libcxx/include/__pstl/backends/libdispatch.h

   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 <__numeric/reduce.h>
  27 #include <__pstl/backend_fwd.h>
  28 #include <__pstl/cpu_algos/any_of.h>
  29 #include <__pstl/cpu_algos/cpu_traits.h>
  30 #include <__pstl/cpu_algos/fill.h>
  31 #include <__pstl/cpu_algos/find_if.h>
  32 #include <__pstl/cpu_algos/for_each.h>
  33 #include <__pstl/cpu_algos/merge.h>
  34 #include <__pstl/cpu_algos/stable_sort.h>
  35 #include <__pstl/cpu_algos/transform.h>
  36 #include <__pstl/cpu_algos/transform_reduce.h>
  37 #include <__utility/empty.h>
  38 #include <__utility/exception_guard.h>
  39 #include <__utility/move.h>
  40 #include <__utility/pair.h>
  41 #include <new>
  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