xref: /llvm-project/libcxx/include/__pstl/backends/libdispatch.h (revision 59890c13343af9e308281b3c76bac425087f4f8a)
1 //===----------------------------------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #ifndef _LIBCPP___PSTL_BACKENDS_LIBDISPATCH_H
10 #define _LIBCPP___PSTL_BACKENDS_LIBDISPATCH_H
11 
12 #include <__algorithm/inplace_merge.h>
13 #include <__algorithm/lower_bound.h>
14 #include <__algorithm/max.h>
15 #include <__algorithm/merge.h>
16 #include <__algorithm/upper_bound.h>
17 #include <__atomic/atomic.h>
18 #include <__config>
19 #include <__cstddef/ptrdiff_t.h>
20 #include <__exception/terminate.h>
21 #include <__iterator/iterator_traits.h>
22 #include <__iterator/move_iterator.h>
23 #include <__memory/allocator.h>
24 #include <__memory/construct_at.h>
25 #include <__memory/unique_ptr.h>
26 #include <__new/exceptions.h>
27 #include <__numeric/reduce.h>
28 #include <__pstl/backend_fwd.h>
29 #include <__pstl/cpu_algos/any_of.h>
30 #include <__pstl/cpu_algos/cpu_traits.h>
31 #include <__pstl/cpu_algos/fill.h>
32 #include <__pstl/cpu_algos/find_if.h>
33 #include <__pstl/cpu_algos/for_each.h>
34 #include <__pstl/cpu_algos/merge.h>
35 #include <__pstl/cpu_algos/stable_sort.h>
36 #include <__pstl/cpu_algos/transform.h>
37 #include <__pstl/cpu_algos/transform_reduce.h>
38 #include <__utility/empty.h>
39 #include <__utility/exception_guard.h>
40 #include <__utility/move.h>
41 #include <__utility/pair.h>
42 #include <optional>
43 
44 _LIBCPP_PUSH_MACROS
45 #include <__undef_macros>
46 
47 #if _LIBCPP_STD_VER >= 17
48 
49 _LIBCPP_BEGIN_NAMESPACE_STD
50 namespace __pstl {
51 
52 namespace __libdispatch {
53 // ::dispatch_apply is marked as __attribute__((nothrow)) because it doesn't let exceptions propagate, and neither do
54 // we.
55 // TODO: Do we want to add [[_Clang::__callback__(__func, __context, __)]]?
56 _LIBCPP_EXPORTED_FROM_ABI void
57 __dispatch_apply(size_t __chunk_count, void* __context, void (*__func)(void* __context, size_t __chunk)) noexcept;
58 
59 template <class _Func>
60 _LIBCPP_HIDE_FROM_ABI void __dispatch_apply(size_t __chunk_count, _Func __func) noexcept {
61   __libdispatch::__dispatch_apply(__chunk_count, &__func, [](void* __context, size_t __chunk) {
62     (*static_cast<_Func*>(__context))(__chunk);
63   });
64 }
65 
66 struct __chunk_partitions {
67   ptrdiff_t __chunk_count_; // includes the first chunk
68   ptrdiff_t __chunk_size_;
69   ptrdiff_t __first_chunk_size_;
70 };
71 
72 [[__gnu__::__const__]] _LIBCPP_EXPORTED_FROM_ABI __chunk_partitions __partition_chunks(ptrdiff_t __size) noexcept;
73 
74 template <class _RandomAccessIterator, class _Functor>
75 _LIBCPP_HIDE_FROM_ABI optional<__empty>
76 __dispatch_parallel_for(__chunk_partitions __partitions, _RandomAccessIterator __first, _Functor __func) {
77   // Perform the chunked execution.
78   __libdispatch::__dispatch_apply(__partitions.__chunk_count_, [&](size_t __chunk) {
79     auto __this_chunk_size = __chunk == 0 ? __partitions.__first_chunk_size_ : __partitions.__chunk_size_;
80     auto __index =
81         __chunk == 0
82             ? 0
83             : (__chunk * __partitions.__chunk_size_) + (__partitions.__first_chunk_size_ - __partitions.__chunk_size_);
84     __func(__first + __index, __first + __index + __this_chunk_size);
85   });
86 
87   return __empty{};
88 }
89 } // namespace __libdispatch
90 
91 template <>
92 struct __cpu_traits<__libdispatch_backend_tag> {
93   template <class _RandomAccessIterator, class _Functor>
94   _LIBCPP_HIDE_FROM_ABI static optional<__empty>
95   __for_each(_RandomAccessIterator __first, _RandomAccessIterator __last, _Functor __func) {
96     return __libdispatch::__dispatch_parallel_for(
97         __libdispatch::__partition_chunks(__last - __first), std::move(__first), std::move(__func));
98   }
99 
100   template <class _RandomAccessIterator1, class _RandomAccessIterator2, class _RandomAccessIteratorOut>
101   struct __merge_range {
102     __merge_range(_RandomAccessIterator1 __mid1, _RandomAccessIterator2 __mid2, _RandomAccessIteratorOut __result)
103         : __mid1_(__mid1), __mid2_(__mid2), __result_(__result) {}
104 
105     _RandomAccessIterator1 __mid1_;
106     _RandomAccessIterator2 __mid2_;
107     _RandomAccessIteratorOut __result_;
108   };
109 
110   template <typename _RandomAccessIterator1,
111             typename _RandomAccessIterator2,
112             typename _RandomAccessIterator3,
113             typename _Compare,
114             typename _LeafMerge>
115   _LIBCPP_HIDE_FROM_ABI static optional<__empty>
116   __merge(_RandomAccessIterator1 __first1,
117           _RandomAccessIterator1 __last1,
118           _RandomAccessIterator2 __first2,
119           _RandomAccessIterator2 __last2,
120           _RandomAccessIterator3 __result,
121           _Compare __comp,
122           _LeafMerge __leaf_merge) noexcept {
123     __libdispatch::__chunk_partitions __partitions =
124         __libdispatch::__partition_chunks(std::max<ptrdiff_t>(__last1 - __first1, __last2 - __first2));
125 
126     if (__partitions.__chunk_count_ == 0)
127       return __empty{};
128 
129     if (__partitions.__chunk_count_ == 1) {
130       __leaf_merge(__first1, __last1, __first2, __last2, __result, __comp);
131       return __empty{};
132     }
133 
134     using __merge_range_t = __merge_range<_RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator3>;
135     auto const __n_ranges = __partitions.__chunk_count_ + 1;
136 
137     // TODO: use __uninitialized_buffer
138     auto __destroy = [=](__merge_range_t* __ptr) {
139       std::destroy_n(__ptr, __n_ranges);
140       std::allocator<__merge_range_t>().deallocate(__ptr, __n_ranges);
141     };
142 
143     unique_ptr<__merge_range_t[], decltype(__destroy)> __ranges(
144         [&]() -> __merge_range_t* {
145 #  if _LIBCPP_HAS_EXCEPTIONS
146           try {
147 #  endif
148             return std::allocator<__merge_range_t>().allocate(__n_ranges);
149 #  if _LIBCPP_HAS_EXCEPTIONS
150           } catch (const std::bad_alloc&) {
151             return nullptr;
152           }
153 #  endif
154         }(),
155         __destroy);
156 
157     if (!__ranges)
158       return nullopt;
159 
160     // TODO: Improve the case where the smaller range is merged into just a few (or even one) chunks of the larger case
161     __merge_range_t* __r = __ranges.get();
162     std::__construct_at(__r++, __first1, __first2, __result);
163 
164     bool __iterate_first_range = __last1 - __first1 > __last2 - __first2;
165 
166     auto __compute_chunk = [&](size_t __chunk_size) -> __merge_range_t {
167       auto [__mid1, __mid2] = [&] {
168         if (__iterate_first_range) {
169           auto __m1 = __first1 + __chunk_size;
170           auto __m2 = std::lower_bound(__first2, __last2, __m1[-1], __comp);
171           return std::make_pair(__m1, __m2);
172         } else {
173           auto __m2 = __first2 + __chunk_size;
174           auto __m1 = std::lower_bound(__first1, __last1, __m2[-1], __comp);
175           return std::make_pair(__m1, __m2);
176         }
177       }();
178 
179       __result += (__mid1 - __first1) + (__mid2 - __first2);
180       __first1 = __mid1;
181       __first2 = __mid2;
182       return {std::move(__mid1), std::move(__mid2), __result};
183     };
184 
185     // handle first chunk
186     std::__construct_at(__r++, __compute_chunk(__partitions.__first_chunk_size_));
187 
188     // handle 2 -> N - 1 chunks
189     for (ptrdiff_t __i = 0; __i != __partitions.__chunk_count_ - 2; ++__i)
190       std::__construct_at(__r++, __compute_chunk(__partitions.__chunk_size_));
191 
192     // handle last chunk
193     std::__construct_at(__r, __last1, __last2, __result);
194 
195     __libdispatch::__dispatch_apply(__partitions.__chunk_count_, [&](size_t __index) {
196       auto __first_iters = __ranges[__index];
197       auto __last_iters  = __ranges[__index + 1];
198       __leaf_merge(
199           __first_iters.__mid1_,
200           __last_iters.__mid1_,
201           __first_iters.__mid2_,
202           __last_iters.__mid2_,
203           __first_iters.__result_,
204           __comp);
205     });
206 
207     return __empty{};
208   }
209 
210   template <class _RandomAccessIterator, class _Transform, class _Value, class _Combiner, class _Reduction>
211   _LIBCPP_HIDE_FROM_ABI static optional<_Value> __transform_reduce(
212       _RandomAccessIterator __first,
213       _RandomAccessIterator __last,
214       _Transform __transform,
215       _Value __init,
216       _Combiner __combiner,
217       _Reduction __reduction) {
218     if (__first == __last)
219       return __init;
220 
221     auto __partitions = __libdispatch::__partition_chunks(__last - __first);
222 
223     auto __destroy = [__count = __partitions.__chunk_count_](_Value* __ptr) {
224       std::destroy_n(__ptr, __count);
225       std::allocator<_Value>().deallocate(__ptr, __count);
226     };
227 
228     // TODO: use __uninitialized_buffer
229     // TODO: allocate one element per worker instead of one element per chunk
230     unique_ptr<_Value[], decltype(__destroy)> __values(
231         std::allocator<_Value>().allocate(__partitions.__chunk_count_), __destroy);
232 
233     // __dispatch_apply is noexcept
234     __libdispatch::__dispatch_apply(__partitions.__chunk_count_, [&](size_t __chunk) {
235       auto __this_chunk_size = __chunk == 0 ? __partitions.__first_chunk_size_ : __partitions.__chunk_size_;
236       auto __index           = __chunk == 0 ? 0
237                                             : (__chunk * __partitions.__chunk_size_) +
238                                         (__partitions.__first_chunk_size_ - __partitions.__chunk_size_);
239       if (__this_chunk_size != 1) {
240         std::__construct_at(
241             __values.get() + __chunk,
242             __reduction(__first + __index + 2,
243                         __first + __index + __this_chunk_size,
244                         __combiner(__transform(__first + __index), __transform(__first + __index + 1))));
245       } else {
246         std::__construct_at(__values.get() + __chunk, __transform(__first + __index));
247       }
248     });
249 
250     return std::reduce(
251         std::make_move_iterator(__values.get()),
252         std::make_move_iterator(__values.get() + __partitions.__chunk_count_),
253         std::move(__init),
254         __combiner);
255   }
256 
257   template <class _RandomAccessIterator, class _Comp, class _LeafSort>
258   _LIBCPP_HIDE_FROM_ABI static optional<__empty>
259   __stable_sort(_RandomAccessIterator __first, _RandomAccessIterator __last, _Comp __comp, _LeafSort __leaf_sort) {
260     const auto __size = __last - __first;
261     auto __partitions = __libdispatch::__partition_chunks(__size);
262 
263     if (__partitions.__chunk_count_ == 0)
264       return __empty{};
265 
266     if (__partitions.__chunk_count_ == 1) {
267       __leaf_sort(__first, __last, __comp);
268       return __empty{};
269     }
270 
271     using _Value = __iter_value_type<_RandomAccessIterator>;
272 
273     auto __destroy = [__size](_Value* __ptr) {
274       std::destroy_n(__ptr, __size);
275       std::allocator<_Value>().deallocate(__ptr, __size);
276     };
277 
278     // TODO: use __uninitialized_buffer
279     unique_ptr<_Value[], decltype(__destroy)> __values(std::allocator<_Value>().allocate(__size), __destroy);
280 
281     // Initialize all elements to a moved-from state
282     // TODO: Don't do this - this can be done in the first merge - see https://llvm.org/PR63928
283     std::__construct_at(__values.get(), std::move(*__first));
284     for (__iter_diff_t<_RandomAccessIterator> __i = 1; __i != __size; ++__i) {
285       std::__construct_at(__values.get() + __i, std::move(__values.get()[__i - 1]));
286     }
287     *__first = std::move(__values.get()[__size - 1]);
288 
289     __libdispatch::__dispatch_parallel_for(
290         __partitions,
291         __first,
292         [&__leaf_sort, &__comp](_RandomAccessIterator __chunk_first, _RandomAccessIterator __chunk_last) {
293           __leaf_sort(std::move(__chunk_first), std::move(__chunk_last), __comp);
294         });
295 
296     bool __objects_are_in_buffer = false;
297     do {
298       const auto __old_chunk_size = __partitions.__chunk_size_;
299       if (__partitions.__chunk_count_ % 2 == 1) {
300         auto __inplace_merge_chunks = [&__comp, &__partitions](auto __first_chunk_begin) {
301           std::inplace_merge(
302               __first_chunk_begin,
303               __first_chunk_begin + __partitions.__first_chunk_size_,
304               __first_chunk_begin + __partitions.__first_chunk_size_ + __partitions.__chunk_size_,
305               __comp);
306         };
307         if (__objects_are_in_buffer)
308           __inplace_merge_chunks(__values.get());
309         else
310           __inplace_merge_chunks(__first);
311         __partitions.__first_chunk_size_ += 2 * __partitions.__chunk_size_;
312       } else {
313         __partitions.__first_chunk_size_ += __partitions.__chunk_size_;
314       }
315 
316       __partitions.__chunk_size_ *= 2;
317       __partitions.__chunk_count_ /= 2;
318 
319       auto __merge_chunks = [__partitions, __old_chunk_size, &__comp](auto __from_first, auto __to_first) {
320         __libdispatch::__dispatch_parallel_for(
321             __partitions,
322             __from_first,
323             [__old_chunk_size, &__from_first, &__to_first, &__comp](auto __chunk_first, auto __chunk_last) {
324               std::merge(std::make_move_iterator(__chunk_first),
325                          std::make_move_iterator(__chunk_last - __old_chunk_size),
326                          std::make_move_iterator(__chunk_last - __old_chunk_size),
327                          std::make_move_iterator(__chunk_last),
328                          __to_first + (__chunk_first - __from_first),
329                          __comp);
330             });
331       };
332 
333       if (__objects_are_in_buffer)
334         __merge_chunks(__values.get(), __first);
335       else
336         __merge_chunks(__first, __values.get());
337       __objects_are_in_buffer = !__objects_are_in_buffer;
338     } while (__partitions.__chunk_count_ > 1);
339 
340     if (__objects_are_in_buffer) {
341       std::move(__values.get(), __values.get() + __size, __first);
342     }
343 
344     return __empty{};
345   }
346 
347   _LIBCPP_HIDE_FROM_ABI static void __cancel_execution() {}
348 
349   static constexpr size_t __lane_size = 64;
350 };
351 
352 // Mandatory implementations of the computational basis
353 template <class _ExecutionPolicy>
354 struct __find_if<__libdispatch_backend_tag, _ExecutionPolicy>
355     : __cpu_parallel_find_if<__libdispatch_backend_tag, _ExecutionPolicy> {};
356 
357 template <class _ExecutionPolicy>
358 struct __for_each<__libdispatch_backend_tag, _ExecutionPolicy>
359     : __cpu_parallel_for_each<__libdispatch_backend_tag, _ExecutionPolicy> {};
360 
361 template <class _ExecutionPolicy>
362 struct __merge<__libdispatch_backend_tag, _ExecutionPolicy>
363     : __cpu_parallel_merge<__libdispatch_backend_tag, _ExecutionPolicy> {};
364 
365 template <class _ExecutionPolicy>
366 struct __stable_sort<__libdispatch_backend_tag, _ExecutionPolicy>
367     : __cpu_parallel_stable_sort<__libdispatch_backend_tag, _ExecutionPolicy> {};
368 
369 template <class _ExecutionPolicy>
370 struct __transform<__libdispatch_backend_tag, _ExecutionPolicy>
371     : __cpu_parallel_transform<__libdispatch_backend_tag, _ExecutionPolicy> {};
372 
373 template <class _ExecutionPolicy>
374 struct __transform_binary<__libdispatch_backend_tag, _ExecutionPolicy>
375     : __cpu_parallel_transform_binary<__libdispatch_backend_tag, _ExecutionPolicy> {};
376 
377 template <class _ExecutionPolicy>
378 struct __transform_reduce<__libdispatch_backend_tag, _ExecutionPolicy>
379     : __cpu_parallel_transform_reduce<__libdispatch_backend_tag, _ExecutionPolicy> {};
380 
381 template <class _ExecutionPolicy>
382 struct __transform_reduce_binary<__libdispatch_backend_tag, _ExecutionPolicy>
383     : __cpu_parallel_transform_reduce_binary<__libdispatch_backend_tag, _ExecutionPolicy> {};
384 
385 // Not mandatory, but better optimized
386 template <class _ExecutionPolicy>
387 struct __any_of<__libdispatch_backend_tag, _ExecutionPolicy>
388     : __cpu_parallel_any_of<__libdispatch_backend_tag, _ExecutionPolicy> {};
389 
390 template <class _ExecutionPolicy>
391 struct __fill<__libdispatch_backend_tag, _ExecutionPolicy>
392     : __cpu_parallel_fill<__libdispatch_backend_tag, _ExecutionPolicy> {};
393 
394 } // namespace __pstl
395 _LIBCPP_END_NAMESPACE_STD
396 
397 #endif // _LIBCPP_STD_VER >= 17
398 
399 _LIBCPP_POP_MACROS
400 
401 #endif // _LIBCPP___PSTL_BACKENDS_LIBDISPATCH_H
402