1 /* 2 Copyright (c) 2005-2021 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_sort_H 18 #define __TBB_parallel_sort_H 19 20 #include "detail/_namespace_injection.h" 21 #include "parallel_for.h" 22 #include "blocked_range.h" 23 #include "profiling.h" 24 25 #include <algorithm> 26 #include <iterator> 27 #include <functional> 28 #include <cstddef> 29 30 namespace tbb { 31 namespace detail { 32 #if __TBB_CPP20_CONCEPTS_PRESENT 33 inline namespace d0 { 34 35 // TODO: consider using std::strict_weak_order concept 36 template <typename Compare, typename Iterator> 37 concept compare = requires( const std::remove_reference_t<Compare>& comp, typename std::iterator_traits<Iterator>::reference value ) { 38 // Forward via iterator_traits::reference 39 { comp(typename std::iterator_traits<Iterator>::reference(value), 40 typename std::iterator_traits<Iterator>::reference(value)) } -> std::convertible_to<bool>; 41 }; 42 43 // Inspired by std::__PartiallyOrderedWith exposition only concept 44 template <typename T> 45 concept less_than_comparable = requires( const std::remove_reference_t<T>& lhs, 46 const std::remove_reference_t<T>& rhs ) { 47 { lhs < rhs } -> boolean_testable; 48 }; 49 50 } // namespace d0 51 #endif // __TBB_CPP20_CONCEPTS_PRESENT 52 namespace d1 { 53 54 //! Range used in quicksort to split elements into subranges based on a value. 55 /** The split operation selects a splitter and places all elements less than or equal 56 to the value in the first range and the remaining elements in the second range. 57 @ingroup algorithms */ 58 template<typename RandomAccessIterator, typename Compare> 59 class quick_sort_range { 60 std::size_t median_of_three( const RandomAccessIterator& array, std::size_t l, std::size_t m, std::size_t r ) const { 61 return comp(array[l], array[m]) ? ( comp(array[m], array[r]) ? m : ( comp(array[l], array[r]) ? r : l ) ) 62 : ( comp(array[r], array[m]) ? m : ( comp(array[r], array[l]) ? r : l ) ); 63 } 64 65 std::size_t pseudo_median_of_nine( const RandomAccessIterator& array, const quick_sort_range& range ) const { 66 std::size_t offset = range.size / 8u; 67 return median_of_three(array, 68 median_of_three(array, 0 , offset, offset * 2), 69 median_of_three(array, offset * 3, offset * 4, offset * 5), 70 median_of_three(array, offset * 6, offset * 7, range.size - 1)); 71 72 } 73 74 std::size_t split_range( quick_sort_range& range ) { 75 RandomAccessIterator array = range.begin; 76 RandomAccessIterator first_element = range.begin; 77 std::size_t m = pseudo_median_of_nine(array, range); 78 if( m != 0 ) std::iter_swap(array, array + m); 79 80 std::size_t i = 0; 81 std::size_t j = range.size; 82 // Partition interval [i + 1,j - 1] with key *first_element. 83 for(;;) { 84 __TBB_ASSERT( i < j, nullptr ); 85 // Loop must terminate since array[l] == *first_element. 86 do { 87 --j; 88 __TBB_ASSERT( i <= j, "bad ordering relation?" ); 89 } while( comp(*first_element, array[j]) ); 90 do { 91 __TBB_ASSERT( i <= j, nullptr ); 92 if( i == j ) goto partition; 93 ++i; 94 } while( comp(array[i], *first_element) ); 95 if( i == j ) goto partition; 96 std::iter_swap(array + i, array + j); 97 } 98 partition: 99 // Put the partition key were it belongs 100 std::iter_swap(array + j, first_element); 101 // array[l..j) is less or equal to key. 102 // array(j..r) is greater or equal to key. 103 // array[j] is equal to key 104 i = j + 1; 105 std::size_t new_range_size = range.size - i; 106 range.size = j; 107 return new_range_size; 108 } 109 110 public: 111 quick_sort_range() = default; 112 quick_sort_range( const quick_sort_range& ) = default; 113 void operator=( const quick_sort_range& ) = delete; 114 115 static constexpr std::size_t grainsize = 500; 116 const Compare& comp; 117 std::size_t size; 118 RandomAccessIterator begin; 119 120 quick_sort_range( RandomAccessIterator begin_, std::size_t size_, const Compare& comp_ ) : 121 comp(comp_), size(size_), begin(begin_) {} 122 123 bool empty() const { return size == 0; } 124 bool is_divisible() const { return size >= grainsize; } 125 126 quick_sort_range( quick_sort_range& range, split ) 127 : comp(range.comp) 128 , size(split_range(range)) 129 // +1 accounts for the pivot element, which is at its correct place 130 // already and, therefore, is not included into subranges. 131 , begin(range.begin + range.size + 1) {} 132 }; 133 134 //! Body class used to test if elements in a range are presorted 135 /** @ingroup algorithms */ 136 template<typename RandomAccessIterator, typename Compare> 137 class quick_sort_pretest_body { 138 const Compare& comp; 139 task_group_context& context; 140 141 public: 142 quick_sort_pretest_body() = default; 143 quick_sort_pretest_body( const quick_sort_pretest_body& ) = default; 144 void operator=( const quick_sort_pretest_body& ) = delete; 145 146 quick_sort_pretest_body( const Compare& _comp, task_group_context& _context ) : comp(_comp), context(_context) {} 147 148 void operator()( const blocked_range<RandomAccessIterator>& range ) const { 149 RandomAccessIterator my_end = range.end(); 150 151 int i = 0; 152 //TODO: consider using std::is_sorted() for each 64 iterations (requires performance measurements) 153 for( RandomAccessIterator k = range.begin(); k != my_end; ++k, ++i ) { 154 if( i % 64 == 0 && context.is_group_execution_cancelled() ) break; 155 156 // The k - 1 is never out-of-range because the first chunk starts at begin+serial_cutoff+1 157 if( comp(*(k), *(k - 1)) ) { 158 context.cancel_group_execution(); 159 break; 160 } 161 } 162 } 163 }; 164 165 //! Body class used to sort elements in a range that is smaller than the grainsize. 166 /** @ingroup algorithms */ 167 template<typename RandomAccessIterator, typename Compare> 168 struct quick_sort_body { 169 void operator()( const quick_sort_range<RandomAccessIterator,Compare>& range ) const { 170 std::sort(range.begin, range.begin + range.size, range.comp); 171 } 172 }; 173 174 //! Method to perform parallel_for based quick sort. 175 /** @ingroup algorithms */ 176 template<typename RandomAccessIterator, typename Compare> 177 void do_parallel_quick_sort( RandomAccessIterator begin, RandomAccessIterator end, const Compare& comp ) { 178 parallel_for(quick_sort_range<RandomAccessIterator,Compare>(begin, end - begin, comp), 179 quick_sort_body<RandomAccessIterator,Compare>(), 180 auto_partitioner()); 181 } 182 183 //! Wrapper method to initiate the sort by calling parallel_for. 184 /** @ingroup algorithms */ 185 template<typename RandomAccessIterator, typename Compare> 186 void parallel_quick_sort( RandomAccessIterator begin, RandomAccessIterator end, const Compare& comp ) { 187 task_group_context my_context(PARALLEL_SORT); 188 constexpr int serial_cutoff = 9; 189 190 __TBB_ASSERT( begin + serial_cutoff < end, "min_parallel_size is smaller than serial cutoff?" ); 191 RandomAccessIterator k = begin; 192 for( ; k != begin + serial_cutoff; ++k ) { 193 if( comp(*(k + 1), *k) ) { 194 do_parallel_quick_sort(begin, end, comp); 195 } 196 } 197 198 // Check is input range already sorted 199 parallel_for(blocked_range<RandomAccessIterator>(k + 1, end), 200 quick_sort_pretest_body<RandomAccessIterator, Compare>(comp, my_context), 201 auto_partitioner(), 202 my_context); 203 204 if( my_context.is_group_execution_cancelled() ) 205 do_parallel_quick_sort(begin, end, comp); 206 } 207 208 /** \page parallel_sort_iter_req Requirements on iterators for parallel_sort 209 Requirements on the iterator type \c It and its value type \c T for \c parallel_sort: 210 211 - \code void iter_swap( It a, It b ) \endcode Swaps the values of the elements the given 212 iterators \c a and \c b are pointing to. \c It should be a random access iterator. 213 214 - \code bool Compare::operator()( const T& x, const T& y ) \endcode True if x comes before y; 215 **/ 216 217 /** \name parallel_sort 218 See also requirements on \ref parallel_sort_iter_req "iterators for parallel_sort". **/ 219 //@{ 220 221 //! Sorts the data in [begin,end) using the given comparator 222 /** The compare function object is used for all comparisons between elements during sorting. 223 The compare object must define a bool operator() function. 224 @ingroup algorithms **/ 225 template<typename RandomAccessIterator, typename Compare> 226 __TBB_requires(std::random_access_iterator<RandomAccessIterator> && 227 compare<Compare, RandomAccessIterator>) 228 void parallel_sort( RandomAccessIterator begin, RandomAccessIterator end, const Compare& comp ) { 229 constexpr int min_parallel_size = 500; 230 if( end > begin ) { 231 if( end - begin < min_parallel_size ) { 232 std::sort(begin, end, comp); 233 } else { 234 parallel_quick_sort(begin, end, comp); 235 } 236 } 237 } 238 239 //! Sorts the data in [begin,end) with a default comparator \c std::less 240 /** @ingroup algorithms **/ 241 template<typename RandomAccessIterator> 242 __TBB_requires(std::random_access_iterator<RandomAccessIterator> && 243 less_than_comparable<typename std::iterator_traits<RandomAccessIterator>::value_type>) 244 void parallel_sort( RandomAccessIterator begin, RandomAccessIterator end ) { 245 parallel_sort(begin, end, std::less<typename std::iterator_traits<RandomAccessIterator>::value_type>()); 246 } 247 248 //! Sorts the data in rng using the given comparator 249 /** @ingroup algorithms **/ 250 template<typename Range, typename Compare> 251 __TBB_requires(container_based_sequence<Range, std::random_access_iterator_tag> && 252 compare<Compare, range_iterator_type<Range>>) 253 void parallel_sort( Range&& rng, const Compare& comp ) { 254 parallel_sort(std::begin(rng), std::end(rng), comp); 255 } 256 257 //! Sorts the data in rng with a default comparator \c std::less 258 /** @ingroup algorithms **/ 259 template<typename Range> 260 __TBB_requires(container_based_sequence<Range, std::random_access_iterator_tag> && 261 less_than_comparable<typename std::iterator_traits<range_iterator_type<Range>>::value_type>) 262 void parallel_sort( Range&& rng ) { 263 parallel_sort(std::begin(rng), std::end(rng)); 264 } 265 //@} 266 267 } // namespace d1 268 } // namespace detail 269 270 inline namespace v1 { 271 using detail::d1::parallel_sort; 272 } // namespace v1 273 } // namespace tbb 274 275 #endif /*__TBB_parallel_sort_H*/ 276