__pstl/backends/libdispatch.h

0001 //===----------------------------------------------------------------------===//
0002 //
0003 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
0004 // See https://llvm.org/LICENSE.txt for license information.
0005 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
0006 //
0007 //===----------------------------------------------------------------------===//
0008
0009 #ifndef _LIBCPP___PSTL_BACKENDS_LIBDISPATCH_H
0010 #define _LIBCPP___PSTL_BACKENDS_LIBDISPATCH_H
0011
0012 #include <__algorithm/inplace_merge.h>
0013 #include <__algorithm/lower_bound.h>
0014 #include <__algorithm/max.h>
0015 #include <__algorithm/merge.h>
0016 #include <__algorithm/upper_bound.h>
0017 #include <__atomic/atomic.h>
0018 #include <__config>
0019 #include <__cstddef/ptrdiff_t.h>
0020 #include <__exception/terminate.h>
0021 #include <__iterator/iterator_traits.h>
0022 #include <__iterator/move_iterator.h>
0023 #include <__memory/allocator.h>
0024 #include <__memory/construct_at.h>
0025 #include <__memory/unique_ptr.h>
0026 #include <__new/exceptions.h>
0027 #include <__numeric/reduce.h>
0028 #include <__pstl/backend_fwd.h>
0029 #include <__pstl/cpu_algos/any_of.h>
0030 #include <__pstl/cpu_algos/cpu_traits.h>
0031 #include <__pstl/cpu_algos/fill.h>
0032 #include <__pstl/cpu_algos/find_if.h>
0033 #include <__pstl/cpu_algos/for_each.h>
0034 #include <__pstl/cpu_algos/merge.h>
0035 #include <__pstl/cpu_algos/stable_sort.h>
0036 #include <__pstl/cpu_algos/transform.h>
0037 #include <__pstl/cpu_algos/transform_reduce.h>
0038 #include <__utility/empty.h>
0039 #include <__utility/exception_guard.h>
0040 #include <__utility/move.h>
0041 #include <__utility/pair.h>
0042 #include <optional>
0043
0044 _LIBCPP_PUSH_MACROS
0045 #include <__undef_macros>
0046
0047 #if _LIBCPP_STD_VER >= 17
0048
0049 _LIBCPP_BEGIN_NAMESPACE_STD
0050 namespace __pstl {
0051
0052 namespace __libdispatch {
0053 // ::dispatch_apply is marked as __attribute__((nothrow)) because it doesn't let exceptions propagate, and neither do
0054 // we.
0055 // TODO: Do we want to add [[_Clang::__callback__(__func, __context, __)]]?
0056 _LIBCPP_EXPORTED_FROM_ABI void
0057 __dispatch_apply(size_t __chunk_count, void* __context, void (*__func)(void* __context, size_t __chunk)) noexcept;
0058
0059 template <class _Func>
0060 _LIBCPP_HIDE_FROM_ABI void __dispatch_apply(size_t __chunk_count, _Func __func) noexcept {
0061   __libdispatch::__dispatch_apply(__chunk_count, &__func, [](void* __context, size_t __chunk) {
0062     (*static_cast<_Func*>(__context))(__chunk);
0063   });
0064 }
0065
0066 struct __chunk_partitions {
0067   ptrdiff_t __chunk_count_; // includes the first chunk
0068   ptrdiff_t __chunk_size_;
0069   ptrdiff_t __first_chunk_size_;
0070 };
0071
0072 [[__gnu__::__const__]] _LIBCPP_EXPORTED_FROM_ABI __chunk_partitions __partition_chunks(ptrdiff_t __size) noexcept;
0073
0074 template <class _RandomAccessIterator, class _Functor>
0075 _LIBCPP_HIDE_FROM_ABI optional<__empty>
0076 __dispatch_parallel_for(__chunk_partitions __partitions, _RandomAccessIterator __first, _Functor __func) {
0077   // Perform the chunked execution.
0078   __libdispatch::__dispatch_apply(__partitions.__chunk_count_, [&](size_t __chunk) {
0079     auto __this_chunk_size = __chunk == 0 ? __partitions.__first_chunk_size_ : __partitions.__chunk_size_;
0080     auto __index =
0081         __chunk == 0
0082             ? 0
0083             : (__chunk * __partitions.__chunk_size_) + (__partitions.__first_chunk_size_ - __partitions.__chunk_size_);
0084     __func(__first + __index, __first + __index + __this_chunk_size);
0085   });
0086
0087   return __empty{};
0088 }
0089 } // namespace __libdispatch
0090
0091 template <>
0092 struct __cpu_traits<__libdispatch_backend_tag> {
0093   template <class _RandomAccessIterator, class _Functor>
0094   _LIBCPP_HIDE_FROM_ABI static optional<__empty>
0095   __for_each(_RandomAccessIterator __first, _RandomAccessIterator __last, _Functor __func) {
0096     return __libdispatch::__dispatch_parallel_for(
0097         __libdispatch::__partition_chunks(__last - __first), std::move(__first), std::move(__func));
0098   }
0099
0100   template <class _RandomAccessIterator1, class _RandomAccessIterator2, class _RandomAccessIteratorOut>
0101   struct __merge_range {
0102     __merge_range(_RandomAccessIterator1 __mid1, _RandomAccessIterator2 __mid2, _RandomAccessIteratorOut __result)
0103         : __mid1_(__mid1), __mid2_(__mid2), __result_(__result) {}
0104
0105     _RandomAccessIterator1 __mid1_;
0106     _RandomAccessIterator2 __mid2_;
0107     _RandomAccessIteratorOut __result_;
0108   };
0109
0110   template <typename _RandomAccessIterator1,
0111             typename _RandomAccessIterator2,
0112             typename _RandomAccessIterator3,
0113             typename _Compare,
0114             typename _LeafMerge>
0115   _LIBCPP_HIDE_FROM_ABI static optional<__empty>
0116   __merge(_RandomAccessIterator1 __first1,
0117           _RandomAccessIterator1 __last1,
0118           _RandomAccessIterator2 __first2,
0119           _RandomAccessIterator2 __last2,
0120           _RandomAccessIterator3 __result,
0121           _Compare __comp,
0122           _LeafMerge __leaf_merge) noexcept {
0123     __libdispatch::__chunk_partitions __partitions =
0124         __libdispatch::__partition_chunks(std::max<ptrdiff_t>(__last1 - __first1, __last2 - __first2));
0125
0126     if (__partitions.__chunk_count_ == 0)
0127       return __empty{};
0128
0129     if (__partitions.__chunk_count_ == 1) {
0130       __leaf_merge(__first1, __last1, __first2, __last2, __result, __comp);
0131       return __empty{};
0132     }
0133
0134     using __merge_range_t = __merge_range<_RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator3>;
0135     auto const __n_ranges = __partitions.__chunk_count_ + 1;
0136
0137     // TODO: use __uninitialized_buffer
0138     auto __destroy = [=](__merge_range_t* __ptr) {
0139       std::destroy_n(__ptr, __n_ranges);
0140       std::allocator<__merge_range_t>().deallocate(__ptr, __n_ranges);
0141     };
0142
0143     unique_ptr<__merge_range_t[], decltype(__destroy)> __ranges(
0144         [&]() -> __merge_range_t* {
0145 #  if _LIBCPP_HAS_EXCEPTIONS
0146           try {
0147 #  endif
0148             return std::allocator<__merge_range_t>().allocate(__n_ranges);
0149 #  if _LIBCPP_HAS_EXCEPTIONS
0150           } catch (const std::bad_alloc&) {
0151             return nullptr;
0152           }
0153 #  endif
0154         }(),
0155         __destroy);
0156
0157     if (!__ranges)
0158       return nullopt;
0159
0160     // TODO: Improve the case where the smaller range is merged into just a few (or even one) chunks of the larger case
0161     __merge_range_t* __r = __ranges.get();
0162     std::__construct_at(__r++, __first1, __first2, __result);
0163
0164     bool __iterate_first_range = __last1 - __first1 > __last2 - __first2;
0165
0166     auto __compute_chunk = [&](size_t __chunk_size) -> __merge_range_t {
0167       auto [__mid1, __mid2] = [&] {
0168         if (__iterate_first_range) {
0169           auto __m1 = __first1 + __chunk_size;
0170           auto __m2 = std::lower_bound(__first2, __last2, __m1[-1], __comp);
0171           return std::make_pair(__m1, __m2);
0172         } else {
0173           auto __m2 = __first2 + __chunk_size;
0174           auto __m1 = std::lower_bound(__first1, __last1, __m2[-1], __comp);
0175           return std::make_pair(__m1, __m2);
0176         }
0177       }();
0178
0179       __result += (__mid1 - __first1) + (__mid2 - __first2);
0180       __first1 = __mid1;
0181       __first2 = __mid2;
0182       return {std::move(__mid1), std::move(__mid2), __result};
0183     };
0184
0185     // handle first chunk
0186     std::__construct_at(__r++, __compute_chunk(__partitions.__first_chunk_size_));
0187
0188     // handle 2 -> N - 1 chunks
0189     for (ptrdiff_t __i = 0; __i != __partitions.__chunk_count_ - 2; ++__i)
0190       std::__construct_at(__r++, __compute_chunk(__partitions.__chunk_size_));
0191
0192     // handle last chunk
0193     std::__construct_at(__r, __last1, __last2, __result);
0194
0195     __libdispatch::__dispatch_apply(__partitions.__chunk_count_, [&](size_t __index) {
0196       auto __first_iters = __ranges[__index];
0197       auto __last_iters  = __ranges[__index + 1];
0198       __leaf_merge(
0199           __first_iters.__mid1_,
0200           __last_iters.__mid1_,
0201           __first_iters.__mid2_,
0202           __last_iters.__mid2_,
0203           __first_iters.__result_,
0204           __comp);
0205     });
0206
0207     return __empty{};
0208   }
0209
0210   template <class _RandomAccessIterator, class _Transform, class _Value, class _Combiner, class _Reduction>
0211   _LIBCPP_HIDE_FROM_ABI static optional<_Value> __transform_reduce(
0212       _RandomAccessIterator __first,
0213       _RandomAccessIterator __last,
0214       _Transform __transform,
0215       _Value __init,
0216       _Combiner __combiner,
0217       _Reduction __reduction) {
0218     if (__first == __last)
0219       return __init;
0220
0221     auto __partitions = __libdispatch::__partition_chunks(__last - __first);
0222
0223     auto __destroy = [__count = __partitions.__chunk_count_](_Value* __ptr) {
0224       std::destroy_n(__ptr, __count);
0225       std::allocator<_Value>().deallocate(__ptr, __count);
0226     };
0227
0228     // TODO: use __uninitialized_buffer
0229     // TODO: allocate one element per worker instead of one element per chunk
0230     unique_ptr<_Value[], decltype(__destroy)> __values(
0231         std::allocator<_Value>().allocate(__partitions.__chunk_count_), __destroy);
0232
0233     // __dispatch_apply is noexcept
0234     __libdispatch::__dispatch_apply(__partitions.__chunk_count_, [&](size_t __chunk) {
0235       auto __this_chunk_size = __chunk == 0 ? __partitions.__first_chunk_size_ : __partitions.__chunk_size_;
0236       auto __index           = __chunk == 0 ? 0
0237                                             : (__chunk * __partitions.__chunk_size_) +
0238                                         (__partitions.__first_chunk_size_ - __partitions.__chunk_size_);
0239       if (__this_chunk_size != 1) {
0240         std::__construct_at(
0241             __values.get() + __chunk,
0242             __reduction(__first + __index + 2,
0243                         __first + __index + __this_chunk_size,
0244                         __combiner(__transform(__first + __index), __transform(__first + __index + 1))));
0245       } else {
0246         std::__construct_at(__values.get() + __chunk, __transform(__first + __index));
0247       }
0248     });
0249
0250     return std::reduce(
0251         std::make_move_iterator(__values.get()),
0252         std::make_move_iterator(__values.get() + __partitions.__chunk_count_),
0253         std::move(__init),
0254         __combiner);
0255   }
0256
0257   template <class _RandomAccessIterator, class _Comp, class _LeafSort>
0258   _LIBCPP_HIDE_FROM_ABI static optional<__empty>
0259   __stable_sort(_RandomAccessIterator __first, _RandomAccessIterator __last, _Comp __comp, _LeafSort __leaf_sort) {
0260     const auto __size = __last - __first;
0261     auto __partitions = __libdispatch::__partition_chunks(__size);
0262
0263     if (__partitions.__chunk_count_ == 0)
0264       return __empty{};
0265
0266     if (__partitions.__chunk_count_ == 1) {
0267       __leaf_sort(__first, __last, __comp);
0268       return __empty{};
0269     }
0270
0271     using _Value = __iter_value_type<_RandomAccessIterator>;
0272
0273     auto __destroy = [__size](_Value* __ptr) {
0274       std::destroy_n(__ptr, __size);
0275       std::allocator<_Value>().deallocate(__ptr, __size);
0276     };
0277
0278     // TODO: use __uninitialized_buffer
0279     unique_ptr<_Value[], decltype(__destroy)> __values(std::allocator<_Value>().allocate(__size), __destroy);
0280
0281     // Initialize all elements to a moved-from state
0282     // TODO: Don't do this - this can be done in the first merge - see https://llvm.org/PR63928
0283     std::__construct_at(__values.get(), std::move(*__first));
0284     for (__iter_diff_t<_RandomAccessIterator> __i = 1; __i != __size; ++__i) {
0285       std::__construct_at(__values.get() + __i, std::move(__values.get()[__i - 1]));
0286     }
0287     *__first = std::move(__values.get()[__size - 1]);
0288
0289     __libdispatch::__dispatch_parallel_for(
0290         __partitions,
0291         __first,
0292         [&__leaf_sort, &__comp](_RandomAccessIterator __chunk_first, _RandomAccessIterator __chunk_last) {
0293           __leaf_sort(std::move(__chunk_first), std::move(__chunk_last), __comp);
0294         });
0295
0296     bool __objects_are_in_buffer = false;
0297     do {
0298       const auto __old_chunk_size = __partitions.__chunk_size_;
0299       if (__partitions.__chunk_count_ % 2 == 1) {
0300         auto __inplace_merge_chunks = [&__comp, &__partitions](auto __first_chunk_begin) {
0301           std::inplace_merge(
0302               __first_chunk_begin,
0303               __first_chunk_begin + __partitions.__first_chunk_size_,
0304               __first_chunk_begin + __partitions.__first_chunk_size_ + __partitions.__chunk_size_,
0305               __comp);
0306         };
0307         if (__objects_are_in_buffer)
0308           __inplace_merge_chunks(__values.get());
0309         else
0310           __inplace_merge_chunks(__first);
0311         __partitions.__first_chunk_size_ += 2 * __partitions.__chunk_size_;
0312       } else {
0313         __partitions.__first_chunk_size_ += __partitions.__chunk_size_;
0314       }
0315
0316       __partitions.__chunk_size_ *= 2;
0317       __partitions.__chunk_count_ /= 2;
0318
0319       auto __merge_chunks = [__partitions, __old_chunk_size, &__comp](auto __from_first, auto __to_first) {
0320         __libdispatch::__dispatch_parallel_for(
0321             __partitions,
0322             __from_first,
0323             [__old_chunk_size, &__from_first, &__to_first, &__comp](auto __chunk_first, auto __chunk_last) {
0324               std::merge(std::make_move_iterator(__chunk_first),
0325                          std::make_move_iterator(__chunk_last - __old_chunk_size),
0326                          std::make_move_iterator(__chunk_last - __old_chunk_size),
0327                          std::make_move_iterator(__chunk_last),
0328                          __to_first + (__chunk_first - __from_first),
0329                          __comp);
0330             });
0331       };
0332
0333       if (__objects_are_in_buffer)
0334         __merge_chunks(__values.get(), __first);
0335       else
0336         __merge_chunks(__first, __values.get());
0337       __objects_are_in_buffer = !__objects_are_in_buffer;
0338     } while (__partitions.__chunk_count_ > 1);
0339
0340     if (__objects_are_in_buffer) {
0341       std::move(__values.get(), __values.get() + __size, __first);
0342     }
0343
0344     return __empty{};
0345   }
0346
0347   _LIBCPP_HIDE_FROM_ABI static void __cancel_execution() {}
0348
0349   static constexpr size_t __lane_size = 64;
0350 };
0351
0352 // Mandatory implementations of the computational basis
0353 template <class _ExecutionPolicy>
0354 struct __find_if<__libdispatch_backend_tag, _ExecutionPolicy>
0355     : __cpu_parallel_find_if<__libdispatch_backend_tag, _ExecutionPolicy> {};
0356
0357 template <class _ExecutionPolicy>
0358 struct __for_each<__libdispatch_backend_tag, _ExecutionPolicy>
0359     : __cpu_parallel_for_each<__libdispatch_backend_tag, _ExecutionPolicy> {};
0360
0361 template <class _ExecutionPolicy>
0362 struct __merge<__libdispatch_backend_tag, _ExecutionPolicy>
0363     : __cpu_parallel_merge<__libdispatch_backend_tag, _ExecutionPolicy> {};
0364
0365 template <class _ExecutionPolicy>
0366 struct __stable_sort<__libdispatch_backend_tag, _ExecutionPolicy>
0367     : __cpu_parallel_stable_sort<__libdispatch_backend_tag, _ExecutionPolicy> {};
0368
0369 template <class _ExecutionPolicy>
0370 struct __transform<__libdispatch_backend_tag, _ExecutionPolicy>
0371     : __cpu_parallel_transform<__libdispatch_backend_tag, _ExecutionPolicy> {};
0372
0373 template <class _ExecutionPolicy>
0374 struct __transform_binary<__libdispatch_backend_tag, _ExecutionPolicy>
0375     : __cpu_parallel_transform_binary<__libdispatch_backend_tag, _ExecutionPolicy> {};
0376
0377 template <class _ExecutionPolicy>
0378 struct __transform_reduce<__libdispatch_backend_tag, _ExecutionPolicy>
0379     : __cpu_parallel_transform_reduce<__libdispatch_backend_tag, _ExecutionPolicy> {};
0380
0381 template <class _ExecutionPolicy>
0382 struct __transform_reduce_binary<__libdispatch_backend_tag, _ExecutionPolicy>
0383     : __cpu_parallel_transform_reduce_binary<__libdispatch_backend_tag, _ExecutionPolicy> {};
0384
0385 // Not mandatory, but better optimized
0386 template <class _ExecutionPolicy>
0387 struct __any_of<__libdispatch_backend_tag, _ExecutionPolicy>
0388     : __cpu_parallel_any_of<__libdispatch_backend_tag, _ExecutionPolicy> {};
0389
0390 template <class _ExecutionPolicy>
0391 struct __fill<__libdispatch_backend_tag, _ExecutionPolicy>
0392     : __cpu_parallel_fill<__libdispatch_backend_tag, _ExecutionPolicy> {};
0393
0394 } // namespace __pstl
0395 _LIBCPP_END_NAMESPACE_STD
0396
0397 #endif // _LIBCPP_STD_VER >= 17
0398
0399 _LIBCPP_POP_MACROS
0400
0401 #endif // _LIBCPP___PSTL_BACKENDS_LIBDISPATCH_H