multiple child tasks and spawns it |
141 | // two arguments |
142 | template <typename function1, typename function2> |
143 | void add_children (const function1& _func1, const function2& _func2) |
144 | { |
145 | // The third argument is dummy, it is ignored actually. |
146 | parallel_invoke_noop noop; |
147 | internal::spawner<2, function1, function2, parallel_invoke_noop>& sub_root = *new(allocate_child())internal::spawner<2, function1, function2, parallel_invoke_noop>(_func1, _func2, noop); |
148 | spawn(sub_root); |
149 | } |
150 | // three arguments |
151 | template <typename function1, typename function2, typename function3> |
152 | void add_children (const function1& _func1, const function2& _func2, const function3& _func3) |
153 | { |
154 | internal::spawner<3, function1, function2, function3>& sub_root = *new(allocate_child())internal::spawner<3, function1, function2, function3>(_func1, _func2, _func3); |
155 | spawn(sub_root); |
156 | } |
157 | #endif // __TBB_VARIADIC_PARALLEL_INVOKE |
158 | |
159 | // Waits for all child tasks |
160 | template <typename F0> |
161 | void run_and_finish(const F0& f0) |
162 | { |
163 | internal::function_invoker<F0>* invoker = new (allocate_child()) internal::function_invoker<F0>(f0); |
164 | __TBB_ASSERT(invoker, "Child task allocation failed" ); |
165 | spawn_and_wait_for_all(*invoker); |
166 | } |
167 | }; |
168 | // The class destroys root if exception occurred as well as in normal case |
169 | class parallel_invoke_cleaner: internal::no_copy { |
170 | public: |
171 | #if __TBB_TASK_GROUP_CONTEXT |
172 | parallel_invoke_cleaner(int number_of_children, tbb::task_group_context& context) |
173 | : root(*new(task::allocate_root(context)) internal::parallel_invoke_helper(number_of_children)) |
174 | #else |
175 | parallel_invoke_cleaner(int number_of_children, tbb::task_group_context&) |
176 | : root(*new(task::allocate_root()) internal::parallel_invoke_helper(number_of_children)) |
177 | #endif /* !__TBB_TASK_GROUP_CONTEXT */ |
178 | {} |
179 | |
180 | ~parallel_invoke_cleaner(){ |
181 | root.destroy(root); |
182 | } |
183 | internal::parallel_invoke_helper& root; |
184 | }; |
185 | |
186 | #if __TBB_VARIADIC_PARALLEL_INVOKE |
187 | // Determine whether the last parameter in a pack is task_group_context |
188 | template<typename... T> struct impl_selector; // to workaround a GCC bug |
189 | |
190 | template<typename T1, typename... T> struct impl_selector<T1, T...> { |
191 | typedef typename impl_selector<T...>::type type; |
192 | }; |
193 | |
194 | template<typename T> struct impl_selector<T> { |
195 | typedef false_type type; |
196 | }; |
197 | template<> struct impl_selector<task_group_context&> { |
198 | typedef true_type type; |
199 | }; |
200 | |
201 | // Select task_group_context parameter from the back of a pack |
202 | inline task_group_context& get_context( task_group_context& tgc ) { return tgc; } |
203 | |
204 | template<typename T1, typename... T> |
205 | task_group_context& get_context( T1&& /*ignored*/, T&&... t ) |
206 | { return get_context( std::forward<T>(t)... ); } |
207 | |
208 | // task_group_context is known to be at the back of the parameter pack |
209 | template<typename F0, typename F1, typename... F> |
210 | void parallel_invoke_impl(true_type, F0&& f0, F1&& f1, F&&... f) { |
211 | __TBB_STATIC_ASSERT(sizeof...(F)>0, "Variadic parallel_invoke implementation broken?" ); |
212 | // # of child tasks: f0, f1, and a task for each two elements of the pack except the last |
213 | const size_t number_of_children = 2 + sizeof...(F)/2; |
214 | parallel_invoke_cleaner cleaner(number_of_children, get_context(std::forward<F>(f)...)); |
215 | parallel_invoke_helper& root = cleaner.root; |
216 | |
217 | root.add_children(std::forward<F>(f)...); |
218 | root.add_children(std::forward<F1>(f1)); |
219 | root.run_and_finish(std::forward<F0>(f0)); |
220 | } |
221 | |
222 | // task_group_context is not in the pack, needs to be added |
223 | template<typename F0, typename F1, typename... F> |
224 | void parallel_invoke_impl(false_type, F0&& f0, F1&& f1, F&&... f) { |
225 | tbb::task_group_context context(PARALLEL_INVOKE); |
226 | // Add context to the arguments, and redirect to the other overload |
227 | parallel_invoke_impl(true_type(), std::forward<F0>(f0), std::forward<F1>(f1), std::forward<F>(f)..., context); |
228 | } |
229 | #endif |
230 | } // namespace internal |
231 | //! @endcond |
232 | |
233 | /** \name parallel_invoke |
234 | **/ |
235 | //@{ |
236 | //! Executes a list of tasks in parallel and waits for all tasks to complete. |
237 | /** @ingroup algorithms */ |
238 | |
239 | #if __TBB_VARIADIC_PARALLEL_INVOKE |
240 | |
241 | // parallel_invoke for two or more arguments via variadic templates |
242 | // presence of task_group_context is defined automatically |
243 | template<typename F0, typename F1, typename... F> |
244 | void parallel_invoke(F0&& f0, F1&& f1, F&&... f) { |
245 | typedef typename internal::impl_selector<internal::false_type, F...>::type selector_type; |
246 | internal::parallel_invoke_impl(selector_type(), std::forward<F0>(f0), std::forward<F1>(f1), std::forward<F>(f)...); |
247 | } |
248 | |
249 | #else |
250 | |
251 | // parallel_invoke with user-defined context |
252 | // two arguments |
253 | template<typename F0, typename F1 > |
254 | void parallel_invoke(const F0& f0, const F1& f1, tbb::task_group_context& context) { |
255 | internal::parallel_invoke_cleaner cleaner(2, context); |
256 | internal::parallel_invoke_helper& root = cleaner.root; |
257 | |
258 | root.add_children(f1); |
259 | |
260 | root.run_and_finish(f0); |
261 | } |
262 | |
263 | // three arguments |
264 | template<typename F0, typename F1, typename F2 > |
265 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, tbb::task_group_context& context) { |
266 | internal::parallel_invoke_cleaner cleaner(3, context); |
267 | internal::parallel_invoke_helper& root = cleaner.root; |
268 | |
269 | root.add_children(f2); |
270 | root.add_children(f1); |
271 | |
272 | root.run_and_finish(f0); |
273 | } |
274 | |
275 | // four arguments |
276 | template<typename F0, typename F1, typename F2, typename F3> |
277 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, |
278 | tbb::task_group_context& context) |
279 | { |
280 | internal::parallel_invoke_cleaner cleaner(4, context); |
281 | internal::parallel_invoke_helper& root = cleaner.root; |
282 | |
283 | root.add_children(f3); |
284 | root.add_children(f2); |
285 | root.add_children(f1); |
286 | |
287 | root.run_and_finish(f0); |
288 | } |
289 | |
290 | // five arguments |
291 | template<typename F0, typename F1, typename F2, typename F3, typename F4 > |
292 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, const F4& f4, |
293 | tbb::task_group_context& context) |
294 | { |
295 | internal::parallel_invoke_cleaner cleaner(3, context); |
296 | internal::parallel_invoke_helper& root = cleaner.root; |
297 | |
298 | root.add_children(f4, f3); |
299 | root.add_children(f2, f1); |
300 | |
301 | root.run_and_finish(f0); |
302 | } |
303 | |
304 | // six arguments |
305 | template<typename F0, typename F1, typename F2, typename F3, typename F4, typename F5> |
306 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, const F4& f4, const F5& f5, |
307 | tbb::task_group_context& context) |
308 | { |
309 | internal::parallel_invoke_cleaner cleaner(3, context); |
310 | internal::parallel_invoke_helper& root = cleaner.root; |
311 | |
312 | root.add_children(f5, f4, f3); |
313 | root.add_children(f2, f1); |
314 | |
315 | root.run_and_finish(f0); |
316 | } |
317 | |
318 | // seven arguments |
319 | template<typename F0, typename F1, typename F2, typename F3, typename F4, typename F5, typename F6> |
320 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, const F4& f4, |
321 | const F5& f5, const F6& f6, |
322 | tbb::task_group_context& context) |
323 | { |
324 | internal::parallel_invoke_cleaner cleaner(3, context); |
325 | internal::parallel_invoke_helper& root = cleaner.root; |
326 | |
327 | root.add_children(f6, f5, f4); |
328 | root.add_children(f3, f2, f1); |
329 | |
330 | root.run_and_finish(f0); |
331 | } |
332 | |
333 | // eight arguments |
334 | template<typename F0, typename F1, typename F2, typename F3, typename F4, |
335 | typename F5, typename F6, typename F7> |
336 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, const F4& f4, |
337 | const F5& f5, const F6& f6, const F7& f7, |
338 | tbb::task_group_context& context) |
339 | { |
340 | internal::parallel_invoke_cleaner cleaner(4, context); |
341 | internal::parallel_invoke_helper& root = cleaner.root; |
342 | |
343 | root.add_children(f7, f6, f5); |
344 | root.add_children(f4, f3); |
345 | root.add_children(f2, f1); |
346 | |
347 | root.run_and_finish(f0); |
348 | } |
349 | |
350 | // nine arguments |
351 | template<typename F0, typename F1, typename F2, typename F3, typename F4, |
352 | typename F5, typename F6, typename F7, typename F8> |
353 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, const F4& f4, |
354 | const F5& f5, const F6& f6, const F7& f7, const F8& f8, |
355 | tbb::task_group_context& context) |
356 | { |
357 | internal::parallel_invoke_cleaner cleaner(4, context); |
358 | internal::parallel_invoke_helper& root = cleaner.root; |
359 | |
360 | root.add_children(f8, f7, f6); |
361 | root.add_children(f5, f4, f3); |
362 | root.add_children(f2, f1); |
363 | |
364 | root.run_and_finish(f0); |
365 | } |
366 | |
367 | // ten arguments |
368 | template<typename F0, typename F1, typename F2, typename F3, typename F4, |
369 | typename F5, typename F6, typename F7, typename F8, typename F9> |
370 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, const F4& f4, |
371 | const F5& f5, const F6& f6, const F7& f7, const F8& f8, const F9& f9, |
372 | tbb::task_group_context& context) |
373 | { |
374 | internal::parallel_invoke_cleaner cleaner(4, context); |
375 | internal::parallel_invoke_helper& root = cleaner.root; |
376 | |
377 | root.add_children(f9, f8, f7); |
378 | root.add_children(f6, f5, f4); |
379 | root.add_children(f3, f2, f1); |
380 | |
381 | root.run_and_finish(f0); |
382 | } |
383 | |
384 | // two arguments |
385 | template<typename F0, typename F1> |
386 | void parallel_invoke(const F0& f0, const F1& f1) { |
387 | task_group_context context(internal::PARALLEL_INVOKE); |
388 | parallel_invoke<F0, F1>(f0, f1, context); |
389 | } |
390 | // three arguments |
391 | template<typename F0, typename F1, typename F2> |
392 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2) { |
393 | task_group_context context(internal::PARALLEL_INVOKE); |
394 | parallel_invoke<F0, F1, F2>(f0, f1, f2, context); |
395 | } |
396 | // four arguments |
397 | template<typename F0, typename F1, typename F2, typename F3 > |
398 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3) { |
399 | task_group_context context(internal::PARALLEL_INVOKE); |
400 | parallel_invoke<F0, F1, F2, F3>(f0, f1, f2, f3, context); |
401 | } |
402 | // five arguments |
403 | template<typename F0, typename F1, typename F2, typename F3, typename F4> |
404 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, const F4& f4) { |
405 | task_group_context context(internal::PARALLEL_INVOKE); |
406 | parallel_invoke<F0, F1, F2, F3, F4>(f0, f1, f2, f3, f4, context); |
407 | } |
408 | // six arguments |
409 | template<typename F0, typename F1, typename F2, typename F3, typename F4, typename F5> |
410 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, const F4& f4, const F5& f5) { |
411 | task_group_context context(internal::PARALLEL_INVOKE); |
412 | parallel_invoke<F0, F1, F2, F3, F4, F5>(f0, f1, f2, f3, f4, f5, context); |
413 | } |
414 | // seven arguments |
415 | template<typename F0, typename F1, typename F2, typename F3, typename F4, typename F5, typename F6> |
416 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, const F4& f4, |
417 | const F5& f5, const F6& f6) |
418 | { |
419 | task_group_context context(internal::PARALLEL_INVOKE); |
420 | parallel_invoke<F0, F1, F2, F3, F4, F5, F6>(f0, f1, f2, f3, f4, f5, f6, context); |
421 | } |
422 | // eight arguments |
423 | template<typename F0, typename F1, typename F2, typename F3, typename F4, |
424 | typename F5, typename F6, typename F7> |
425 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, const F4& f4, |
426 | const F5& f5, const F6& f6, const F7& f7) |
427 | { |
428 | task_group_context context(internal::PARALLEL_INVOKE); |
429 | parallel_invoke<F0, F1, F2, F3, F4, F5, F6, F7>(f0, f1, f2, f3, f4, f5, f6, f7, context); |
430 | } |
431 | // nine arguments |
432 | template<typename F0, typename F1, typename F2, typename F3, typename F4, |
433 | typename F5, typename F6, typename F7, typename F8> |
434 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, const F4& f4, |
435 | const F5& f5, const F6& f6, const F7& f7, const F8& f8) |
436 | { |
437 | task_group_context context(internal::PARALLEL_INVOKE); |
438 | parallel_invoke<F0, F1, F2, F3, F4, F5, F6, F7, F8>(f0, f1, f2, f3, f4, f5, f6, f7, f8, context); |
439 | } |
440 | // ten arguments |
441 | template<typename F0, typename F1, typename F2, typename F3, typename F4, |
442 | typename F5, typename F6, typename F7, typename F8, typename F9> |
443 | void parallel_invoke(const F0& f0, const F1& f1, const F2& f2, const F3& f3, const F4& f4, |
444 | const F5& f5, const F6& f6, const F7& f7, const F8& f8, const F9& f9) |
445 | { |
446 | task_group_context context(internal::PARALLEL_INVOKE); |
447 | parallel_invoke<F0, F1, F2, F3, F4, F5, F6, F7, F8, F9>(f0, f1, f2, f3, f4, f5, f6, f7, f8, f9, context); |
448 | } |
449 | #endif // __TBB_VARIADIC_PARALLEL_INVOKE |
450 | //@} |
451 | |
452 | } // namespace |
453 | |
454 | #endif /* __TBB_parallel_invoke_H */ |
455 | |