From 2b6d1facbba1b3e7a95562574c50064f74560480 Mon Sep 17 00:00:00 2001 From: Paolo Carlini Date: Fri, 6 Nov 2009 11:39:35 +0000 Subject: 2009-11-06 Paolo Carlini * include/parallel/multiway_merge.h: Simple formatting and uglification fixes. * include/parallel/find_selectors.h: Likewise. * include/parallel/losertree.h: Likewise. * include/parallel/list_partition.h: Likewise. * include/parallel/for_each.h: Likewise. * include/parallel/multiseq_selection.h: Likewise. * include/parallel/workstealing.h: Likewise. * include/parallel/par_loop.h: Likewise. * include/parallel/numeric: Likewise. * include/parallel/quicksort.h: Likewise. * include/parallel/equally_split.h: Likewise. * include/parallel/omp_loop_static.h: Likewise. * include/parallel/random_shuffle.h: Likewise. * include/parallel/balanced_quicksort.h: Likewise. * include/parallel/tags.h: Likewise. * include/parallel/set_operations.h: Likewise. * include/parallel/merge.h: Likewise. * include/parallel/unique_copy.h: Likewise. * include/parallel/multiway_mergesort.h: Likewise. * include/parallel/search.h: Likewise. * include/parallel/partition.h: Likewise. * include/parallel/partial_sum.h: Likewise. * include/parallel/find.h: Likewise. * include/parallel/queue.h: Likewise. * include/parallel/omp_loop.h: Likewise. * include/parallel/checkers.h: Likewise. * include/parallel/sort.h: Likewise. git-svn-id: https://gcc.gnu.org/svn/gcc/trunk@153966 138bc75d-0d04-0410-961f-82ee72b054a4 --- libstdc++-v3/ChangeLog | 31 + libstdc++-v3/include/parallel/balanced_quicksort.h | 753 ++-- libstdc++-v3/include/parallel/checkers.h | 1 + libstdc++-v3/include/parallel/equally_split.h | 99 +- libstdc++-v3/include/parallel/find.h | 633 ++-- libstdc++-v3/include/parallel/find_selectors.h | 97 +- libstdc++-v3/include/parallel/for_each.h | 34 +- libstdc++-v3/include/parallel/list_partition.h | 12 +- libstdc++-v3/include/parallel/losertree.h | 1799 +++++----- libstdc++-v3/include/parallel/merge.h | 78 +- libstdc++-v3/include/parallel/multiseq_selection.h | 24 +- libstdc++-v3/include/parallel/multiway_merge.h | 3699 ++++++++++---------- libstdc++-v3/include/parallel/multiway_mergesort.h | 84 +- libstdc++-v3/include/parallel/numeric | 2 +- libstdc++-v3/include/parallel/omp_loop.h | 8 +- libstdc++-v3/include/parallel/omp_loop_static.h | 4 +- libstdc++-v3/include/parallel/par_loop.h | 13 +- libstdc++-v3/include/parallel/partial_sum.h | 9 +- libstdc++-v3/include/parallel/partition.h | 726 ++-- libstdc++-v3/include/parallel/queue.h | 30 +- libstdc++-v3/include/parallel/quicksort.h | 56 +- libstdc++-v3/include/parallel/random_shuffle.h | 844 +++-- libstdc++-v3/include/parallel/search.h | 194 +- libstdc++-v3/include/parallel/set_operations.h | 841 +++-- libstdc++-v3/include/parallel/sort.h | 120 +- libstdc++-v3/include/parallel/tags.h | 64 +- libstdc++-v3/include/parallel/unique_copy.h | 279 +- libstdc++-v3/include/parallel/workstealing.h | 435 +-- 28 files changed, 5472 insertions(+), 5497 deletions(-) diff --git a/libstdc++-v3/ChangeLog b/libstdc++-v3/ChangeLog index 6650087a3df..ec39612f8f5 100644 --- a/libstdc++-v3/ChangeLog +++ b/libstdc++-v3/ChangeLog @@ -1,3 +1,34 @@ +2009-11-06 Paolo Carlini + + * include/parallel/multiway_merge.h: Simple formatting and + uglification fixes. + * include/parallel/find_selectors.h: Likewise. + * include/parallel/losertree.h: Likewise. + * include/parallel/list_partition.h: Likewise. + * include/parallel/for_each.h: Likewise. + * include/parallel/multiseq_selection.h: Likewise. + * include/parallel/workstealing.h: Likewise. + * include/parallel/par_loop.h: Likewise. + * include/parallel/numeric: Likewise. + * include/parallel/quicksort.h: Likewise. + * include/parallel/equally_split.h: Likewise. + * include/parallel/omp_loop_static.h: Likewise. + * include/parallel/random_shuffle.h: Likewise. + * include/parallel/balanced_quicksort.h: Likewise. + * include/parallel/tags.h: Likewise. + * include/parallel/set_operations.h: Likewise. + * include/parallel/merge.h: Likewise. + * include/parallel/unique_copy.h: Likewise. + * include/parallel/multiway_mergesort.h: Likewise. + * include/parallel/search.h: Likewise. + * include/parallel/partition.h: Likewise. + * include/parallel/partial_sum.h: Likewise. + * include/parallel/find.h: Likewise. + * include/parallel/queue.h: Likewise. + * include/parallel/omp_loop.h: Likewise. + * include/parallel/checkers.h: Likewise. + * include/parallel/sort.h: Likewise. + 2009-11-06 Jonathan Wakely PR libstdc++/41949 diff --git a/libstdc++-v3/include/parallel/balanced_quicksort.h b/libstdc++-v3/include/parallel/balanced_quicksort.h index 85e4d699e26..3107d97a45b 100644 --- a/libstdc++-v3/include/parallel/balanced_quicksort.h +++ b/libstdc++-v3/include/parallel/balanced_quicksort.h @@ -57,436 +57,435 @@ namespace __gnu_parallel { -/** @brief Information local to one thread in the parallel quicksort run. */ -template - struct _QSBThreadLocal - { - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::difference_type _DifferenceType; - - /** @brief Continuous part of the sequence, described by an - iterator pair. */ - typedef std::pair<_RAIter, _RAIter> _Piece; - - /** @brief Initial piece to work on. */ - _Piece _M_initial; - - /** @brief Work-stealing queue. */ - _RestrictedBoundedConcurrentQueue<_Piece> _M_leftover_parts; - - /** @brief Number of threads involved in this algorithm. */ - _ThreadIndex _M_num_threads; - - /** @brief Pointer to a counter of elements left over to sort. */ - volatile _DifferenceType* _M_elements_leftover; - - /** @brief The complete sequence to sort. */ - _Piece _M_global; - - /** @brief Constructor. - * @param __queue_size size of the work-stealing queue. */ - _QSBThreadLocal(int __queue_size) : _M_leftover_parts(__queue_size) { } - }; - -/** @brief Balanced quicksort divide step. - * @param __begin Begin iterator of subsequence. - * @param __end End iterator of subsequence. - * @param __comp Comparator. - * @param __num_threads Number of threads that are allowed to work on - * this part. - * @pre @__c (__end-__begin)>=1 */ -template - typename std::iterator_traits<_RAIter>::difference_type - __qsb_divide(_RAIter __begin, _RAIter __end, - _Compare __comp, _ThreadIndex __num_threads) - { - _GLIBCXX_PARALLEL_ASSERT(__num_threads > 0); - - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::value_type _ValueType; - typedef typename _TraitsType::difference_type _DifferenceType; - - _RAIter __pivot_pos = - __median_of_three_iterators(__begin, __begin + (__end - __begin) / 2, - __end - 1, __comp); + /** @brief Information local to one thread in the parallel quicksort run. */ + template + struct _QSBThreadLocal + { + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::difference_type _DifferenceType; + + /** @brief Continuous part of the sequence, described by an + iterator pair. */ + typedef std::pair<_RAIter, _RAIter> _Piece; + + /** @brief Initial piece to work on. */ + _Piece _M_initial; + + /** @brief Work-stealing queue. */ + _RestrictedBoundedConcurrentQueue<_Piece> _M_leftover_parts; + + /** @brief Number of threads involved in this algorithm. */ + _ThreadIndex _M_num_threads; + + /** @brief Pointer to a counter of elements left over to sort. */ + volatile _DifferenceType* _M_elements_leftover; + + /** @brief The complete sequence to sort. */ + _Piece _M_global; + + /** @brief Constructor. + * @param __queue_size size of the work-stealing queue. */ + _QSBThreadLocal(int __queue_size) : _M_leftover_parts(__queue_size) { } + }; + + /** @brief Balanced quicksort divide step. + * @param __begin Begin iterator of subsequence. + * @param __end End iterator of subsequence. + * @param __comp Comparator. + * @param __num_threads Number of threads that are allowed to work on + * this part. + * @pre @__c (__end-__begin)>=1 */ + template + typename std::iterator_traits<_RAIter>::difference_type + __qsb_divide(_RAIter __begin, _RAIter __end, + _Compare __comp, _ThreadIndex __num_threads) + { + _GLIBCXX_PARALLEL_ASSERT(__num_threads > 0); + + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::value_type _ValueType; + typedef typename _TraitsType::difference_type _DifferenceType; + + _RAIter __pivot_pos = + __median_of_three_iterators(__begin, __begin + (__end - __begin) / 2, + __end - 1, __comp); #if defined(_GLIBCXX_ASSERTIONS) - // Must be in between somewhere. - _DifferenceType __n = __end - __begin; - - _GLIBCXX_PARALLEL_ASSERT( - (!__comp(*__pivot_pos, *__begin) && - !__comp(*(__begin + __n / 2), *__pivot_pos)) - || (!__comp(*__pivot_pos, *__begin) && - !__comp(*(__end - 1), *__pivot_pos)) - || (!__comp(*__pivot_pos, *(__begin + __n / 2)) && - !__comp(*__begin, *__pivot_pos)) - || (!__comp(*__pivot_pos, *(__begin + __n / 2)) && - !__comp(*(__end - 1), *__pivot_pos)) - || (!__comp(*__pivot_pos, *(__end - 1)) && - !__comp(*__begin, *__pivot_pos)) - || (!__comp(*__pivot_pos, *(__end - 1)) && - !__comp(*(__begin + __n / 2), *__pivot_pos))); + // Must be in between somewhere. + _DifferenceType __n = __end - __begin; + + _GLIBCXX_PARALLEL_ASSERT((!__comp(*__pivot_pos, *__begin) + && !__comp(*(__begin + __n / 2), + *__pivot_pos)) + || (!__comp(*__pivot_pos, *__begin) + && !__comp(*(__end - 1), *__pivot_pos)) + || (!__comp(*__pivot_pos, *(__begin + __n / 2)) + && !__comp(*__begin, *__pivot_pos)) + || (!__comp(*__pivot_pos, *(__begin + __n / 2)) + && !__comp(*(__end - 1), *__pivot_pos)) + || (!__comp(*__pivot_pos, *(__end - 1)) + && !__comp(*__begin, *__pivot_pos)) + || (!__comp(*__pivot_pos, *(__end - 1)) + && !__comp(*(__begin + __n / 2), + *__pivot_pos))); #endif - // Swap pivot value to end. - if (__pivot_pos != (__end - 1)) - std::swap(*__pivot_pos, *(__end - 1)); - __pivot_pos = __end - 1; + // Swap pivot value to end. + if (__pivot_pos != (__end - 1)) + std::swap(*__pivot_pos, *(__end - 1)); + __pivot_pos = __end - 1; - __gnu_parallel::binder2nd<_Compare, _ValueType, _ValueType, bool> - __pred(__comp, *__pivot_pos); + __gnu_parallel::binder2nd<_Compare, _ValueType, _ValueType, bool> + __pred(__comp, *__pivot_pos); - // Divide, returning __end - __begin - 1 in the worst case. - _DifferenceType __split_pos = __parallel_partition( - __begin, __end - 1, __pred, __num_threads); + // Divide, returning __end - __begin - 1 in the worst case. + _DifferenceType __split_pos = __parallel_partition(__begin, __end - 1, + __pred, + __num_threads); - // Swap back pivot to middle. - std::swap(*(__begin + __split_pos), *__pivot_pos); - __pivot_pos = __begin + __split_pos; + // Swap back pivot to middle. + std::swap(*(__begin + __split_pos), *__pivot_pos); + __pivot_pos = __begin + __split_pos; #if _GLIBCXX_ASSERTIONS - _RAIter __r; - for (__r = __begin; __r != __pivot_pos; ++__r) - _GLIBCXX_PARALLEL_ASSERT(__comp(*__r, *__pivot_pos)); - for (; __r != __end; ++__r) - _GLIBCXX_PARALLEL_ASSERT(!__comp(*__r, *__pivot_pos)); + _RAIter __r; + for (__r = __begin; __r != __pivot_pos; ++__r) + _GLIBCXX_PARALLEL_ASSERT(__comp(*__r, *__pivot_pos)); + for (; __r != __end; ++__r) + _GLIBCXX_PARALLEL_ASSERT(!__comp(*__r, *__pivot_pos)); #endif - return __split_pos; - } - -/** @brief Quicksort conquer step. - * @param __tls Array of thread-local storages. - * @param __begin Begin iterator of subsequence. - * @param __end End iterator of subsequence. - * @param __comp Comparator. - * @param __iam Number of the thread processing this function. - * @param __num_threads - * Number of threads that are allowed to work on this part. */ -template - void - __qsb_conquer(_QSBThreadLocal<_RAIter>** __tls, - _RAIter __begin, _RAIter __end, - _Compare __comp, - _ThreadIndex __iam, _ThreadIndex __num_threads, - bool __parent_wait) - { - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::value_type _ValueType; - typedef typename _TraitsType::difference_type _DifferenceType; - - _DifferenceType __n = __end - __begin; - - if (__num_threads <= 1 || __n <= 1) - { - __tls[__iam]->_M_initial.first = __begin; - __tls[__iam]->_M_initial.second = __end; + return __split_pos; + } - __qsb_local_sort_with_helping(__tls, __comp, __iam, __parent_wait); + /** @brief Quicksort conquer step. + * @param __tls Array of thread-local storages. + * @param __begin Begin iterator of subsequence. + * @param __end End iterator of subsequence. + * @param __comp Comparator. + * @param __iam Number of the thread processing this function. + * @param __num_threads + * Number of threads that are allowed to work on this part. */ + template + void + __qsb_conquer(_QSBThreadLocal<_RAIter>** __tls, + _RAIter __begin, _RAIter __end, + _Compare __comp, + _ThreadIndex __iam, _ThreadIndex __num_threads, + bool __parent_wait) + { + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::value_type _ValueType; + typedef typename _TraitsType::difference_type _DifferenceType; - return; - } + _DifferenceType __n = __end - __begin; + + if (__num_threads <= 1 || __n <= 1) + { + __tls[__iam]->_M_initial.first = __begin; + __tls[__iam]->_M_initial.second = __end; + + __qsb_local_sort_with_helping(__tls, __comp, __iam, __parent_wait); + + return; + } - // Divide step. - _DifferenceType __split_pos = - __qsb_divide(__begin, __end, __comp, __num_threads); + // Divide step. + _DifferenceType __split_pos = + __qsb_divide(__begin, __end, __comp, __num_threads); #if _GLIBCXX_ASSERTIONS - _GLIBCXX_PARALLEL_ASSERT(0 <= __split_pos && - __split_pos < (__end - __begin)); + _GLIBCXX_PARALLEL_ASSERT(0 <= __split_pos && + __split_pos < (__end - __begin)); #endif - _ThreadIndex __num_threads_leftside = - std::max<_ThreadIndex>(1, std::min<_ThreadIndex>( - __num_threads - 1, __split_pos * __num_threads / __n)); + _ThreadIndex + __num_threads_leftside = std::max<_ThreadIndex> + (1, std::min<_ThreadIndex>(__num_threads - 1, __split_pos + * __num_threads / __n)); -# pragma omp atomic - *__tls[__iam]->_M_elements_leftover -= (_DifferenceType)1; +# pragma omp atomic + *__tls[__iam]->_M_elements_leftover -= (_DifferenceType)1; - // Conquer step. -# pragma omp parallel num_threads(2) - { - bool __wait; - if(omp_get_num_threads() < 2) - __wait = false; - else - __wait = __parent_wait; - -# pragma omp sections - { + // Conquer step. +# pragma omp parallel num_threads(2) + { + bool __wait; + if(omp_get_num_threads() < 2) + __wait = false; + else + __wait = __parent_wait; + +# pragma omp sections + { # pragma omp section - { - __qsb_conquer(__tls, __begin, __begin + __split_pos, __comp, - __iam, - __num_threads_leftside, - __wait); - __wait = __parent_wait; - } - // The pivot_pos is left in place, to ensure termination. + { + __qsb_conquer(__tls, __begin, __begin + __split_pos, __comp, + __iam, __num_threads_leftside, __wait); + __wait = __parent_wait; + } + // The pivot_pos is left in place, to ensure termination. # pragma omp section - { - __qsb_conquer(__tls, __begin + __split_pos + 1, __end, __comp, - __iam + __num_threads_leftside, - __num_threads - __num_threads_leftside, - __wait); - __wait = __parent_wait; - } - } + { + __qsb_conquer(__tls, __begin + __split_pos + 1, __end, __comp, + __iam + __num_threads_leftside, + __num_threads - __num_threads_leftside, __wait); + __wait = __parent_wait; + } + } + } } - } - -/** - * @brief Quicksort step doing load-balanced local sort. - * @param __tls Array of thread-local storages. - * @param __comp Comparator. - * @param __iam Number of the thread processing this function. - */ -template - void - __qsb_local_sort_with_helping(_QSBThreadLocal<_RAIter>** __tls, - _Compare& __comp, int __iam, bool __wait) - { - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::value_type _ValueType; - typedef typename _TraitsType::difference_type _DifferenceType; - typedef std::pair<_RAIter, _RAIter> _Piece; - - _QSBThreadLocal<_RAIter>& __tl = *__tls[__iam]; - - _DifferenceType __base_case_n = - _Settings::get().sort_qsb_base_case_maximal_n; - if (__base_case_n < 2) - __base_case_n = 2; - _ThreadIndex __num_threads = __tl._M_num_threads; - - // Every thread has its own random number generator. - _RandomNumber __rng(__iam + 1); - - _Piece __current = __tl._M_initial; - - _DifferenceType __elements_done = 0; + + /** + * @brief Quicksort step doing load-balanced local sort. + * @param __tls Array of thread-local storages. + * @param __comp Comparator. + * @param __iam Number of the thread processing this function. + */ + template + void + __qsb_local_sort_with_helping(_QSBThreadLocal<_RAIter>** __tls, + _Compare& __comp, int __iam, bool __wait) + { + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::value_type _ValueType; + typedef typename _TraitsType::difference_type _DifferenceType; + typedef std::pair<_RAIter, _RAIter> _Piece; + + _QSBThreadLocal<_RAIter>& __tl = *__tls[__iam]; + + _DifferenceType + __base_case_n = _Settings::get().sort_qsb_base_case_maximal_n; + if (__base_case_n < 2) + __base_case_n = 2; + _ThreadIndex __num_threads = __tl._M_num_threads; + + // Every thread has its own random number generator. + _RandomNumber __rng(__iam + 1); + + _Piece __current = __tl._M_initial; + + _DifferenceType __elements_done = 0; #if _GLIBCXX_ASSERTIONS - _DifferenceType __total_elements_done = 0; + _DifferenceType __total_elements_done = 0; #endif - for (;;) - { - // Invariant: __current must be a valid (maybe empty) range. - _RAIter __begin = __current.first, __end = __current.second; - _DifferenceType __n = __end - __begin; - - if (__n > __base_case_n) - { - // Divide. - _RAIter __pivot_pos = __begin + __rng(__n); - - // Swap __pivot_pos value to end. - if (__pivot_pos != (__end - 1)) - std::swap(*__pivot_pos, *(__end - 1)); - __pivot_pos = __end - 1; - - __gnu_parallel::binder2nd - <_Compare, _ValueType, _ValueType, bool> - __pred(__comp, *__pivot_pos); - - // Divide, leave pivot unchanged in last place. - _RAIter __split_pos1, __split_pos2; - __split_pos1 = - __gnu_sequential::partition(__begin, __end - 1, __pred); - - // Left side: < __pivot_pos; __right side: >= __pivot_pos. + for (;;) + { + // Invariant: __current must be a valid (maybe empty) range. + _RAIter __begin = __current.first, __end = __current.second; + _DifferenceType __n = __end - __begin; + + if (__n > __base_case_n) + { + // Divide. + _RAIter __pivot_pos = __begin + __rng(__n); + + // Swap __pivot_pos value to end. + if (__pivot_pos != (__end - 1)) + std::swap(*__pivot_pos, *(__end - 1)); + __pivot_pos = __end - 1; + + __gnu_parallel::binder2nd + <_Compare, _ValueType, _ValueType, bool> + __pred(__comp, *__pivot_pos); + + // Divide, leave pivot unchanged in last place. + _RAIter __split_pos1, __split_pos2; + __split_pos1 = __gnu_sequential::partition(__begin, __end - 1, + __pred); + + // Left side: < __pivot_pos; __right side: >= __pivot_pos. #if _GLIBCXX_ASSERTIONS - _GLIBCXX_PARALLEL_ASSERT(__begin <= __split_pos1 - && __split_pos1 < __end); + _GLIBCXX_PARALLEL_ASSERT(__begin <= __split_pos1 + && __split_pos1 < __end); #endif - // Swap pivot back to middle. - if (__split_pos1 != __pivot_pos) - std::swap(*__split_pos1, *__pivot_pos); - __pivot_pos = __split_pos1; - - // In case all elements are equal, __split_pos1 == 0. - if ((__split_pos1 + 1 - __begin) < (__n >> 7) - || (__end - __split_pos1) < (__n >> 7)) - { - // Very unequal split, one part smaller than one 128th - // elements not strictly larger than the pivot. - __gnu_parallel::__unary_negate<__gnu_parallel::__binder1st - <_Compare, _ValueType, _ValueType, bool>, _ValueType> - __pred(__gnu_parallel::__binder1st - <_Compare, _ValueType, _ValueType, bool>( - __comp, *__pivot_pos)); - - // Find other end of pivot-equal range. - __split_pos2 = __gnu_sequential::partition(__split_pos1 + 1, - __end, __pred); - } - else - // Only skip the pivot. - __split_pos2 = __split_pos1 + 1; - - // Elements equal to pivot are done. - __elements_done += (__split_pos2 - __split_pos1); + // Swap pivot back to middle. + if (__split_pos1 != __pivot_pos) + std::swap(*__split_pos1, *__pivot_pos); + __pivot_pos = __split_pos1; + + // In case all elements are equal, __split_pos1 == 0. + if ((__split_pos1 + 1 - __begin) < (__n >> 7) + || (__end - __split_pos1) < (__n >> 7)) + { + // Very unequal split, one part smaller than one 128th + // elements not strictly larger than the pivot. + __gnu_parallel::__unary_negate<__gnu_parallel::__binder1st + <_Compare, _ValueType, _ValueType, bool>, _ValueType> + __pred(__gnu_parallel::__binder1st + <_Compare, _ValueType, _ValueType, bool> + (__comp, *__pivot_pos)); + + // Find other end of pivot-equal range. + __split_pos2 = __gnu_sequential::partition(__split_pos1 + 1, + __end, __pred); + } + else + // Only skip the pivot. + __split_pos2 = __split_pos1 + 1; + + // Elements equal to pivot are done. + __elements_done += (__split_pos2 - __split_pos1); #if _GLIBCXX_ASSERTIONS - __total_elements_done += (__split_pos2 - __split_pos1); + __total_elements_done += (__split_pos2 - __split_pos1); #endif - // Always push larger part onto stack. - if (((__split_pos1 + 1) - __begin) < (__end - (__split_pos2))) - { - // Right side larger. - if ((__split_pos2) != __end) - __tl._M_leftover_parts.push_front( - std::make_pair(__split_pos2, __end)); - - //__current.first = __begin; //already set anyway - __current.second = __split_pos1; - continue; - } - else - { - // Left side larger. - if (__begin != __split_pos1) - __tl._M_leftover_parts.push_front(std::make_pair(__begin, - __split_pos1)); - - __current.first = __split_pos2; - //__current.second = __end; //already set anyway - continue; - } - } - else - { - __gnu_sequential::sort(__begin, __end, __comp); - __elements_done += __n; + // Always push larger part onto stack. + if (((__split_pos1 + 1) - __begin) < (__end - (__split_pos2))) + { + // Right side larger. + if ((__split_pos2) != __end) + __tl._M_leftover_parts.push_front + (std::make_pair(__split_pos2, __end)); + + //__current.first = __begin; //already set anyway + __current.second = __split_pos1; + continue; + } + else + { + // Left side larger. + if (__begin != __split_pos1) + __tl._M_leftover_parts.push_front(std::make_pair + (__begin, __split_pos1)); + + __current.first = __split_pos2; + //__current.second = __end; //already set anyway + continue; + } + } + else + { + __gnu_sequential::sort(__begin, __end, __comp); + __elements_done += __n; #if _GLIBCXX_ASSERTIONS - __total_elements_done += __n; + __total_elements_done += __n; #endif - // Prefer own stack, small pieces. - if (__tl._M_leftover_parts.pop_front(__current)) - continue; + // Prefer own stack, small pieces. + if (__tl._M_leftover_parts.pop_front(__current)) + continue; -# pragma omp atomic - *__tl._M_elements_leftover -= __elements_done; +# pragma omp atomic + *__tl._M_elements_leftover -= __elements_done; - __elements_done = 0; + __elements_done = 0; #if _GLIBCXX_ASSERTIONS - double __search_start = omp_get_wtime(); + double __search_start = omp_get_wtime(); #endif - // Look for new work. - bool __successfully_stolen = false; - while (__wait && *__tl._M_elements_leftover > 0 - && !__successfully_stolen + // Look for new work. + bool __successfully_stolen = false; + while (__wait && *__tl._M_elements_leftover > 0 + && !__successfully_stolen #if _GLIBCXX_ASSERTIONS - // Possible dead-lock. - && (omp_get_wtime() < (__search_start + 1.0)) + // Possible dead-lock. + && (omp_get_wtime() < (__search_start + 1.0)) #endif - ) - { - _ThreadIndex __victim; - __victim = __rng(__num_threads); - - // Large pieces. - __successfully_stolen = (__victim != __iam) - && __tls[__victim]->_M_leftover_parts.pop_back(__current); - if (!__successfully_stolen) - __yield(); + ) + { + _ThreadIndex __victim; + __victim = __rng(__num_threads); + + // Large pieces. + __successfully_stolen = (__victim != __iam) + && __tls[__victim]->_M_leftover_parts.pop_back(__current); + if (!__successfully_stolen) + __yield(); #if !defined(__ICC) && !defined(__ECC) -# pragma omp flush +# pragma omp flush #endif - } + } #if _GLIBCXX_ASSERTIONS - if (omp_get_wtime() >= (__search_start + 1.0)) - { - sleep(1); - _GLIBCXX_PARALLEL_ASSERT(omp_get_wtime() - < (__search_start + 1.0)); - } + if (omp_get_wtime() >= (__search_start + 1.0)) + { + sleep(1); + _GLIBCXX_PARALLEL_ASSERT(omp_get_wtime() + < (__search_start + 1.0)); + } #endif - if (!__successfully_stolen) - { + if (!__successfully_stolen) + { #if _GLIBCXX_ASSERTIONS - _GLIBCXX_PARALLEL_ASSERT(*__tl._M_elements_leftover == 0); + _GLIBCXX_PARALLEL_ASSERT(*__tl._M_elements_leftover == 0); #endif - return; - } - } - } - } - -/** @brief Top-level quicksort routine. - * @param __begin Begin iterator of sequence. - * @param __end End iterator of sequence. - * @param __comp Comparator. - * @param __num_threads Number of threads that are allowed to work on - * this part. - */ -template - void - __parallel_sort_qsb(_RAIter __begin, _RAIter __end, - _Compare __comp, - _ThreadIndex __num_threads) - { - _GLIBCXX_CALL(__end - __begin) - - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::value_type _ValueType; - typedef typename _TraitsType::difference_type _DifferenceType; - typedef std::pair<_RAIter, _RAIter> _Piece; - - typedef _QSBThreadLocal<_RAIter> _TLSType; - - _DifferenceType __n = __end - __begin; - - if (__n <= 1) - return; - - // At least one element per processor. - if (__num_threads > __n) - __num_threads = static_cast<_ThreadIndex>(__n); - - // Initialize thread local storage - _TLSType** __tls = new _TLSType*[__num_threads]; - _DifferenceType __queue_size = - __num_threads * (_ThreadIndex)(log2(__n) + 1); - for (_ThreadIndex __t = 0; __t < __num_threads; ++__t) - __tls[__t] = new _QSBThreadLocal<_RAIter>(__queue_size); - - // There can never be more than ceil(log2(__n)) ranges on the stack, - // because - // 1. Only one processor pushes onto the stack - // 2. The largest range has at most length __n - // 3. Each range is larger than half of the range remaining - volatile _DifferenceType _M_elements_leftover = __n; - for (int __i = 0; __i < __num_threads; ++__i) - { - __tls[__i]->_M_elements_leftover = &_M_elements_leftover; - __tls[__i]->_M_num_threads = __num_threads; - __tls[__i]->_M_global = std::make_pair(__begin, __end); - - // Just in case nothing is left to assign. - __tls[__i]->_M_initial = std::make_pair(__end, __end); - } + return; + } + } + } + } - // Main recursion call. - __qsb_conquer( - __tls, __begin, __begin + __n, __comp, 0, __num_threads, true); + /** @brief Top-level quicksort routine. + * @param __begin Begin iterator of sequence. + * @param __end End iterator of sequence. + * @param __comp Comparator. + * @param __num_threads Number of threads that are allowed to work on + * this part. + */ + template + void + __parallel_sort_qsb(_RAIter __begin, _RAIter __end, + _Compare __comp, _ThreadIndex __num_threads) + { + _GLIBCXX_CALL(__end - __begin) + + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::value_type _ValueType; + typedef typename _TraitsType::difference_type _DifferenceType; + typedef std::pair<_RAIter, _RAIter> _Piece; + + typedef _QSBThreadLocal<_RAIter> _TLSType; + + _DifferenceType __n = __end - __begin; + + if (__n <= 1) + return; + + // At least one element per processor. + if (__num_threads > __n) + __num_threads = static_cast<_ThreadIndex>(__n); + + // Initialize thread local storage + _TLSType** __tls = new _TLSType*[__num_threads]; + _DifferenceType __queue_size = (__num_threads + * (_ThreadIndex)(__rd_log2(__n) + 1)); + for (_ThreadIndex __t = 0; __t < __num_threads; ++__t) + __tls[__t] = new _QSBThreadLocal<_RAIter>(__queue_size); + + // There can never be more than ceil(__rd_log2(__n)) ranges on the + // stack, because + // 1. Only one processor pushes onto the stack + // 2. The largest range has at most length __n + // 3. Each range is larger than half of the range remaining + volatile _DifferenceType __elements_leftover = __n; + for (int __i = 0; __i < __num_threads; ++__i) + { + __tls[__i]->_M_elements_leftover = &__elements_leftover; + __tls[__i]->_M_num_threads = __num_threads; + __tls[__i]->_M_global = std::make_pair(__begin, __end); + + // Just in case nothing is left to assign. + __tls[__i]->_M_initial = std::make_pair(__end, __end); + } + + // Main recursion call. + __qsb_conquer(__tls, __begin, __begin + __n, __comp, 0, + __num_threads, true); #if _GLIBCXX_ASSERTIONS - // All stack must be empty. - _Piece __dummy; - for (int __i = 1; __i < __num_threads; ++__i) - _GLIBCXX_PARALLEL_ASSERT( - !__tls[__i]->_M_leftover_parts.pop_back(__dummy)); + // All stack must be empty. + _Piece __dummy; + for (int __i = 1; __i < __num_threads; ++__i) + _GLIBCXX_PARALLEL_ASSERT( + !__tls[__i]->_M_leftover_parts.pop_back(__dummy)); #endif - for (int __i = 0; __i < __num_threads; ++__i) - delete __tls[__i]; - delete[] __tls; - } + for (int __i = 0; __i < __num_threads; ++__i) + delete __tls[__i]; + delete[] __tls; + } } // namespace __gnu_parallel #endif /* _GLIBCXX_PARALLEL_BALANCED_QUICKSORT_H */ diff --git a/libstdc++-v3/include/parallel/checkers.h b/libstdc++-v3/include/parallel/checkers.h index 70cc949e3fe..8abfc991e67 100644 --- a/libstdc++-v3/include/parallel/checkers.h +++ b/libstdc++-v3/include/parallel/checkers.h @@ -68,5 +68,6 @@ namespace __gnu_parallel return true; } +} #endif /* _GLIBCXX_PARALLEL_CHECKERS_H */ diff --git a/libstdc++-v3/include/parallel/equally_split.h b/libstdc++-v3/include/parallel/equally_split.h index 31ed87d16f3..307be0b52f0 100644 --- a/libstdc++-v3/include/parallel/equally_split.h +++ b/libstdc++-v3/include/parallel/equally_split.h @@ -1,6 +1,6 @@ // -*- C++ -*- -// Copyright (C) 2007, 2009 Free Software Foundation, Inc. +// Copyright (C) 2007, 2008, 2009 Free Software Foundation, Inc. // // This file is part of the GNU ISO C++ Library. This library is free // software; you can redistribute it and/or modify it under the terms @@ -33,57 +33,56 @@ namespace __gnu_parallel { -/** @brief function to split a sequence into parts of almost equal size. - * - * The resulting sequence __s of length __num_threads+1 contains the splitting - * positions when splitting the range [0,__n) into parts of almost - * equal size (plus minus 1). The first entry is 0, the last one -* n. There may result empty parts. - * @param __n Number of elements - * @param __num_threads Number of parts - * @param __s Splitters - * @returns End of __splitter sequence, i.e. @__c __s+__num_threads+1 */ -template - _OutputIterator - equally_split(_DifferenceType __n, _ThreadIndex __num_threads, - _OutputIterator __s) - { - _DifferenceType __chunk_length = __n / __num_threads; - _DifferenceType __num_longer_chunks = __n % __num_threads; - _DifferenceType __pos = 0; - for (_ThreadIndex __i = 0; __i < __num_threads; ++__i) - { - *__s++ = __pos; - __pos += (__i < __num_longer_chunks) ? - (__chunk_length + 1) : __chunk_length; - } - *__s++ = __n; - return __s; - } + /** @brief function to split a sequence into parts of almost equal size. + * + * The resulting sequence __s of length __num_threads+1 contains the + * splitting positions when splitting the range [0,__n) into parts of + * almost equal size (plus minus 1). The first entry is 0, the last + * one n. There may result empty parts. + * @param __n Number of elements + * @param __num_threads Number of parts + * @param __s Splitters + * @returns End of __splitter sequence, i.e. @__c __s+__num_threads+1 */ + template + _OutputIterator + equally_split(_DifferenceType __n, _ThreadIndex __num_threads, + _OutputIterator __s) + { + _DifferenceType __chunk_length = __n / __num_threads; + _DifferenceType __num_longer_chunks = __n % __num_threads; + _DifferenceType __pos = 0; + for (_ThreadIndex __i = 0; __i < __num_threads; ++__i) + { + *__s++ = __pos; + __pos += ((__i < __num_longer_chunks) + ? (__chunk_length + 1) : __chunk_length); + } + *__s++ = __n; + return __s; + } - -/** @brief function to split a sequence into parts of almost equal size. - * - * Returns the position of the splitting point between - * thread number __thread_no (included) and - * thread number __thread_no+1 (excluded). - * @param __n Number of elements - * @param __num_threads Number of parts - * @returns splitting point */ -template - _DifferenceType - equally_split_point(_DifferenceType __n, - _ThreadIndex __num_threads, - _ThreadIndex __thread_no) - { - _DifferenceType __chunk_length = __n / __num_threads; - _DifferenceType __num_longer_chunks = __n % __num_threads; - if (__thread_no < __num_longer_chunks) - return __thread_no * (__chunk_length + 1); - else - return __num_longer_chunks * (__chunk_length + 1) + /** @brief function to split a sequence into parts of almost equal size. + * + * Returns the position of the splitting point between + * thread number __thread_no (included) and + * thread number __thread_no+1 (excluded). + * @param __n Number of elements + * @param __num_threads Number of parts + * @returns splitting point */ + template + _DifferenceType + equally_split_point(_DifferenceType __n, + _ThreadIndex __num_threads, + _ThreadIndex __thread_no) + { + _DifferenceType __chunk_length = __n / __num_threads; + _DifferenceType __num_longer_chunks = __n % __num_threads; + if (__thread_no < __num_longer_chunks) + return __thread_no * (__chunk_length + 1); + else + return __num_longer_chunks * (__chunk_length + 1) + (__thread_no - __num_longer_chunks) * __chunk_length; - } + } } #endif /* _GLIBCXX_PARALLEL_EQUALLY_SPLIT_H */ diff --git a/libstdc++-v3/include/parallel/find.h b/libstdc++-v3/include/parallel/find.h index 050a80d8af0..b4e581488ae 100644 --- a/libstdc++-v3/include/parallel/find.h +++ b/libstdc++-v3/include/parallel/find.h @@ -42,360 +42,363 @@ namespace __gnu_parallel { -/** - * @brief Parallel std::find, switch for different algorithms. - * @param __begin1 Begin iterator of first sequence. - * @param __end1 End iterator of first sequence. - * @param __begin2 Begin iterator of second sequence. Must have same - * length as first sequence. - * @param __pred Find predicate. - * @param __selector _Functionality (e. g. std::find_if (), std::equal(),...) - * @return Place of finding in both sequences. - */ -template - inline std::pair<_RAIter1, _RAIter2> - __find_template(_RAIter1 __begin1, _RAIter1 __end1, - _RAIter2 __begin2, _Pred __pred, _Selector __selector) - { - switch (_Settings::get().find_algorithm) - { - case GROWING_BLOCKS: - return __find_template(__begin1, __end1, __begin2, __pred, __selector, - growing_blocks_tag()); - case CONSTANT_SIZE_BLOCKS: - return __find_template(__begin1, __end1, __begin2, __pred, __selector, - constant_size_blocks_tag()); - case EQUAL_SPLIT: - return __find_template(__begin1, __end1, __begin2, __pred, __selector, - equal_split_tag()); - default: - _GLIBCXX_PARALLEL_ASSERT(false); - return std::make_pair(__begin1, __begin2); - } - } + /** + * @brief Parallel std::find, switch for different algorithms. + * @param __begin1 Begin iterator of first sequence. + * @param __end1 End iterator of first sequence. + * @param __begin2 Begin iterator of second sequence. Must have same + * length as first sequence. + * @param __pred Find predicate. + * @param __selector _Functionality (e. g. std::find_if(), std::equal(),...) + * @return Place of finding in both sequences. + */ + template + inline std::pair<_RAIter1, _RAIter2> + __find_template(_RAIter1 __begin1, _RAIter1 __end1, + _RAIter2 __begin2, _Pred __pred, _Selector __selector) + { + switch (_Settings::get().find_algorithm) + { + case GROWING_BLOCKS: + return __find_template(__begin1, __end1, __begin2, __pred, + __selector, growing_blocks_tag()); + case CONSTANT_SIZE_BLOCKS: + return __find_template(__begin1, __end1, __begin2, __pred, + __selector, constant_size_blocks_tag()); + case EQUAL_SPLIT: + return __find_template(__begin1, __end1, __begin2, __pred, + __selector, equal_split_tag()); + default: + _GLIBCXX_PARALLEL_ASSERT(false); + return std::make_pair(__begin1, __begin2); + } + } #if _GLIBCXX_FIND_EQUAL_SPLIT -/** - * @brief Parallel std::find, equal splitting variant. - * @param __begin1 Begin iterator of first sequence. - * @param __end1 End iterator of first sequence. - * @param __begin2 Begin iterator of second sequence. Second __sequence - * must have same length as first sequence. - * @param __pred Find predicate. - * @param __selector _Functionality (e. g. std::find_if (), std::equal(),...) - * @return Place of finding in both sequences. - */ -template - std::pair<_RAIter1, _RAIter2> - __find_template(_RAIter1 __begin1, - _RAIter1 __end1, - _RAIter2 __begin2, - _Pred __pred, - _Selector __selector, - equal_split_tag) - { - _GLIBCXX_CALL(__end1 - __begin1) - - typedef std::iterator_traits<_RAIter1> _TraitsType; - typedef typename _TraitsType::difference_type _DifferenceType; - typedef typename _TraitsType::value_type _ValueType; - - _DifferenceType __length = __end1 - __begin1; - _DifferenceType __result = __length; - _DifferenceType* __borders; - - omp_lock_t __result_lock; - omp_init_lock(&__result_lock); - - _ThreadIndex __num_threads = __get_max_threads(); -# pragma omp parallel num_threads(__num_threads) + /** + * @brief Parallel std::find, equal splitting variant. + * @param __begin1 Begin iterator of first sequence. + * @param __end1 End iterator of first sequence. + * @param __begin2 Begin iterator of second sequence. Second __sequence + * must have same length as first sequence. + * @param __pred Find predicate. + * @param __selector _Functionality (e. g. std::find_if(), std::equal(),...) + * @return Place of finding in both sequences. + */ + template + std::pair<_RAIter1, _RAIter2> + __find_template(_RAIter1 __begin1, _RAIter1 __end1, + _RAIter2 __begin2, _Pred __pred, + _Selector __selector, equal_split_tag) + { + _GLIBCXX_CALL(__end1 - __begin1) + + typedef std::iterator_traits<_RAIter1> _TraitsType; + typedef typename _TraitsType::difference_type _DifferenceType; + typedef typename _TraitsType::value_type _ValueType; + + _DifferenceType __length = __end1 - __begin1; + _DifferenceType __result = __length; + _DifferenceType* __borders; + + omp_lock_t __result_lock; + omp_init_lock(&__result_lock); + + _ThreadIndex __num_threads = __get_max_threads(); +# pragma omp parallel num_threads(__num_threads) { -# pragma omp single - { - __num_threads = omp_get_num_threads(); - __borders = new _DifferenceType[__num_threads + 1]; - equally_split(__length, __num_threads, __borders); - } //single - - _ThreadIndex __iam = omp_get_thread_num(); - _DifferenceType __start = __borders[__iam], - __stop = __borders[__iam + 1]; - - _RAIter1 __i1 = __begin1 + __start; - _RAIter2 __i2 = __begin2 + __start; - for (_DifferenceType __pos = __start; __pos < __stop; ++__pos) - { - #pragma omp flush(__result) - // Result has been set to something lower. - if (__result < __pos) - break; - - if (__selector(__i1, __i2, __pred)) - { - omp_set_lock(&__result_lock); - if (__pos < __result) - __result = __pos; - omp_unset_lock(&__result_lock); - break; - } - ++__i1; - ++__i2; - } +# pragma omp single + { + __num_threads = omp_get_num_threads(); + __borders = new _DifferenceType[__num_threads + 1]; + equally_split(__length, __num_threads, __borders); + } //single + + _ThreadIndex __iam = omp_get_thread_num(); + _DifferenceType __start = __borders[__iam], + __stop = __borders[__iam + 1]; + + _RAIter1 __i1 = __begin1 + __start; + _RAIter2 __i2 = __begin2 + __start; + for (_DifferenceType __pos = __start; __pos < __stop; ++__pos) + { +# pragma omp flush(__result) + // Result has been set to something lower. + if (__result < __pos) + break; + + if (__selector(__i1, __i2, __pred)) + { + omp_set_lock(&__result_lock); + if (__pos < __result) + __result = __pos; + omp_unset_lock(&__result_lock); + break; + } + ++__i1; + ++__i2; + } } //parallel - omp_destroy_lock(&__result_lock); - delete[] __borders; + omp_destroy_lock(&__result_lock); + delete[] __borders; - return - std::pair<_RAIter1, _RAIter2>(__begin1 + __result, __begin2 + __result); - } + return std::pair<_RAIter1, _RAIter2>(__begin1 + __result, + __begin2 + __result); + } #endif #if _GLIBCXX_FIND_GROWING_BLOCKS -/** - * @brief Parallel std::find, growing block size variant. - * @param __begin1 Begin iterator of first sequence. - * @param __end1 End iterator of first sequence. - * @param __begin2 Begin iterator of second sequence. Second __sequence - * must have same length as first sequence. - * @param __pred Find predicate. - * @param __selector _Functionality (e. g. std::find_if (), std::equal(),...) - * @return Place of finding in both sequences. - * @see __gnu_parallel::_Settings::find_sequential_search_size - * @see __gnu_parallel::_Settings::find_initial_block_size - * @see __gnu_parallel::_Settings::find_maximum_block_size - * @see __gnu_parallel::_Settings::find_increasing_factor - * - * There are two main differences between the growing blocks and - * the constant-size blocks variants. - * 1. For GB, the block size grows; for CSB, the block size is fixed. - * 2. For GB, the blocks are allocated dynamically; - * for CSB, the blocks are allocated in a predetermined manner, - * namely spacial round-robin. - */ -template - std::pair<_RAIter1, _RAIter2> - __find_template(_RAIter1 __begin1, _RAIter1 __end1, - _RAIter2 __begin2, _Pred __pred, _Selector __selector, - growing_blocks_tag) - { - _GLIBCXX_CALL(__end1 - __begin1) - - typedef std::iterator_traits<_RAIter1> _TraitsType; - typedef typename _TraitsType::difference_type _DifferenceType; - typedef typename _TraitsType::value_type _ValueType; - - const _Settings& __s = _Settings::get(); - - _DifferenceType __length = __end1 - __begin1; - - _DifferenceType __sequential_search_size = - std::min<_DifferenceType>(__length, __s.find_sequential_search_size); - - // Try it sequentially first. - std::pair<_RAIter1, _RAIter2> __find_seq_result = - __selector._M_sequential_algorithm( - __begin1, __begin1 + __sequential_search_size, __begin2, __pred); - - if (__find_seq_result.first != (__begin1 + __sequential_search_size)) - return __find_seq_result; - - // Index of beginning of next free block (after sequential find). - _DifferenceType __next_block_start = __sequential_search_size; - _DifferenceType __result = __length; - - omp_lock_t __result_lock; - omp_init_lock(&__result_lock); - - _ThreadIndex __num_threads = __get_max_threads(); -# pragma omp parallel shared(__result) num_threads(__num_threads) + /** + * @brief Parallel std::find, growing block size variant. + * @param __begin1 Begin iterator of first sequence. + * @param __end1 End iterator of first sequence. + * @param __begin2 Begin iterator of second sequence. Second __sequence + * must have same length as first sequence. + * @param __pred Find predicate. + * @param __selector _Functionality (e. g. std::find_if(), std::equal(),...) + * @return Place of finding in both sequences. + * @see __gnu_parallel::_Settings::find_sequential_search_size + * @see __gnu_parallel::_Settings::find_initial_block_size + * @see __gnu_parallel::_Settings::find_maximum_block_size + * @see __gnu_parallel::_Settings::find_increasing_factor + * + * There are two main differences between the growing blocks and + * the constant-size blocks variants. + * 1. For GB, the block size grows; for CSB, the block size is fixed. + * 2. For GB, the blocks are allocated dynamically; + * for CSB, the blocks are allocated in a predetermined manner, + * namely spacial round-robin. + */ + template + std::pair<_RAIter1, _RAIter2> + __find_template(_RAIter1 __begin1, _RAIter1 __end1, + _RAIter2 __begin2, _Pred __pred, _Selector __selector, + growing_blocks_tag) + { + _GLIBCXX_CALL(__end1 - __begin1) + + typedef std::iterator_traits<_RAIter1> _TraitsType; + typedef typename _TraitsType::difference_type _DifferenceType; + typedef typename _TraitsType::value_type _ValueType; + + const _Settings& __s = _Settings::get(); + + _DifferenceType __length = __end1 - __begin1; + + _DifferenceType + __sequential_search_size = std::min<_DifferenceType> + (__length, __s.find_sequential_search_size); + + // Try it sequentially first. + std::pair<_RAIter1, _RAIter2> + __find_seq_result = __selector._M_sequential_algorithm + (__begin1, __begin1 + __sequential_search_size, + __begin2, __pred); + + if (__find_seq_result.first != (__begin1 + __sequential_search_size)) + return __find_seq_result; + + // Index of beginning of next free block (after sequential find). + _DifferenceType __next_block_start = __sequential_search_size; + _DifferenceType __result = __length; + + omp_lock_t __result_lock; + omp_init_lock(&__result_lock); + + _ThreadIndex __num_threads = __get_max_threads(); +# pragma omp parallel shared(__result) num_threads(__num_threads) { # pragma omp single - __num_threads = omp_get_num_threads(); + __num_threads = omp_get_num_threads(); - // Not within first __k elements -> start parallel. - _ThreadIndex __iam = omp_get_thread_num(); + // Not within first __k elements -> start parallel. + _ThreadIndex __iam = omp_get_thread_num(); - _DifferenceType __block_size = __s.find_initial_block_size; - _DifferenceType __start = - __fetch_and_add<_DifferenceType>(&__next_block_start, __block_size); + _DifferenceType __block_size = __s.find_initial_block_size; + _DifferenceType __start = __fetch_and_add<_DifferenceType> + (&__next_block_start, __block_size); - // Get new block, update pointer to next block. - _DifferenceType __stop = - std::min<_DifferenceType>(__length, __start + __block_size); + // Get new block, update pointer to next block. + _DifferenceType __stop = + std::min<_DifferenceType>(__length, __start + __block_size); - std::pair<_RAIter1, _RAIter2> __local_result; + std::pair<_RAIter1, _RAIter2> __local_result; - while (__start < __length) - { + while (__start < __length) + { # pragma omp flush(__result) - // Get new value of result. - if (__result < __start) - { - // No chance to find first element. - break; - } - - __local_result = __selector._M_sequential_algorithm( - __begin1 + __start, __begin1 + __stop, - __begin2 + __start, __pred); - if (__local_result.first != (__begin1 + __stop)) - { - omp_set_lock(&__result_lock); - if ((__local_result.first - __begin1) < __result) - { - __result = __local_result.first - __begin1; - - // Result cannot be in future blocks, stop algorithm. - __fetch_and_add<_DifferenceType>( - &__next_block_start, __length); - } - omp_unset_lock(&__result_lock); - } - - __block_size = std::min<_DifferenceType>( - __block_size * __s.find_increasing_factor, - __s.find_maximum_block_size); - - // Get new block, update pointer to next block. - __start = - __fetch_and_add<_DifferenceType>( - &__next_block_start, __block_size); - __stop = ((__length < (__start + __block_size)) - ? __length : (__start + __block_size)); - } + // Get new value of result. + if (__result < __start) + { + // No chance to find first element. + break; + } + + __local_result = __selector._M_sequential_algorithm + (__begin1 + __start, __begin1 + __stop, + __begin2 + __start, __pred); + + if (__local_result.first != (__begin1 + __stop)) + { + omp_set_lock(&__result_lock); + if ((__local_result.first - __begin1) < __result) + { + __result = __local_result.first - __begin1; + + // Result cannot be in future blocks, stop algorithm. + __fetch_and_add<_DifferenceType>(&__next_block_start, + __length); + } + omp_unset_lock(&__result_lock); + } + + __block_size = std::min<_DifferenceType> + (__block_size * __s.find_increasing_factor, + __s.find_maximum_block_size); + + // Get new block, update pointer to next block. + __start = __fetch_and_add<_DifferenceType>(&__next_block_start, + __block_size); + __stop = (__length < (__start + __block_size) + ? __length : (__start + __block_size)); + } } //parallel - omp_destroy_lock(&__result_lock); + omp_destroy_lock(&__result_lock); - // Return iterator on found element. - return - std::pair<_RAIter1, _RAIter2>(__begin1 + __result, __begin2 + __result); - } + // Return iterator on found element. + return + std::pair<_RAIter1, _RAIter2>(__begin1 + __result, + __begin2 + __result); + } #endif #if _GLIBCXX_FIND_CONSTANT_SIZE_BLOCKS -/** - * @brief Parallel std::find, constant block size variant. - * @param __begin1 Begin iterator of first sequence. - * @param __end1 End iterator of first sequence. - * @param __begin2 Begin iterator of second sequence. Second __sequence - * must have same length as first sequence. - * @param __pred Find predicate. - * @param __selector _Functionality (e. g. std::find_if (), std::equal(),...) - * @return Place of finding in both sequences. - * @see __gnu_parallel::_Settings::find_sequential_search_size - * @see __gnu_parallel::_Settings::find_block_size - * There are two main differences between the growing blocks and the - * constant-size blocks variants. - * 1. For GB, the block size grows; for CSB, the block size is fixed. - * 2. For GB, the blocks are allocated dynamically; for CSB, the - * blocks are allocated in a predetermined manner, namely spacial - * round-robin. - */ -template - std::pair<_RAIter1, _RAIter2> - __find_template(_RAIter1 __begin1, _RAIter1 __end1, - _RAIter2 __begin2, _Pred __pred, _Selector __selector, - constant_size_blocks_tag) - { - _GLIBCXX_CALL(__end1 - __begin1) - typedef std::iterator_traits<_RAIter1> _TraitsType; - typedef typename _TraitsType::difference_type _DifferenceType; - typedef typename _TraitsType::value_type _ValueType; - - const _Settings& __s = _Settings::get(); - - _DifferenceType __length = __end1 - __begin1; - - _DifferenceType __sequential_search_size = std::min<_DifferenceType>( - __length, __s.find_sequential_search_size); - - // Try it sequentially first. - std::pair<_RAIter1, _RAIter2> __find_seq_result = - __selector._M_sequential_algorithm( - __begin1, __begin1 + __sequential_search_size, __begin2, __pred); - - if (__find_seq_result.first != (__begin1 + __sequential_search_size)) - return __find_seq_result; - - _DifferenceType __result = __length; - omp_lock_t __result_lock; - omp_init_lock(&__result_lock); - - // Not within first __sequential_search_size elements -> start parallel. - - _ThreadIndex __num_threads = __get_max_threads(); -# pragma omp parallel shared(__result) num_threads(__num_threads) + /** + * @brief Parallel std::find, constant block size variant. + * @param __begin1 Begin iterator of first sequence. + * @param __end1 End iterator of first sequence. + * @param __begin2 Begin iterator of second sequence. Second __sequence + * must have same length as first sequence. + * @param __pred Find predicate. + * @param __selector _Functionality (e. g. std::find_if(), std::equal(),...) + * @return Place of finding in both sequences. + * @see __gnu_parallel::_Settings::find_sequential_search_size + * @see __gnu_parallel::_Settings::find_block_size + * There are two main differences between the growing blocks and the + * constant-size blocks variants. + * 1. For GB, the block size grows; for CSB, the block size is fixed. + * 2. For GB, the blocks are allocated dynamically; for CSB, the + * blocks are allocated in a predetermined manner, namely spacial + * round-robin. + */ + template + std::pair<_RAIter1, _RAIter2> + __find_template(_RAIter1 __begin1, _RAIter1 __end1, + _RAIter2 __begin2, _Pred __pred, _Selector __selector, + constant_size_blocks_tag) + { + _GLIBCXX_CALL(__end1 - __begin1) + typedef std::iterator_traits<_RAIter1> _TraitsType; + typedef typename _TraitsType::difference_type _DifferenceType; + typedef typename _TraitsType::value_type _ValueType; + + const _Settings& __s = _Settings::get(); + + _DifferenceType __length = __end1 - __begin1; + + _DifferenceType __sequential_search_size = std::min<_DifferenceType> + (__length, __s.find_sequential_search_size); + + // Try it sequentially first. + std::pair<_RAIter1, _RAIter2> + __find_seq_result = __selector._M_sequential_algorithm + (__begin1, __begin1 + __sequential_search_size, __begin2, __pred); + + if (__find_seq_result.first != (__begin1 + __sequential_search_size)) + return __find_seq_result; + + _DifferenceType __result = __length; + omp_lock_t __result_lock; + omp_init_lock(&__result_lock); + + // Not within first __sequential_search_size elements -> start parallel. + + _ThreadIndex __num_threads = __get_max_threads(); +# pragma omp parallel shared(__result) num_threads(__num_threads) { # pragma omp single - __num_threads = omp_get_num_threads(); + __num_threads = omp_get_num_threads(); - _ThreadIndex __iam = omp_get_thread_num(); - _DifferenceType __block_size = __s.find_initial_block_size; + _ThreadIndex __iam = omp_get_thread_num(); + _DifferenceType __block_size = __s.find_initial_block_size; - // First element of thread's current iteration. - _DifferenceType __iteration_start = __sequential_search_size; + // First element of thread's current iteration. + _DifferenceType __iteration_start = __sequential_search_size; - // Where to work (initialization). - _DifferenceType __start = __iteration_start + __iam * __block_size; - _DifferenceType __stop = - std::min<_DifferenceType>(__length, __start + __block_size); + // Where to work (initialization). + _DifferenceType __start = __iteration_start + __iam * __block_size; + _DifferenceType __stop = std::min<_DifferenceType>(__length, + __start + + __block_size); - std::pair<_RAIter1, _RAIter2> __local_result; + std::pair<_RAIter1, _RAIter2> __local_result; - while (__start < __length) - { - // Get new value of result. + while (__start < __length) + { + // Get new value of result. # pragma omp flush(__result) - // No chance to find first element. - if (__result < __start) - break; - __local_result = __selector._M_sequential_algorithm( - __begin1 + __start, __begin1 + __stop, - __begin2 + __start, __pred); - if (__local_result.first != (__begin1 + __stop)) - { - omp_set_lock(&__result_lock); - if ((__local_result.first - __begin1) < __result) - __result = __local_result.first - __begin1; - omp_unset_lock(&__result_lock); - // Will not find better value in its interval. - break; - } - - __iteration_start += __num_threads * __block_size; - - // Where to work. - __start = __iteration_start + __iam * __block_size; - __stop = std::min<_DifferenceType>( - __length, __start + __block_size); - } + // No chance to find first element. + if (__result < __start) + break; + + __local_result = __selector._M_sequential_algorithm + (__begin1 + __start, __begin1 + __stop, + __begin2 + __start, __pred); + + if (__local_result.first != (__begin1 + __stop)) + { + omp_set_lock(&__result_lock); + if ((__local_result.first - __begin1) < __result) + __result = __local_result.first - __begin1; + omp_unset_lock(&__result_lock); + // Will not find better value in its interval. + break; + } + + __iteration_start += __num_threads * __block_size; + + // Where to work. + __start = __iteration_start + __iam * __block_size; + __stop = std::min<_DifferenceType>(__length, + __start + __block_size); + } } //parallel - omp_destroy_lock(&__result_lock); + omp_destroy_lock(&__result_lock); - // Return iterator on found element. - return - std::pair<_RAIter1, _RAIter2>(__begin1 + __result, __begin2 + __result); - } + // Return iterator on found element. + return std::pair<_RAIter1, _RAIter2>(__begin1 + __result, + __begin2 + __result); + } #endif } // end namespace diff --git a/libstdc++-v3/include/parallel/find_selectors.h b/libstdc++-v3/include/parallel/find_selectors.h index 0d385bc6c94..df77978a9dd 100644 --- a/libstdc++-v3/include/parallel/find_selectors.h +++ b/libstdc++-v3/include/parallel/find_selectors.h @@ -103,12 +103,12 @@ namespace __gnu_parallel typename _Pred> std::pair<_RAIter1, _RAIter2> _M_sequential_algorithm(_RAIter1 __begin1, - _RAIter1 __end1, - _RAIter2 __begin2, _Pred __pred) + _RAIter1 __end1, + _RAIter2 __begin2, _Pred __pred) { // Passed end iterator is one short. _RAIter1 __spot = adjacent_find(__begin1, __end1 + 1, - __pred, sequential_tag()); + __pred, sequential_tag()); if (__spot == (__end1 + 1)) __spot = __end1; return std::make_pair(__spot, __begin2); @@ -141,56 +141,57 @@ namespace __gnu_parallel typename _Pred> std::pair<_RAIter1, _RAIter2> _M_sequential_algorithm(_RAIter1 __begin1, - _RAIter1 __end1, - _RAIter2 __begin2, _Pred __pred) - { return mismatch(__begin1, __end1, __begin2, __pred, sequential_tag()); - } + _RAIter1 __end1, + _RAIter2 __begin2, _Pred __pred) + { return mismatch(__begin1, __end1, __begin2, + __pred, sequential_tag()); } }; /** @brief Test predicate on several elements. */ template - struct __find_first_of_selector : public __generic_find_selector - { - _FIterator _M_begin; - _FIterator _M_end; - - explicit __find_first_of_selector(_FIterator __begin, _FIterator __end) - : _M_begin(__begin), _M_end(__end) { } - - /** @brief Test on one position. - * @param __i1 _Iterator on first sequence. - * @param __i2 _Iterator on second sequence (unused). - * @param __pred Find predicate. */ - template - bool - operator()(_RAIter1 __i1, _RAIter2 __i2, _Pred __pred) - { - for (_FIterator __pos_in_candidates = _M_begin; - __pos_in_candidates != _M_end; ++__pos_in_candidates) - if (__pred(*__i1, *__pos_in_candidates)) - return true; - return false; - } - - /** @brief Corresponding sequential algorithm on a sequence. - * @param __begin1 Begin iterator of first sequence. - * @param __end1 End iterator of first sequence. - * @param __begin2 Begin iterator of second sequence. - * @param __pred Find predicate. */ - template - std::pair<_RAIter1, _RAIter2> - _M_sequential_algorithm(_RAIter1 __begin1, - _RAIter1 __end1, - _RAIter2 __begin2, _Pred __pred) - { - return std::make_pair( - find_first_of(__begin1, __end1, _M_begin, _M_end, __pred, - sequential_tag()), __begin2); - } - }; + struct __find_first_of_selector : public __generic_find_selector + { + _FIterator _M_begin; + _FIterator _M_end; + + explicit __find_first_of_selector(_FIterator __begin, + _FIterator __end) + : _M_begin(__begin), _M_end(__end) { } + + /** @brief Test on one position. + * @param __i1 _Iterator on first sequence. + * @param __i2 _Iterator on second sequence (unused). + * @param __pred Find predicate. */ + template + bool + operator()(_RAIter1 __i1, _RAIter2 __i2, _Pred __pred) + { + for (_FIterator __pos_in_candidates = _M_begin; + __pos_in_candidates != _M_end; ++__pos_in_candidates) + if (__pred(*__i1, *__pos_in_candidates)) + return true; + return false; + } + + /** @brief Corresponding sequential algorithm on a sequence. + * @param __begin1 Begin iterator of first sequence. + * @param __end1 End iterator of first sequence. + * @param __begin2 Begin iterator of second sequence. + * @param __pred Find predicate. */ + template + std::pair<_RAIter1, _RAIter2> + _M_sequential_algorithm(_RAIter1 __begin1, + _RAIter1 __end1, + _RAIter2 __begin2, _Pred __pred) + { + return std::make_pair(find_first_of(__begin1, __end1, + _M_begin, _M_end, __pred, + sequential_tag()), __begin2); + } + }; } #endif /* _GLIBCXX_PARALLEL_FIND_SELECTORS_H */ diff --git a/libstdc++-v3/include/parallel/for_each.h b/libstdc++-v3/include/parallel/for_each.h index b585999f514..f758b7ab78a 100644 --- a/libstdc++-v3/include/parallel/for_each.h +++ b/libstdc++-v3/include/parallel/for_each.h @@ -69,31 +69,21 @@ namespace __gnu_parallel _Parallelism __parallelism_tag) { if (__parallelism_tag == parallel_unbalanced) - return __for_each_template_random_access_ed(__begin, __end, __user_op, - __functionality, __reduction, - __reduction_start, - __output, __bound); + return __for_each_template_random_access_ed + (__begin, __end, __user_op, __functionality, __reduction, + __reduction_start, __output, __bound); else if (__parallelism_tag == parallel_omp_loop) - return __for_each_template_random_access_omp_loop( - __begin, __end, __user_op, - __functionality, - __reduction, - __reduction_start, - __output, __bound); + return __for_each_template_random_access_omp_loop + (__begin, __end, __user_op, __functionality, __reduction, + __reduction_start, __output, __bound); else if (__parallelism_tag == parallel_omp_loop_static) - return __for_each_template_random_access_omp_loop( - __begin, __end, __user_op, - __functionality, - __reduction, - __reduction_start, - __output, __bound); + return __for_each_template_random_access_omp_loop + (__begin, __end, __user_op, __functionality, __reduction, + __reduction_start, __output, __bound); else //e. g. parallel_balanced - return __for_each_template_random_access_workstealing(__begin, __end, - __user_op, - __functionality, - __reduction, - __reduction_start, - __output, __bound); + return __for_each_template_random_access_workstealing + (__begin, __end, __user_op, __functionality, __reduction, + __reduction_start, __output, __bound); } } diff --git a/libstdc++-v3/include/parallel/list_partition.h b/libstdc++-v3/include/parallel/list_partition.h index 9c4f48187f0..573085aa872 100644 --- a/libstdc++-v3/include/parallel/list_partition.h +++ b/libstdc++-v3/include/parallel/list_partition.h @@ -48,11 +48,11 @@ namespace __gnu_parallel template void __shrink_and_double(std::vector<_IIter>& __os_starts, - size_t& __count_to_two, size_t& __range_length, - const bool __make_twice) + size_t& __count_to_two, size_t& __range_length, + const bool __make_twice) { ++__count_to_two; - if (not __make_twice or __count_to_two < 2) + if (!__make_twice || __count_to_two < 2) __shrink(__os_starts, __count_to_two, __range_length); else { @@ -68,7 +68,7 @@ namespace __gnu_parallel template void __shrink(std::vector<_IIter>& __os_starts, size_t& __count_to_two, - size_t& __range_length) + size_t& __range_length) { for (typename std::vector<_IIter>::size_type __i = 0; __i <= (__os_starts.size() / 2); ++__i) @@ -112,8 +112,8 @@ namespace __gnu_parallel std::vector<_IIter> __os_starts(2 * __oversampling * __num_parts + 1); - __os_starts[0]= __begin; - _IIter __prev = __begin, __it = __begin; + __os_starts[0] = __begin; + _IIter __prev = __begin, __it = __begin; size_t __dist_limit = 0, __dist = 0; size_t __cur = 1, __next = 1; size_t __range_length = 1; diff --git a/libstdc++-v3/include/parallel/losertree.h b/libstdc++-v3/include/parallel/losertree.h index 9b9914ff1f0..425cf6d4b6f 100644 --- a/libstdc++-v3/include/parallel/losertree.h +++ b/libstdc++-v3/include/parallel/losertree.h @@ -40,993 +40,1000 @@ namespace __gnu_parallel { -/** - * @brief Guarded loser/tournament tree. - * - * The smallest element is at the top. - * - * Guarding is done explicitly through one flag _M_sup per element, - * inf is not needed due to a better initialization routine. This - * is a well-performing variant. - * - * @param _Tp the element type - * @param _Compare the comparator to use, defaults to std::less<_Tp> - */ -template - class _LoserTreeBase - { - protected: - /** @brief Internal representation of a _LoserTree element. */ - struct _Loser + /** + * @brief Guarded loser/tournament tree. + * + * The smallest element is at the top. + * + * Guarding is done explicitly through one flag _M_sup per element, + * inf is not needed due to a better initialization routine. This + * is a well-performing variant. + * + * @param _Tp the element type + * @param _Compare the comparator to use, defaults to std::less<_Tp> + */ + template + class _LoserTreeBase { - /** @brief flag, true iff this is a "maximum" __sentinel. */ - bool _M_sup; - /** @brief __index of the __source __sequence. */ - int _M_source; - /** @brief _M_key of the element in the _LoserTree. */ - _Tp _M_key; + protected: + /** @brief Internal representation of a _LoserTree element. */ + struct _Loser + { + /** @brief flag, true iff this is a "maximum" __sentinel. */ + bool _M_sup; + /** @brief __index of the __source __sequence. */ + int _M_source; + /** @brief _M_key of the element in the _LoserTree. */ + _Tp _M_key; + }; + + unsigned int _M_ik, _M_k, _M_offset; + + /** log_2{_M_k} */ + unsigned int _M_log_k; + + /** @brief _LoserTree __elements. */ + _Loser* _M_losers; + + /** @brief _Compare to use. */ + _Compare _M_comp; + + /** + * @brief State flag that determines whether the _LoserTree is empty. + * + * Only used for building the _LoserTree. + */ + bool _M_first_insert; + + public: + /** + * @brief The constructor. + * + * @param __k The number of sequences to merge. + * @param __comp The comparator to use. + */ + _LoserTreeBase(unsigned int __k, _Compare __comp) + : _M_comp(__comp) + { + _M_ik = __k; + + // Compute log_2{_M_k} for the _Loser Tree + _M_log_k = __rd_log2(_M_ik - 1) + 1; + + // Next greater power of 2. + _M_k = 1 << _M_log_k; + _M_offset = _M_k; + + // Avoid default-constructing _M_losers[]._M_key + _M_losers = static_cast<_Loser*>(::operator new(2 * _M_k + * sizeof(_Loser))); + for (unsigned int __i = _M_ik - 1; __i < _M_k; ++__i) + _M_losers[__i + _M_k]._M_sup = true; + + _M_first_insert = true; + } + + /** + * @brief The destructor. + */ + ~_LoserTreeBase() + { ::operator delete(_M_losers); } + + /** + * @brief Initializes the sequence "_M_source" with the element "__key". + * + * @param __key the element to insert + * @param __source __index of the __source __sequence + * @param __sup flag that determines whether the value to insert is an + * explicit __supremum. + */ + void + __insert_start(const _Tp& __key, int __source, bool __sup) + { + unsigned int __pos = _M_k + __source; + + if(_M_first_insert) + { + // Construct all keys, so we can easily deconstruct them. + for (unsigned int __i = 0; __i < (2 * _M_k); ++__i) + new(&(_M_losers[__i]._M_key)) _Tp(__key); + _M_first_insert = false; + } + else + new(&(_M_losers[__pos]._M_key)) _Tp(__key); + + _M_losers[__pos]._M_sup = __sup; + _M_losers[__pos]._M_source = __source; + } + + /** + * @return the index of the sequence with the smallest element. + */ + int __get_min_source() + { return _M_losers[0]._M_source; } }; - unsigned int _M_ik, _M_k, _M_offset; - - /** log_2{_M_k} */ - unsigned int _M_log_k; - - /** @brief _LoserTree __elements. */ - _Loser* _M_losers; - - /** @brief _Compare to use. */ - _Compare _M_comp; - /** - * @brief State flag that determines whether the _LoserTree is empty. + * @brief Stable _LoserTree variant. * - * Only used for building the _LoserTree. - */ - bool _M_first_insert; - - public: - /** - * @brief The constructor. + * Provides the stable implementations of insert_start, __init_winner, + * __init and __delete_min_insert. * - * @param __k The number of sequences to merge. - * @param __comp The comparator to use. + * Unstable variant is done using partial specialisation below. */ - _LoserTreeBase(unsigned int __k, _Compare __comp) - : _M_comp(__comp) + template + class _LoserTree + : public _LoserTreeBase<_Tp, _Compare> { - _M_ik = __k; - - // Compute log_2{_M_k} for the _Loser Tree - _M_log_k = __rd_log2(_M_ik - 1) + 1; - - // Next greater power of 2. - _M_k = 1 << _M_log_k; - _M_offset = _M_k; - - // Avoid default-constructing _M_losers[]._M_key - _M_losers - = static_cast<_Loser*>(::operator new(2 * _M_k * sizeof(_Loser))); - for (unsigned int __i = _M_ik - 1; __i < _M_k; ++__i) - _M_losers[__i + _M_k]._M_sup = true; - - _M_first_insert = true; - } + typedef _LoserTreeBase<_Tp, _Compare> _Base; + using _Base::_M_k; + using _Base::_M_losers; + using _Base::_M_first_insert; + + public: + _LoserTree(unsigned int __k, _Compare __comp) + : _Base::_LoserTreeBase(__k, __comp) + { } + + unsigned int + __init_winner(unsigned int __root) + { + if (__root >= _M_k) + return __root; + else + { + unsigned int __left = __init_winner(2 * __root); + unsigned int __right = __init_winner(2 * __root + 1); + if (_M_losers[__right]._M_sup + || (!_M_losers[__left]._M_sup + && !_M_comp(_M_losers[__right]._M_key, + _M_losers[__left]._M_key))) + { + // Left one is less or equal. + _M_losers[__root] = _M_losers[__right]; + return __left; + } + else + { + // Right one is less. + _M_losers[__root] = _M_losers[__left]; + return __right; + } + } + } + + void __init() + { _M_losers[0] = _M_losers[__init_winner(1)]; } + + /** + * @brief Delete the smallest element and insert a new element from + * the previously smallest element's sequence. + * + * This implementation is stable. + */ + // Do not pass a const reference since __key will be used as + // local variable. + void + __delete_min_insert(_Tp __key, bool __sup) + { +#if _GLIBCXX_ASSERTIONS + // no dummy sequence can ever be at the top! + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); +#endif - /** - * @brief The destructor. - */ - ~_LoserTreeBase() - { ::operator delete(_M_losers); } + int __source = _M_losers[0]._M_source; + for (unsigned int __pos = (_M_k + __source) / 2; __pos > 0; + __pos /= 2) + { + // The smaller one gets promoted, ties are broken by _M_source. + if ((__sup && (!_M_losers[__pos]._M_sup + || _M_losers[__pos]._M_source < __source)) + || (!__sup && !_M_losers[__pos]._M_sup + && ((_M_comp(_M_losers[__pos]._M_key, __key)) + || (!_M_comp(__key, _M_losers[__pos]._M_key) + && _M_losers[__pos]._M_source < __source)))) + { + // The other one is smaller. + std::swap(_M_losers[__pos]._M_sup, __sup); + std::swap(_M_losers[__pos]._M_source, __source); + std::swap(_M_losers[__pos]._M_key, __key); + } + } + + _M_losers[0]._M_sup = __sup; + _M_losers[0]._M_source = __source; + _M_losers[0]._M_key = __key; + } + }; /** - * @brief Initializes the sequence "_M_source" with the element "_M_key". + * @brief Unstable _LoserTree variant. * - * @param _M_key the element to insert - * @param _M_source __index of the __source __sequence - * @param _M_sup flag that determines whether the value to insert is an - * explicit __supremum. + * Stability (non-stable here) is selected with partial specialization. */ - void - __insert_start(const _Tp& _M_key, int _M_source, bool _M_sup) + template + class _LoserTree + : public _LoserTreeBase<_Tp, _Compare> { - unsigned int __pos = _M_k + _M_source; - - if(_M_first_insert) - { - // Construct all keys, so we can easily deconstruct them. - for (unsigned int __i = 0; __i < (2 * _M_k); ++__i) - new(&(_M_losers[__i]._M_key)) _Tp(_M_key); - _M_first_insert = false; - } - else - new(&(_M_losers[__pos]._M_key)) _Tp(_M_key); - - _M_losers[__pos]._M_sup = _M_sup; - _M_losers[__pos]._M_source = _M_source; - } + typedef _LoserTreeBase<_Tp, _Compare> _Base; + using _Base::_M_log_k; + using _Base::_M_k; + using _Base::_M_losers; + using _Base::_M_first_insert; + + public: + _LoserTree(unsigned int __k, _Compare __comp) + : _Base::_LoserTreeBase(__k, __comp) + { } + + /** + * Computes the winner of the competition at position "__root". + * + * Called recursively (starting at 0) to build the initial tree. + * + * @param __root __index of the "game" to start. + */ + unsigned int + __init_winner(unsigned int __root) + { + if (__root >= _M_k) + return __root; + else + { + unsigned int __left = __init_winner(2 * __root); + unsigned int __right = __init_winner(2 * __root + 1); + if (_M_losers[__right]._M_sup + || (!_M_losers[__left]._M_sup + && !_M_comp(_M_losers[__right]._M_key, + _M_losers[__left]._M_key))) + { + // Left one is less or equal. + _M_losers[__root] = _M_losers[__right]; + return __left; + } + else + { + // Right one is less. + _M_losers[__root] = _M_losers[__left]; + return __right; + } + } + } + + void + __init() + { _M_losers[0] = _M_losers[__init_winner(1)]; } + + /** + * Delete the _M_key smallest element and insert the element __key + * instead. + * + * @param __key the _M_key to insert + * @param __sup true iff __key is an explicitly marked supremum + */ + // Do not pass a const reference since __key will be used as local + // variable. + void + __delete_min_insert(_Tp __key, bool __sup) + { +#if _GLIBCXX_ASSERTIONS + // no dummy sequence can ever be at the top! + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); +#endif - /** - * @return the index of the sequence with the smallest element. - */ - int __get_min_source() - { return _M_losers[0]._M_source; } - }; + int __source = _M_losers[0]._M_source; + for (unsigned int __pos = (_M_k + __source) / 2; __pos > 0; + __pos /= 2) + { + // The smaller one gets promoted. + if (__sup || (!_M_losers[__pos]._M_sup + && _M_comp(_M_losers[__pos]._M_key, __key))) + { + // The other one is smaller. + std::swap(_M_losers[__pos]._M_sup, __sup); + std::swap(_M_losers[__pos]._M_source, __source); + std::swap(_M_losers[__pos]._M_key, __key); + } + } + + _M_losers[0]._M_sup = __sup; + _M_losers[0]._M_source = __source; + _M_losers[0]._M_key = __key; + } + }; /** - * @brief Stable _LoserTree variant. - * - * Provides the stable implementations of insert_start, __init_winner, - * __init and __delete_min_insert. - * - * Unstable variant is done using partial specialisation below. + * @brief Base class of _Loser Tree implementation using pointers. */ -template - class _LoserTree - : public _LoserTreeBase<_Tp, _Compare> - { - typedef _LoserTreeBase<_Tp, _Compare> _Base; - using _Base::_M_k; - using _Base::_M_losers; - using _Base::_M_first_insert; - - public: - _LoserTree(unsigned int __k, _Compare __comp) - : _Base::_LoserTreeBase(__k, __comp) - { } - - unsigned int - __init_winner(unsigned int __root) + template + class _LoserTreePointerBase { - if (__root >= _M_k) - return __root; - else - { - unsigned int __left = __init_winner (2 * __root); - unsigned int __right = __init_winner (2 * __root + 1); - if (_M_losers[__right]._M_sup - || (!_M_losers[__left]._M_sup - && !_M_comp(_M_losers[__right]._M_key, - _M_losers[__left]._M_key))) - { - // Left one is less or equal. - _M_losers[__root] = _M_losers[__right]; - return __left; - } - else - { - // Right one is less. - _M_losers[__root] = _M_losers[__left]; - return __right; - } - } - } - - void __init() - { _M_losers[0] = _M_losers[__init_winner(1)]; } + protected: + /** @brief Internal representation of _LoserTree __elements. */ + struct _Loser + { + bool _M_sup; + int _M_source; + const _Tp* _M_keyp; + }; + + unsigned int _M_ik, _M_k, _M_offset; + _Loser* _M_losers; + _Compare _M_comp; + + public: + _LoserTreePointerBase(unsigned int __k, + _Compare __comp = std::less<_Tp>()) + : _M_comp(__comp) + { + _M_ik = __k; + + // Next greater power of 2. + _M_k = 1 << (__rd_log2(_M_ik - 1) + 1); + _M_offset = _M_k; + _M_losers = new _Loser[_M_k * 2]; + for (unsigned int __i = _M_ik - 1; __i < _M_k; __i++) + _M_losers[__i + _M_k]._M_sup = true; + } + + ~_LoserTreePointerBase() + { ::operator delete[](_M_losers); } + + int __get_min_source() + { return _M_losers[0]._M_source; } + + void __insert_start(const _Tp& __key, int __source, bool __sup) + { + unsigned int __pos = _M_k + __source; + + _M_losers[__pos]._M_sup = __sup; + _M_losers[__pos]._M_source = __source; + _M_losers[__pos]._M_keyp = &__key; + } + }; - /** - * @brief Delete the smallest element and insert a new element from - * the previously smallest element's sequence. - * - * This implementation is stable. - */ - // Do not pass a const reference since _M_key will be used as - // local variable. - void - __delete_min_insert(_Tp _M_key, bool _M_sup) + /** + * @brief Stable _LoserTree implementation. + * + * The unstable variant is implemented using partial instantiation below. + */ + template + class _LoserTreePointer + : public _LoserTreePointerBase<_Tp, _Compare> { + typedef _LoserTreePointerBase<_Tp, _Compare> _Base; + using _Base::_M_k; + using _Base::_M_losers; + + public: + _LoserTreePointer(unsigned int __k, _Compare __comp = std::less<_Tp>()) + : _Base::_LoserTreePointerBase(__k, __comp) + { } + + unsigned int + __init_winner(unsigned int __root) + { + if (__root >= _M_k) + return __root; + else + { + unsigned int __left = __init_winner(2 * __root); + unsigned int __right = __init_winner(2 * __root + 1); + if (_M_losers[__right]._M_sup + || (!_M_losers[__left]._M_sup + && !_M_comp(*_M_losers[__right]._M_keyp, + *_M_losers[__left]._M_keyp))) + { + // Left one is less or equal. + _M_losers[__root] = _M_losers[__right]; + return __left; + } + else + { + // Right one is less. + _M_losers[__root] = _M_losers[__left]; + return __right; + } + } + } + + void __init() + { _M_losers[0] = _M_losers[__init_winner(1)]; } + + void __delete_min_insert(const _Tp& __key, bool __sup) + { #if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top! - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); + // no dummy sequence can ever be at the top! + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); #endif - int _M_source = _M_losers[0]._M_source; - for (unsigned int __pos = (_M_k + _M_source) / 2; __pos > 0; - __pos /= 2) - { - // The smaller one gets promoted, ties are broken by _M_source. - if ((_M_sup && (!_M_losers[__pos]._M_sup - || _M_losers[__pos]._M_source < _M_source)) - || (!_M_sup && !_M_losers[__pos]._M_sup - && ((_M_comp(_M_losers[__pos]._M_key, _M_key)) - || (!_M_comp(_M_key, _M_losers[__pos]._M_key) - && _M_losers[__pos]._M_source < _M_source)))) - { - // The other one is smaller. - std::swap(_M_losers[__pos]._M_sup, _M_sup); - std::swap(_M_losers[__pos]._M_source, _M_source); - std::swap(_M_losers[__pos]._M_key, _M_key); - } - } - - _M_losers[0]._M_sup = _M_sup; - _M_losers[0]._M_source = _M_source; - _M_losers[0]._M_key = _M_key; - } - }; + const _Tp* __keyp = &__key; + int __source = _M_losers[0]._M_source; + for (unsigned int __pos = (_M_k + __source) / 2; __pos > 0; + __pos /= 2) + { + // The smaller one gets promoted, ties are broken by __source. + if ((__sup && (!_M_losers[__pos]._M_sup + || _M_losers[__pos]._M_source < __source)) + || (!__sup && !_M_losers[__pos]._M_sup && + ((_M_comp(*_M_losers[__pos]._M_keyp, *__keyp)) + || (!_M_comp(*__keyp, *_M_losers[__pos]._M_keyp) + && _M_losers[__pos]._M_source < __source)))) + { + // The other one is smaller. + std::swap(_M_losers[__pos]._M_sup, __sup); + std::swap(_M_losers[__pos]._M_source, __source); + std::swap(_M_losers[__pos]._M_keyp, __keyp); + } + } + + _M_losers[0]._M_sup = __sup; + _M_losers[0]._M_source = __source; + _M_losers[0]._M_keyp = __keyp; + } + }; /** - * @brief Unstable _LoserTree variant. + * @brief Unstable _LoserTree implementation. * - * Stability (non-stable here) is selected with partial specialization. + * The stable variant is above. */ -template - class _LoserTree - : public _LoserTreeBase<_Tp, _Compare> - { - typedef _LoserTreeBase<_Tp, _Compare> _Base; - using _Base::_M_log_k; - using _Base::_M_k; - using _Base::_M_losers; - using _Base::_M_first_insert; - - public: - _LoserTree(unsigned int __k, _Compare __comp) - : _Base::_LoserTreeBase(__k, __comp) - { } - - /** - * Computes the winner of the competition at position "__root". - * - * Called recursively (starting at 0) to build the initial tree. - * - * @param __root __index of the "game" to start. - */ - unsigned int - __init_winner(unsigned int __root) - { - if (__root >= _M_k) - return __root; - else - { - unsigned int __left = __init_winner (2 * __root); - unsigned int __right = __init_winner (2 * __root + 1); - if (_M_losers[__right]._M_sup - || (!_M_losers[__left]._M_sup - && !_M_comp(_M_losers[__right]._M_key, - _M_losers[__left]._M_key))) - { - // Left one is less or equal. - _M_losers[__root] = _M_losers[__right]; - return __left; - } - else - { - // Right one is less. - _M_losers[__root] = _M_losers[__left]; - return __right; - } - } - } - - void - __init() - { _M_losers[0] = _M_losers[__init_winner(1)]; } - - /** - * Delete the _M_key smallest element and insert the element _M_key - * instead. - * - * @param _M_key the _M_key to insert - * @param _M_sup true iff _M_key is an explicitly marked supremum - */ - // Do not pass a const reference since _M_key will be used as local - // variable. - void - __delete_min_insert(_Tp _M_key, bool _M_sup) + template + class _LoserTreePointer + : public _LoserTreePointerBase<_Tp, _Compare> { + typedef _LoserTreePointerBase<_Tp, _Compare> _Base; + using _Base::_M_k; + using _Base::_M_losers; + + public: + _LoserTreePointer(unsigned int __k, _Compare __comp = std::less<_Tp>()) + : _Base::_LoserTreePointerBase(__k, __comp) + { } + + unsigned int + __init_winner(unsigned int __root) + { + if (__root >= _M_k) + return __root; + else + { + unsigned int __left = __init_winner(2 * __root); + unsigned int __right = __init_winner(2 * __root + 1); + if (_M_losers[__right]._M_sup + || (!_M_losers[__left]._M_sup + && !_M_comp(*_M_losers[__right]._M_keyp, + *_M_losers[__left]._M_keyp))) + { + // Left one is less or equal. + _M_losers[__root] = _M_losers[__right]; + return __left; + } + else + { + // Right one is less. + _M_losers[__root] = _M_losers[__left]; + return __right; + } + } + } + + void __init() + { _M_losers[0] = _M_losers[__init_winner(1)]; } + + void __delete_min_insert(const _Tp& __key, bool __sup) + { #if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top! - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); + // no dummy sequence can ever be at the top! + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); #endif - int _M_source = _M_losers[0]._M_source; - for (unsigned int __pos = (_M_k + _M_source) / 2; __pos > 0; - __pos /= 2) - { - // The smaller one gets promoted. - if (_M_sup || (!_M_losers[__pos]._M_sup - && _M_comp(_M_losers[__pos]._M_key, _M_key))) - { - // The other one is smaller. - std::swap(_M_losers[__pos]._M_sup, _M_sup); - std::swap(_M_losers[__pos]._M_source, _M_source); - std::swap(_M_losers[__pos]._M_key, _M_key); - } - } - - _M_losers[0]._M_sup = _M_sup; - _M_losers[0]._M_source = _M_source; - _M_losers[0]._M_key = _M_key; - } - }; - -/** - * @brief Base class of _Loser Tree implementation using pointers. - */ -template - class _LoserTreePointerBase - { - protected: - /** @brief Internal representation of _LoserTree __elements. */ - struct _Loser - { - bool _M_sup; - int _M_source; - const _Tp* _M_keyp; + const _Tp* __keyp = &__key; + int __source = _M_losers[0]._M_source; + for (unsigned int __pos = (_M_k + __source) / 2; __pos > 0; + __pos /= 2) + { + // The smaller one gets promoted. + if (__sup || (!_M_losers[__pos]._M_sup + && _M_comp(*_M_losers[__pos]._M_keyp, *__keyp))) + { + // The other one is smaller. + std::swap(_M_losers[__pos]._M_sup, __sup); + std::swap(_M_losers[__pos]._M_source, __source); + std::swap(_M_losers[__pos]._M_keyp, __keyp); + } + } + + _M_losers[0]._M_sup = __sup; + _M_losers[0]._M_source = __source; + _M_losers[0]._M_keyp = __keyp; + } }; - unsigned int _M_ik, _M_k, _M_offset; - _Loser* _M_losers; - _Compare _M_comp; - - public: - _LoserTreePointerBase(unsigned int __k, - _Compare __comp = std::less<_Tp>()) - : _M_comp(__comp) + /** @brief Base class for unguarded _LoserTree implementation. + * + * The whole element is copied into the tree structure. + * + * No guarding is done, therefore not a single input sequence must + * run empty. Unused __sequence heads are marked with a sentinel which + * is > all elements that are to be merged. + * + * This is a very fast variant. + */ + template + class _LoserTreeUnguardedBase { - _M_ik = __k; - - // Next greater power of 2. - _M_k = 1 << (__rd_log2(_M_ik - 1) + 1); - _M_offset = _M_k; - _M_losers = new _Loser[_M_k * 2]; - for (unsigned int __i = _M_ik - 1; __i < _M_k; __i++) - _M_losers[__i + _M_k]._M_sup = true; - } + protected: + struct _Loser + { + int _M_source; + _Tp _M_key; + }; + + unsigned int _M_ik, _M_k, _M_offset; + _Loser* _M_losers; + _Compare _M_comp; + + public: + _LoserTreeUnguardedBase(unsigned int __k, const _Tp __sentinel, + _Compare __comp = std::less<_Tp>()) + : _M_comp(__comp) + { + _M_ik = __k; + + // Next greater power of 2. + _M_k = 1 << (__rd_log2(_M_ik - 1) + 1); + _M_offset = _M_k; + // Avoid default-constructing _M_losers[]._M_key + _M_losers = static_cast<_Loser*>(::operator new(2 * _M_k + * sizeof(_Loser))); + + for (unsigned int __i = _M_k + _M_ik - 1; __i < (2 * _M_k); ++__i) + { + _M_losers[__i]._M_key = __sentinel; + _M_losers[__i]._M_source = -1; + } + } + + ~_LoserTreeUnguardedBase() + { ::operator delete(_M_losers); } + + int + __get_min_source() + { +#if _GLIBCXX_ASSERTIONS + // no dummy sequence can ever be at the top! + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); +#endif + return _M_losers[0]._M_source; + } - ~_LoserTreePointerBase() - { ::operator delete[](_M_losers); } + void + __insert_start(const _Tp& __key, int __source, bool) + { + unsigned int __pos = _M_k + __source; - int __get_min_source() - { return _M_losers[0]._M_source; } + new(&(_M_losers[__pos]._M_key)) _Tp(__key); + _M_losers[__pos]._M_source = __source; + } + }; - void __insert_start(const _Tp& _M_key, int _M_source, bool _M_sup) + /** + * @brief Stable implementation of unguarded _LoserTree. + * + * Unstable variant is selected below with partial specialization. + */ + template + class _LoserTreeUnguarded + : public _LoserTreeUnguardedBase<_Tp, _Compare> { - unsigned int __pos = _M_k + _M_source; - - _M_losers[__pos]._M_sup = _M_sup; - _M_losers[__pos]._M_source = _M_source; - _M_losers[__pos]._M_keyp = &_M_key; - } - }; - -/** - * @brief Stable _LoserTree implementation. - * - * The unstable variant is implemented using partial instantiation below. - */ -template - class _LoserTreePointer - : public _LoserTreePointerBase<_Tp, _Compare> - { - typedef _LoserTreePointerBase<_Tp, _Compare> _Base; - using _Base::_M_k; - using _Base::_M_losers; + typedef _LoserTreeUnguardedBase<_Tp, _Compare> _Base; + using _Base::_M_k; + using _Base::_M_losers; public: - _LoserTreePointer(unsigned int __k, _Compare __comp = std::less<_Tp>()) - : _Base::_LoserTreePointerBase(__k, __comp) - { } + _LoserTreeUnguarded(unsigned int __k, const _Tp __sentinel, + _Compare __comp = std::less<_Tp>()) + : _Base::_LoserTreeUnguardedBase(__k, __sentinel, __comp) + { } + + unsigned int + __init_winner(unsigned int __root) + { + if (__root >= _M_k) + return __root; + else + { + unsigned int __left = __init_winner(2 * __root); + unsigned int __right = __init_winner(2 * __root + 1); + if (!_M_comp(_M_losers[__right]._M_key, + _M_losers[__left]._M_key)) + { + // Left one is less or equal. + _M_losers[__root] = _M_losers[__right]; + return __left; + } + else + { + // Right one is less. + _M_losers[__root] = _M_losers[__left]; + return __right; + } + } + } + + void + __init() + { + _M_losers[0] = _M_losers[__init_winner(1)]; - unsigned int - __init_winner(unsigned int __root) - { - if (__root >= _M_k) - return __root; - else - { - unsigned int __left = __init_winner (2 * __root); - unsigned int __right = __init_winner (2 * __root + 1); - if (_M_losers[__right]._M_sup - || (!_M_losers[__left]._M_sup - && !_M_comp(*_M_losers[__right]._M_keyp, - *_M_losers[__left]._M_keyp))) - { - // Left one is less or equal. - _M_losers[__root] = _M_losers[__right]; - return __left; - } - else - { - // Right one is less. - _M_losers[__root] = _M_losers[__left]; - return __right; - } - } - } - - void __init() - { _M_losers[0] = _M_losers[__init_winner(1)]; } - - void __delete_min_insert(const _Tp& _M_key, bool _M_sup) - { #if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top! - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); + // no dummy sequence can ever be at the top at the beginning + // (0 sequences!) + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); #endif + } - const _Tp* _M_keyp = &_M_key; - int _M_source = _M_losers[0]._M_source; - for (unsigned int __pos = (_M_k + _M_source) / 2; __pos > 0; __pos /= 2) - { - // The smaller one gets promoted, ties are broken by _M_source. - if ((_M_sup && (!_M_losers[__pos]._M_sup || - _M_losers[__pos]._M_source < _M_source)) || - (!_M_sup && !_M_losers[__pos]._M_sup && - ((_M_comp(*_M_losers[__pos]._M_keyp, *_M_keyp)) || - (!_M_comp(*_M_keyp, *_M_losers[__pos]._M_keyp) - && _M_losers[__pos]._M_source < _M_source)))) - { - // The other one is smaller. - std::swap(_M_losers[__pos]._M_sup, _M_sup); - std::swap(_M_losers[__pos]._M_source, _M_source); - std::swap(_M_losers[__pos]._M_keyp, _M_keyp); - } - } - - _M_losers[0]._M_sup = _M_sup; - _M_losers[0]._M_source = _M_source; - _M_losers[0]._M_keyp = _M_keyp; - } - }; - -/** - * @brief Unstable _LoserTree implementation. - * - * The stable variant is above. - */ -template - class _LoserTreePointer - : public _LoserTreePointerBase<_Tp, _Compare> - { - typedef _LoserTreePointerBase<_Tp, _Compare> _Base; - using _Base::_M_k; - using _Base::_M_losers; - - public: - _LoserTreePointer(unsigned int __k, _Compare __comp = std::less<_Tp>()) - : _Base::_LoserTreePointerBase(__k, __comp) - { } - - unsigned int - __init_winner(unsigned int __root) - { - if (__root >= _M_k) - return __root; - else - { - unsigned int __left = __init_winner (2 * __root); - unsigned int __right = __init_winner (2 * __root + 1); - if (_M_losers[__right]._M_sup - || (!_M_losers[__left]._M_sup - && !_M_comp(*_M_losers[__right]._M_keyp, - *_M_losers[__left]._M_keyp))) - { - // Left one is less or equal. - _M_losers[__root] = _M_losers[__right]; - return __left; - } - else - { - // Right one is less. - _M_losers[__root] = _M_losers[__left]; - return __right; - } - } - } - - void __init() - { _M_losers[0] = _M_losers[__init_winner(1)]; } - - void __delete_min_insert(const _Tp& _M_key, bool _M_sup) - { + // Do not pass a const reference since __key will be used as + // local variable. + void + __delete_min_insert(_Tp __key, bool) + { #if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top! - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); + // no dummy sequence can ever be at the top! + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); #endif - const _Tp* _M_keyp = &_M_key; - int _M_source = _M_losers[0]._M_source; - for (unsigned int __pos = (_M_k + _M_source) / 2; __pos > 0; - __pos /= 2) - { - // The smaller one gets promoted. - if (_M_sup || (!_M_losers[__pos]._M_sup - && _M_comp(*_M_losers[__pos]._M_keyp, *_M_keyp))) - { - // The other one is smaller. - std::swap(_M_losers[__pos]._M_sup, _M_sup); - std::swap(_M_losers[__pos]._M_source, _M_source); - std::swap(_M_losers[__pos]._M_keyp, _M_keyp); - } - } - - _M_losers[0]._M_sup = _M_sup; - _M_losers[0]._M_source = _M_source; - _M_losers[0]._M_keyp = _M_keyp; - } - }; - -/** @brief Base class for unguarded _LoserTree implementation. - * - * The whole element is copied into the tree structure. - * - * No guarding is done, therefore not a single input sequence must - * run empty. Unused __sequence heads are marked with a sentinel which - * is > all elements that are to be merged. - * - * This is a very fast variant. - */ -template - class _LoserTreeUnguardedBase - { - protected: - struct _Loser - { - int _M_source; - _Tp _M_key; + int __source = _M_losers[0]._M_source; + for (unsigned int __pos = (_M_k + __source) / 2; __pos > 0; + __pos /= 2) + { + // The smaller one gets promoted, ties are broken by _M_source. + if (_M_comp(_M_losers[__pos]._M_key, __key) + || (!_M_comp(__key, _M_losers[__pos]._M_key) + && _M_losers[__pos]._M_source < __source)) + { + // The other one is smaller. + std::swap(_M_losers[__pos]._M_source, __source); + std::swap(_M_losers[__pos]._M_key, __key); + } + } + + _M_losers[0]._M_source = __source; + _M_losers[0]._M_key = __key; + } }; - unsigned int _M_ik, _M_k, _M_offset; - _Loser* _M_losers; - _Compare _M_comp; - - public: - _LoserTreeUnguardedBase(unsigned int __k, const _Tp _sentinel, - _Compare __comp = std::less<_Tp>()) - : _M_comp(__comp) - { - _M_ik = __k; - - // Next greater power of 2. - _M_k = 1 << (__rd_log2(_M_ik - 1) + 1); - _M_offset = _M_k; - // Avoid default-constructing _M_losers[]._M_key - _M_losers - = static_cast<_Loser*>(::operator new(2 * _M_k * sizeof(_Loser))); - - for (unsigned int __i = _M_k + _M_ik - 1; __i < (2 * _M_k); ++__i) - { - _M_losers[__i]._M_key = _sentinel; - _M_losers[__i]._M_source = -1; - } - } - - ~_LoserTreeUnguardedBase() - { ::operator delete(_M_losers); } - - int - __get_min_source() - { -#if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top! - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); -#endif - return _M_losers[0]._M_source; - } - - void - __insert_start(const _Tp& _M_key, int _M_source, bool) - { - unsigned int __pos = _M_k + _M_source; - - new(&(_M_losers[__pos]._M_key)) _Tp(_M_key); - _M_losers[__pos]._M_source = _M_source; - } - }; - -/** - * @brief Stable implementation of unguarded _LoserTree. - * - * Unstable variant is selected below with partial specialization. - */ -template - class _LoserTreeUnguarded - : public _LoserTreeUnguardedBase<_Tp, _Compare> - { - typedef _LoserTreeUnguardedBase<_Tp, _Compare> _Base; - using _Base::_M_k; - using _Base::_M_losers; - -public: - _LoserTreeUnguarded(unsigned int __k, const _Tp _sentinel, - _Compare __comp = std::less<_Tp>()) - : _Base::_LoserTreeUnguardedBase(__k, _sentinel, __comp) - { } - - unsigned int - __init_winner(unsigned int __root) - { - if (__root >= _M_k) - return __root; - else - { - unsigned int __left = __init_winner (2 * __root); - unsigned int __right = __init_winner (2 * __root + 1); - if (!_M_comp(_M_losers[__right]._M_key, _M_losers[__left]._M_key)) - { - // Left one is less or equal. - _M_losers[__root] = _M_losers[__right]; - return __left; - } - else - { - // Right one is less. - _M_losers[__root] = _M_losers[__left]; - return __right; - } - } - } - - void - __init() + /** + * @brief Non-Stable implementation of unguarded _LoserTree. + * + * Stable implementation is above. + */ + template + class _LoserTreeUnguarded + : public _LoserTreeUnguardedBase<_Tp, _Compare> { - _M_losers[0] = _M_losers[__init_winner(1)]; - -#if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top at the beginning - // (0 sequences!) - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); -#endif - } + typedef _LoserTreeUnguardedBase<_Tp, _Compare> _Base; + using _Base::_M_k; + using _Base::_M_losers; - // Do not pass a const reference since _M_key will be used as - // local variable. - void - __delete_min_insert(_Tp _M_key, bool) - { -#if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top! - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); -#endif - - int _M_source = _M_losers[0]._M_source; - for (unsigned int __pos = (_M_k + _M_source) / 2; __pos > 0; __pos /= 2) - { - // The smaller one gets promoted, ties are broken by _M_source. - if (_M_comp(_M_losers[__pos]._M_key, _M_key) - || (!_M_comp(_M_key, _M_losers[__pos]._M_key) - && _M_losers[__pos]._M_source < _M_source)) - { - // The other one is smaller. - std::swap(_M_losers[__pos]._M_source, _M_source); - std::swap(_M_losers[__pos]._M_key, _M_key); - } - } - - _M_losers[0]._M_source = _M_source; - _M_losers[0]._M_key = _M_key; - } - }; - -/** - * @brief Non-Stable implementation of unguarded _LoserTree. - * - * Stable implementation is above. - */ -template - class _LoserTreeUnguarded - : public _LoserTreeUnguardedBase<_Tp, _Compare> - { - typedef _LoserTreeUnguardedBase<_Tp, _Compare> _Base; - using _Base::_M_k; - using _Base::_M_losers; - -public: - _LoserTreeUnguarded(unsigned int __k, const _Tp _sentinel, - _Compare __comp = std::less<_Tp>()) - : _Base::_LoserTreeUnguardedBase(__k, _sentinel, __comp) - { } - - unsigned int - __init_winner(unsigned int __root) - { - if (__root >= _M_k) - return __root; - else - { - unsigned int __left = __init_winner (2 * __root); - unsigned int __right = __init_winner (2 * __root + 1); + public: + _LoserTreeUnguarded(unsigned int __k, const _Tp __sentinel, + _Compare __comp = std::less<_Tp>()) + : _Base::_LoserTreeUnguardedBase(__k, __sentinel, __comp) + { } + + unsigned int + __init_winner(unsigned int __root) + { + if (__root >= _M_k) + return __root; + else + { + unsigned int __left = __init_winner(2 * __root); + unsigned int __right = __init_winner(2 * __root + 1); #if _GLIBCXX_ASSERTIONS - // If __left one is sentinel then __right one must be, too. - if (_M_losers[__left]._M_source == -1) - _GLIBCXX_PARALLEL_ASSERT(_M_losers[__right]._M_source == -1); + // If __left one is sentinel then __right one must be, too. + if (_M_losers[__left]._M_source == -1) + _GLIBCXX_PARALLEL_ASSERT(_M_losers[__right]._M_source == -1); #endif - if (!_M_comp(_M_losers[__right]._M_key, _M_losers[__left]._M_key)) - { - // Left one is less or equal. - _M_losers[__root] = _M_losers[__right]; - return __left; - } - else - { - // Right one is less. - _M_losers[__root] = _M_losers[__left]; - return __right; - } - } - } - - void - __init() - { - _M_losers[0] = _M_losers[__init_winner(1)]; + if (!_M_comp(_M_losers[__right]._M_key, + _M_losers[__left]._M_key)) + { + // Left one is less or equal. + _M_losers[__root] = _M_losers[__right]; + return __left; + } + else + { + // Right one is less. + _M_losers[__root] = _M_losers[__left]; + return __right; + } + } + } + + void + __init() + { + _M_losers[0] = _M_losers[__init_winner(1)]; #if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top at the beginning - // (0 sequences!) - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); + // no dummy sequence can ever be at the top at the beginning + // (0 sequences!) + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); #endif - } + } - // Do not pass a const reference since _M_key will be used as - // local variable. - void - __delete_min_insert(_Tp _M_key, bool) - { + // Do not pass a const reference since __key will be used as + // local variable. + void + __delete_min_insert(_Tp __key, bool) + { #if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top! - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); + // no dummy sequence can ever be at the top! + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); #endif - int _M_source = _M_losers[0]._M_source; - for (unsigned int __pos = (_M_k + _M_source) / 2; __pos > 0; __pos /= 2) - { - // The smaller one gets promoted. - if (_M_comp(_M_losers[__pos]._M_key, _M_key)) - { - // The other one is smaller. - std::swap(_M_losers[__pos]._M_source, _M_source); - std::swap(_M_losers[__pos]._M_key, _M_key); - } - } - - _M_losers[0]._M_source = _M_source; - _M_losers[0]._M_key = _M_key; - } - }; - -/** @brief Unguarded loser tree, keeping only pointers to the -* elements in the tree structure. -* -* No guarding is done, therefore not a single input sequence must -* run empty. This is a very fast variant. -*/ -template - class _LoserTreePointerUnguardedBase - { - protected: - struct _Loser - { - int _M_source; - const _Tp* _M_keyp; + int __source = _M_losers[0]._M_source; + for (unsigned int __pos = (_M_k + __source) / 2; __pos > 0; + __pos /= 2) + { + // The smaller one gets promoted. + if (_M_comp(_M_losers[__pos]._M_key, __key)) + { + // The other one is smaller. + std::swap(_M_losers[__pos]._M_source, __source); + std::swap(_M_losers[__pos]._M_key, __key); + } + } + + _M_losers[0]._M_source = __source; + _M_losers[0]._M_key = __key; + } }; - unsigned int _M_ik, _M_k, _M_offset; - _Loser* _M_losers; - _Compare _M_comp; - - public: - - _LoserTreePointerUnguardedBase(unsigned int __k, const _Tp& _sentinel, - _Compare __comp = std::less<_Tp>()) - : _M_comp(__comp) - { - _M_ik = __k; - - // Next greater power of 2. - _M_k = 1 << (__rd_log2(_M_ik - 1) + 1); - _M_offset = _M_k; - // Avoid default-constructing _M_losers[]._M_key - _M_losers = new _Loser[2 * _M_k]; - - for (unsigned int __i = _M_k + _M_ik - 1; __i < (2 * _M_k); ++__i) - { - _M_losers[__i]._M_keyp = &_sentinel; - _M_losers[__i]._M_source = -1; - } - } - - ~_LoserTreePointerUnguardedBase() - { delete[] _M_losers; } - - int - __get_min_source() + /** @brief Unguarded loser tree, keeping only pointers to the + * elements in the tree structure. + * + * No guarding is done, therefore not a single input sequence must + * run empty. This is a very fast variant. + */ + template + class _LoserTreePointerUnguardedBase { + protected: + struct _Loser + { + int _M_source; + const _Tp* _M_keyp; + }; + + unsigned int _M_ik, _M_k, _M_offset; + _Loser* _M_losers; + _Compare _M_comp; + + public: + + _LoserTreePointerUnguardedBase(unsigned int __k, const _Tp& __sentinel, + _Compare __comp = std::less<_Tp>()) + : _M_comp(__comp) + { + _M_ik = __k; + + // Next greater power of 2. + _M_k = 1 << (__rd_log2(_M_ik - 1) + 1); + _M_offset = _M_k; + // Avoid default-constructing _M_losers[]._M_key + _M_losers = new _Loser[2 * _M_k]; + + for (unsigned int __i = _M_k + _M_ik - 1; __i < (2 * _M_k); ++__i) + { + _M_losers[__i]._M_keyp = &__sentinel; + _M_losers[__i]._M_source = -1; + } + } + + ~_LoserTreePointerUnguardedBase() + { delete[] _M_losers; } + + int + __get_min_source() + { #if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top! - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); + // no dummy sequence can ever be at the top! + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); #endif - return _M_losers[0]._M_source; - } + return _M_losers[0]._M_source; + } - void - __insert_start(const _Tp& _M_key, int _M_source, bool) - { - unsigned int __pos = _M_k + _M_source; - - _M_losers[__pos]._M_keyp = &_M_key; - _M_losers[__pos]._M_source = _M_source; - } - }; - -/** - * @brief Stable unguarded _LoserTree variant storing pointers. - * - * Unstable variant is implemented below using partial specialization. - */ -template - class _LoserTreePointerUnguarded - : public _LoserTreePointerUnguardedBase<_Tp, _Compare> - { - typedef _LoserTreePointerUnguardedBase<_Tp, _Compare> _Base; - using _Base::_M_k; - using _Base::_M_losers; + void + __insert_start(const _Tp& __key, int __source, bool) + { + unsigned int __pos = _M_k + __source; - public: - _LoserTreePointerUnguarded(unsigned int __k, const _Tp& _sentinel, - _Compare __comp = std::less<_Tp>()) - : _Base::_LoserTreePointerUnguardedBase(__k, _sentinel, __comp) - { } + _M_losers[__pos]._M_keyp = &__key; + _M_losers[__pos]._M_source = __source; + } + }; - unsigned int - __init_winner(unsigned int __root) - { - if (__root >= _M_k) - return __root; - else - { - unsigned int __left = __init_winner (2 * __root); - unsigned int __right = __init_winner (2 * __root + 1); - if (!_M_comp(*_M_losers[__right]._M_keyp, - *_M_losers[__left]._M_keyp)) - { - // Left one is less or equal. - _M_losers[__root] = _M_losers[__right]; - return __left; - } - else - { - // Right one is less. - _M_losers[__root] = _M_losers[__left]; - return __right; - } - } - } - - void - __init() + /** + * @brief Stable unguarded _LoserTree variant storing pointers. + * + * Unstable variant is implemented below using partial specialization. + */ + template + class _LoserTreePointerUnguarded + : public _LoserTreePointerUnguardedBase<_Tp, _Compare> { - _M_losers[0] = _M_losers[__init_winner(1)]; + typedef _LoserTreePointerUnguardedBase<_Tp, _Compare> _Base; + using _Base::_M_k; + using _Base::_M_losers; + + public: + _LoserTreePointerUnguarded(unsigned int __k, const _Tp& __sentinel, + _Compare __comp = std::less<_Tp>()) + : _Base::_LoserTreePointerUnguardedBase(__k, __sentinel, __comp) + { } + + unsigned int + __init_winner(unsigned int __root) + { + if (__root >= _M_k) + return __root; + else + { + unsigned int __left = __init_winner(2 * __root); + unsigned int __right = __init_winner(2 * __root + 1); + if (!_M_comp(*_M_losers[__right]._M_keyp, + *_M_losers[__left]._M_keyp)) + { + // Left one is less or equal. + _M_losers[__root] = _M_losers[__right]; + return __left; + } + else + { + // Right one is less. + _M_losers[__root] = _M_losers[__left]; + return __right; + } + } + } + + void + __init() + { + _M_losers[0] = _M_losers[__init_winner(1)]; #if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top at the beginning - // (0 sequences!) - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); + // no dummy sequence can ever be at the top at the beginning + // (0 sequences!) + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); #endif - } + } - void - __delete_min_insert(const _Tp& _M_key, bool _M_sup) - { + void + __delete_min_insert(const _Tp& __key, bool __sup) + { #if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top! - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); + // no dummy sequence can ever be at the top! + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); #endif - const _Tp* _M_keyp = &_M_key; - int _M_source = _M_losers[0]._M_source; - for (unsigned int __pos = (_M_k + _M_source) / 2; __pos > 0; __pos /= 2) - { - // The smaller one gets promoted, ties are broken by _M_source. - if (_M_comp(*_M_losers[__pos]._M_keyp, *_M_keyp) - || (!_M_comp(*_M_keyp, *_M_losers[__pos]._M_keyp) - && _M_losers[__pos]._M_source < _M_source)) - { - // The other one is smaller. - std::swap(_M_losers[__pos]._M_source, _M_source); - std::swap(_M_losers[__pos]._M_keyp, _M_keyp); - } - } - - _M_losers[0]._M_source = _M_source; - _M_losers[0]._M_keyp = _M_keyp; - } - }; - -/** - * @brief Unstable unguarded _LoserTree variant storing pointers. - * - * Stable variant is above. - */ -template - class _LoserTreePointerUnguarded - : public _LoserTreePointerUnguardedBase<_Tp, _Compare> - { - typedef _LoserTreePointerUnguardedBase<_Tp, _Compare> _Base; - using _Base::_M_k; - using _Base::_M_losers; - -public: - _LoserTreePointerUnguarded(unsigned int __k, const _Tp& _sentinel, - _Compare __comp = std::less<_Tp>()) - : _Base::_LoserTreePointerUnguardedBase(__k, _sentinel, __comp) - { } - - unsigned int - __init_winner(unsigned int __root) + const _Tp* __keyp = &__key; + int __source = _M_losers[0]._M_source; + for (unsigned int __pos = (_M_k + __source) / 2; __pos > 0; + __pos /= 2) + { + // The smaller one gets promoted, ties are broken by _M_source. + if (_M_comp(*_M_losers[__pos]._M_keyp, *__keyp) + || (!_M_comp(*__keyp, *_M_losers[__pos]._M_keyp) + && _M_losers[__pos]._M_source < __source)) + { + // The other one is smaller. + std::swap(_M_losers[__pos]._M_source, __source); + std::swap(_M_losers[__pos]._M_keyp, __keyp); + } + } + + _M_losers[0]._M_source = __source; + _M_losers[0]._M_keyp = __keyp; + } + }; + + /** + * @brief Unstable unguarded _LoserTree variant storing pointers. + * + * Stable variant is above. + */ + template + class _LoserTreePointerUnguarded + : public _LoserTreePointerUnguardedBase<_Tp, _Compare> { - if (__root >= _M_k) - return __root; - else - { - unsigned int __left = __init_winner (2 * __root); - unsigned int __right = __init_winner (2 * __root + 1); + typedef _LoserTreePointerUnguardedBase<_Tp, _Compare> _Base; + using _Base::_M_k; + using _Base::_M_losers; + + public: + _LoserTreePointerUnguarded(unsigned int __k, const _Tp& __sentinel, + _Compare __comp = std::less<_Tp>()) + : _Base::_LoserTreePointerUnguardedBase(__k, __sentinel, __comp) + { } + + unsigned int + __init_winner(unsigned int __root) + { + if (__root >= _M_k) + return __root; + else + { + unsigned int __left = __init_winner(2 * __root); + unsigned int __right = __init_winner(2 * __root + 1); #if _GLIBCXX_ASSERTIONS - // If __left one is sentinel then __right one must be, too. - if (_M_losers[__left]._M_source == -1) - _GLIBCXX_PARALLEL_ASSERT(_M_losers[__right]._M_source == -1); + // If __left one is sentinel then __right one must be, too. + if (_M_losers[__left]._M_source == -1) + _GLIBCXX_PARALLEL_ASSERT(_M_losers[__right]._M_source == -1); #endif - if (!_M_comp(*_M_losers[__right]._M_keyp, - *_M_losers[__left]._M_keyp)) - { - // Left one is less or equal. - _M_losers[__root] = _M_losers[__right]; - return __left; - } - else - { - // Right one is less. - _M_losers[__root] = _M_losers[__left]; - return __right; - } - } - } - - void - __init() - { - _M_losers[0] = _M_losers[__init_winner(1)]; + if (!_M_comp(*_M_losers[__right]._M_keyp, + *_M_losers[__left]._M_keyp)) + { + // Left one is less or equal. + _M_losers[__root] = _M_losers[__right]; + return __left; + } + else + { + // Right one is less. + _M_losers[__root] = _M_losers[__left]; + return __right; + } + } + } + + void + __init() + { + _M_losers[0] = _M_losers[__init_winner(1)]; #if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top at the beginning - // (0 sequences!) - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); + // no dummy sequence can ever be at the top at the beginning + // (0 sequences!) + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); #endif - } + } - void - __delete_min_insert(const _Tp& _M_key, bool _M_sup) - { + void + __delete_min_insert(const _Tp& __key, bool __sup) + { #if _GLIBCXX_ASSERTIONS - // no dummy sequence can ever be at the top! - _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); + // no dummy sequence can ever be at the top! + _GLIBCXX_PARALLEL_ASSERT(_M_losers[0]._M_source != -1); #endif - const _Tp* _M_keyp = &_M_key; - int _M_source = _M_losers[0]._M_source; - for (unsigned int __pos = (_M_k + _M_source) / 2; __pos > 0; __pos /= 2) - { - // The smaller one gets promoted. - if (_M_comp(*(_M_losers[__pos]._M_keyp), *_M_keyp)) - { - // The other one is smaller. - std::swap(_M_losers[__pos]._M_source, _M_source); - std::swap(_M_losers[__pos]._M_keyp, _M_keyp); - } - } - - _M_losers[0]._M_source = _M_source; - _M_losers[0]._M_keyp = _M_keyp; - } - }; + const _Tp* __keyp = &__key; + int __source = _M_losers[0]._M_source; + for (unsigned int __pos = (_M_k + __source) / 2; __pos > 0; + __pos /= 2) + { + // The smaller one gets promoted. + if (_M_comp(*(_M_losers[__pos]._M_keyp), *__keyp)) + { + // The other one is smaller. + std::swap(_M_losers[__pos]._M_source, __source); + std::swap(_M_losers[__pos]._M_keyp, __keyp); + } + } + + _M_losers[0]._M_source = __source; + _M_losers[0]._M_keyp = __keyp; + } + }; } // namespace __gnu_parallel #endif /* _GLIBCXX_PARALLEL_LOSERTREE_H */ diff --git a/libstdc++-v3/include/parallel/merge.h b/libstdc++-v3/include/parallel/merge.h index c323c629893..2343d7326e8 100644 --- a/libstdc++-v3/include/parallel/merge.h +++ b/libstdc++-v3/include/parallel/merge.h @@ -54,11 +54,10 @@ namespace __gnu_parallel typename _OutputIterator, typename _DifferenceTp, typename _Compare> _OutputIterator - __merge_advance_usual(_RAIter1& __begin1, - _RAIter1 __end1, - _RAIter2& __begin2, - _RAIter2 __end2, _OutputIterator __target, - _DifferenceTp __max_length, _Compare __comp) + __merge_advance_usual(_RAIter1& __begin1, _RAIter1 __end1, + _RAIter2& __begin2, _RAIter2 __end2, + _OutputIterator __target, + _DifferenceTp __max_length, _Compare __comp) { typedef _DifferenceTp _DifferenceType; while (__begin1 != __end1 && __begin2 != __end2 && __max_length > 0) @@ -103,12 +102,10 @@ namespace __gnu_parallel typename _OutputIterator, typename _DifferenceTp, typename _Compare> _OutputIterator - __merge_advance_movc(_RAIter1& __begin1, - _RAIter1 __end1, - _RAIter2& __begin2, - _RAIter2 __end2, - _OutputIterator __target, - _DifferenceTp __max_length, _Compare __comp) + __merge_advance_movc(_RAIter1& __begin1, _RAIter1 __end1, + _RAIter2& __begin2, _RAIter2 __end2, + _OutputIterator __target, + _DifferenceTp __max_length, _Compare __comp) { typedef _DifferenceTp _DifferenceType; typedef typename std::iterator_traits<_RAIter1>::value_type @@ -172,14 +169,14 @@ namespace __gnu_parallel typename _Compare> inline _OutputIterator __merge_advance(_RAIter1& __begin1, _RAIter1 __end1, - _RAIter2& __begin2, _RAIter2 __end2, - _OutputIterator __target, _DifferenceTp __max_length, - _Compare __comp) + _RAIter2& __begin2, _RAIter2 __end2, + _OutputIterator __target, _DifferenceTp __max_length, + _Compare __comp) { _GLIBCXX_CALL(__max_length) - return __merge_advance_movc(__begin1, __end1, __begin2, __end2, __target, - __max_length, __comp); + return __merge_advance_movc(__begin1, __end1, __begin2, __end2, + __target, __max_length, __comp); } /** @brief Merge routine fallback to sequential in case the @@ -195,17 +192,15 @@ namespace __gnu_parallel template inline _RAIter3 - __parallel_merge_advance(_RAIter1& __begin1, - _RAIter1 __end1, - _RAIter2& __begin2, - // different iterators, parallel implementation - // not available - _RAIter2 __end2, - _RAIter3 __target, typename - std::iterator_traits<_RAIter1>:: - difference_type __max_length, _Compare __comp) + __parallel_merge_advance(_RAIter1& __begin1, _RAIter1 __end1, + _RAIter2& __begin2, + // different iterators, parallel implementation + // not available + _RAIter2 __end2, _RAIter3 __target, typename + std::iterator_traits<_RAIter1>:: + difference_type __max_length, _Compare __comp) { return __merge_advance(__begin1, __end1, __begin2, __end2, __target, - __max_length, __comp); } + __max_length, __comp); } /** @brief Parallel merge routine being able to merge only the @__c * __max_length smallest elements. @@ -225,13 +220,11 @@ namespace __gnu_parallel template inline _RAIter3 - __parallel_merge_advance(_RAIter1& __begin1, - _RAIter1 __end1, - _RAIter1& __begin2, - _RAIter1 __end2, - _RAIter3 __target, typename - std::iterator_traits<_RAIter1>:: - difference_type __max_length, _Compare __comp) + __parallel_merge_advance(_RAIter1& __begin1, _RAIter1 __end1, + _RAIter1& __begin2, _RAIter1 __end2, + _RAIter3 __target, typename + std::iterator_traits<_RAIter1>:: + difference_type __max_length, _Compare __comp) { typedef typename std::iterator_traits<_RAIter1>::value_type _ValueType; @@ -242,17 +235,14 @@ namespace __gnu_parallel typedef typename std::pair<_RAIter1, _RAIter1> _IteratorPair; - _IteratorPair - seqs[2] = { std::make_pair(__begin1, __end1), - std::make_pair(__begin2, __end2) }; - _RAIter3 - __target_end = parallel_multiway_merge - < /* __stable = */ true, /* __sentinels = */ false>( - seqs, seqs + 2, __target, - multiway_merge_exact_splitting - < /* __stable = */ true, _IteratorPair*, - _Compare, _DifferenceType1>, - __max_length, __comp, omp_get_max_threads()); + _IteratorPair __seqs[2] = { std::make_pair(__begin1, __end1), + std::make_pair(__begin2, __end2) }; + _RAIter3 __target_end = parallel_multiway_merge + < /* __stable = */ true, /* __sentinels = */ false> + (__seqs, __seqs + 2, __target, multiway_merge_exact_splitting + < /* __stable = */ true, _IteratorPair*, + _Compare, _DifferenceType1>, __max_length, __comp, + omp_get_max_threads()); return __target_end; } diff --git a/libstdc++-v3/include/parallel/multiseq_selection.h b/libstdc++-v3/include/parallel/multiseq_selection.h index ac06385b44b..e434eee624c 100644 --- a/libstdc++-v3/include/parallel/multiseq_selection.h +++ b/libstdc++-v3/include/parallel/multiseq_selection.h @@ -53,8 +53,8 @@ namespace __gnu_parallel /** @brief Compare __a pair of types lexicographically, ascending. */ template class _Lexicographic - : public std::binary_function< - std::pair<_T1, _T2>, std::pair<_T1, _T2>, bool> + : public std::binary_function, + std::pair<_T1, _T2>, bool> { private: _Compare& _M_comp; @@ -142,19 +142,19 @@ namespace __gnu_parallel // Number of sequences, number of elements in total (possibly // including padding). - _DifferenceType __m = std::distance(__begin_seqs, __end_seqs), __N = 0, + _DifferenceType __m = std::distance(__begin_seqs, __end_seqs), __nn = 0, __nmax, __n, __r; for (int __i = 0; __i < __m; __i++) { - __N += std::distance(__begin_seqs[__i].first, + __nn += std::distance(__begin_seqs[__i].first, __begin_seqs[__i].second); _GLIBCXX_PARALLEL_ASSERT( std::distance(__begin_seqs[__i].first, __begin_seqs[__i].second) > 0); } - if (__rank == __N) + if (__rank == __nn) { for (int __i = 0; __i < __m; __i++) __begin_offsets[__i] = __begin_seqs[__i].second; // Very end. @@ -163,9 +163,9 @@ namespace __gnu_parallel } _GLIBCXX_PARALLEL_ASSERT(__m != 0); - _GLIBCXX_PARALLEL_ASSERT(__N != 0); + _GLIBCXX_PARALLEL_ASSERT(__nn != 0); _GLIBCXX_PARALLEL_ASSERT(__rank >= 0); - _GLIBCXX_PARALLEL_ASSERT(__rank < __N); + _GLIBCXX_PARALLEL_ASSERT(__rank < __nn); _DifferenceType* __ns = new _DifferenceType[__m]; _DifferenceType* __a = new _DifferenceType[__m]; @@ -401,14 +401,14 @@ namespace __gnu_parallel // Number of sequences, number of elements in total (possibly // including padding). _DifferenceType __m = std::distance(__begin_seqs, __end_seqs); - _DifferenceType __N = 0; + _DifferenceType __nn = 0; _DifferenceType __nmax, __n, __r; for (int __i = 0; __i < __m; __i++) - __N += std::distance(__begin_seqs[__i].first, - __begin_seqs[__i].second); + __nn += std::distance(__begin_seqs[__i].first, + __begin_seqs[__i].second); - if (__m == 0 || __N == 0 || __rank < 0 || __rank >= __N) + if (__m == 0 || __nn == 0 || __rank < 0 || __rank >= __nn) { // result undefined if there is no data or __rank is outside bounds throw std::exception(); @@ -433,7 +433,7 @@ namespace __gnu_parallel // Pad all lists to this length, at least as long as any ns[__i], // equality iff __nmax = 2^__k - 1 - __l = pow2(__r) - 1; + __l = __round_up_to_pow2(__r) - 1; for (int __i = 0; __i < __m; ++__i) { diff --git a/libstdc++-v3/include/parallel/multiway_merge.h b/libstdc++-v3/include/parallel/multiway_merge.h index a5bb2d74adc..4238a1c6923 100644 --- a/libstdc++-v3/include/parallel/multiway_merge.h +++ b/libstdc++-v3/include/parallel/multiway_merge.h @@ -54,2023 +54,1998 @@ namespace __gnu_parallel { -/** @brief _Iterator wrapper supporting an implicit supremum at the end - * of the sequence, dominating all comparisons. - * - * The implicit supremum comes with a performance cost. - * - * Deriving from _RAIter is not possible since - * _RAIter need not be a class. - */ -template - class _GuardedIterator - { - private: - /** @brief Current iterator __position. */ - _RAIter _M_current; - - /** @brief End iterator of the sequence. */ - _RAIter _M_end; - - /** @brief _Compare. */ - _Compare& __comp; - - public: - /** @brief Constructor. Sets iterator to beginning of sequence. - * @param __begin Begin iterator of sequence. - * @param __end End iterator of sequence. - * @param __comp Comparator provided for associated overloaded - * compare operators. */ - _GuardedIterator(_RAIter __begin, _RAIter __end, _Compare& __comp) - : _M_current(__begin), _M_end(__end), __comp(__comp) - { } - - /** @brief Pre-increment operator. - * @return This. */ - _GuardedIterator<_RAIter, _Compare>& - operator++() + /** @brief _Iterator wrapper supporting an implicit supremum at the end + * of the sequence, dominating all comparisons. + * + * The implicit supremum comes with a performance cost. + * + * Deriving from _RAIter is not possible since + * _RAIter need not be a class. + */ + template + class _GuardedIterator { - ++_M_current; - return *this; - } + private: + /** @brief Current iterator __position. */ + _RAIter _M_current; + + /** @brief End iterator of the sequence. */ + _RAIter _M_end; + + /** @brief _Compare. */ + _Compare& __comp; + + public: + /** @brief Constructor. Sets iterator to beginning of sequence. + * @param __begin Begin iterator of sequence. + * @param __end End iterator of sequence. + * @param __comp Comparator provided for associated overloaded + * compare operators. */ + _GuardedIterator(_RAIter __begin, _RAIter __end, _Compare& __comp) + : _M_current(__begin), _M_end(__end), __comp(__comp) + { } + + /** @brief Pre-increment operator. + * @return This. */ + _GuardedIterator<_RAIter, _Compare>& + operator++() + { + ++_M_current; + return *this; + } - /** @brief Dereference operator. - * @return Referenced element. */ - typename std::iterator_traits<_RAIter>::value_type& - operator*() - { return *_M_current; } - - /** @brief Convert to wrapped iterator. - * @return Wrapped iterator. */ - operator _RAIter() - { return _M_current; } - - /** @brief Compare two elements referenced by guarded iterators. - * @param __bi1 First iterator. - * @param __bi2 Second iterator. - * @return @__c true if less. */ - friend bool - operator<(_GuardedIterator<_RAIter, _Compare>& __bi1, - _GuardedIterator<_RAIter, _Compare>& __bi2) - { - if (__bi1._M_current == __bi1._M_end) //__bi1 is sup - return __bi2._M_current == __bi2._M_end; //__bi2 is not sup - if (__bi2._M_current == __bi2._M_end) //__bi2 is sup - return true; - return (__bi1.__comp)(*__bi1, *__bi2); //normal compare - } + /** @brief Dereference operator. + * @return Referenced element. */ + typename std::iterator_traits<_RAIter>::value_type& + operator*() + { return *_M_current; } + + /** @brief Convert to wrapped iterator. + * @return Wrapped iterator. */ + operator _RAIter() + { return _M_current; } + + /** @brief Compare two elements referenced by guarded iterators. + * @param __bi1 First iterator. + * @param __bi2 Second iterator. + * @return @__c true if less. */ + friend bool + operator<(_GuardedIterator<_RAIter, _Compare>& __bi1, + _GuardedIterator<_RAIter, _Compare>& __bi2) + { + if (__bi1._M_current == __bi1._M_end) // __bi1 is sup + return __bi2._M_current == __bi2._M_end; // __bi2 is not sup + if (__bi2._M_current == __bi2._M_end) // __bi2 is sup + return true; + return (__bi1.__comp)(*__bi1, *__bi2); // normal compare + } - /** @brief Compare two elements referenced by guarded iterators. - * @param __bi1 First iterator. - * @param __bi2 Second iterator. - * @return @__c True if less equal. */ - friend bool - operator<=(_GuardedIterator<_RAIter, _Compare>& __bi1, - _GuardedIterator<_RAIter, _Compare>& __bi2) - { - if (__bi2._M_current == __bi2._M_end) //__bi1 is sup - return __bi1._M_current != __bi1._M_end; //__bi2 is not sup - if (__bi1._M_current == __bi1._M_end) //__bi2 is sup - return false; - return !(__bi1.__comp)(*__bi2, *__bi1); //normal compare - } - }; - -template - class _UnguardedIterator - { - private: - /** @brief Current iterator __position. */ - _RAIter _M_current; - /** @brief _Compare. */ - mutable _Compare& __comp; - - public: - /** @brief Constructor. Sets iterator to beginning of sequence. - * @param __begin Begin iterator of sequence. - * @param __end Unused, only for compatibility. - * @param __comp Unused, only for compatibility. */ - _UnguardedIterator(_RAIter __begin, - _RAIter /* __end */, _Compare& __comp) - : _M_current(__begin), __comp(__comp) - { } - - /** @brief Pre-increment operator. - * @return This. */ - _UnguardedIterator<_RAIter, _Compare>& - operator++() + /** @brief Compare two elements referenced by guarded iterators. + * @param __bi1 First iterator. + * @param __bi2 Second iterator. + * @return @__c True if less equal. */ + friend bool + operator<=(_GuardedIterator<_RAIter, _Compare>& __bi1, + _GuardedIterator<_RAIter, _Compare>& __bi2) + { + if (__bi2._M_current == __bi2._M_end) // __bi1 is sup + return __bi1._M_current != __bi1._M_end; // __bi2 is not sup + if (__bi1._M_current == __bi1._M_end) // __bi2 is sup + return false; + return !(__bi1.__comp)(*__bi2, *__bi1); // normal compare + } + }; + + template + class _UnguardedIterator { - ++_M_current; - return *this; - } + private: + /** @brief Current iterator __position. */ + _RAIter _M_current; + /** @brief _Compare. */ + mutable _Compare& __comp; + + public: + /** @brief Constructor. Sets iterator to beginning of sequence. + * @param __begin Begin iterator of sequence. + * @param __end Unused, only for compatibility. + * @param __comp Unused, only for compatibility. */ + _UnguardedIterator(_RAIter __begin, + _RAIter /* __end */, _Compare& __comp) + : _M_current(__begin), __comp(__comp) + { } + + /** @brief Pre-increment operator. + * @return This. */ + _UnguardedIterator<_RAIter, _Compare>& + operator++() + { + ++_M_current; + return *this; + } - /** @brief Dereference operator. - * @return Referenced element. */ - typename std::iterator_traits<_RAIter>::value_type& - operator*() - { return *_M_current; } - - /** @brief Convert to wrapped iterator. - * @return Wrapped iterator. */ - operator _RAIter() - { return _M_current; } - - /** @brief Compare two elements referenced by unguarded iterators. - * @param __bi1 First iterator. - * @param __bi2 Second iterator. - * @return @__c true if less. */ - friend bool - operator<(_UnguardedIterator<_RAIter, _Compare>& __bi1, - _UnguardedIterator<_RAIter, _Compare>& __bi2) - { - // Normal compare. - return (__bi1.__comp)(*__bi1, *__bi2); - } + /** @brief Dereference operator. + * @return Referenced element. */ + typename std::iterator_traits<_RAIter>::value_type& + operator*() + { return *_M_current; } + + /** @brief Convert to wrapped iterator. + * @return Wrapped iterator. */ + operator _RAIter() + { return _M_current; } + + /** @brief Compare two elements referenced by unguarded iterators. + * @param __bi1 First iterator. + * @param __bi2 Second iterator. + * @return @__c true if less. */ + friend bool + operator<(_UnguardedIterator<_RAIter, _Compare>& __bi1, + _UnguardedIterator<_RAIter, _Compare>& __bi2) + { + // Normal compare. + return (__bi1.__comp)(*__bi1, *__bi2); + } - /** @brief Compare two elements referenced by unguarded iterators. - * @param __bi1 First iterator. - * @param __bi2 Second iterator. - * @return @__c True if less equal. */ - friend bool - operator<=(_UnguardedIterator<_RAIter, _Compare>& __bi1, - _UnguardedIterator<_RAIter, _Compare>& __bi2) + /** @brief Compare two elements referenced by unguarded iterators. + * @param __bi1 First iterator. + * @param __bi2 Second iterator. + * @return @__c True if less equal. */ + friend bool + operator<=(_UnguardedIterator<_RAIter, _Compare>& __bi1, + _UnguardedIterator<_RAIter, _Compare>& __bi2) + { + // Normal compare. + return !(__bi1.__comp)(*__bi2, *__bi1); + } + }; + + /** @brief Highly efficient 3-way merging procedure. + * + * Merging is done with the algorithm implementation described by Peter + * Sanders. Basically, the idea is to minimize the number of necessary + * comparison after merging an element. The implementation trick + * that makes this fast is that the order of the sequences is stored + * in the instruction pointer (translated into labels in C++). + * + * This works well for merging up to 4 sequences. + * + * Note that making the merging stable does not come at a + * performance hit. + * + * Whether the merging is done guarded or unguarded is selected by the + * used iterator class. + * + * @param __seqs_begin Begin iterator of iterator pair input sequence. + * @param __seqs_end End iterator of iterator pair input sequence. + * @param __target Begin iterator of output sequence. + * @param __comp Comparator. + * @param __length Maximum length to merge, less equal than the + * total number of elements available. + * + * @return End iterator of output sequence. + */ + template class iterator, + typename _RAIterIterator, + typename _RAIter3, + typename _DifferenceTp, + typename _Compare> + _RAIter3 + multiway_merge_3_variant(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + _DifferenceTp __length, _Compare __comp) { - // Normal compare. - return !(__bi1.__comp)(*__bi2, *__bi1); - } - }; - -/** @brief Highly efficient 3-way merging procedure. - * - * Merging is done with the algorithm implementation described by Peter - * Sanders. Basically, the idea is to minimize the number of necessary - * comparison after merging an element. The implementation trick - * that makes this fast is that the order of the sequences is stored - * in the instruction pointer (translated into labels in C++). - * - * This works well for merging up to 4 sequences. - * - * Note that making the merging stable does not come at a - * performance hit. - * - * Whether the merging is done guarded or unguarded is selected by the - * used iterator class. - * - * @param __seqs_begin Begin iterator of iterator pair input sequence. - * @param __seqs_end End iterator of iterator pair input sequence. - * @param __target Begin iterator of output sequence. - * @param __comp Comparator. - * @param __length Maximum length to merge, less equal than the - * total number of elements available. - * - * @return End iterator of output sequence. - */ -template class iterator, - typename _RAIterIterator, - typename _RAIter3, - typename _DifferenceTp, - typename _Compare> - _RAIter3 - multiway_merge_3_variant(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - _DifferenceTp __length, _Compare __comp) - { - _GLIBCXX_CALL(__length); - - typedef _DifferenceTp _DifferenceType; - - typedef typename std::iterator_traits<_RAIterIterator> - ::value_type::first_type - _RAIter1; - typedef typename std::iterator_traits<_RAIter1>::value_type - _ValueType; - - if (__length == 0) - return __target; + _GLIBCXX_CALL(__length); + + typedef _DifferenceTp _DifferenceType; + + typedef typename std::iterator_traits<_RAIterIterator> + ::value_type::first_type + _RAIter1; + typedef typename std::iterator_traits<_RAIter1>::value_type + _ValueType; + + if (__length == 0) + return __target; #if _GLIBCXX_ASSERTIONS - _DifferenceTp orig_length = __length; + _DifferenceTp __orig_length = __length; #endif - iterator<_RAIter1, _Compare> - __seq0(__seqs_begin[0].first, __seqs_begin[0].second, __comp), - __seq1(__seqs_begin[1].first, __seqs_begin[1].second, __comp), - __seq2(__seqs_begin[2].first, __seqs_begin[2].second, __comp); + iterator<_RAIter1, _Compare> + __seq0(__seqs_begin[0].first, __seqs_begin[0].second, __comp), + __seq1(__seqs_begin[1].first, __seqs_begin[1].second, __comp), + __seq2(__seqs_begin[2].first, __seqs_begin[2].second, __comp); - if (__seq0 <= __seq1) - { - if (__seq1 <= __seq2) - goto __s012; - else - if (__seq2 < __seq0) - goto __s201; + if (__seq0 <= __seq1) + { + if (__seq1 <= __seq2) + goto __s012; else - goto __s021; - } - else - { - if (__seq1 <= __seq2) - { - if (__seq0 <= __seq2) - goto __s102; + if (__seq2 < __seq0) + goto __s201; else - goto __s120; - } - else - goto __s210; - } -#define _GLIBCXX_PARALLEL_MERGE_3_CASE(__a,__b,__c,c0,c1) \ - __s ## __a ## __b ## __c : \ - *__target = *__seq ## __a; \ - ++__target; \ - --__length; \ - ++__seq ## __a; \ - if (__length == 0) goto finish; \ - if (__seq ## __a c0 __seq ## __b) goto __s ## __a ## __b ## __c; \ - if (__seq ## __a c1 __seq ## __c) goto __s ## __b ## __a ## __c; \ - goto __s ## __b ## __c ## __a; - - _GLIBCXX_PARALLEL_MERGE_3_CASE(0, 1, 2, <=, <=); - _GLIBCXX_PARALLEL_MERGE_3_CASE(1, 2, 0, <=, < ); - _GLIBCXX_PARALLEL_MERGE_3_CASE(2, 0, 1, < , < ); - _GLIBCXX_PARALLEL_MERGE_3_CASE(1, 0, 2, < , <=); - _GLIBCXX_PARALLEL_MERGE_3_CASE(0, 2, 1, <=, <=); - _GLIBCXX_PARALLEL_MERGE_3_CASE(2, 1, 0, < , < ); + goto __s021; + } + else + { + if (__seq1 <= __seq2) + { + if (__seq0 <= __seq2) + goto __s102; + else + goto __s120; + } + else + goto __s210; + } +#define _GLIBCXX_PARALLEL_MERGE_3_CASE(__a, __b, __c, __c0, __c1) \ + __s ## __a ## __b ## __c : \ + *__target = *__seq ## __a; \ + ++__target; \ + --__length; \ + ++__seq ## __a; \ + if (__length == 0) goto __finish; \ + if (__seq ## __a __c0 __seq ## __b) goto __s ## __a ## __b ## __c; \ + if (__seq ## __a __c1 __seq ## __c) goto __s ## __b ## __a ## __c; \ + goto __s ## __b ## __c ## __a; + + _GLIBCXX_PARALLEL_MERGE_3_CASE(0, 1, 2, <=, <=); + _GLIBCXX_PARALLEL_MERGE_3_CASE(1, 2, 0, <=, < ); + _GLIBCXX_PARALLEL_MERGE_3_CASE(2, 0, 1, < , < ); + _GLIBCXX_PARALLEL_MERGE_3_CASE(1, 0, 2, < , <=); + _GLIBCXX_PARALLEL_MERGE_3_CASE(0, 2, 1, <=, <=); + _GLIBCXX_PARALLEL_MERGE_3_CASE(2, 1, 0, < , < ); #undef _GLIBCXX_PARALLEL_MERGE_3_CASE - finish: - ; + __finish: + ; #if _GLIBCXX_ASSERTIONS - _GLIBCXX_PARALLEL_ASSERT( - ((_RAIter1)__seq0 - __seqs_begin[0].first) + - ((_RAIter1)__seq1 - __seqs_begin[1].first) + - ((_RAIter1)__seq2 - __seqs_begin[2].first) - == orig_length); + _GLIBCXX_PARALLEL_ASSERT( + ((_RAIter1)__seq0 - __seqs_begin[0].first) + + ((_RAIter1)__seq1 - __seqs_begin[1].first) + + ((_RAIter1)__seq2 - __seqs_begin[2].first) + == __orig_length); #endif - __seqs_begin[0].first = __seq0; - __seqs_begin[1].first = __seq1; - __seqs_begin[2].first = __seq2; - - return __target; - } - -/** - * @brief Highly efficient 4-way merging procedure. - * - * Merging is done with the algorithm implementation described by Peter - * Sanders. Basically, the idea is to minimize the number of necessary - * comparison after merging an element. The implementation trick - * that makes this fast is that the order of the sequences is stored - * in the instruction pointer (translated into goto labels in C++). - * - * This works well for merging up to 4 sequences. - * - * Note that making the merging stable does not come at a - * performance hit. - * - * Whether the merging is done guarded or unguarded is selected by the - * used iterator class. - * - * @param __seqs_begin Begin iterator of iterator pair input sequence. - * @param __seqs_end End iterator of iterator pair input sequence. - * @param __target Begin iterator of output sequence. - * @param __comp Comparator. - * @param __length Maximum length to merge, less equal than the - * total number of elements available. - * - * @return End iterator of output sequence. - */ -template class iterator, - typename _RAIterIterator, - typename _RAIter3, - typename _DifferenceTp, - typename _Compare> - _RAIter3 - multiway_merge_4_variant(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - _DifferenceTp __length, _Compare __comp) - { - _GLIBCXX_CALL(__length); - typedef _DifferenceTp _DifferenceType; - - typedef typename std::iterator_traits<_RAIterIterator> - ::value_type::first_type - _RAIter1; - typedef typename std::iterator_traits<_RAIter1>::value_type - _ValueType; - - iterator<_RAIter1, _Compare> - __seq0(__seqs_begin[0].first, __seqs_begin[0].second, __comp), - __seq1(__seqs_begin[1].first, __seqs_begin[1].second, __comp), - __seq2(__seqs_begin[2].first, __seqs_begin[2].second, __comp), - __seq3(__seqs_begin[3].first, __seqs_begin[3].second, __comp); - -#define _GLIBCXX_PARALLEL_DECISION(__a,__b,__c,d) { \ - if (__seq ## d < __seq ## __a) goto __s ## d ## __a ## __b ## __c; \ - if (__seq ## d < __seq ## __b) goto __s ## __a ## d ## __b ## __c; \ - if (__seq ## d < __seq ## __c) goto __s ## __a ## __b ## d ## __c; \ - goto __s ## __a ## __b ## __c ## d; } - - if (__seq0 <= __seq1) - { - if (__seq1 <= __seq2) - _GLIBCXX_PARALLEL_DECISION(0,1,2,3) - else - if (__seq2 < __seq0) - _GLIBCXX_PARALLEL_DECISION(2,0,1,3) - else - _GLIBCXX_PARALLEL_DECISION(0,2,1,3) - } - else - { - if (__seq1 <= __seq2) - { - if (__seq0 <= __seq2) - _GLIBCXX_PARALLEL_DECISION(1,0,2,3) - else - _GLIBCXX_PARALLEL_DECISION(1,2,0,3) - } - else - _GLIBCXX_PARALLEL_DECISION(2,1,0,3) - } + __seqs_begin[0].first = __seq0; + __seqs_begin[1].first = __seq1; + __seqs_begin[2].first = __seq2; + + return __target; + } + + /** + * @brief Highly efficient 4-way merging procedure. + * + * Merging is done with the algorithm implementation described by Peter + * Sanders. Basically, the idea is to minimize the number of necessary + * comparison after merging an element. The implementation trick + * that makes this fast is that the order of the sequences is stored + * in the instruction pointer (translated into goto labels in C++). + * + * This works well for merging up to 4 sequences. + * + * Note that making the merging stable does not come at a + * performance hit. + * + * Whether the merging is done guarded or unguarded is selected by the + * used iterator class. + * + * @param __seqs_begin Begin iterator of iterator pair input sequence. + * @param __seqs_end End iterator of iterator pair input sequence. + * @param __target Begin iterator of output sequence. + * @param __comp Comparator. + * @param __length Maximum length to merge, less equal than the + * total number of elements available. + * + * @return End iterator of output sequence. + */ + template class iterator, + typename _RAIterIterator, + typename _RAIter3, + typename _DifferenceTp, + typename _Compare> + _RAIter3 + multiway_merge_4_variant(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + _DifferenceTp __length, _Compare __comp) + { + _GLIBCXX_CALL(__length); + typedef _DifferenceTp _DifferenceType; + + typedef typename std::iterator_traits<_RAIterIterator> + ::value_type::first_type + _RAIter1; + typedef typename std::iterator_traits<_RAIter1>::value_type + _ValueType; -#define _GLIBCXX_PARALLEL_MERGE_4_CASE(__a,__b,__c,d,c0,c1,c2) \ - __s ## __a ## __b ## __c ## d: \ - if (__length == 0) goto finish; \ - *__target = *__seq ## __a; \ - ++__target; \ - --__length; \ - ++__seq ## __a; \ - if (__seq ## __a c0 __seq ## __b) goto __s ## __a ## __b ## __c ## d; \ - if (__seq ## __a c1 __seq ## __c) goto __s ## __b ## __a ## __c ## d; \ - if (__seq ## __a c2 __seq ## d) goto __s ## __b ## __c ## __a ## d; \ - goto __s ## __b ## __c ## d ## __a; - - _GLIBCXX_PARALLEL_MERGE_4_CASE(0, 1, 2, 3, <=, <=, <=); - _GLIBCXX_PARALLEL_MERGE_4_CASE(0, 1, 3, 2, <=, <=, <=); - _GLIBCXX_PARALLEL_MERGE_4_CASE(0, 2, 1, 3, <=, <=, <=); - _GLIBCXX_PARALLEL_MERGE_4_CASE(0, 2, 3, 1, <=, <=, <=); - _GLIBCXX_PARALLEL_MERGE_4_CASE(0, 3, 1, 2, <=, <=, <=); - _GLIBCXX_PARALLEL_MERGE_4_CASE(0, 3, 2, 1, <=, <=, <=); - _GLIBCXX_PARALLEL_MERGE_4_CASE(1, 0, 2, 3, < , <=, <=); - _GLIBCXX_PARALLEL_MERGE_4_CASE(1, 0, 3, 2, < , <=, <=); - _GLIBCXX_PARALLEL_MERGE_4_CASE(1, 2, 0, 3, <=, < , <=); - _GLIBCXX_PARALLEL_MERGE_4_CASE(1, 2, 3, 0, <=, <=, < ); - _GLIBCXX_PARALLEL_MERGE_4_CASE(1, 3, 0, 2, <=, < , <=); - _GLIBCXX_PARALLEL_MERGE_4_CASE(1, 3, 2, 0, <=, <=, < ); - _GLIBCXX_PARALLEL_MERGE_4_CASE(2, 0, 1, 3, < , < , <=); - _GLIBCXX_PARALLEL_MERGE_4_CASE(2, 0, 3, 1, < , <=, < ); - _GLIBCXX_PARALLEL_MERGE_4_CASE(2, 1, 0, 3, < , < , <=); - _GLIBCXX_PARALLEL_MERGE_4_CASE(2, 1, 3, 0, < , <=, < ); - _GLIBCXX_PARALLEL_MERGE_4_CASE(2, 3, 0, 1, <=, < , < ); - _GLIBCXX_PARALLEL_MERGE_4_CASE(2, 3, 1, 0, <=, < , < ); - _GLIBCXX_PARALLEL_MERGE_4_CASE(3, 0, 1, 2, < , < , < ); - _GLIBCXX_PARALLEL_MERGE_4_CASE(3, 0, 2, 1, < , < , < ); - _GLIBCXX_PARALLEL_MERGE_4_CASE(3, 1, 0, 2, < , < , < ); - _GLIBCXX_PARALLEL_MERGE_4_CASE(3, 1, 2, 0, < , < , < ); - _GLIBCXX_PARALLEL_MERGE_4_CASE(3, 2, 0, 1, < , < , < ); - _GLIBCXX_PARALLEL_MERGE_4_CASE(3, 2, 1, 0, < , < , < ); + iterator<_RAIter1, _Compare> + __seq0(__seqs_begin[0].first, __seqs_begin[0].second, __comp), + __seq1(__seqs_begin[1].first, __seqs_begin[1].second, __comp), + __seq2(__seqs_begin[2].first, __seqs_begin[2].second, __comp), + __seq3(__seqs_begin[3].first, __seqs_begin[3].second, __comp); + +#define _GLIBCXX_PARALLEL_DECISION(__a, __b, __c, __d) { \ + if (__seq ## __d < __seq ## __a) \ + goto __s ## __d ## __a ## __b ## __c; \ + if (__seq ## __d < __seq ## __b) \ + goto __s ## __a ## __d ## __b ## __c; \ + if (__seq ## __d < __seq ## __c) \ + goto __s ## __a ## __b ## __d ## __c; \ + goto __s ## __a ## __b ## __c ## __d; } + + if (__seq0 <= __seq1) + { + if (__seq1 <= __seq2) + _GLIBCXX_PARALLEL_DECISION(0,1,2,3) + else + if (__seq2 < __seq0) + _GLIBCXX_PARALLEL_DECISION(2,0,1,3) + else + _GLIBCXX_PARALLEL_DECISION(0,2,1,3) + } + else + { + if (__seq1 <= __seq2) + { + if (__seq0 <= __seq2) + _GLIBCXX_PARALLEL_DECISION(1,0,2,3) + else + _GLIBCXX_PARALLEL_DECISION(1,2,0,3) + } + else + _GLIBCXX_PARALLEL_DECISION(2,1,0,3) + } + +#define _GLIBCXX_PARALLEL_MERGE_4_CASE(__a, __b, __c, __d, \ + __c0, __c1, __c2) \ + __s ## __a ## __b ## __c ## __d: \ + if (__length == 0) goto __finish; \ + *__target = *__seq ## __a; \ + ++__target; \ + --__length; \ + ++__seq ## __a; \ + if (__seq ## __a __c0 __seq ## __b) \ + goto __s ## __a ## __b ## __c ## __d; \ + if (__seq ## __a __c1 __seq ## __c) \ + goto __s ## __b ## __a ## __c ## __d; \ + if (__seq ## __a __c2 __seq ## __d) \ + goto __s ## __b ## __c ## __a ## __d; \ + goto __s ## __b ## __c ## __d ## __a; + + _GLIBCXX_PARALLEL_MERGE_4_CASE(0, 1, 2, 3, <=, <=, <=); + _GLIBCXX_PARALLEL_MERGE_4_CASE(0, 1, 3, 2, <=, <=, <=); + _GLIBCXX_PARALLEL_MERGE_4_CASE(0, 2, 1, 3, <=, <=, <=); + _GLIBCXX_PARALLEL_MERGE_4_CASE(0, 2, 3, 1, <=, <=, <=); + _GLIBCXX_PARALLEL_MERGE_4_CASE(0, 3, 1, 2, <=, <=, <=); + _GLIBCXX_PARALLEL_MERGE_4_CASE(0, 3, 2, 1, <=, <=, <=); + _GLIBCXX_PARALLEL_MERGE_4_CASE(1, 0, 2, 3, < , <=, <=); + _GLIBCXX_PARALLEL_MERGE_4_CASE(1, 0, 3, 2, < , <=, <=); + _GLIBCXX_PARALLEL_MERGE_4_CASE(1, 2, 0, 3, <=, < , <=); + _GLIBCXX_PARALLEL_MERGE_4_CASE(1, 2, 3, 0, <=, <=, < ); + _GLIBCXX_PARALLEL_MERGE_4_CASE(1, 3, 0, 2, <=, < , <=); + _GLIBCXX_PARALLEL_MERGE_4_CASE(1, 3, 2, 0, <=, <=, < ); + _GLIBCXX_PARALLEL_MERGE_4_CASE(2, 0, 1, 3, < , < , <=); + _GLIBCXX_PARALLEL_MERGE_4_CASE(2, 0, 3, 1, < , <=, < ); + _GLIBCXX_PARALLEL_MERGE_4_CASE(2, 1, 0, 3, < , < , <=); + _GLIBCXX_PARALLEL_MERGE_4_CASE(2, 1, 3, 0, < , <=, < ); + _GLIBCXX_PARALLEL_MERGE_4_CASE(2, 3, 0, 1, <=, < , < ); + _GLIBCXX_PARALLEL_MERGE_4_CASE(2, 3, 1, 0, <=, < , < ); + _GLIBCXX_PARALLEL_MERGE_4_CASE(3, 0, 1, 2, < , < , < ); + _GLIBCXX_PARALLEL_MERGE_4_CASE(3, 0, 2, 1, < , < , < ); + _GLIBCXX_PARALLEL_MERGE_4_CASE(3, 1, 0, 2, < , < , < ); + _GLIBCXX_PARALLEL_MERGE_4_CASE(3, 1, 2, 0, < , < , < ); + _GLIBCXX_PARALLEL_MERGE_4_CASE(3, 2, 0, 1, < , < , < ); + _GLIBCXX_PARALLEL_MERGE_4_CASE(3, 2, 1, 0, < , < , < ); #undef _GLIBCXX_PARALLEL_MERGE_4_CASE #undef _GLIBCXX_PARALLEL_DECISION - finish: - ; - - __seqs_begin[0].first = __seq0; - __seqs_begin[1].first = __seq1; - __seqs_begin[2].first = __seq2; - __seqs_begin[3].first = __seq3; - - return __target; - } - -/** @brief Multi-way merging procedure for a high branching factor, - * guarded case. - * - * This merging variant uses a LoserTree class as selected by _LT. - * - * Stability is selected through the used LoserTree class _LT. - * - * At least one non-empty sequence is required. - * - * @param __seqs_begin Begin iterator of iterator pair input sequence. - * @param __seqs_end End iterator of iterator pair input sequence. - * @param __target Begin iterator of output sequence. - * @param __comp Comparator. - * @param __length Maximum length to merge, less equal than the - * total number of elements available. - * - * @return End iterator of output sequence. - */ -template - _RAIter3 - multiway_merge_loser_tree(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - _DifferenceTp __length, _Compare __comp) - { - _GLIBCXX_CALL(__length) + __finish: + ; - typedef _DifferenceTp _DifferenceType; - typedef typename std::iterator_traits<_RAIterIterator> - ::value_type::first_type - _RAIter1; - typedef typename std::iterator_traits<_RAIter1>::value_type - _ValueType; + __seqs_begin[0].first = __seq0; + __seqs_begin[1].first = __seq1; + __seqs_begin[2].first = __seq2; + __seqs_begin[3].first = __seq3; - int __k = static_cast(__seqs_end - __seqs_begin); + return __target; + } - _LT __lt(__k, __comp); + /** @brief Multi-way merging procedure for a high branching factor, + * guarded case. + * + * This merging variant uses a LoserTree class as selected by _LT. + * + * Stability is selected through the used LoserTree class _LT. + * + * At least one non-empty sequence is required. + * + * @param __seqs_begin Begin iterator of iterator pair input sequence. + * @param __seqs_end End iterator of iterator pair input sequence. + * @param __target Begin iterator of output sequence. + * @param __comp Comparator. + * @param __length Maximum length to merge, less equal than the + * total number of elements available. + * + * @return End iterator of output sequence. + */ + template + _RAIter3 + multiway_merge_loser_tree(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + _DifferenceTp __length, _Compare __comp) + { + _GLIBCXX_CALL(__length) - // Default value for potentially non-default-constructible types. - _ValueType* __arbitrary_element = NULL; + typedef _DifferenceTp _DifferenceType; + typedef typename std::iterator_traits<_RAIterIterator> + ::value_type::first_type + _RAIter1; + typedef typename std::iterator_traits<_RAIter1>::value_type + _ValueType; - for (int __t = 0; __t < __k; ++__t) - { - if(__arbitrary_element == NULL - && _GLIBCXX_PARALLEL_LENGTH(__seqs_begin[__t]) > 0) - __arbitrary_element = &(*__seqs_begin[__t].first); - } + int __k = static_cast(__seqs_end - __seqs_begin); - for (int __t = 0; __t < __k; ++__t) - { - if (__seqs_begin[__t].first == __seqs_begin[__t].second) - __lt.__insert_start(*__arbitrary_element, __t, true); - else - __lt.__insert_start(*__seqs_begin[__t].first, __t, false); - } + _LT __lt(__k, __comp); - __lt.__init(); + // Default value for potentially non-default-constructible types. + _ValueType* __arbitrary_element = NULL; - int __source; + for (int __t = 0; __t < __k; ++__t) + { + if(__arbitrary_element == NULL + && _GLIBCXX_PARALLEL_LENGTH(__seqs_begin[__t]) > 0) + __arbitrary_element = &(*__seqs_begin[__t].first); + } - for (_DifferenceType __i = 0; __i < __length; ++__i) - { - //take out - __source = __lt.__get_min_source(); + for (int __t = 0; __t < __k; ++__t) + { + if (__seqs_begin[__t].first == __seqs_begin[__t].second) + __lt.__insert_start(*__arbitrary_element, __t, true); + else + __lt.__insert_start(*__seqs_begin[__t].first, __t, false); + } - *(__target++) = *(__seqs_begin[__source].first++); + __lt.__init(); - // Feed. - if (__seqs_begin[__source].first == __seqs_begin[__source].second) - __lt.__delete_min_insert(*__arbitrary_element, true); - else - // Replace from same __source. - __lt.__delete_min_insert(*__seqs_begin[__source].first, false); - } + int __source; - return __target; - } - -/** @brief Multi-way merging procedure for a high branching factor, - * unguarded case. - * - * Merging is done using the LoserTree class _LT. - * - * Stability is selected by the used LoserTrees. - * - * @pre No input will run out of elements during the merge. - * - * @param __seqs_begin Begin iterator of iterator pair input sequence. - * @param __seqs_end End iterator of iterator pair input sequence. - * @param __target Begin iterator of output sequence. - * @param __comp Comparator. - * @param __length Maximum length to merge, less equal than the - * total number of elements available. - * - * @return End iterator of output sequence. - */ -template - _RAIter3 - multiway_merge_loser_tree_unguarded(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - const typename std::iterator_traits::value_type::first_type>::value_type& - __sentinel, - _DifferenceTp __length, - _Compare __comp) - { - _GLIBCXX_CALL(__length) - typedef _DifferenceTp _DifferenceType; + for (_DifferenceType __i = 0; __i < __length; ++__i) + { + //take out + __source = __lt.__get_min_source(); - typedef typename std::iterator_traits<_RAIterIterator> - ::value_type::first_type - _RAIter1; - typedef typename std::iterator_traits<_RAIter1>::value_type - _ValueType; + *(__target++) = *(__seqs_begin[__source].first++); - int __k = __seqs_end - __seqs_begin; + // Feed. + if (__seqs_begin[__source].first == __seqs_begin[__source].second) + __lt.__delete_min_insert(*__arbitrary_element, true); + else + // Replace from same __source. + __lt.__delete_min_insert(*__seqs_begin[__source].first, false); + } - _LT __lt(__k, __sentinel, __comp); + return __target; + } - for (int __t = 0; __t < __k; ++__t) - { -#if _GLIBCXX_ASSERTIONS - _GLIBCXX_PARALLEL_ASSERT(__seqs_begin[__t].first - != __seqs_begin[__t].second); -#endif - __lt.__insert_start(*__seqs_begin[__t].first, __t, false); - } + /** @brief Multi-way merging procedure for a high branching factor, + * unguarded case. + * + * Merging is done using the LoserTree class _LT. + * + * Stability is selected by the used LoserTrees. + * + * @pre No input will run out of elements during the merge. + * + * @param __seqs_begin Begin iterator of iterator pair input sequence. + * @param __seqs_end End iterator of iterator pair input sequence. + * @param __target Begin iterator of output sequence. + * @param __comp Comparator. + * @param __length Maximum length to merge, less equal than the + * total number of elements available. + * + * @return End iterator of output sequence. + */ + template + _RAIter3 + multiway_merge_loser_tree_unguarded(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + const typename std::iterator_traits::value_type::first_type>::value_type& + __sentinel, + _DifferenceTp __length, + _Compare __comp) + { + _GLIBCXX_CALL(__length) + typedef _DifferenceTp _DifferenceType; - __lt.__init(); + typedef typename std::iterator_traits<_RAIterIterator> + ::value_type::first_type + _RAIter1; + typedef typename std::iterator_traits<_RAIter1>::value_type + _ValueType; + + int __k = __seqs_end - __seqs_begin; - int __source; + _LT __lt(__k, __sentinel, __comp); + for (int __t = 0; __t < __k; ++__t) + { #if _GLIBCXX_ASSERTIONS - _DifferenceType __i = 0; + _GLIBCXX_PARALLEL_ASSERT(__seqs_begin[__t].first + != __seqs_begin[__t].second); #endif + __lt.__insert_start(*__seqs_begin[__t].first, __t, false); + } - _RAIter3 __target_end = __target + __length; - while (__target < __target_end) - { - // Take out. - __source = __lt.__get_min_source(); + __lt.__init(); + + int __source; #if _GLIBCXX_ASSERTIONS - _GLIBCXX_PARALLEL_ASSERT(0 <= __source && __source < __k); - _GLIBCXX_PARALLEL_ASSERT(__i == 0 - || !__comp(*(__seqs_begin[__source].first), *(__target - 1))); + _DifferenceType __i = 0; #endif - // Feed. - *(__target++) = *(__seqs_begin[__source].first++); + _RAIter3 __target_end = __target + __length; + while (__target < __target_end) + { + // Take out. + __source = __lt.__get_min_source(); #if _GLIBCXX_ASSERTIONS - ++__i; + _GLIBCXX_PARALLEL_ASSERT(0 <= __source && __source < __k); + _GLIBCXX_PARALLEL_ASSERT(__i == 0 + || !__comp(*(__seqs_begin[__source].first), *(__target - 1))); #endif - // Replace from same __source. - __lt.__delete_min_insert(*__seqs_begin[__source].first, false); - } - return __target; - } - - -/** @brief Multi-way merging procedure for a high branching factor, - * requiring sentinels to exist. - * - * @param __stable The value must the same as for the used LoserTrees. - * @param UnguardedLoserTree _Loser Tree variant to use for the unguarded - * merging. - * @param GuardedLoserTree _Loser Tree variant to use for the guarded - * merging. - * - * @param __seqs_begin Begin iterator of iterator pair input sequence. - * @param __seqs_end End iterator of iterator pair input sequence. - * @param __target Begin iterator of output sequence. - * @param __comp Comparator. - * @param __length Maximum length to merge, less equal than the - * total number of elements available. - * - * @return End iterator of output sequence. - */ -template - _RAIter3 - multiway_merge_loser_tree_sentinel(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - const typename std::iterator_traits::value_type::first_type>::value_type& - __sentinel, - _DifferenceTp __length, - _Compare __comp) - { - _GLIBCXX_CALL(__length) - - typedef _DifferenceTp _DifferenceType; - typedef std::iterator_traits<_RAIterIterator> _TraitsType; - typedef typename std::iterator_traits<_RAIterIterator> - ::value_type::first_type - _RAIter1; - typedef typename std::iterator_traits<_RAIter1>::value_type - _ValueType; - - _RAIter3 __target_end; - - for (_RAIterIterator __s = __seqs_begin; __s != __seqs_end; ++__s) - // Move the sequence ends to the sentinel. This has the - // effect that the sentinel appears to be within the sequence. Then, - // we can use the unguarded variant if we merge out as many - // non-sentinel elements as we have. - ++((*__s).second); - - __target_end = multiway_merge_loser_tree_unguarded - - (__seqs_begin, __seqs_end, __target, __sentinel, __length, __comp); + // Feed. + *(__target++) = *(__seqs_begin[__source].first++); #if _GLIBCXX_ASSERTIONS - _GLIBCXX_PARALLEL_ASSERT(__target_end == __target + __length); - _GLIBCXX_PARALLEL_ASSERT(__is_sorted(__target, __target_end, __comp)); + ++__i; #endif + // Replace from same __source. + __lt.__delete_min_insert(*__seqs_begin[__source].first, false); + } - // Restore the sequence ends so the sentinels are not contained in the - // sequence any more (see comment in loop above). - for (_RAIterIterator __s = __seqs_begin; __s != __seqs_end; ++__s) - --((*__s).second); - - return __target_end; - } - -/** - * @brief Traits for determining whether the loser tree should - * use pointers or copies. - * - * The field "_M_use_pointer" is used to determine whether to use pointers in - * the loser trees or whether to copy the values into the loser tree. - * - * The default behavior is to use pointers if the data type is 4 times as - * big as the pointer to it. - * - * Specialize for your data type to customize the behavior. - * - * Example: - * - * template<> - * struct _LoserTreeTraits - * { static const bool _M_use_pointer = false; }; - * - * template<> - * struct _LoserTreeTraits - * { static const bool _M_use_pointer = true; }; - * - * @param _Tp type to give the loser tree traits for. - */ -template - struct _LoserTreeTraits - { - /** - * @brief True iff to use pointers instead of values in loser trees. - * - * The default behavior is to use pointers if the data type is four - * times as big as the pointer to it. - */ - static const bool _M_use_pointer = (sizeof(_Tp) > 4 * sizeof(_Tp*)); - }; - -/** - * @brief Switch for 3-way merging with __sentinels turned off. - * - * Note that 3-way merging is always stable! - */ -template - struct __multiway_merge_3_variant_sentinel_switch - { - _RAIter3 - operator()(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - _DifferenceTp __length, _Compare __comp) - { - return multiway_merge_3_variant<_GuardedIterator> - (__seqs_begin, __seqs_end, __target, __length, __comp); - } - }; - -/** - * @brief Switch for 3-way merging with __sentinels turned on. - * - * Note that 3-way merging is always stable! - */ -template - struct __multiway_merge_3_variant_sentinel_switch - { - _RAIter3 - operator()(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - _DifferenceTp __length, _Compare __comp) - { - return multiway_merge_3_variant<_UnguardedIterator> - (__seqs_begin, __seqs_end, __target, __length, __comp); - } - }; - -/** - * @brief Switch for 4-way merging with __sentinels turned off. - * - * Note that 4-way merging is always stable! - */ -template - struct __multiway_merge_4_variant_sentinel_switch - { - _RAIter3 - operator()(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - _DifferenceTp __length, _Compare __comp) - { - return multiway_merge_4_variant<_GuardedIterator> - (__seqs_begin, __seqs_end, __target, __length, __comp); - } - }; - -/** - * @brief Switch for 4-way merging with __sentinels turned on. - * - * Note that 4-way merging is always stable! - */ -template - struct __multiway_merge_4_variant_sentinel_switch - { - _RAIter3 - operator()(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - _DifferenceTp __length, _Compare __comp) - { - return multiway_merge_4_variant<_UnguardedIterator> - (__seqs_begin, __seqs_end, __target, __length, __comp); + return __target; } - }; - -/** - * @brief Switch for k-way merging with __sentinels turned on. - */ -template - struct __multiway_merge_k_variant_sentinel_switch - { + + + /** @brief Multi-way merging procedure for a high branching factor, + * requiring sentinels to exist. + * + * @param __stable The value must the same as for the used LoserTrees. + * @param UnguardedLoserTree _Loser Tree variant to use for the unguarded + * merging. + * @param GuardedLoserTree _Loser Tree variant to use for the guarded + * merging. + * + * @param __seqs_begin Begin iterator of iterator pair input sequence. + * @param __seqs_end End iterator of iterator pair input sequence. + * @param __target Begin iterator of output sequence. + * @param __comp Comparator. + * @param __length Maximum length to merge, less equal than the + * total number of elements available. + * + * @return End iterator of output sequence. + */ + template _RAIter3 - operator()(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - const typename std::iterator_traits::value_type::first_type>::value_type& - __sentinel, - _DifferenceTp __length, _Compare __comp) + multiway_merge_loser_tree_sentinel(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + const typename std::iterator_traits::value_type::first_type>::value_type& + __sentinel, + _DifferenceTp __length, + _Compare __comp) { + _GLIBCXX_CALL(__length) + + typedef _DifferenceTp _DifferenceType; + typedef std::iterator_traits<_RAIterIterator> _TraitsType; typedef typename std::iterator_traits<_RAIterIterator> ::value_type::first_type _RAIter1; typedef typename std::iterator_traits<_RAIter1>::value_type _ValueType; - return multiway_merge_loser_tree_sentinel< - typename __gnu_cxx::__conditional_type< - _LoserTreeTraits<_ValueType>::_M_use_pointer, - _LoserTreePointerUnguarded<__stable, _ValueType, _Compare>, - _LoserTreeUnguarded<__stable, _ValueType, _Compare> - >::__type> + _RAIter3 __target_end; + + for (_RAIterIterator __s = __seqs_begin; __s != __seqs_end; ++__s) + // Move the sequence ends to the sentinel. This has the + // effect that the sentinel appears to be within the sequence. Then, + // we can use the unguarded variant if we merge out as many + // non-sentinel elements as we have. + ++((*__s).second); + + __target_end = multiway_merge_loser_tree_unguarded (__seqs_begin, __seqs_end, __target, __sentinel, __length, __comp); + +#if _GLIBCXX_ASSERTIONS + _GLIBCXX_PARALLEL_ASSERT(__target_end == __target + __length); + _GLIBCXX_PARALLEL_ASSERT(__is_sorted(__target, __target_end, __comp)); +#endif + + // Restore the sequence ends so the sentinels are not contained in the + // sequence any more (see comment in loop above). + for (_RAIterIterator __s = __seqs_begin; __s != __seqs_end; ++__s) + --((*__s).second); + + return __target_end; } - }; - -/** - * @brief Switch for k-way merging with __sentinels turned off. - */ -template - struct __multiway_merge_k_variant_sentinel_switch - { + + /** + * @brief Traits for determining whether the loser tree should + * use pointers or copies. + * + * The field "_M_use_pointer" is used to determine whether to use pointers + * in he loser trees or whether to copy the values into the loser tree. + * + * The default behavior is to use pointers if the data type is 4 times as + * big as the pointer to it. + * + * Specialize for your data type to customize the behavior. + * + * Example: + * + * template<> + * struct _LoserTreeTraits + * { static const bool _M_use_pointer = false; }; + * + * template<> + * struct _LoserTreeTraits + * { static const bool _M_use_pointer = true; }; + * + * @param _Tp type to give the loser tree traits for. + */ + template + struct _LoserTreeTraits + { + /** + * @brief True iff to use pointers instead of values in loser trees. + * + * The default behavior is to use pointers if the data type is four + * times as big as the pointer to it. + */ + static const bool _M_use_pointer = (sizeof(_Tp) > 4 * sizeof(_Tp*)); + }; + + /** + * @brief Switch for 3-way merging with __sentinels turned off. + * + * Note that 3-way merging is always stable! + */ + template + struct __multiway_merge_3_variant_sentinel_switch + { + _RAIter3 + operator()(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + _DifferenceTp __length, _Compare __comp) + { return multiway_merge_3_variant<_GuardedIterator> + (__seqs_begin, __seqs_end, __target, __length, __comp); } + }; + + /** + * @brief Switch for 3-way merging with __sentinels turned on. + * + * Note that 3-way merging is always stable! + */ + template + struct __multiway_merge_3_variant_sentinel_switch + { + _RAIter3 + operator()(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + _DifferenceTp __length, _Compare __comp) + { return multiway_merge_3_variant<_UnguardedIterator> + (__seqs_begin, __seqs_end, __target, __length, __comp); } + }; + + /** + * @brief Switch for 4-way merging with __sentinels turned off. + * + * Note that 4-way merging is always stable! + */ + template + struct __multiway_merge_4_variant_sentinel_switch + { + _RAIter3 + operator()(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + _DifferenceTp __length, _Compare __comp) + { return multiway_merge_4_variant<_GuardedIterator> + (__seqs_begin, __seqs_end, __target, __length, __comp); } + }; + + /** + * @brief Switch for 4-way merging with __sentinels turned on. + * + * Note that 4-way merging is always stable! + */ + template + struct __multiway_merge_4_variant_sentinel_switch + { + _RAIter3 + operator()(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + _DifferenceTp __length, _Compare __comp) + { return multiway_merge_4_variant<_UnguardedIterator> + (__seqs_begin, __seqs_end, __target, __length, __comp); } + }; + + /** + * @brief Switch for k-way merging with __sentinels turned on. + */ + template + struct __multiway_merge_k_variant_sentinel_switch + { + _RAIter3 + operator()(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + const typename std::iterator_traits::value_type::first_type>::value_type& + __sentinel, + _DifferenceTp __length, _Compare __comp) + { + typedef typename std::iterator_traits<_RAIterIterator> + ::value_type::first_type + _RAIter1; + typedef typename std::iterator_traits<_RAIter1>::value_type + _ValueType; + + return multiway_merge_loser_tree_sentinel< + typename __gnu_cxx::__conditional_type< + _LoserTreeTraits<_ValueType>::_M_use_pointer, + _LoserTreePointerUnguarded<__stable, _ValueType, _Compare>, + _LoserTreeUnguarded<__stable, _ValueType, _Compare> + >::__type> + (__seqs_begin, __seqs_end, __target, __sentinel, __length, __comp); + } + }; + + /** + * @brief Switch for k-way merging with __sentinels turned off. + */ + template + struct __multiway_merge_k_variant_sentinel_switch + { + _RAIter3 + operator()(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + const typename std::iterator_traits::value_type::first_type>::value_type& + __sentinel, + _DifferenceTp __length, _Compare __comp) + { + typedef typename std::iterator_traits<_RAIterIterator> + ::value_type::first_type + _RAIter1; + typedef typename std::iterator_traits<_RAIter1>::value_type + _ValueType; + + return multiway_merge_loser_tree< + typename __gnu_cxx::__conditional_type< + _LoserTreeTraits<_ValueType>::_M_use_pointer, + _LoserTreePointer<__stable, _ValueType, _Compare>, + _LoserTree<__stable, _ValueType, _Compare> + >::__type >(__seqs_begin, __seqs_end, __target, __length, __comp); + } + }; + + /** @brief Sequential multi-way merging switch. + * + * The _GLIBCXX_PARALLEL_DECISION is based on the branching factor and + * runtime settings. + * @param __seqs_begin Begin iterator of iterator pair input sequence. + * @param __seqs_end End iterator of iterator pair input sequence. + * @param __target Begin iterator of output sequence. + * @param __comp Comparator. + * @param __length Maximum length to merge, possibly larger than the + * number of elements available. + * @param __stable Stable merging incurs a performance penalty. + * @param __sentinel The sequences have __a __sentinel element. + * @return End iterator of output sequence. */ + template _RAIter3 - operator()(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - const typename std::iterator_traits::value_type::first_type>::value_type& - __sentinel, - _DifferenceTp __length, _Compare __comp) + __sequential_multiway_merge(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + const typename std::iterator_traits::value_type::first_type>::value_type& + __sentinel, + _DifferenceTp __length, _Compare __comp) { + _GLIBCXX_CALL(__length) + + typedef _DifferenceTp _DifferenceType; typedef typename std::iterator_traits<_RAIterIterator> ::value_type::first_type _RAIter1; typedef typename std::iterator_traits<_RAIter1>::value_type _ValueType; - return multiway_merge_loser_tree< - typename __gnu_cxx::__conditional_type< - _LoserTreeTraits<_ValueType>::_M_use_pointer, - _LoserTreePointer<__stable, _ValueType, _Compare>, - _LoserTree<__stable, _ValueType, _Compare> - >::__type >(__seqs_begin, __seqs_end, __target, __length, __comp); - } - }; - -/** @brief Sequential multi-way merging switch. - * - * The _GLIBCXX_PARALLEL_DECISION is based on the branching factor and - * runtime settings. - * @param __seqs_begin Begin iterator of iterator pair input sequence. - * @param __seqs_end End iterator of iterator pair input sequence. - * @param __target Begin iterator of output sequence. - * @param __comp Comparator. - * @param __length Maximum length to merge, possibly larger than the - * number of elements available. - * @param __stable Stable merging incurs a performance penalty. - * @param __sentinel The sequences have __a __sentinel element. - * @return End iterator of output sequence. */ -template - _RAIter3 - __sequential_multiway_merge(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - const typename std::iterator_traits::value_type::first_type>::value_type& - __sentinel, - _DifferenceTp __length, _Compare __comp) - { - _GLIBCXX_CALL(__length) - - typedef _DifferenceTp _DifferenceType; - typedef typename std::iterator_traits<_RAIterIterator> - ::value_type::first_type - _RAIter1; - typedef typename std::iterator_traits<_RAIter1>::value_type - _ValueType; - #if _GLIBCXX_ASSERTIONS - for (_RAIterIterator __s = __seqs_begin; __s != __seqs_end; ++__s) - { - _GLIBCXX_PARALLEL_ASSERT( - __is_sorted((*__s).first, (*__s).second, __comp)); - } + for (_RAIterIterator __s = __seqs_begin; __s != __seqs_end; ++__s) + { + _GLIBCXX_PARALLEL_ASSERT(__is_sorted((*__s).first, + (*__s).second, __comp)); + } #endif - _DifferenceTp __total_length = 0; - for (_RAIterIterator __s = __seqs_begin; __s != __seqs_end; ++__s) - __total_length += _GLIBCXX_PARALLEL_LENGTH(*__s); + _DifferenceTp __total_length = 0; + for (_RAIterIterator __s = __seqs_begin; __s != __seqs_end; ++__s) + __total_length += _GLIBCXX_PARALLEL_LENGTH(*__s); - __length = std::min<_DifferenceTp>(__length, __total_length); + __length = std::min<_DifferenceTp>(__length, __total_length); - if(__length == 0) - return __target; + if(__length == 0) + return __target; - _RAIter3 __return_target = __target; - int __k = static_cast(__seqs_end - __seqs_begin); + _RAIter3 __return_target = __target; + int __k = static_cast(__seqs_end - __seqs_begin); - switch (__k) - { - case 0: - break; - case 1: - __return_target = std::copy(__seqs_begin[0].first, - __seqs_begin[0].first + __length, - __target); - __seqs_begin[0].first += __length; - break; - case 2: - __return_target = __merge_advance(__seqs_begin[0].first, - __seqs_begin[0].second, - __seqs_begin[1].first, - __seqs_begin[1].second, - __target, __length, __comp); - break; - case 3: - __return_target = __multiway_merge_3_variant_sentinel_switch< - __sentinels, - _RAIterIterator, - _RAIter3, - _DifferenceTp, - _Compare>()(__seqs_begin, __seqs_end, __target, __length, __comp); - break; - case 4: - __return_target = __multiway_merge_4_variant_sentinel_switch< - __sentinels, - _RAIterIterator, - _RAIter3, - _DifferenceTp, - _Compare>()(__seqs_begin, __seqs_end, __target, __length, __comp); - break; - default: - __return_target = __multiway_merge_k_variant_sentinel_switch< - __sentinels, - __stable, - _RAIterIterator, - _RAIter3, - _DifferenceTp, - _Compare>()(__seqs_begin, __seqs_end, __target, __sentinel, - __length, __comp); - break; - } -#if _GLIBCXX_ASSERTIONS - _GLIBCXX_PARALLEL_ASSERT( - __is_sorted(__target, __target + __length, __comp)); -#endif - - return __return_target; - } - -/** - * @brief Stable sorting functor. - * - * Used to reduce code instanciation in multiway_merge_sampling_splitting. - */ -template - struct _SamplingSorter - { - void - operator()(_RAIter __first, _RAIter __last, _StrictWeakOrdering __comp) - { __gnu_sequential::stable_sort(__first, __last, __comp); } - }; - -/** - * @brief Non-__stable sorting functor. - * - * Used to reduce code instantiation in multiway_merge_sampling_splitting. - */ -template - struct _SamplingSorter - { - void - operator()(_RAIter __first, _RAIter __last, _StrictWeakOrdering __comp) - { __gnu_sequential::sort(__first, __last, __comp); } - }; - -/** - * @brief Sampling based splitting for parallel multiway-merge routine. - */ -template - void - multiway_merge_sampling_splitting(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _DifferenceType __length, - _DifferenceType __total_length, - _Compare __comp, - std::vector > *__pieces) - { - typedef typename std::iterator_traits<_RAIterIterator> - ::value_type::first_type - _RAIter1; - typedef typename std::iterator_traits<_RAIter1>::value_type - _ValueType; - - // __k sequences. - int __k = static_cast(__seqs_end - __seqs_begin); - - int __num_threads = omp_get_num_threads(); - - _DifferenceType __num_samples = - __gnu_parallel::_Settings::get().merge_oversampling * __num_threads; - - _ValueType* __samples = static_cast<_ValueType*> - (::operator new(sizeof(_ValueType) * __k * __num_samples)); - // Sample. - for (int __s = 0; __s < __k; ++__s) - for (_DifferenceType __i = 0; __i < __num_samples; ++__i) + switch (__k) { - _DifferenceType sample_index = - static_cast<_DifferenceType> - (_GLIBCXX_PARALLEL_LENGTH(__seqs_begin[__s]) - * (double(__i + 1) / (__num_samples + 1)) - * (double(__length) / __total_length)); - new(&(__samples[__s * __num_samples + __i])) - _ValueType(__seqs_begin[__s].first[sample_index]); + case 0: + break; + case 1: + __return_target = std::copy(__seqs_begin[0].first, + __seqs_begin[0].first + __length, + __target); + __seqs_begin[0].first += __length; + break; + case 2: + __return_target = __merge_advance(__seqs_begin[0].first, + __seqs_begin[0].second, + __seqs_begin[1].first, + __seqs_begin[1].second, + __target, __length, __comp); + break; + case 3: + __return_target = __multiway_merge_3_variant_sentinel_switch + <__sentinels, _RAIterIterator, _RAIter3, _DifferenceTp, _Compare>() + (__seqs_begin, __seqs_end, __target, __length, __comp); + break; + case 4: + __return_target = __multiway_merge_4_variant_sentinel_switch + <__sentinels, _RAIterIterator, _RAIter3, _DifferenceTp, _Compare>() + (__seqs_begin, __seqs_end, __target, __length, __comp); + break; + default: + __return_target = __multiway_merge_k_variant_sentinel_switch + <__sentinels, __stable, _RAIterIterator, _RAIter3, _DifferenceTp, + _Compare>() + (__seqs_begin, __seqs_end, __target, __sentinel, __length, __comp); + break; } +#if _GLIBCXX_ASSERTIONS + _GLIBCXX_PARALLEL_ASSERT( + __is_sorted(__target, __target + __length, __comp)); +#endif - // Sort stable or non-stable, depending on value of template parameter - // "__stable". - _SamplingSorter<__stable, _ValueType*, _Compare>() - (__samples, __samples + (__num_samples * __k), __comp); + return __return_target; + } - for (int __slab = 0; __slab < __num_threads; ++__slab) - // For each slab / processor. - for (int __seq = 0; __seq < __k; ++__seq) - { - // For each sequence. - if (__slab > 0) - __pieces[__slab][__seq].first = - std::upper_bound - (__seqs_begin[__seq].first, - __seqs_begin[__seq].second, - __samples[__num_samples * __k * __slab / __num_threads], - __comp) - __seqs_begin[__seq].first; - else - // Absolute beginning. - __pieces[__slab][__seq].first = 0; - if ((__slab + 1) < __num_threads) - __pieces[__slab][__seq].second = - std::upper_bound - (__seqs_begin[__seq].first, - __seqs_begin[__seq].second, - __samples[__num_samples * __k * (__slab + 1) / - __num_threads], __comp) - - __seqs_begin[__seq].first; - else - // Absolute end. - __pieces[__slab][__seq].second - = _GLIBCXX_PARALLEL_LENGTH(__seqs_begin[__seq]); - } - ::operator delete(__samples); - } - -/** - * @brief Exact splitting for parallel multiway-merge routine. - * - * None of the passed sequences may be empty. - */ -template - void - multiway_merge_exact_splitting(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _DifferenceType __length, - _DifferenceType __total_length, - _Compare __comp, + /** + * @brief Stable sorting functor. + * + * Used to reduce code instanciation in multiway_merge_sampling_splitting. + */ + template + struct _SamplingSorter + { + void + operator()(_RAIter __first, _RAIter __last, _StrictWeakOrdering __comp) + { __gnu_sequential::stable_sort(__first, __last, __comp); } + }; + + /** + * @brief Non-__stable sorting functor. + * + * Used to reduce code instantiation in multiway_merge_sampling_splitting. + */ + template + struct _SamplingSorter + { + void + operator()(_RAIter __first, _RAIter __last, _StrictWeakOrdering __comp) + { __gnu_sequential::sort(__first, __last, __comp); } + }; + + /** + * @brief Sampling based splitting for parallel multiway-merge routine. + */ + template + void + multiway_merge_sampling_splitting(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _DifferenceType __length, + _DifferenceType __total_length, + _Compare __comp, std::vector > *__pieces) - { - typedef typename std::iterator_traits<_RAIterIterator> - ::value_type::first_type - _RAIter1; - - const bool __tight = (__total_length == __length); - - // __k sequences. - const int __k = static_cast(__seqs_end - __seqs_begin); - - const int __num_threads = omp_get_num_threads(); - - // (Settings::multiway_merge_splitting == __gnu_parallel::_Settings::EXACT). - std::vector<_RAIter1>* __offsets = - new std::vector<_RAIter1>[__num_threads]; - std::vector > __se(__k); + { + typedef typename std::iterator_traits<_RAIterIterator> + ::value_type::first_type + _RAIter1; + typedef typename std::iterator_traits<_RAIter1>::value_type + _ValueType; - copy(__seqs_begin, __seqs_end, __se.begin()); + // __k sequences. + int __k = static_cast(__seqs_end - __seqs_begin); - _DifferenceType* __borders = - new _DifferenceType[__num_threads + 1]; - equally_split(__length, __num_threads, __borders); + int __num_threads = omp_get_num_threads(); - for (int __s = 0; __s < (__num_threads - 1); ++__s) - { - __offsets[__s].resize(__k); - multiseq_partition - (__se.begin(), __se.end(), __borders[__s + 1], - __offsets[__s].begin(), __comp); + _DifferenceType __num_samples = + __gnu_parallel::_Settings::get().merge_oversampling * __num_threads; - // Last one also needed and available. - if (!__tight) + _ValueType* __samples = static_cast<_ValueType*> + (::operator new(sizeof(_ValueType) * __k * __num_samples)); + // Sample. + for (int __s = 0; __s < __k; ++__s) + for (_DifferenceType __i = 0; __i < __num_samples; ++__i) { - __offsets[__num_threads - 1].resize(__k); - multiseq_partition(__se.begin(), __se.end(), - _DifferenceType(__length), - __offsets[__num_threads - 1].begin(), __comp); + _DifferenceType sample_index = static_cast<_DifferenceType> + (_GLIBCXX_PARALLEL_LENGTH(__seqs_begin[__s]) + * (double(__i + 1) / (__num_samples + 1)) + * (double(__length) / __total_length)); + new(&(__samples[__s * __num_samples + __i])) + _ValueType(__seqs_begin[__s].first[sample_index]); } - } - delete[] __borders; - for (int __slab = 0; __slab < __num_threads; ++__slab) - { + // Sort stable or non-stable, depending on value of template parameter + // "__stable". + _SamplingSorter<__stable, _ValueType*, _Compare>() + (__samples, __samples + (__num_samples * __k), __comp); + + for (int __slab = 0; __slab < __num_threads; ++__slab) // For each slab / processor. for (int __seq = 0; __seq < __k; ++__seq) { // For each sequence. - if (__slab == 0) - { - // Absolute beginning. - __pieces[__slab][__seq].first = 0; - } + if (__slab > 0) + __pieces[__slab][__seq].first = std::upper_bound + (__seqs_begin[__seq].first, __seqs_begin[__seq].second, + __samples[__num_samples * __k * __slab / __num_threads], + __comp) + - __seqs_begin[__seq].first; else - __pieces[__slab][__seq].first = - __pieces[__slab - 1][__seq].second; - if (!__tight || __slab < (__num_threads - 1)) - __pieces[__slab][__seq].second = - __offsets[__slab][__seq] - __seqs_begin[__seq].first; + // Absolute beginning. + __pieces[__slab][__seq].first = 0; + if ((__slab + 1) < __num_threads) + __pieces[__slab][__seq].second = std::upper_bound + (__seqs_begin[__seq].first, __seqs_begin[__seq].second, + __samples[__num_samples * __k * (__slab + 1) / __num_threads], + __comp) + - __seqs_begin[__seq].first; else - { - // __slab == __num_threads - 1 - __pieces[__slab][__seq].second = - _GLIBCXX_PARALLEL_LENGTH(__seqs_begin[__seq]); - } + // Absolute end. + __pieces[__slab][__seq].second = + _GLIBCXX_PARALLEL_LENGTH(__seqs_begin[__seq]); } - } - delete[] __offsets; - } - -/** @brief Parallel multi-way merge routine. - * - * The _GLIBCXX_PARALLEL_DECISION is based on the branching factor - * and runtime settings. - * - * Must not be called if the number of sequences is 1. - * - * @param _Splitter functor to split input (either __exact or sampling based) - * - * @param __seqs_begin Begin iterator of iterator pair input sequence. - * @param __seqs_end End iterator of iterator pair input sequence. - * @param __target Begin iterator of output sequence. - * @param __comp Comparator. - * @param __length Maximum length to merge, possibly larger than the - * number of elements available. - * @param __stable Stable merging incurs a performance penalty. - * @param __sentinel Ignored. - * @return End iterator of output sequence. - */ -template - _RAIter3 - parallel_multiway_merge(_RAIterIterator __seqs_begin, - _RAIterIterator __seqs_end, - _RAIter3 __target, - _Splitter __splitter, - _DifferenceTp __length, - _Compare __comp, - _ThreadIndex __num_threads) + ::operator delete(__samples); + } + + /** + * @brief Exact splitting for parallel multiway-merge routine. + * + * None of the passed sequences may be empty. + */ + template + void + multiway_merge_exact_splitting(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _DifferenceType __length, + _DifferenceType __total_length, + _Compare __comp, + std::vector > *__pieces) { -#if _GLIBCXX_ASSERTIONS - _GLIBCXX_PARALLEL_ASSERT(__seqs_end - __seqs_begin > 1); -#endif + typedef typename std::iterator_traits<_RAIterIterator> + ::value_type::first_type + _RAIter1; - _GLIBCXX_CALL(__length) + const bool __tight = (__total_length == __length); - typedef _DifferenceTp _DifferenceType; - typedef typename std::iterator_traits<_RAIterIterator> - ::value_type::first_type - _RAIter1; - typedef typename - std::iterator_traits<_RAIter1>::value_type _ValueType; - - // Leave only non-empty sequences. - typedef std::pair<_RAIter1, _RAIter1> seq_type; - seq_type* __ne_seqs = new seq_type[__seqs_end - __seqs_begin]; - int __k = 0; - _DifferenceType __total_length = 0; - for (_RAIterIterator __raii = __seqs_begin; - __raii != __seqs_end; ++__raii) - { - _DifferenceTp __seq_length = _GLIBCXX_PARALLEL_LENGTH(*__raii); - if(__seq_length > 0) - { - __total_length += __seq_length; - __ne_seqs[__k++] = *__raii; - } - } + // __k sequences. + const int __k = static_cast(__seqs_end - __seqs_begin); - _GLIBCXX_CALL(__total_length) + const int __num_threads = omp_get_num_threads(); - __length = std::min<_DifferenceTp>(__length, __total_length); + // (Settings::multiway_merge_splitting + // == __gnu_parallel::_Settings::EXACT). + std::vector<_RAIter1>* __offsets = + new std::vector<_RAIter1>[__num_threads]; + std::vector > __se(__k); + + copy(__seqs_begin, __seqs_end, __se.begin()); - if (__total_length == 0 || __k == 0) + _DifferenceType* __borders = + new _DifferenceType[__num_threads + 1]; + equally_split(__length, __num_threads, __borders); + + for (int __s = 0; __s < (__num_threads - 1); ++__s) + { + __offsets[__s].resize(__k); + multiseq_partition(__se.begin(), __se.end(), __borders[__s + 1], + __offsets[__s].begin(), __comp); + + // Last one also needed and available. + if (!__tight) + { + __offsets[__num_threads - 1].resize(__k); + multiseq_partition(__se.begin(), __se.end(), + _DifferenceType(__length), + __offsets[__num_threads - 1].begin(), + __comp); + } + } + delete[] __borders; + + for (int __slab = 0; __slab < __num_threads; ++__slab) + { + // For each slab / processor. + for (int __seq = 0; __seq < __k; ++__seq) + { + // For each sequence. + if (__slab == 0) + { + // Absolute beginning. + __pieces[__slab][__seq].first = 0; + } + else + __pieces[__slab][__seq].first = + __pieces[__slab - 1][__seq].second; + if (!__tight || __slab < (__num_threads - 1)) + __pieces[__slab][__seq].second = + __offsets[__slab][__seq] - __seqs_begin[__seq].first; + else + { + // __slab == __num_threads - 1 + __pieces[__slab][__seq].second = + _GLIBCXX_PARALLEL_LENGTH(__seqs_begin[__seq]); + } + } + } + delete[] __offsets; + } + + /** @brief Parallel multi-way merge routine. + * + * The _GLIBCXX_PARALLEL_DECISION is based on the branching factor + * and runtime settings. + * + * Must not be called if the number of sequences is 1. + * + * @param _Splitter functor to split input (either __exact or sampling based) + * + * @param __seqs_begin Begin iterator of iterator pair input sequence. + * @param __seqs_end End iterator of iterator pair input sequence. + * @param __target Begin iterator of output sequence. + * @param __comp Comparator. + * @param __length Maximum length to merge, possibly larger than the + * number of elements available. + * @param __stable Stable merging incurs a performance penalty. + * @param __sentinel Ignored. + * @return End iterator of output sequence. + */ + template + _RAIter3 + parallel_multiway_merge(_RAIterIterator __seqs_begin, + _RAIterIterator __seqs_end, + _RAIter3 __target, + _Splitter __splitter, + _DifferenceTp __length, + _Compare __comp, + _ThreadIndex __num_threads) { - delete[] __ne_seqs; - return __target; - } +#if _GLIBCXX_ASSERTIONS + _GLIBCXX_PARALLEL_ASSERT(__seqs_end - __seqs_begin > 1); +#endif + + _GLIBCXX_CALL(__length) + + typedef _DifferenceTp _DifferenceType; + typedef typename std::iterator_traits<_RAIterIterator> + ::value_type::first_type + _RAIter1; + typedef typename + std::iterator_traits<_RAIter1>::value_type _ValueType; + + // Leave only non-empty sequences. + typedef std::pair<_RAIter1, _RAIter1> seq_type; + seq_type* __ne_seqs = new seq_type[__seqs_end - __seqs_begin]; + int __k = 0; + _DifferenceType __total_length = 0; + for (_RAIterIterator __raii = __seqs_begin; + __raii != __seqs_end; ++__raii) + { + _DifferenceTp __seq_length = _GLIBCXX_PARALLEL_LENGTH(*__raii); + if(__seq_length > 0) + { + __total_length += __seq_length; + __ne_seqs[__k++] = *__raii; + } + } - std::vector >* __pieces; + _GLIBCXX_CALL(__total_length) - __num_threads = static_cast<_ThreadIndex> - (std::min<_DifferenceType>(__num_threads, __total_length)); + __length = std::min<_DifferenceTp>(__length, __total_length); -# pragma omp parallel num_threads (__num_threads) - { + if (__total_length == 0 || __k == 0) + { + delete[] __ne_seqs; + return __target; + } + + std::vector >* __pieces; + + __num_threads = static_cast<_ThreadIndex> + (std::min<_DifferenceType>(__num_threads, __total_length)); + +# pragma omp parallel num_threads (__num_threads) + { # pragma omp single - { - __num_threads = omp_get_num_threads(); - // Thread __t will have to merge pieces[__iam][0..__k - 1] - __pieces = new std::vector< - std::pair<_DifferenceType, _DifferenceType> >[__num_threads]; - for (int __s = 0; __s < __num_threads; ++__s) - __pieces[__s].resize(__k); + { + __num_threads = omp_get_num_threads(); + // Thread __t will have to merge pieces[__iam][0..__k - 1] + __pieces = new std::vector< + std::pair<_DifferenceType, _DifferenceType> >[__num_threads]; + for (int __s = 0; __s < __num_threads; ++__s) + __pieces[__s].resize(__k); - _DifferenceType __num_samples = - __gnu_parallel::_Settings::get().merge_oversampling * - __num_threads; + _DifferenceType __num_samples = + __gnu_parallel::_Settings::get().merge_oversampling + * __num_threads; - __splitter(__ne_seqs, __ne_seqs + __k, __length, __total_length, - __comp, __pieces); - } //single + __splitter(__ne_seqs, __ne_seqs + __k, __length, __total_length, + __comp, __pieces); + } //single - _ThreadIndex __iam = omp_get_thread_num(); + _ThreadIndex __iam = omp_get_thread_num(); - _DifferenceType __target_position = 0; + _DifferenceType __target_position = 0; - for (int __c = 0; __c < __k; ++__c) - __target_position += __pieces[__iam][__c].first; + for (int __c = 0; __c < __k; ++__c) + __target_position += __pieces[__iam][__c].first; - seq_type* __chunks = new seq_type[__k]; + seq_type* __chunks = new seq_type[__k]; - for (int __s = 0; __s < __k; ++__s) - { - __chunks[__s] = std::make_pair( - __ne_seqs[__s].first + __pieces[__iam][__s].first, - __ne_seqs[__s].first + __pieces[__iam][__s].second); - } + for (int __s = 0; __s < __k; ++__s) + __chunks[__s] = std::make_pair(__ne_seqs[__s].first + + __pieces[__iam][__s].first, + __ne_seqs[__s].first + + __pieces[__iam][__s].second); - if(__length > __target_position) - __sequential_multiway_merge<__stable, __sentinels>( - __chunks, __chunks + __k, __target + __target_position, - *(__seqs_begin->second), __length - __target_position, __comp); + if(__length > __target_position) + __sequential_multiway_merge<__stable, __sentinels> + (__chunks, __chunks + __k, __target + __target_position, + *(__seqs_begin->second), __length - __target_position, __comp); - delete[] __chunks; - } // parallel + delete[] __chunks; + } // parallel #if _GLIBCXX_ASSERTIONS - _GLIBCXX_PARALLEL_ASSERT( - __is_sorted(__target, __target + __length, __comp)); + _GLIBCXX_PARALLEL_ASSERT( + __is_sorted(__target, __target + __length, __comp)); #endif - __k = 0; - // Update ends of sequences. - for (_RAIterIterator __raii = __seqs_begin; - __raii != __seqs_end; ++__raii) - { - _DifferenceTp __length = _GLIBCXX_PARALLEL_LENGTH(*__raii); - if(__length > 0) - (*__raii).first += __pieces[__num_threads - 1][__k++].second; - } + __k = 0; + // Update ends of sequences. + for (_RAIterIterator __raii = __seqs_begin; + __raii != __seqs_end; ++__raii) + { + _DifferenceTp __length = _GLIBCXX_PARALLEL_LENGTH(*__raii); + if(__length > 0) + (*__raii).first += __pieces[__num_threads - 1][__k++].second; + } - delete[] __pieces; - delete[] __ne_seqs; + delete[] __pieces; + delete[] __ne_seqs; - return __target + __length; + return __target + __length; + } + + /** + * @brief Multiway Merge Frontend. + * + * Merge the sequences specified by seqs_begin and __seqs_end into + * __target. __seqs_begin and __seqs_end must point to a sequence of + * pairs. These pairs must contain an iterator to the beginning + * of a sequence in their first entry and an iterator the _M_end of + * the same sequence in their second entry. + * + * Ties are broken arbitrarily. See stable_multiway_merge for a variant + * that breaks ties by sequence number but is slower. + * + * The first entries of the pairs (i.e. the begin iterators) will be moved + * forward. + * + * The output sequence has to provide enough space for all elements + * that are written to it. + * + * This function will merge the input sequences: + * + * - not stable + * - parallel, depending on the input size and Settings + * - using sampling for splitting + * - not using sentinels + * + * Example: + * + * 
+   *   int sequences[10][10];
+   *   for (int __i = 0; __i < 10; ++__i)
+   *     for (int __j = 0; __i < 10; ++__j)
+   *       sequences[__i][__j] = __j;
+   *
+   *   int __out[33];
+   *   std::vector > seqs;
+   *   for (int __i = 0; __i < 10; ++__i)
+   *     { seqs.push(std::make_pair(sequences[__i],
+   *                                      sequences[__i] + 10)) }
+   *
+   *   multiway_merge(seqs.begin(), seqs.end(), __target, std::less(), 33);
+   *
+ * + * @see stable_multiway_merge + * + * @pre All input sequences must be sorted. + * @pre Target must provide enough space to merge out length elements or + * the number of elements in all sequences, whichever is smaller. + * + * @post [__target, return __value) contains merged __elements from the + * input sequences. + * @post return __value - __target = min(__length, number of elements in all + * sequences). + * + * @param _RAIterPairIterator iterator over sequence + * of pairs of iterators + * @param _RAIterOut iterator over target sequence + * @param _DifferenceTp difference type for the sequence + * @param _Compare strict weak ordering type to compare elements + * in sequences + * + * @param __seqs_begin __begin of sequence __sequence + * @param __seqs_end _M_end of sequence __sequence + * @param __target target sequence to merge to. + * @param __comp strict weak ordering to use for element comparison. + * @param __length Maximum length to merge, possibly larger than the + * number of elements available. + * + * @return _M_end iterator of output sequence + */ + // multiway_merge + // public interface + template + _RAIterOut + multiway_merge(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + __gnu_parallel::sequential_tag) + { + typedef _DifferenceTp _DifferenceType; + _GLIBCXX_CALL(__seqs_end - __seqs_begin) + + // catch special case: no sequences + if (__seqs_begin == __seqs_end) + return __target; + + // Execute multiway merge *sequentially*. + return __sequential_multiway_merge + + (__seqs_begin, __seqs_end, __target, + *(__seqs_begin->second), __length, __comp); } -/** - * @brief Multiway Merge Frontend. - * - * Merge the sequences specified by seqs_begin and __seqs_end into - * __target. __seqs_begin and __seqs_end must point to a sequence of - * pairs. These pairs must contain an iterator to the beginning - * of a sequence in their first entry and an iterator the _M_end of - * the same sequence in their second entry. - * - * Ties are broken arbitrarily. See stable_multiway_merge for a variant - * that breaks ties by sequence number but is slower. - * - * The first entries of the pairs (i.e. the begin iterators) will be moved - * forward. - * - * The output sequence has to provide enough space for all elements - * that are written to it. - * - * This function will merge the input sequences: - * - * - not stable - * - parallel, depending on the input size and Settings - * - using sampling for splitting - * - not using sentinels - * - * Example: - * - * 
- *   int sequences[10][10];
- *   for (int __i = 0; __i < 10; ++__i)
- *     for (int __j = 0; __i < 10; ++__j)
- *       sequences[__i][__j] = __j;
- *
- *   int __out[33];
- *   std::vector > seqs;
- *   for (int __i = 0; __i < 10; ++__i)
- *     { seqs.push(std::make_pair(sequences[__i], sequences[__i] + 10)) }
- *
- *   multiway_merge(seqs.begin(), seqs.end(), __target, std::less(), 33);
- *
- * - * @see stable_multiway_merge - * - * @pre All input sequences must be sorted. - * @pre Target must provide enough space to merge out length elements or - * the number of elements in all sequences, whichever is smaller. - * - * @post [__target, return __value) contains merged __elements from the - * input sequences. - * @post return __value - __target = min(__length, number of elements in all - * sequences). - * - * @param _RAIterPairIterator iterator over sequence - * of pairs of iterators - * @param _RAIterOut iterator over target sequence - * @param _DifferenceTp difference type for the sequence - * @param _Compare strict weak ordering type to compare elements - * in sequences - * - * @param __seqs_begin __begin of sequence __sequence - * @param __seqs_end _M_end of sequence __sequence - * @param __target target sequence to merge to. - * @param __comp strict weak ordering to use for element comparison. - * @param __length Maximum length to merge, possibly larger than the - * number of elements available. - * - * @return _M_end iterator of output sequence - */ -// multiway_merge -// public interface -template - _RAIterOut - multiway_merge(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - __gnu_parallel::sequential_tag) - { - typedef _DifferenceTp _DifferenceType; - _GLIBCXX_CALL(__seqs_end - __seqs_begin) + // public interface + template + _RAIterOut + multiway_merge(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + __gnu_parallel::exact_tag __tag) + { + typedef _DifferenceTp _DifferenceType; + _GLIBCXX_CALL(__seqs_end - __seqs_begin) // catch special case: no sequences if (__seqs_begin == __seqs_end) return __target; - // Execute multiway merge *sequentially*. - return __sequential_multiway_merge - - (__seqs_begin, __seqs_end, __target, - *(__seqs_begin->second), __length, __comp); - } - -// public interface -template - _RAIterOut - multiway_merge(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - __gnu_parallel::exact_tag __tag) - { - typedef _DifferenceTp _DifferenceType; - _GLIBCXX_CALL(__seqs_end - __seqs_begin) - - // catch special case: no sequences - if (__seqs_begin == __seqs_end) - return __target; + // Execute merge; maybe parallel, depending on the number of merged + // elements and the number of sequences and global thresholds in + // Settings. + if ((__seqs_end - __seqs_begin > 1) + && _GLIBCXX_PARALLEL_CONDITION( + ((__seqs_end - __seqs_begin) >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_k) + && ((_SequenceIndex)__length >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) + return parallel_multiway_merge + + (__seqs_begin, __seqs_end, __target, + multiway_merge_exact_splitting + ::value_type*, _Compare, _DifferenceTp>, + static_cast<_DifferenceType>(__length), __comp, + __tag.__get_num_threads()); + else + return __sequential_multiway_merge + + (__seqs_begin, __seqs_end, __target, + *(__seqs_begin->second), __length, __comp); + } - // Execute merge; maybe parallel, depending on the number of merged - // elements and the number of sequences and global thresholds in - // Settings. - if ((__seqs_end - __seqs_begin > 1) && - _GLIBCXX_PARALLEL_CONDITION( - ((__seqs_end - __seqs_begin) >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_k) - && ((_SequenceIndex)__length >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) - return parallel_multiway_merge - ( - __seqs_begin, __seqs_end, __target, - multiway_merge_exact_splitting - ::value_type*, _Compare, _DifferenceTp>, - static_cast<_DifferenceType>(__length), __comp, - __tag.__get_num_threads()); - else - return __sequential_multiway_merge - ( - __seqs_begin, __seqs_end, __target, *(__seqs_begin->second), - __length, __comp); - } - -// public interface -template - _RAIterOut - multiway_merge(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - __gnu_parallel::sampling_tag __tag) - { - typedef _DifferenceTp _DifferenceType; - _GLIBCXX_CALL(__seqs_end - __seqs_begin) - - // catch special case: no sequences - if (__seqs_begin == __seqs_end) - return __target; + // public interface + template + _RAIterOut + multiway_merge(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + __gnu_parallel::sampling_tag __tag) + { + typedef _DifferenceTp _DifferenceType; + _GLIBCXX_CALL(__seqs_end - __seqs_begin) - // Execute merge; maybe parallel, depending on the number of merged - // elements and the number of sequences and global thresholds in - // Settings. - if ((__seqs_end - __seqs_begin > 1) && - _GLIBCXX_PARALLEL_CONDITION( - ((__seqs_end - __seqs_begin) >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_k) - && ((_SequenceIndex)__length >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) - return parallel_multiway_merge - ( - __seqs_begin, __seqs_end, - __target, - multiway_merge_exact_splitting - ::value_type*, _Compare, _DifferenceTp>, - static_cast<_DifferenceType>(__length), __comp, - __tag.__get_num_threads()); - else - return __sequential_multiway_merge - ( - __seqs_begin, __seqs_end, - __target, *(__seqs_begin->second), __length, __comp); - } - -// public interface -template - _RAIterOut - multiway_merge(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - parallel_tag __tag = parallel_tag(0)) - { - return multiway_merge(__seqs_begin, __seqs_end, __target, __length, - __comp, exact_tag(__tag.__get_num_threads())); - } - -// public interface -template - _RAIterOut - multiway_merge(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - default_parallel_tag __tag) - { - return multiway_merge(__seqs_begin, __seqs_end, __target, __length, - __comp, exact_tag(__tag.__get_num_threads())); - } - -// stable_multiway_merge -// public interface -template - _RAIterOut - stable_multiway_merge(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - __gnu_parallel::sequential_tag) - { - typedef _DifferenceTp _DifferenceType; - _GLIBCXX_CALL(__seqs_end - __seqs_begin) - - // catch special case: no sequences - if (__seqs_begin == __seqs_end) - return __target; + // catch special case: no sequences + if (__seqs_begin == __seqs_end) + return __target; - // Execute multiway merge *sequentially*. - return __sequential_multiway_merge - - (__seqs_begin, __seqs_end, __target, *(__seqs_begin->second), __length, - __comp); - } - -// public interface -template - _RAIterOut - stable_multiway_merge(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - __gnu_parallel::exact_tag __tag) - { - typedef _DifferenceTp _DifferenceType; - _GLIBCXX_CALL(__seqs_end - __seqs_begin) - - // catch special case: no sequences - if (__seqs_begin == __seqs_end) - return __target; + // Execute merge; maybe parallel, depending on the number of merged + // elements and the number of sequences and global thresholds in + // Settings. + if ((__seqs_end - __seqs_begin > 1) + && _GLIBCXX_PARALLEL_CONDITION( + ((__seqs_end - __seqs_begin) >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_k) + && ((_SequenceIndex)__length >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) + return parallel_multiway_merge + + (__seqs_begin, __seqs_end, __target, + multiway_merge_exact_splitting + ::value_type*, _Compare, _DifferenceTp>, + static_cast<_DifferenceType>(__length), __comp, + __tag.__get_num_threads()); + else + return __sequential_multiway_merge + + (__seqs_begin, __seqs_end, __target, + *(__seqs_begin->second), __length, __comp); + } - // Execute merge; maybe parallel, depending on the number of merged - // elements and the number of sequences and global thresholds in - // Settings. - if ((__seqs_end - __seqs_begin > 1) && - _GLIBCXX_PARALLEL_CONDITION( - ((__seqs_end - __seqs_begin) >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_k) - && ((_SequenceIndex)__length >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) - return parallel_multiway_merge - ( - __seqs_begin, __seqs_end, - __target, - multiway_merge_exact_splitting - ::value_type*, _Compare, _DifferenceTp>, - static_cast<_DifferenceType>(__length), __comp, - __tag.__get_num_threads()); - else - return __sequential_multiway_merge( - __seqs_begin, __seqs_end, - __target, *(__seqs_begin->second), __length, __comp); - } - -// public interface -template - _RAIterOut - stable_multiway_merge(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - sampling_tag __tag) - { - typedef _DifferenceTp _DifferenceType; - _GLIBCXX_CALL(__seqs_end - __seqs_begin) - - // catch special case: no sequences - if (__seqs_begin == __seqs_end) - return __target; + // public interface + template + _RAIterOut + multiway_merge(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + parallel_tag __tag = parallel_tag(0)) + { return multiway_merge(__seqs_begin, __seqs_end, __target, __length, + __comp, exact_tag(__tag.__get_num_threads())); } + + // public interface + template + _RAIterOut + multiway_merge(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + default_parallel_tag __tag) + { return multiway_merge(__seqs_begin, __seqs_end, __target, __length, + __comp, exact_tag(__tag.__get_num_threads())); } + + // stable_multiway_merge + // public interface + template + _RAIterOut + stable_multiway_merge(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + __gnu_parallel::sequential_tag) + { + typedef _DifferenceTp _DifferenceType; + _GLIBCXX_CALL(__seqs_end - __seqs_begin) - // Execute merge; maybe parallel, depending on the number of merged - // elements and the number of sequences and global thresholds in - // Settings. - if ((__seqs_end - __seqs_begin > 1) && - _GLIBCXX_PARALLEL_CONDITION( - ((__seqs_end - __seqs_begin) >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_k) - && ((_SequenceIndex)__length >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) - return parallel_multiway_merge - ( - __seqs_begin, __seqs_end, - __target, - multiway_merge_sampling_splitting - ::value_type*, _Compare, _DifferenceTp>, - static_cast<_DifferenceType>(__length), __comp, - __tag.__get_num_threads()); - else + // catch special case: no sequences + if (__seqs_begin == __seqs_end) + return __target; + + // Execute multiway merge *sequentially*. return __sequential_multiway_merge - ( - __seqs_begin, __seqs_end, - __target, *(__seqs_begin->second), __length, __comp); - } - -// public interface -template - _RAIterOut - stable_multiway_merge(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - parallel_tag __tag = parallel_tag(0)) - { - return stable_multiway_merge - (__seqs_begin, __seqs_end, __target, __length, __comp, - exact_tag(__tag.__get_num_threads())); - } - -// public interface -template - _RAIterOut - stable_multiway_merge(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - default_parallel_tag __tag) - { - return stable_multiway_merge - (__seqs_begin, __seqs_end, __target, __length, __comp, - exact_tag(__tag.__get_num_threads())); - } - -/** - * @brief Multiway Merge Frontend. - * - * Merge the sequences specified by seqs_begin and __seqs_end into - * __target. __seqs_begin and __seqs_end must point to a sequence of - * pairs. These pairs must contain an iterator to the beginning - * of a sequence in their first entry and an iterator the _M_end of - * the same sequence in their second entry. - * - * Ties are broken arbitrarily. See stable_multiway_merge for a variant - * that breaks ties by sequence number but is slower. - * - * The first entries of the pairs (i.e. the begin iterators) will be moved - * forward accordingly. - * - * The output sequence has to provide enough space for all elements - * that are written to it. - * - * This function will merge the input sequences: - * - * - not stable - * - parallel, depending on the input size and Settings - * - using sampling for splitting - * - using sentinels - * - * You have to take care that the element the _M_end iterator points to is - * readable and contains a value that is greater than any other non-sentinel - * value in all sequences. - * - * Example: - * - * 
- *   int sequences[10][11];
- *   for (int __i = 0; __i < 10; ++__i)
- *     for (int __j = 0; __i < 11; ++__j)
- *       sequences[__i][__j] = __j; // __last one is sentinel!
- *
- *   int __out[33];
- *   std::vector > seqs;
- *   for (int __i = 0; __i < 10; ++__i)
- *     { seqs.push(std::make_pair(sequences[__i], sequences[__i] + 10)) }
- *
- *   multiway_merge(seqs.begin(), seqs.end(), __target, std::less(), 33);
- *
- * - * @pre All input sequences must be sorted. - * @pre Target must provide enough space to merge out length elements or - * the number of elements in all sequences, whichever is smaller. - * @pre For each @__c __i, @__c __seqs_begin[__i].second must be the end - * marker of the sequence, but also reference the one more __sentinel - * element. - * - * @post [__target, return __value) contains merged __elements from the - * input sequences. - * @post return __value - __target = min(__length, number of elements in all - * sequences). - * - * @see stable_multiway_merge_sentinels - * - * @param _RAIterPairIterator iterator over sequence - * of pairs of iterators - * @param _RAIterOut iterator over target sequence - * @param _DifferenceTp difference type for the sequence - * @param _Compare strict weak ordering type to compare elements - * in sequences - * - * @param __seqs_begin __begin of sequence __sequence - * @param __seqs_end _M_end of sequence __sequence - * @param __target target sequence to merge to. - * @param __comp strict weak ordering to use for element comparison. - * @param __length Maximum length to merge, possibly larger than the - * number of elements available. - * - * @return _M_end iterator of output sequence - */ -// multiway_merge_sentinels -// public interface -template - _RAIterOut - multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - __gnu_parallel::sequential_tag) - { - typedef _DifferenceTp _DifferenceType; - _GLIBCXX_CALL(__seqs_end - __seqs_begin) - - // catch special case: no sequences - if (__seqs_begin == __seqs_end) - return __target; + + (__seqs_begin, __seqs_end, __target, + *(__seqs_begin->second), __length, __comp); + } - // Execute multiway merge *sequentially*. - return __sequential_multiway_merge - - (__seqs_begin, __seqs_end, - __target, *(__seqs_begin->second), __length, __comp); - } - -// public interface -template - _RAIterOut - multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - __gnu_parallel::exact_tag __tag) - { - typedef _DifferenceTp _DifferenceType; - _GLIBCXX_CALL(__seqs_end - __seqs_begin) - - // catch special case: no sequences - if (__seqs_begin == __seqs_end) - return __target; + // public interface + template + _RAIterOut + stable_multiway_merge(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + __gnu_parallel::exact_tag __tag) + { + typedef _DifferenceTp _DifferenceType; + _GLIBCXX_CALL(__seqs_end - __seqs_begin) - // Execute merge; maybe parallel, depending on the number of merged - // elements and the number of sequences and global thresholds in - // Settings. - if ((__seqs_end - __seqs_begin > 1) && - _GLIBCXX_PARALLEL_CONDITION( - ((__seqs_end - __seqs_begin) >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_k) - && ((_SequenceIndex)__length >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) - return parallel_multiway_merge - ( - __seqs_begin, __seqs_end, - __target, - multiway_merge_exact_splitting - ::value_type*, _Compare, _DifferenceTp>, - static_cast<_DifferenceType>(__length), __comp, - __tag.__get_num_threads()); - else - return __sequential_multiway_merge - ( - __seqs_begin, __seqs_end, - __target, *(__seqs_begin->second), __length, __comp); - } - -// public interface -template - _RAIterOut - multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - sampling_tag __tag) - { - typedef _DifferenceTp _DifferenceType; - _GLIBCXX_CALL(__seqs_end - __seqs_begin) - - // catch special case: no sequences - if (__seqs_begin == __seqs_end) - return __target; + // catch special case: no sequences + if (__seqs_begin == __seqs_end) + return __target; - // Execute merge; maybe parallel, depending on the number of merged - // elements and the number of sequences and global thresholds in - // Settings. - if ((__seqs_end - __seqs_begin > 1) && - _GLIBCXX_PARALLEL_CONDITION( - ((__seqs_end - __seqs_begin) >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_k) - && ((_SequenceIndex)__length >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) - return parallel_multiway_merge - - (__seqs_begin, __seqs_end, __target, - multiway_merge_sampling_splitting - ::value_type*, _Compare, _DifferenceTp>, - static_cast<_DifferenceType>(__length), __comp, - __tag.__get_num_threads()); - else + // Execute merge; maybe parallel, depending on the number of merged + // elements and the number of sequences and global thresholds in + // Settings. + if ((__seqs_end - __seqs_begin > 1) + && _GLIBCXX_PARALLEL_CONDITION( + ((__seqs_end - __seqs_begin) >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_k) + && ((_SequenceIndex)__length >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) + return parallel_multiway_merge + + (__seqs_begin, __seqs_end, __target, + multiway_merge_exact_splitting + ::value_type*, _Compare, _DifferenceTp>, + static_cast<_DifferenceType>(__length), __comp, + __tag.__get_num_threads()); + else + return __sequential_multiway_merge + + (__seqs_begin, __seqs_end, __target, + *(__seqs_begin->second), __length, __comp); + } + + // public interface + template + _RAIterOut + stable_multiway_merge(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + sampling_tag __tag) + { + typedef _DifferenceTp _DifferenceType; + _GLIBCXX_CALL(__seqs_end - __seqs_begin) + + // catch special case: no sequences + if (__seqs_begin == __seqs_end) + return __target; + + // Execute merge; maybe parallel, depending on the number of merged + // elements and the number of sequences and global thresholds in + // Settings. + if ((__seqs_end - __seqs_begin > 1) + && _GLIBCXX_PARALLEL_CONDITION( + ((__seqs_end - __seqs_begin) >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_k) + && ((_SequenceIndex)__length >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) + return parallel_multiway_merge + + (__seqs_begin, __seqs_end, __target, + multiway_merge_sampling_splitting + ::value_type*, _Compare, _DifferenceTp>, + static_cast<_DifferenceType>(__length), __comp, + __tag.__get_num_threads()); + else + return __sequential_multiway_merge + + (__seqs_begin, __seqs_end, __target, + *(__seqs_begin->second), __length, __comp); + } + + // public interface + template + _RAIterOut + stable_multiway_merge(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + parallel_tag __tag = parallel_tag(0)) + { + return stable_multiway_merge + (__seqs_begin, __seqs_end, __target, __length, __comp, + exact_tag(__tag.__get_num_threads())); + } + + // public interface + template + _RAIterOut + stable_multiway_merge(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + default_parallel_tag __tag) + { + return stable_multiway_merge + (__seqs_begin, __seqs_end, __target, __length, __comp, + exact_tag(__tag.__get_num_threads())); + } + + /** + * @brief Multiway Merge Frontend. + * + * Merge the sequences specified by seqs_begin and __seqs_end into + * __target. __seqs_begin and __seqs_end must point to a sequence of + * pairs. These pairs must contain an iterator to the beginning + * of a sequence in their first entry and an iterator the _M_end of + * the same sequence in their second entry. + * + * Ties are broken arbitrarily. See stable_multiway_merge for a variant + * that breaks ties by sequence number but is slower. + * + * The first entries of the pairs (i.e. the begin iterators) will be moved + * forward accordingly. + * + * The output sequence has to provide enough space for all elements + * that are written to it. + * + * This function will merge the input sequences: + * + * - not stable + * - parallel, depending on the input size and Settings + * - using sampling for splitting + * - using sentinels + * + * You have to take care that the element the _M_end iterator points to is + * readable and contains a value that is greater than any other non-sentinel + * value in all sequences. + * + * Example: + * + * 
+   *   int sequences[10][11];
+   *   for (int __i = 0; __i < 10; ++__i)
+   *     for (int __j = 0; __i < 11; ++__j)
+   *       sequences[__i][__j] = __j; // __last one is sentinel!
+   *
+   *   int __out[33];
+   *   std::vector > seqs;
+   *   for (int __i = 0; __i < 10; ++__i)
+   *     { seqs.push(std::make_pair(sequences[__i],
+   *                                      sequences[__i] + 10)) }
+   *
+   *   multiway_merge(seqs.begin(), seqs.end(), __target, std::less(), 33);
+   *
+ * + * @pre All input sequences must be sorted. + * @pre Target must provide enough space to merge out length elements or + * the number of elements in all sequences, whichever is smaller. + * @pre For each @__c __i, @__c __seqs_begin[__i].second must be the end + * marker of the sequence, but also reference the one more __sentinel + * element. + * + * @post [__target, return __value) contains merged __elements from the + * input sequences. + * @post return __value - __target = min(__length, number of elements in all + * sequences). + * + * @see stable_multiway_merge_sentinels + * + * @param _RAIterPairIterator iterator over sequence + * of pairs of iterators + * @param _RAIterOut iterator over target sequence + * @param _DifferenceTp difference type for the sequence + * @param _Compare strict weak ordering type to compare elements + * in sequences + * + * @param __seqs_begin __begin of sequence __sequence + * @param __seqs_end _M_end of sequence __sequence + * @param __target target sequence to merge to. + * @param __comp strict weak ordering to use for element comparison. + * @param __length Maximum length to merge, possibly larger than the + * number of elements available. + * + * @return _M_end iterator of output sequence + */ + // multiway_merge_sentinels + // public interface + template + _RAIterOut + multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + __gnu_parallel::sequential_tag) + { + typedef _DifferenceTp _DifferenceType; + _GLIBCXX_CALL(__seqs_end - __seqs_begin) + + // catch special case: no sequences + if (__seqs_begin == __seqs_end) + return __target; + + // Execute multiway merge *sequentially*. return __sequential_multiway_merge - ( - __seqs_begin, __seqs_end, - __target, *(__seqs_begin->second), __length, __comp); - } - -// public interface -template - _RAIterOut - multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - parallel_tag __tag = parallel_tag(0)) - { - return multiway_merge_sentinels - (__seqs_begin, __seqs_end, __target, __length, __comp, - exact_tag(__tag.__get_num_threads())); - } - -// public interface -template - _RAIterOut - multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - default_parallel_tag __tag) - { - return multiway_merge_sentinels - (__seqs_begin, __seqs_end, __target, __length, __comp, - exact_tag(__tag.__get_num_threads())); - } - -// stable_multiway_merge_sentinels -// public interface -template - _RAIterOut - stable_multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - __gnu_parallel::sequential_tag) - { - typedef _DifferenceTp _DifferenceType; - _GLIBCXX_CALL(__seqs_end - __seqs_begin) - - // catch special case: no sequences - if (__seqs_begin == __seqs_end) - return __target; + + (__seqs_begin, __seqs_end, + __target, *(__seqs_begin->second), __length, __comp); + } - // Execute multiway merge *sequentially*. - return __sequential_multiway_merge - - (__seqs_begin, __seqs_end, __target, *(__seqs_begin->second), - __length, __comp); - } - -// public interface -template - _RAIterOut - stable_multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - __gnu_parallel::exact_tag __tag) - { - typedef _DifferenceTp _DifferenceType; - _GLIBCXX_CALL(__seqs_end - __seqs_begin) - - // catch special case: no sequences - if (__seqs_begin == __seqs_end) - return __target; + // public interface + template + _RAIterOut + multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + __gnu_parallel::exact_tag __tag) + { + typedef _DifferenceTp _DifferenceType; + _GLIBCXX_CALL(__seqs_end - __seqs_begin) + + // catch special case: no sequences + if (__seqs_begin == __seqs_end) + return __target; - // Execute merge; maybe parallel, depending on the number of merged - // elements and the number of sequences and global thresholds in - // Settings. - if ((__seqs_end - __seqs_begin > 1) && - _GLIBCXX_PARALLEL_CONDITION( - ((__seqs_end - __seqs_begin) >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_k) - && ((_SequenceIndex)__length >= - __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) - return parallel_multiway_merge - ( - __seqs_begin, __seqs_end, - __target, - multiway_merge_exact_splitting - ::value_type*, _Compare, _DifferenceTp>, - static_cast<_DifferenceType>(__length), __comp, - __tag.__get_num_threads()); - else + // Execute merge; maybe parallel, depending on the number of merged + // elements and the number of sequences and global thresholds in + // Settings. + if ((__seqs_end - __seqs_begin > 1) + && _GLIBCXX_PARALLEL_CONDITION( + ((__seqs_end - __seqs_begin) >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_k) + && ((_SequenceIndex)__length >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) + return parallel_multiway_merge + + (__seqs_begin, __seqs_end, __target, + multiway_merge_exact_splitting + ::value_type*, _Compare, _DifferenceTp>, + static_cast<_DifferenceType>(__length), __comp, + __tag.__get_num_threads()); + else + return __sequential_multiway_merge + + (__seqs_begin, __seqs_end, __target, + *(__seqs_begin->second), __length, __comp); + } + + // public interface + template + _RAIterOut + multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + sampling_tag __tag) + { + typedef _DifferenceTp _DifferenceType; + _GLIBCXX_CALL(__seqs_end - __seqs_begin) + + // catch special case: no sequences + if (__seqs_begin == __seqs_end) + return __target; + + // Execute merge; maybe parallel, depending on the number of merged + // elements and the number of sequences and global thresholds in + // Settings. + if ((__seqs_end - __seqs_begin > 1) + && _GLIBCXX_PARALLEL_CONDITION( + ((__seqs_end - __seqs_begin) >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_k) + && ((_SequenceIndex)__length >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) + return parallel_multiway_merge + + (__seqs_begin, __seqs_end, __target, + multiway_merge_sampling_splitting + ::value_type*, _Compare, _DifferenceTp>, + static_cast<_DifferenceType>(__length), __comp, + __tag.__get_num_threads()); + else + return __sequential_multiway_merge + ( + __seqs_begin, __seqs_end, __target, + *(__seqs_begin->second), __length, __comp); + } + + // public interface + template + _RAIterOut + multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + parallel_tag __tag = parallel_tag(0)) + { + return multiway_merge_sentinels + (__seqs_begin, __seqs_end, __target, __length, __comp, + exact_tag(__tag.__get_num_threads())); + } + + // public interface + template + _RAIterOut + multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + default_parallel_tag __tag) + { + return multiway_merge_sentinels + (__seqs_begin, __seqs_end, __target, __length, __comp, + exact_tag(__tag.__get_num_threads())); + } + + // stable_multiway_merge_sentinels + // public interface + template + _RAIterOut + stable_multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + __gnu_parallel::sequential_tag) + { + typedef _DifferenceTp _DifferenceType; + _GLIBCXX_CALL(__seqs_end - __seqs_begin) + + // catch special case: no sequences + if (__seqs_begin == __seqs_end) + return __target; + + // Execute multiway merge *sequentially*. return __sequential_multiway_merge - ( - __seqs_begin, __seqs_end, __target, *(__seqs_begin->second), - __length, __comp); - } - -// public interface -template - _RAIterOut - stable_multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, - _Compare __comp, - sampling_tag __tag) - { - typedef _DifferenceTp _DifferenceType; - _GLIBCXX_CALL(__seqs_end - __seqs_begin) - - // catch special case: no sequences - if (__seqs_begin == __seqs_end) - return __target; + + (__seqs_begin, __seqs_end, __target, + *(__seqs_begin->second), __length, __comp); + } - // Execute merge; maybe parallel, depending on the number of merged - // elements and the number of sequences and global thresholds in - // Settings. - if ((__seqs_end - __seqs_begin > 1) && - _GLIBCXX_PARALLEL_CONDITION( - ((__seqs_end - __seqs_begin) >= + // public interface + template + _RAIterOut + stable_multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + __gnu_parallel::exact_tag __tag) + { + typedef _DifferenceTp _DifferenceType; + _GLIBCXX_CALL(__seqs_end - __seqs_begin) + + // catch special case: no sequences + if (__seqs_begin == __seqs_end) + return __target; + + // Execute merge; maybe parallel, depending on the number of merged + // elements and the number of sequences and global thresholds in + // Settings. + if ((__seqs_end - __seqs_begin > 1) + && _GLIBCXX_PARALLEL_CONDITION( + ((__seqs_end - __seqs_begin) >= __gnu_parallel::_Settings::get().multiway_merge_minimal_k) - && ((_SequenceIndex)__length >= + && ((_SequenceIndex)__length >= __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) - return parallel_multiway_merge - ( - __seqs_begin, __seqs_end, - __target, - multiway_merge_sampling_splitting - ::value_type*, _Compare, _DifferenceTp>, - static_cast<_DifferenceType>(__length), __comp, - __tag.__get_num_threads()); - else - return __sequential_multiway_merge - ( - __seqs_begin, __seqs_end, - __target, *(__seqs_begin->second), __length, __comp); - } - -// public interface -template - _RAIterOut - stable_multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, - _Compare __comp, - parallel_tag __tag = parallel_tag(0)) - { - return stable_multiway_merge_sentinels - (__seqs_begin, __seqs_end, __target, __length, __comp, - exact_tag(__tag.__get_num_threads())); - } - -// public interface -template - _RAIterOut - stable_multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, - _RAIterPairIterator __seqs_end, - _RAIterOut __target, - _DifferenceTp __length, _Compare __comp, - default_parallel_tag __tag) - { - return stable_multiway_merge_sentinels - (__seqs_begin, __seqs_end, __target, __length, __comp, - exact_tag(__tag.__get_num_threads())); - } + return parallel_multiway_merge + + (__seqs_begin, __seqs_end, __target, + multiway_merge_exact_splitting + ::value_type*, _Compare, _DifferenceTp>, + static_cast<_DifferenceType>(__length), __comp, + __tag.__get_num_threads()); + else + return __sequential_multiway_merge + + (__seqs_begin, __seqs_end, __target, + *(__seqs_begin->second), __length, __comp); + } + + // public interface + template + _RAIterOut + stable_multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, + _Compare __comp, + sampling_tag __tag) + { + typedef _DifferenceTp _DifferenceType; + _GLIBCXX_CALL(__seqs_end - __seqs_begin) + // catch special case: no sequences + if (__seqs_begin == __seqs_end) + return __target; + + // Execute merge; maybe parallel, depending on the number of merged + // elements and the number of sequences and global thresholds in + // Settings. + if ((__seqs_end - __seqs_begin > 1) + && _GLIBCXX_PARALLEL_CONDITION( + ((__seqs_end - __seqs_begin) >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_k) + && ((_SequenceIndex)__length >= + __gnu_parallel::_Settings::get().multiway_merge_minimal_n))) + return parallel_multiway_merge + + (__seqs_begin, __seqs_end, __target, + multiway_merge_sampling_splitting + ::value_type*, _Compare, _DifferenceTp>, + static_cast<_DifferenceType>(__length), __comp, + __tag.__get_num_threads()); + else + return __sequential_multiway_merge + + (__seqs_begin, __seqs_end, __target, + *(__seqs_begin->second), __length, __comp); + } + + // public interface + template + _RAIterOut + stable_multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, + _Compare __comp, + parallel_tag __tag = parallel_tag(0)) + { + return stable_multiway_merge_sentinels + (__seqs_begin, __seqs_end, __target, __length, __comp, + exact_tag(__tag.__get_num_threads())); + } + + // public interface + template + _RAIterOut + stable_multiway_merge_sentinels(_RAIterPairIterator __seqs_begin, + _RAIterPairIterator __seqs_end, + _RAIterOut __target, + _DifferenceTp __length, _Compare __comp, + default_parallel_tag __tag) + { + return stable_multiway_merge_sentinels + (__seqs_begin, __seqs_end, __target, __length, __comp, + exact_tag(__tag.__get_num_threads())); + } }; // namespace __gnu_parallel #endif /* _GLIBCXX_PARALLEL_MULTIWAY_MERGE_H */ diff --git a/libstdc++-v3/include/parallel/multiway_mergesort.h b/libstdc++-v3/include/parallel/multiway_mergesort.h index 3b047b40815..1f2c43db61f 100644 --- a/libstdc++-v3/include/parallel/multiway_mergesort.h +++ b/libstdc++-v3/include/parallel/multiway_mergesort.h @@ -125,8 +125,7 @@ namespace __gnu_parallel /** @brief Split by exact splitting. */ template - struct _SplitConsistently + struct _SplitConsistently { void operator()(const _ThreadIndex __iam, @@ -140,19 +139,19 @@ namespace __gnu_parallel std::vector > - seqs(__sd->_M_num_threads); + __seqs(__sd->_M_num_threads); for (_ThreadIndex __s = 0; __s < __sd->_M_num_threads; __s++) - seqs[__s] = std::make_pair(__sd->_M_temporary[__s], - __sd->_M_temporary[__s] - + (__sd->_M_starts[__s + 1] - - __sd->_M_starts[__s])); + __seqs[__s] = std::make_pair(__sd->_M_temporary[__s], + __sd->_M_temporary[__s] + + (__sd->_M_starts[__s + 1] + - __sd->_M_starts[__s])); - std::vector<_SortingPlacesIterator> _M_offsets(__sd->_M_num_threads); + std::vector<_SortingPlacesIterator> __offsets(__sd->_M_num_threads); // if not last thread if (__iam < __sd->_M_num_threads - 1) - multiseq_partition(seqs.begin(), seqs.end(), - __sd->_M_starts[__iam + 1], _M_offsets.begin(), + multiseq_partition(__seqs.begin(), __seqs.end(), + __sd->_M_starts[__iam + 1], __offsets.begin(), __comp); for (int __seq = 0; __seq < __sd->_M_num_threads; __seq++) @@ -160,7 +159,7 @@ namespace __gnu_parallel // for each sequence if (__iam < (__sd->_M_num_threads - 1)) __sd->_M_pieces[__iam][__seq]._M_end - = _M_offsets[__seq] - seqs[__seq].first; + = __offsets[__seq] - __seqs[__seq].first; else // very end of this sequence __sd->_M_pieces[__iam][__seq]._M_end = @@ -185,8 +184,7 @@ namespace __gnu_parallel /** @brief Split by sampling. */ template - struct _SplitConsistently + struct _SplitConsistently { void operator()(const _ThreadIndex __iam, @@ -282,10 +280,8 @@ namespace __gnu_parallel const _RAIter& __target, _Compare& __comp, _DiffType __length_am) const - { - stable_multiway_merge(__seqs_begin, __seqs_end, __target, __length_am, - __comp, sequential_tag()); - } + { stable_multiway_merge(__seqs_begin, __seqs_end, __target, + __length_am, __comp, sequential_tag()); } }; template_M_starts[__iam + 1] - __sd->_M_starts[__iam]; + _DifferenceType __length_local = + __sd->_M_starts[__iam + 1] - __sd->_M_starts[__iam]; // Sort in temporary storage, leave space for sentinel. @@ -350,8 +344,7 @@ namespace __gnu_parallel _DifferenceType __num_samples = _Settings::get().sort_mwms_oversampling * __sd->_M_num_threads - 1; - _SplitConsistently - <__exact, _RAIter, _Compare, _SortingPlacesIterator>() + _SplitConsistently<__exact, _RAIter, _Compare, _SortingPlacesIterator>() (__iam, __sd, __comp, __num_samples); // Offset from __target __begin, __length after merging. @@ -364,26 +357,24 @@ namespace __gnu_parallel } typedef std::vector< - std::pair<_SortingPlacesIterator, _SortingPlacesIterator> > + std::pair<_SortingPlacesIterator, _SortingPlacesIterator> > _SeqVector; - _SeqVector seqs(__sd->_M_num_threads); + _SeqVector __seqs(__sd->_M_num_threads); for (int __s = 0; __s < __sd->_M_num_threads; ++__s) { - seqs[__s] = - std::make_pair - (__sd->_M_temporary[__s] + __sd->_M_pieces[__iam][__s]._M_begin, - __sd->_M_temporary[__s] + __sd->_M_pieces[__iam][__s]._M_end); + __seqs[__s] = + std::make_pair(__sd->_M_temporary[__s] + + __sd->_M_pieces[__iam][__s]._M_begin, + __sd->_M_temporary[__s] + + __sd->_M_pieces[__iam][__s]._M_end); } __possibly_stable_multiway_merge< - __stable, - typename _SeqVector::iterator, - _RAIter, - _Compare, _DifferenceType>() - (seqs.begin(), seqs.end(), - __sd->_M_source + __offset, __comp, - __length_am); + __stable, typename _SeqVector::iterator, + _RAIter, _Compare, _DifferenceType>()(__seqs.begin(), __seqs.end(), + __sd->_M_source + __offset, __comp, + __length_am); # pragma omp barrier @@ -421,7 +412,7 @@ namespace __gnu_parallel // shared variables _PMWMSSortingData<_RAIter> __sd; - _DifferenceType* _M_starts; + _DifferenceType* __starts; # pragma omp parallel num_threads(__num_threads) { @@ -450,30 +441,29 @@ namespace __gnu_parallel = new std::vector<_Piece<_DifferenceType> >[__num_threads]; for (int __s = 0; __s < __num_threads; ++__s) __sd._M_pieces[__s].resize(__num_threads); - _M_starts = __sd._M_starts - = new _DifferenceType[__num_threads + 1]; + __starts = __sd._M_starts = new _DifferenceType[__num_threads + 1]; _DifferenceType __chunk_length = __n / __num_threads; _DifferenceType __split = __n % __num_threads; _DifferenceType __pos = 0; for (int __i = 0; __i < __num_threads; ++__i) { - _M_starts[__i] = __pos; - __pos += (__i < __split) - ? (__chunk_length + 1) : __chunk_length; + __starts[__i] = __pos; + __pos += ((__i < __split) + ? (__chunk_length + 1) : __chunk_length); } - _M_starts[__num_threads] = __pos; + __starts[__num_threads] = __pos; } //single // Now sort in parallel. parallel_sort_mwms_pu<__stable, __exact>(&__sd, __comp); } //parallel - delete[] _M_starts; + delete[] __starts; delete[] __sd._M_temporary; if (!__exact) - ::operator delete(__sd._M_samples); + ::operator delete(__sd._M_samples); delete[] __sd._M_offsets; delete[] __sd._M_pieces; diff --git a/libstdc++-v3/include/parallel/numeric b/libstdc++-v3/include/parallel/numeric index dc4c96761c8..6c5607cedc5 100644 --- a/libstdc++-v3/include/parallel/numeric +++ b/libstdc++-v3/include/parallel/numeric @@ -69,7 +69,7 @@ namespace __parallel __accumulate_switch(_IIter __begin, _IIter __end, _Tp __init, _IteratorTag) { return accumulate(__begin, __end, __init, -__gnu_parallel::sequential_tag()); } + __gnu_parallel::sequential_tag()); } template diff --git a/libstdc++-v3/include/parallel/omp_loop.h b/libstdc++-v3/include/parallel/omp_loop.h index 136b5c8cbef..d07fd7a9307 100644 --- a/libstdc++-v3/include/parallel/omp_loop.h +++ b/libstdc++-v3/include/parallel/omp_loop.h @@ -74,8 +74,8 @@ namespace __gnu_parallel _DifferenceType; _DifferenceType __length = __end - __begin; - _ThreadIndex __num_threads = - __gnu_parallel::min<_DifferenceType>(__get_max_threads(), __length); + _ThreadIndex __num_threads = __gnu_parallel::min<_DifferenceType> + (__get_max_threads(), __length); _Result *__thread_results; @@ -94,8 +94,8 @@ namespace __gnu_parallel #pragma omp for schedule(dynamic, _Settings::get().workstealing_chunk_size) for (_DifferenceType __pos = 0; __pos < __length; ++__pos) - __thread_results[__iam] = - __r(__thread_results[__iam], __f(__o, __begin+__pos)); + __thread_results[__iam] = __r(__thread_results[__iam], + __f(__o, __begin+__pos)); } //parallel for (_ThreadIndex __i = 0; __i < __num_threads; ++__i) diff --git a/libstdc++-v3/include/parallel/omp_loop_static.h b/libstdc++-v3/include/parallel/omp_loop_static.h index e7ca267c92c..f43726e1973 100644 --- a/libstdc++-v3/include/parallel/omp_loop_static.h +++ b/libstdc++-v3/include/parallel/omp_loop_static.h @@ -74,8 +74,8 @@ namespace __gnu_parallel _DifferenceType; _DifferenceType __length = __end - __begin; - _ThreadIndex __num_threads = - std::min<_DifferenceType>(__get_max_threads(), __length); + _ThreadIndex __num_threads = std::min<_DifferenceType> + (__get_max_threads(), __length); _Result *__thread_results; diff --git a/libstdc++-v3/include/parallel/par_loop.h b/libstdc++-v3/include/parallel/par_loop.h index f0a463ebba5..1e21d3ad042 100644 --- a/libstdc++-v3/include/parallel/par_loop.h +++ b/libstdc++-v3/include/parallel/par_loop.h @@ -75,25 +75,24 @@ namespace __gnu_parallel _Result *__thread_results; bool* __constructed; - _ThreadIndex __num_threads = - __gnu_parallel::min<_DifferenceType>(__get_max_threads(), __length); + _ThreadIndex __num_threads = __gnu_parallel::min<_DifferenceType> + (__get_max_threads(), __length); # pragma omp parallel num_threads(__num_threads) { # pragma omp single { __num_threads = omp_get_num_threads(); - __thread_results = - static_cast<_Result*>(::operator new(__num_threads - * sizeof(_Result))); + __thread_results = static_cast<_Result*> + (::operator new(__num_threads * sizeof(_Result))); __constructed = new bool[__num_threads]; } _ThreadIndex __iam = omp_get_thread_num(); // Neutral element. - _Result* __reduct = - static_cast<_Result*>(::operator new(sizeof(_Result))); + _Result* __reduct = static_cast<_Result*> + (::operator new(sizeof(_Result))); _DifferenceType __start = equally_split_point(__length, __num_threads, __iam), diff --git a/libstdc++-v3/include/parallel/partial_sum.h b/libstdc++-v3/include/parallel/partial_sum.h index 10673af04ba..02404203f4e 100644 --- a/libstdc++-v3/include/parallel/partial_sum.h +++ b/libstdc++-v3/include/parallel/partial_sum.h @@ -149,9 +149,10 @@ namespace __gnu_parallel if (__iam == 0) { *__result = *__begin; - __parallel_partial_sum_basecase( - __begin + 1, __begin + __borders[1], __result + 1, - __bin_op, *__begin); + __parallel_partial_sum_basecase(__begin + 1, + __begin + __borders[1], + __result + 1, + __bin_op, *__begin); ::new(&(__sums[__iam])) _ValueType(*(__result + __borders[1] - 1)); } else @@ -168,7 +169,7 @@ namespace __gnu_parallel # pragma omp single __parallel_partial_sum_basecase(__sums + 1, __sums + __num_threads, - __sums + 1, __bin_op, __sums[0]); + __sums + 1, __bin_op, __sums[0]); # pragma omp barrier diff --git a/libstdc++-v3/include/parallel/partition.h b/libstdc++-v3/include/parallel/partition.h index 11362939fc8..86723c51270 100644 --- a/libstdc++-v3/include/parallel/partition.h +++ b/libstdc++-v3/include/parallel/partition.h @@ -44,387 +44,391 @@ namespace __gnu_parallel { -/** @brief Parallel implementation of std::partition. - * @param __begin Begin iterator of input sequence to split. - * @param __end End iterator of input sequence to split. - * @param __pred Partition predicate, possibly including some kind of pivot. - * @param __num_threads Maximum number of threads to use for this task. - * @return Number of elements not fulfilling the predicate. */ -template - typename std::iterator_traits<_RAIter>::difference_type - __parallel_partition(_RAIter __begin, _RAIter __end, - _Predicate __pred, _ThreadIndex __num_threads) - { - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::value_type _ValueType; - typedef typename _TraitsType::difference_type _DifferenceType; - - _DifferenceType __n = __end - __begin; - - _GLIBCXX_CALL(__n) - - const _Settings& __s = _Settings::get(); - - // Shared. - _GLIBCXX_VOLATILE _DifferenceType __left = 0, __right = __n - 1; - _GLIBCXX_VOLATILE _DifferenceType __leftover_left, __leftover_right; - _GLIBCXX_VOLATILE _DifferenceType __leftnew, __rightnew; - - bool* __reserved_left = NULL, * __reserved_right = NULL; - - _DifferenceType __chunk_size; - - omp_lock_t __result_lock; - omp_init_lock(&__result_lock); - - //at least two chunks per thread - if(__right - __left + 1 >= 2 * __num_threads * __chunk_size) -# pragma omp parallel num_threads(__num_threads) - { -# pragma omp single - { - __num_threads = omp_get_num_threads(); - __reserved_left = new bool[__num_threads]; - __reserved_right = new bool[__num_threads]; - - if (__s.partition_chunk_share > 0.0) - __chunk_size = std::max<_DifferenceType>( - __s.partition_chunk_size, - (double)__n * __s.partition_chunk_share / - (double)__num_threads); - else - __chunk_size = __s.partition_chunk_size; - } - - while (__right - __left + 1 >= 2 * __num_threads * __chunk_size) - { -# pragma omp single - { - _DifferenceType __num_chunks - = (__right - __left + 1) / __chunk_size; - - for (int __r = 0; __r < __num_threads; ++__r) - { - __reserved_left[__r] = false; - __reserved_right[__r] = false; - } - __leftover_left = 0; - __leftover_right = 0; - } //implicit barrier - - // Private. - _DifferenceType __thread_left, __thread_left_border, - __thread_right, __thread_right_border; - __thread_left = __left + 1; - - // Just to satisfy the condition below. - __thread_left_border = __thread_left - 1; - __thread_right = __n - 1; - __thread_right_border = __thread_right + 1; - - bool __iam_finished = false; - while (!__iam_finished) - { - if (__thread_left > __thread_left_border) - { - omp_set_lock(&__result_lock); - if (__left + (__chunk_size - 1) > __right) - __iam_finished = true; - else - { - __thread_left = __left; - __thread_left_border = __left + (__chunk_size - 1); - __left += __chunk_size; - } - omp_unset_lock(&__result_lock); - } - - if (__thread_right < __thread_right_border) - { - omp_set_lock(&__result_lock); - if (__left > __right - (__chunk_size - 1)) - __iam_finished = true; - else - { - __thread_right = __right; - __thread_right_border = __right - (__chunk_size - 1); - __right -= __chunk_size; - } - omp_unset_lock(&__result_lock); - } - - if (__iam_finished) - break; - - // Swap as usual. - while (__thread_left < __thread_right) - { - while (__pred(__begin[__thread_left]) - && __thread_left <= __thread_left_border) - ++__thread_left; - while (!__pred(__begin[__thread_right]) - && __thread_right >= __thread_right_border) - --__thread_right; - - if (__thread_left > __thread_left_border - || __thread_right < __thread_right_border) - // Fetch new chunk(__s). - break; - - std::swap(__begin[__thread_left], __begin[__thread_right]); - ++__thread_left; - --__thread_right; - } - } - - // Now swap the leftover chunks to the right places. - if (__thread_left <= __thread_left_border) -# pragma omp atomic - ++__leftover_left; - if (__thread_right >= __thread_right_border) -# pragma omp atomic - ++__leftover_right; - -# pragma omp barrier - -# pragma omp single - { - __leftnew = __left - __leftover_left * __chunk_size; - __rightnew = __right + __leftover_right * __chunk_size; - } - -# pragma omp barrier - - // <=> __thread_left_border + (__chunk_size - 1) >= __leftnew - if (__thread_left <= __thread_left_border - && __thread_left_border >= __leftnew) - { - // Chunk already in place, reserve spot. - __reserved_left - [(__left - (__thread_left_border + 1)) / __chunk_size] - = true; - } - - // <=> __thread_right_border - (__chunk_size - 1) <= __rightnew - if (__thread_right >= __thread_right_border - && __thread_right_border <= __rightnew) - { - // Chunk already in place, reserve spot. - __reserved_right[((__thread_right_border - 1) - __right) - / __chunk_size] = true; - } - -# pragma omp barrier - - if (__thread_left <= __thread_left_border - && __thread_left_border < __leftnew) - { - // Find spot and swap. - _DifferenceType __swapstart = -1; - omp_set_lock(&__result_lock); - for (int __r = 0; __r < __leftover_left; ++__r) - if (!__reserved_left[__r]) - { - __reserved_left[__r] = true; - __swapstart = __left - (__r + 1) * __chunk_size; - break; - } - omp_unset_lock(&__result_lock); + /** @brief Parallel implementation of std::partition. + * @param __begin Begin iterator of input sequence to split. + * @param __end End iterator of input sequence to split. + * @param __pred Partition predicate, possibly including some kind + * of pivot. + * @param __num_threads Maximum number of threads to use for this task. + * @return Number of elements not fulfilling the predicate. */ + template + typename std::iterator_traits<_RAIter>::difference_type + __parallel_partition(_RAIter __begin, _RAIter __end, + _Predicate __pred, _ThreadIndex __num_threads) + { + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::value_type _ValueType; + typedef typename _TraitsType::difference_type _DifferenceType; + + _DifferenceType __n = __end - __begin; + + _GLIBCXX_CALL(__n) + + const _Settings& __s = _Settings::get(); + + // Shared. + _GLIBCXX_VOLATILE _DifferenceType __left = 0, __right = __n - 1; + _GLIBCXX_VOLATILE _DifferenceType __leftover_left, __leftover_right; + _GLIBCXX_VOLATILE _DifferenceType __leftnew, __rightnew; + + bool* __reserved_left = NULL, * __reserved_right = NULL; + + _DifferenceType __chunk_size; + + omp_lock_t __result_lock; + omp_init_lock(&__result_lock); + + //at least two chunks per thread + if (__right - __left + 1 >= 2 * __num_threads * __chunk_size) +# pragma omp parallel num_threads(__num_threads) + { +# pragma omp single + { + __num_threads = omp_get_num_threads(); + __reserved_left = new bool[__num_threads]; + __reserved_right = new bool[__num_threads]; + + if (__s.partition_chunk_share > 0.0) + __chunk_size = std::max<_DifferenceType> + (__s.partition_chunk_size, (double)__n + * __s.partition_chunk_share / (double)__num_threads); + else + __chunk_size = __s.partition_chunk_size; + } + + while (__right - __left + 1 >= 2 * __num_threads * __chunk_size) + { +# pragma omp single + { + _DifferenceType __num_chunks = ((__right - __left + 1) + / __chunk_size); + + for (int __r = 0; __r < __num_threads; ++__r) + { + __reserved_left[__r] = false; + __reserved_right[__r] = false; + } + __leftover_left = 0; + __leftover_right = 0; + } //implicit barrier + + // Private. + _DifferenceType __thread_left, __thread_left_border, + __thread_right, __thread_right_border; + __thread_left = __left + 1; + + // Just to satisfy the condition below. + __thread_left_border = __thread_left - 1; + __thread_right = __n - 1; + __thread_right_border = __thread_right + 1; + + bool __iam_finished = false; + while (!__iam_finished) + { + if (__thread_left > __thread_left_border) + { + omp_set_lock(&__result_lock); + if (__left + (__chunk_size - 1) > __right) + __iam_finished = true; + else + { + __thread_left = __left; + __thread_left_border = __left + (__chunk_size - 1); + __left += __chunk_size; + } + omp_unset_lock(&__result_lock); + } + + if (__thread_right < __thread_right_border) + { + omp_set_lock(&__result_lock); + if (__left > __right - (__chunk_size - 1)) + __iam_finished = true; + else + { + __thread_right = __right; + __thread_right_border = __right - (__chunk_size - 1); + __right -= __chunk_size; + } + omp_unset_lock(&__result_lock); + } + + if (__iam_finished) + break; + + // Swap as usual. + while (__thread_left < __thread_right) + { + while (__pred(__begin[__thread_left]) + && __thread_left <= __thread_left_border) + ++__thread_left; + while (!__pred(__begin[__thread_right]) + && __thread_right >= __thread_right_border) + --__thread_right; + + if (__thread_left > __thread_left_border + || __thread_right < __thread_right_border) + // Fetch new chunk(__s). + break; + + std::swap(__begin[__thread_left], + __begin[__thread_right]); + ++__thread_left; + --__thread_right; + } + } + + // Now swap the leftover chunks to the right places. + if (__thread_left <= __thread_left_border) +# pragma omp atomic + ++__leftover_left; + if (__thread_right >= __thread_right_border) +# pragma omp atomic + ++__leftover_right; + +# pragma omp barrier + +# pragma omp single + { + __leftnew = __left - __leftover_left * __chunk_size; + __rightnew = __right + __leftover_right * __chunk_size; + } + +# pragma omp barrier + + // <=> __thread_left_border + (__chunk_size - 1) >= __leftnew + if (__thread_left <= __thread_left_border + && __thread_left_border >= __leftnew) + { + // Chunk already in place, reserve spot. + __reserved_left[(__left - (__thread_left_border + 1)) + / __chunk_size] = true; + } + + // <=> __thread_right_border - (__chunk_size - 1) <= __rightnew + if (__thread_right >= __thread_right_border + && __thread_right_border <= __rightnew) + { + // Chunk already in place, reserve spot. + __reserved_right[((__thread_right_border - 1) - __right) + / __chunk_size] = true; + } + +# pragma omp barrier + + if (__thread_left <= __thread_left_border + && __thread_left_border < __leftnew) + { + // Find spot and swap. + _DifferenceType __swapstart = -1; + omp_set_lock(&__result_lock); + for (int __r = 0; __r < __leftover_left; ++__r) + if (!__reserved_left[__r]) + { + __reserved_left[__r] = true; + __swapstart = __left - (__r + 1) * __chunk_size; + break; + } + omp_unset_lock(&__result_lock); #if _GLIBCXX_ASSERTIONS - _GLIBCXX_PARALLEL_ASSERT(__swapstart != -1); + _GLIBCXX_PARALLEL_ASSERT(__swapstart != -1); #endif - std::swap_ranges(__begin + __thread_left_border - - (__chunk_size - 1), - __begin + __thread_left_border + 1, - __begin + __swapstart); - } - - if (__thread_right >= __thread_right_border - && __thread_right_border > __rightnew) - { - // Find spot and swap - _DifferenceType __swapstart = -1; - omp_set_lock(&__result_lock); - for (int __r = 0; __r < __leftover_right; ++__r) - if (!__reserved_right[__r]) - { - __reserved_right[__r] = true; - __swapstart = __right + __r * __chunk_size + 1; - break; - } - omp_unset_lock(&__result_lock); + std::swap_ranges(__begin + __thread_left_border + - (__chunk_size - 1), + __begin + __thread_left_border + 1, + __begin + __swapstart); + } + + if (__thread_right >= __thread_right_border + && __thread_right_border > __rightnew) + { + // Find spot and swap + _DifferenceType __swapstart = -1; + omp_set_lock(&__result_lock); + for (int __r = 0; __r < __leftover_right; ++__r) + if (!__reserved_right[__r]) + { + __reserved_right[__r] = true; + __swapstart = __right + __r * __chunk_size + 1; + break; + } + omp_unset_lock(&__result_lock); #if _GLIBCXX_ASSERTIONS - _GLIBCXX_PARALLEL_ASSERT(__swapstart != -1); + _GLIBCXX_PARALLEL_ASSERT(__swapstart != -1); #endif - std::swap_ranges( - __begin + __thread_right_border, - __begin + __thread_right_border + __chunk_size, - __begin + __swapstart); - } + std::swap_ranges(__begin + __thread_right_border, + __begin + __thread_right_border + + __chunk_size, __begin + __swapstart); + } #if _GLIBCXX_ASSERTIONS # pragma omp barrier # pragma omp single - { - for (int __r = 0; __r < __leftover_left; ++__r) - _GLIBCXX_PARALLEL_ASSERT(__reserved_left[__r]); - for (int __r = 0; __r < __leftover_right; ++__r) - _GLIBCXX_PARALLEL_ASSERT(__reserved_right[__r]); - } + { + for (int __r = 0; __r < __leftover_left; ++__r) + _GLIBCXX_PARALLEL_ASSERT(__reserved_left[__r]); + for (int __r = 0; __r < __leftover_right; ++__r) + _GLIBCXX_PARALLEL_ASSERT(__reserved_right[__r]); + } # pragma omp barrier #endif # pragma omp barrier - __left = __leftnew; - __right = __rightnew; - } -# pragma omp flush(__left, __right) - } // end "recursion" //parallel - - _DifferenceType __final_left = __left, __final_right = __right; - - while (__final_left < __final_right) - { - // Go right until key is geq than pivot. - while (__pred(__begin[__final_left]) && __final_left < __final_right) - ++__final_left; - - // Go left until key is less than pivot. - while (!__pred(__begin[__final_right]) && __final_left < __final_right) - --__final_right; - - if (__final_left == __final_right) - break; - std::swap(__begin[__final_left], __begin[__final_right]); - ++__final_left; - --__final_right; - } - - // All elements on the left side are < piv, all elements on the - // right are >= piv - delete[] __reserved_left; - delete[] __reserved_right; - - omp_destroy_lock(&__result_lock); - - // Element "between" __final_left and __final_right might not have - // been regarded yet - if (__final_left < __n && !__pred(__begin[__final_left])) - // Really swapped. - return __final_left; - else - return __final_left + 1; - } - -/** - * @brief Parallel implementation of std::nth_element(). + __left = __leftnew; + __right = __rightnew; + } + +# pragma omp flush(__left, __right) + } // end "recursion" //parallel + + _DifferenceType __final_left = __left, __final_right = __right; + + while (__final_left < __final_right) + { + // Go right until key is geq than pivot. + while (__pred(__begin[__final_left]) + && __final_left < __final_right) + ++__final_left; + + // Go left until key is less than pivot. + while (!__pred(__begin[__final_right]) + && __final_left < __final_right) + --__final_right; + + if (__final_left == __final_right) + break; + std::swap(__begin[__final_left], __begin[__final_right]); + ++__final_left; + --__final_right; + } + + // All elements on the left side are < piv, all elements on the + // right are >= piv + delete[] __reserved_left; + delete[] __reserved_right; + + omp_destroy_lock(&__result_lock); + + // Element "between" __final_left and __final_right might not have + // been regarded yet + if (__final_left < __n && !__pred(__begin[__final_left])) + // Really swapped. + return __final_left; + else + return __final_left + 1; + } + + /** + * @brief Parallel implementation of std::nth_element(). + * @param __begin Begin iterator of input sequence. + * @param __nth _Iterator of element that must be in position afterwards. + * @param __end End iterator of input sequence. + * @param __comp Comparator. + */ + template + void + __parallel_nth_element(_RAIter __begin, _RAIter __nth, + _RAIter __end, _Compare __comp) + { + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::value_type _ValueType; + typedef typename _TraitsType::difference_type _DifferenceType; + + _GLIBCXX_CALL(__end - __begin) + + _RAIter __split; + _RandomNumber __rng; + + _DifferenceType __minimum_length = + std::max<_DifferenceType>(2, _Settings::get().partition_minimal_n); + + // Break if input range to small. + while (static_cast<_SequenceIndex>(__end - __begin) >= __minimum_length) + { + _DifferenceType __n = __end - __begin; + + _RAIter __pivot_pos = __begin + __rng(__n); + + // Swap __pivot_pos value to end. + if (__pivot_pos != (__end - 1)) + std::swap(*__pivot_pos, *(__end - 1)); + __pivot_pos = __end - 1; + + // _Compare must have first_value_type, second_value_type, + // result_type + // _Compare == + // __gnu_parallel::_Lexicographic > + // __pivot_pos == std::pair* + __gnu_parallel::binder2nd<_Compare, _ValueType, _ValueType, bool> + __pred(__comp, *__pivot_pos); + + // Divide, leave pivot unchanged in last place. + _RAIter __split_pos1, __split_pos2; + __split_pos1 = __begin + __parallel_partition(__begin, __end - 1, + __pred, + __get_max_threads()); + + // Left side: < __pivot_pos; __right side: >= __pivot_pos + + // Swap pivot back to middle. + if (__split_pos1 != __pivot_pos) + std::swap(*__split_pos1, *__pivot_pos); + __pivot_pos = __split_pos1; + + // In case all elements are equal, __split_pos1 == 0 + if ((__split_pos1 + 1 - __begin) < (__n >> 7) + || (__end - __split_pos1) < (__n >> 7)) + { + // Very unequal split, one part smaller than one 128th + // elements not strictly larger than the pivot. + __gnu_parallel::__unary_negate<__gnu_parallel:: + __binder1st<_Compare, _ValueType, + _ValueType, bool>, _ValueType> + __pred(__gnu_parallel::__binder1st<_Compare, _ValueType, + _ValueType, bool>(__comp, *__pivot_pos)); + + // Find other end of pivot-equal range. + __split_pos2 = __gnu_sequential::partition(__split_pos1 + 1, + __end, __pred); + } + else + // Only skip the pivot. + __split_pos2 = __split_pos1 + 1; + + // Compare iterators. + if (__split_pos2 <= __nth) + __begin = __split_pos2; + else if (__nth < __split_pos1) + __end = __split_pos1; + else + break; + } + + // Only at most _Settings::partition_minimal_n __elements __left. + __gnu_sequential::sort(__begin, __end, __comp); + } + + /** @brief Parallel implementation of std::partial_sort(). * @param __begin Begin iterator of input sequence. - * @param __nth _Iterator of element that must be in position afterwards. + * @param __middle Sort until this position. * @param __end End iterator of input sequence. - * @param __comp Comparator. - */ -template - void - __parallel_nth_element(_RAIter __begin, _RAIter __nth, - _RAIter __end, _Compare __comp) - { - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::value_type _ValueType; - typedef typename _TraitsType::difference_type _DifferenceType; - - _GLIBCXX_CALL(__end - __begin) - - _RAIter __split; - _RandomNumber __rng; - - _DifferenceType __minimum_length = - std::max<_DifferenceType>(2, _Settings::get().partition_minimal_n); - - // Break if input range to small. - while (static_cast<_SequenceIndex>(__end - __begin) >= __minimum_length) - { - _DifferenceType __n = __end - __begin; - - _RAIter __pivot_pos = __begin + __rng(__n); - - // Swap __pivot_pos value to end. - if (__pivot_pos != (__end - 1)) - std::swap(*__pivot_pos, *(__end - 1)); - __pivot_pos = __end - 1; - - // _Compare must have first_value_type, second_value_type, - // result_type - // _Compare == - // __gnu_parallel::_Lexicographic > - // __pivot_pos == std::pair* - __gnu_parallel::binder2nd<_Compare, _ValueType, _ValueType, bool> - __pred(__comp, *__pivot_pos); - - // Divide, leave pivot unchanged in last place. - _RAIter __split_pos1, __split_pos2; - __split_pos1 = __begin - + __parallel_partition(__begin, __end - 1, __pred, - __get_max_threads()); - - // Left side: < __pivot_pos; __right side: >= __pivot_pos - - // Swap pivot back to middle. - if (__split_pos1 != __pivot_pos) - std::swap(*__split_pos1, *__pivot_pos); - __pivot_pos = __split_pos1; - - // In case all elements are equal, __split_pos1 == 0 - if ((__split_pos1 + 1 - __begin) < (__n >> 7) - || (__end - __split_pos1) < (__n >> 7)) - { - // Very unequal split, one part smaller than one 128th - // elements not strictly larger than the pivot. - __gnu_parallel::__unary_negate<__gnu_parallel:: - __binder1st<_Compare, _ValueType, _ValueType, bool>, _ValueType> - __pred(__gnu_parallel::__binder1st<_Compare, _ValueType, - _ValueType, bool>(__comp, *__pivot_pos)); - - // Find other end of pivot-equal range. - __split_pos2 = __gnu_sequential::partition(__split_pos1 + 1, - __end, __pred); - } - else - // Only skip the pivot. - __split_pos2 = __split_pos1 + 1; - - // Compare iterators. - if (__split_pos2 <= __nth) - __begin = __split_pos2; - else if (__nth < __split_pos1) - __end = __split_pos1; - else - break; - } - - // Only at most _Settings::partition_minimal_n __elements __left. - __gnu_sequential::sort(__begin, __end, __comp); - } - -/** @brief Parallel implementation of std::partial_sort(). -* @param __begin Begin iterator of input sequence. -* @param __middle Sort until this position. -* @param __end End iterator of input sequence. -* @param __comp Comparator. */ -template - void - __parallel_partial_sort(_RAIter __begin, - _RAIter __middle, - _RAIter __end, _Compare __comp) - { - __parallel_nth_element(__begin, __middle, __end, __comp); - std::sort(__begin, __middle, __comp); - } + * @param __comp Comparator. */ + template + void + __parallel_partial_sort(_RAIter __begin, + _RAIter __middle, + _RAIter __end, _Compare __comp) + { + __parallel_nth_element(__begin, __middle, __end, __comp); + std::sort(__begin, __middle, __comp); + } } //namespace __gnu_parallel diff --git a/libstdc++-v3/include/parallel/queue.h b/libstdc++-v3/include/parallel/queue.h index 348bb1ac900..ff5deb09383 100644 --- a/libstdc++-v3/include/parallel/queue.h +++ b/libstdc++-v3/include/parallel/queue.h @@ -65,10 +65,10 @@ namespace __gnu_parallel public: /** @brief Constructor. Not to be called concurrent, of course. * @param _M_max_size Maximal number of elements to be contained. */ - _RestrictedBoundedConcurrentQueue(_SequenceIndex _M_max_size) + _RestrictedBoundedConcurrentQueue(_SequenceIndex __max_size) { - this->_M_max_size = _M_max_size; - _M_base = new _Tp[_M_max_size]; + _M_max_size = __max_size; + _M_base = new _Tp[__max_size]; _M_borders = __encode2(0, 0); #pragma omp flush } @@ -105,12 +105,12 @@ namespace __gnu_parallel while (__former_front > __former_back) { // Chance. - _CASable - __former_borders = __encode2(__former_front, __former_back); - _CASable - __new_borders = __encode2(__former_front - 1, __former_back); - if (__compare_and_swap( - &_M_borders, __former_borders, __new_borders)) + _CASable __former_borders = __encode2(__former_front, + __former_back); + _CASable __new_borders = __encode2(__former_front - 1, + __former_back); + if (__compare_and_swap(&_M_borders, __former_borders, + __new_borders)) { __t = *(_M_base + (__former_front - 1) % _M_max_size); return true; @@ -132,12 +132,12 @@ namespace __gnu_parallel while (__former_front > __former_back) { // Chance. - _CASable - __former_borders = __encode2(__former_front, __former_back); - _CASable - __new_borders = __encode2(__former_front, __former_back + 1); - if (__compare_and_swap( - &_M_borders, __former_borders, __new_borders)) + _CASable __former_borders = __encode2(__former_front, + __former_back); + _CASable __new_borders = __encode2(__former_front, + __former_back + 1); + if (__compare_and_swap(&_M_borders, __former_borders, + __new_borders)) { __t = *(_M_base + __former_back % _M_max_size); return true; diff --git a/libstdc++-v3/include/parallel/quicksort.h b/libstdc++-v3/include/parallel/quicksort.h index 1ed46b4a77f..508c3c1763c 100644 --- a/libstdc++-v3/include/parallel/quicksort.h +++ b/libstdc++-v3/include/parallel/quicksort.h @@ -48,13 +48,12 @@ namespace __gnu_parallel */ template typename std::iterator_traits<_RAIter>::difference_type - __parallel_sort_qs_divide(_RAIter __begin, - _RAIter __end, - _Compare __comp, typename std::iterator_traits - <_RAIter>::difference_type __pivot_rank, - typename std::iterator_traits - <_RAIter>::difference_type - __num_samples, _ThreadIndex __num_threads) + __parallel_sort_qs_divide(_RAIter __begin, _RAIter __end, + _Compare __comp, typename std::iterator_traits + <_RAIter>::difference_type __pivot_rank, + typename std::iterator_traits + <_RAIter>::difference_type + __num_samples, _ThreadIndex __num_threads) { typedef std::iterator_traits<_RAIter> _TraitsType; typedef typename _TraitsType::value_type _ValueType; @@ -64,25 +63,24 @@ namespace __gnu_parallel __num_samples = std::min(__num_samples, __n); // Allocate uninitialized, to avoid default constructor. - _ValueType* __samples = - static_cast<_ValueType*>(::operator new(__num_samples - * sizeof(_ValueType))); + _ValueType* __samples = static_cast<_ValueType*> + (::operator new(__num_samples * sizeof(_ValueType))); for (_DifferenceType __s = 0; __s < __num_samples; ++__s) { - const unsigned long long __index - = static_cast(__s) * __n / __num_samples; + const unsigned long long __index = static_cast + (__s) * __n / __num_samples; ::new(&(__samples[__s])) _ValueType(__begin[__index]); } __gnu_sequential::sort(__samples, __samples + __num_samples, __comp); - _ValueType& pivot = __samples[__pivot_rank * __num_samples / __n]; + _ValueType& __pivot = __samples[__pivot_rank * __num_samples / __n]; __gnu_parallel::binder2nd<_Compare, _ValueType, _ValueType, bool> - __pred(__comp, pivot); - _DifferenceType __split = - __parallel_partition(__begin, __end, __pred, __num_threads); + __pred(__comp, __pivot); + _DifferenceType __split = __parallel_partition(__begin, __end, + __pred, __num_threads); ::operator delete(__samples); @@ -98,10 +96,9 @@ namespace __gnu_parallel */ template void - __parallel_sort_qs_conquer(_RAIter __begin, - _RAIter __end, - _Compare __comp, - _ThreadIndex __num_threads) + __parallel_sort_qs_conquer(_RAIter __begin, _RAIter __end, + _Compare __comp, + _ThreadIndex __num_threads) { typedef std::iterator_traits<_RAIter> _TraitsType; typedef typename _TraitsType::value_type _ValueType; @@ -127,24 +124,22 @@ namespace __gnu_parallel __pivot_rank = __n * __num_threads_left / __num_threads; - _DifferenceType __split = - __parallel_sort_qs_divide(__begin, __end, __comp, __pivot_rank, - _Settings::get().sort_qs_num_samples_preset, - __num_threads); + _DifferenceType __split = __parallel_sort_qs_divide + (__begin, __end, __comp, __pivot_rank, + _Settings::get().sort_qs_num_samples_preset, __num_threads); #pragma omp parallel sections num_threads(2) { #pragma omp section __parallel_sort_qs_conquer(__begin, __begin + __split, - __comp, __num_threads_left); + __comp, __num_threads_left); #pragma omp section __parallel_sort_qs_conquer(__begin + __split, __end, - __comp, __num_threads - __num_threads_left); + __comp, __num_threads - __num_threads_left); } } - /** @brief Unbalanced quicksort main call. * @param __begin Begin iterator of input sequence. * @param __end End iterator input sequence, ignored. @@ -154,10 +149,9 @@ namespace __gnu_parallel */ template void - __parallel_sort_qs(_RAIter __begin, - _RAIter __end, - _Compare __comp, - _ThreadIndex __num_threads) + __parallel_sort_qs(_RAIter __begin, _RAIter __end, + _Compare __comp, + _ThreadIndex __num_threads) { _GLIBCXX_CALL(__n) diff --git a/libstdc++-v3/include/parallel/random_shuffle.h b/libstdc++-v3/include/parallel/random_shuffle.h index a74b9ac82ac..137414f20f9 100644 --- a/libstdc++-v3/include/parallel/random_shuffle.h +++ b/libstdc++-v3/include/parallel/random_shuffle.h @@ -39,492 +39,484 @@ namespace __gnu_parallel { -/** @brief Type to hold the index of a bin. - * - * Since many variables of this type are allocated, it should be - * chosen as small as possible. - */ -typedef unsigned short _BinIndex; - -/** @brief Data known to every thread participating in - __gnu_parallel::__parallel_random_shuffle(). */ -template - struct _DRandomShufflingGlobalData - { - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::value_type _ValueType; - typedef typename _TraitsType::difference_type _DifferenceType; - - /** @brief Begin iterator of the __source. */ - _RAIter& _M_source; - - /** @brief Temporary arrays for each thread. */ - _ValueType** _M_temporaries; - - /** @brief Two-dimensional array to hold the thread-bin distribution. - * - * Dimensions (_M_num_threads + 1) __x (_M_num_bins + 1). */ - _DifferenceType** _M_dist; - - /** @brief Start indexes of the threads' __chunks. */ - _DifferenceType* _M_starts; - - /** @brief Number of the thread that will further process the - corresponding bin. */ - _ThreadIndex* _M_bin_proc; - - /** @brief Number of bins to distribute to. */ - int _M_num_bins; - - /** @brief Number of bits needed to address the bins. */ - int _M_num_bits; - - /** @brief Constructor. */ - _DRandomShufflingGlobalData(_RAIter& __source) - : _M_source(__source) { } - }; - -/** @brief Local data for a thread participating in - __gnu_parallel::__parallel_random_shuffle(). - */ -template - struct _DRSSorterPU - { - /** @brief Number of threads participating in total. */ - int _M_num_threads; - - /** @brief Begin index for bins taken care of by this thread. */ - _BinIndex _M_bins_begin; - - /** @brief End index for bins taken care of by this thread. */ - _BinIndex __bins_end; - - /** @brief Random _M_seed for this thread. */ - uint32_t _M_seed; - - /** @brief Pointer to global data. */ - _DRandomShufflingGlobalData<_RAIter>* _M_sd; - }; - -/** @brief Generate a random number in @__c [0,2^logp). - * @param logp Logarithm (basis 2) of the upper range __bound. - * @param __rng Random number generator to use. - */ -template - inline int - __random_number_pow2(int logp, _RandomNumberGenerator& __rng) - { return __rng.__genrand_bits(logp); } - -/** @brief Random shuffle code executed by each thread. - * @param __pus Array of thread-local data records. */ -template - void - __parallel_random_shuffle_drs_pu(_DRSSorterPU<_RAIter, - _RandomNumberGenerator>* __pus) - { - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::value_type _ValueType; - typedef typename _TraitsType::difference_type _DifferenceType; - - _ThreadIndex __iam = omp_get_thread_num(); - _DRSSorterPU<_RAIter, _RandomNumberGenerator>* d = &__pus[__iam]; - _DRandomShufflingGlobalData<_RAIter>* _M_sd = d->_M_sd; - - // Indexing: _M_dist[bin][processor] - _DifferenceType __length = _M_sd->_M_starts[__iam + 1] - - _M_sd->_M_starts[__iam]; - _BinIndex* __oracles = new _BinIndex[__length]; - _DifferenceType* _M_dist = new _DifferenceType[_M_sd->_M_num_bins + 1]; - _BinIndex* _M_bin_proc = new _BinIndex[_M_sd->_M_num_bins]; - _ValueType** _M_temporaries = new _ValueType*[d->_M_num_threads]; - - // Compute oracles and count appearances. - for (_BinIndex __b = 0; __b < _M_sd->_M_num_bins + 1; ++__b) - _M_dist[__b] = 0; - int _M_num_bits = _M_sd->_M_num_bits; - - _RandomNumber __rng(d->_M_seed); - - // First main loop. - for (_DifferenceType __i = 0; __i < __length; ++__i) - { - _BinIndex __oracle = __random_number_pow2(_M_num_bits, __rng); - __oracles[__i] = __oracle; + /** @brief Type to hold the index of a bin. + * + * Since many variables of this type are allocated, it should be + * chosen as small as possible. + */ + typedef unsigned short _BinIndex; + + /** @brief Data known to every thread participating in + __gnu_parallel::__parallel_random_shuffle(). */ + template + struct _DRandomShufflingGlobalData + { + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::value_type _ValueType; + typedef typename _TraitsType::difference_type _DifferenceType; - // To allow prefix (partial) sum. - ++(_M_dist[__oracle + 1]); - } + /** @brief Begin iterator of the __source. */ + _RAIter& _M_source; - for (_BinIndex __b = 0; __b < _M_sd->_M_num_bins + 1; ++__b) - _M_sd->_M_dist[__b][__iam + 1] = _M_dist[__b]; + /** @brief Temporary arrays for each thread. */ + _ValueType** _M_temporaries; -# pragma omp barrier + /** @brief Two-dimensional array to hold the thread-bin distribution. + * + * Dimensions (_M_num_threads + 1) __x (_M_num_bins + 1). */ + _DifferenceType** _M_dist; -# pragma omp single - { - // Sum up bins, _M_sd->_M_dist[__s + 1][d->_M_num_threads] now contains - // the total number of items in bin __s - for (_BinIndex __s = 0; __s < _M_sd->_M_num_bins; ++__s) - __gnu_sequential::partial_sum( - _M_sd->_M_dist[__s + 1], - _M_sd->_M_dist[__s + 1] + d->_M_num_threads + 1, - _M_sd->_M_dist[__s + 1]); - } + /** @brief Start indexes of the threads' __chunks. */ + _DifferenceType* _M_starts; -# pragma omp barrier + /** @brief Number of the thread that will further process the + corresponding bin. */ + _ThreadIndex* _M_bin_proc; - _SequenceIndex __offset = 0, __global_offset = 0; - for (_BinIndex __s = 0; __s < d->_M_bins_begin; ++__s) - __global_offset += _M_sd->_M_dist[__s + 1][d->_M_num_threads]; + /** @brief Number of bins to distribute to. */ + int _M_num_bins; -# pragma omp barrier + /** @brief Number of bits needed to address the bins. */ + int _M_num_bits; - for (_BinIndex __s = d->_M_bins_begin; __s < d->__bins_end; ++__s) - { - for (int __t = 0; __t < d->_M_num_threads + 1; ++__t) - _M_sd->_M_dist[__s + 1][__t] += __offset; - __offset = _M_sd->_M_dist[__s + 1][d->_M_num_threads]; - } + /** @brief Constructor. */ + _DRandomShufflingGlobalData(_RAIter& __source) + : _M_source(__source) { } + }; + + /** @brief Local data for a thread participating in + __gnu_parallel::__parallel_random_shuffle(). + */ + template + struct _DRSSorterPU + { + /** @brief Number of threads participating in total. */ + int _M_num_threads; + + /** @brief Begin index for bins taken care of by this thread. */ + _BinIndex _M_bins_begin; + + /** @brief End index for bins taken care of by this thread. */ + _BinIndex __bins_end; + + /** @brief Random _M_seed for this thread. */ + uint32_t _M_seed; + + /** @brief Pointer to global data. */ + _DRandomShufflingGlobalData<_RAIter>* _M_sd; + }; + + /** @brief Generate a random number in @__c [0,2^__logp). + * @param __logp Logarithm (basis 2) of the upper range __bound. + * @param __rng Random number generator to use. + */ + template + inline int + __random_number_pow2(int __logp, _RandomNumberGenerator& __rng) + { return __rng.__genrand_bits(__logp); } + + /** @brief Random shuffle code executed by each thread. + * @param __pus Array of thread-local data records. */ + template + void + __parallel_random_shuffle_drs_pu(_DRSSorterPU<_RAIter, + _RandomNumberGenerator>* __pus) + { + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::value_type _ValueType; + typedef typename _TraitsType::difference_type _DifferenceType; - _M_sd->_M_temporaries[__iam] = static_cast<_ValueType*>( - ::operator new(sizeof(_ValueType) * __offset)); + _ThreadIndex __iam = omp_get_thread_num(); + _DRSSorterPU<_RAIter, _RandomNumberGenerator>* __d = &__pus[__iam]; + _DRandomShufflingGlobalData<_RAIter>* __sd = __d->_M_sd; -# pragma omp barrier + // Indexing: _M_dist[bin][processor] + _DifferenceType __length = (__sd->_M_starts[__iam + 1] + - __sd->_M_starts[__iam]); + _BinIndex* __oracles = new _BinIndex[__length]; + _DifferenceType* __dist = new _DifferenceType[__sd->_M_num_bins + 1]; + _BinIndex* __bin_proc = new _BinIndex[__sd->_M_num_bins]; + _ValueType** __temporaries = new _ValueType*[__d->_M_num_threads]; - // Draw local copies to avoid false sharing. - for (_BinIndex __b = 0; __b < _M_sd->_M_num_bins + 1; ++__b) - _M_dist[__b] = _M_sd->_M_dist[__b][__iam]; - for (_BinIndex __b = 0; __b < _M_sd->_M_num_bins; ++__b) - _M_bin_proc[__b] = _M_sd->_M_bin_proc[__b]; - for (_ThreadIndex __t = 0; __t < d->_M_num_threads; ++__t) - _M_temporaries[__t] = _M_sd->_M_temporaries[__t]; + // Compute oracles and count appearances. + for (_BinIndex __b = 0; __b < __sd->_M_num_bins + 1; ++__b) + __dist[__b] = 0; + int __num_bits = __sd->_M_num_bits; - _RAIter _M_source = _M_sd->_M_source; - _DifferenceType __start = _M_sd->_M_starts[__iam]; + _RandomNumber __rng(__d->_M_seed); - // Distribute according to oracles, second main loop. - for (_DifferenceType __i = 0; __i < __length; ++__i) - { - _BinIndex target_bin = __oracles[__i]; - _ThreadIndex target_p = _M_bin_proc[target_bin]; + // First main loop. + for (_DifferenceType __i = 0; __i < __length; ++__i) + { + _BinIndex __oracle = __random_number_pow2(__num_bits, __rng); + __oracles[__i] = __oracle; - // Last column [d->_M_num_threads] stays unchanged. - ::new(&(_M_temporaries[target_p][_M_dist[target_bin + 1]++])) - _ValueType(*(_M_source + __i + __start)); - } + // To allow prefix (partial) sum. + ++(__dist[__oracle + 1]); + } - delete[] __oracles; - delete[] _M_dist; - delete[] _M_bin_proc; - delete[] _M_temporaries; + for (_BinIndex __b = 0; __b < __sd->_M_num_bins + 1; ++__b) + __sd->_M_dist[__b][__iam + 1] = __dist[__b]; -# pragma omp barrier +# pragma omp barrier - // Shuffle bins internally. - for (_BinIndex __b = d->_M_bins_begin; __b < d->__bins_end; ++__b) +# pragma omp single { - _ValueType* __begin = - _M_sd->_M_temporaries[__iam] + - ((__b == d->_M_bins_begin) - ? 0 : _M_sd->_M_dist[__b][d->_M_num_threads]), - * __end = - _M_sd->_M_temporaries[__iam] + - _M_sd->_M_dist[__b + 1][d->_M_num_threads]; - __sequential_random_shuffle(__begin, __end, __rng); - std::copy(__begin, __end, _M_sd->_M_source + __global_offset + - ((__b == d->_M_bins_begin) - ? 0 : _M_sd->_M_dist[__b][d->_M_num_threads])); + // Sum up bins, __sd->_M_dist[__s + 1][__d->_M_num_threads] now + // contains the total number of items in bin __s + for (_BinIndex __s = 0; __s < __sd->_M_num_bins; ++__s) + __gnu_sequential::partial_sum(__sd->_M_dist[__s + 1], + __sd->_M_dist[__s + 1] + + __d->_M_num_threads + 1, + __sd->_M_dist[__s + 1]); } - ::operator delete(_M_sd->_M_temporaries[__iam]); - } - -/** @brief Round up to the next greater power of 2. - * @param __x _Integer to round up */ -template - _Tp - __round_up_to_pow2(_Tp __x) - { - if (__x <= 1) - return 1; - else - return (_Tp)1 << (__rd_log2(__x - 1) + 1); - } - -/** @brief Main parallel random shuffle step. - * @param __begin Begin iterator of sequence. - * @param __end End iterator of sequence. - * @param __n Length of sequence. - * @param __num_threads Number of threads to use. - * @param __rng Random number generator to use. - */ -template - void - __parallel_random_shuffle_drs(_RAIter __begin, - _RAIter __end, - typename std::iterator_traits - <_RAIter>::difference_type __n, - _ThreadIndex __num_threads, - _RandomNumberGenerator& __rng) - { - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::value_type _ValueType; - typedef typename _TraitsType::difference_type _DifferenceType; - - _GLIBCXX_CALL(__n) - - const _Settings& __s = _Settings::get(); - - if (__num_threads > __n) - __num_threads = static_cast<_ThreadIndex>(__n); - - _BinIndex _M_num_bins, __num_bins_cache; +# pragma omp barrier + + _SequenceIndex __offset = 0, __global_offset = 0; + for (_BinIndex __s = 0; __s < __d->_M_bins_begin; ++__s) + __global_offset += __sd->_M_dist[__s + 1][__d->_M_num_threads]; + +# pragma omp barrier + + for (_BinIndex __s = __d->_M_bins_begin; __s < __d->__bins_end; ++__s) + { + for (int __t = 0; __t < __d->_M_num_threads + 1; ++__t) + __sd->_M_dist[__s + 1][__t] += __offset; + __offset = __sd->_M_dist[__s + 1][__d->_M_num_threads]; + } + + __sd->_M_temporaries[__iam] = static_cast<_ValueType*> + (::operator new(sizeof(_ValueType) * __offset)); + +# pragma omp barrier + + // Draw local copies to avoid false sharing. + for (_BinIndex __b = 0; __b < __sd->_M_num_bins + 1; ++__b) + __dist[__b] = __sd->_M_dist[__b][__iam]; + for (_BinIndex __b = 0; __b < __sd->_M_num_bins; ++__b) + __bin_proc[__b] = __sd->_M_bin_proc[__b]; + for (_ThreadIndex __t = 0; __t < __d->_M_num_threads; ++__t) + __temporaries[__t] = __sd->_M_temporaries[__t]; + + _RAIter __source = __sd->_M_source; + _DifferenceType __start = __sd->_M_starts[__iam]; + + // Distribute according to oracles, second main loop. + for (_DifferenceType __i = 0; __i < __length; ++__i) + { + _BinIndex __target_bin = __oracles[__i]; + _ThreadIndex __target_p = __bin_proc[__target_bin]; + + // Last column [__d->_M_num_threads] stays unchanged. + ::new(&(__temporaries[__target_p][__dist[__target_bin + 1]++])) + _ValueType(*(__source + __i + __start)); + } + + delete[] __oracles; + delete[] __dist; + delete[] __bin_proc; + delete[] __temporaries; + +# pragma omp barrier + + // Shuffle bins internally. + for (_BinIndex __b = __d->_M_bins_begin; __b < __d->__bins_end; ++__b) + { + _ValueType* __begin = + (__sd->_M_temporaries[__iam] + + (__b == __d->_M_bins_begin + ? 0 : __sd->_M_dist[__b][__d->_M_num_threads])), + * __end = (__sd->_M_temporaries[__iam] + + __sd->_M_dist[__b + 1][__d->_M_num_threads]); + + __sequential_random_shuffle(__begin, __end, __rng); + std::copy(__begin, __end, __sd->_M_source + __global_offset + + (__b == __d->_M_bins_begin + ? 0 : __sd->_M_dist[__b][__d->_M_num_threads])); + } + + ::operator delete(__sd->_M_temporaries[__iam]); + } + + /** @brief Round up to the next greater power of 2. + * @param __x _Integer to round up */ + template + _Tp + __round_up_to_pow2(_Tp __x) + { + if (__x <= 1) + return 1; + else + return (_Tp)1 << (__rd_log2(__x - 1) + 1); + } + + /** @brief Main parallel random shuffle step. + * @param __begin Begin iterator of sequence. + * @param __end End iterator of sequence. + * @param __n Length of sequence. + * @param __num_threads Number of threads to use. + * @param __rng Random number generator to use. + */ + template + void + __parallel_random_shuffle_drs(_RAIter __begin, _RAIter __end, + typename std::iterator_traits + <_RAIter>::difference_type __n, + _ThreadIndex __num_threads, + _RandomNumberGenerator& __rng) + { + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::value_type _ValueType; + typedef typename _TraitsType::difference_type _DifferenceType; + + _GLIBCXX_CALL(__n) + + const _Settings& __s = _Settings::get(); + + if (__num_threads > __n) + __num_threads = static_cast<_ThreadIndex>(__n); + + _BinIndex __num_bins, __num_bins_cache; #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1 - // Try the L1 cache first. + // Try the L1 cache first. - // Must fit into L1. - __num_bins_cache = std::max<_DifferenceType>( - 1, __n / (__s.L1_cache_size_lb / sizeof(_ValueType))); - __num_bins_cache = __round_up_to_pow2(__num_bins_cache); + // Must fit into L1. + __num_bins_cache = + std::max<_DifferenceType>(1, __n / (__s.L1_cache_size_lb + / sizeof(_ValueType))); + __num_bins_cache = __round_up_to_pow2(__num_bins_cache); - // No more buckets than TLB entries, power of 2 - // Power of 2 and at least one element per bin, at most the TLB size. - _M_num_bins = std::min<_DifferenceType>(__n, __num_bins_cache); + // No more buckets than TLB entries, power of 2 + // Power of 2 and at least one element per bin, at most the TLB size. + __num_bins = std::min<_DifferenceType>(__n, __num_bins_cache); #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB - // 2 TLB entries needed per bin. - _M_num_bins = std::min<_DifferenceType>(__s.TLB_size / 2, _M_num_bins); + // 2 TLB entries needed per bin. + __num_bins = std::min<_DifferenceType>(__s.TLB_size / 2, __num_bins); #endif - _M_num_bins = __round_up_to_pow2(_M_num_bins); + __num_bins = __round_up_to_pow2(__num_bins); - if (_M_num_bins < __num_bins_cache) - { + if (__num_bins < __num_bins_cache) + { #endif - // Now try the L2 cache - // Must fit into L2 - __num_bins_cache = static_cast<_BinIndex>(std::max<_DifferenceType>( - 1, __n / (__s.L2_cache_size / sizeof(_ValueType)))); - __num_bins_cache = __round_up_to_pow2(__num_bins_cache); - - // No more buckets than TLB entries, power of 2. - _M_num_bins = static_cast<_BinIndex>( - std::min(__n, static_cast<_DifferenceType>(__num_bins_cache))); - // Power of 2 and at least one element per bin, at most the TLB size. + // Now try the L2 cache + // Must fit into L2 + __num_bins_cache = static_cast<_BinIndex> + (std::max<_DifferenceType>(1, __n / (__s.L2_cache_size + / sizeof(_ValueType)))); + __num_bins_cache = __round_up_to_pow2(__num_bins_cache); + + // No more buckets than TLB entries, power of 2. + __num_bins = static_cast<_BinIndex> + (std::min(__n, static_cast<_DifferenceType>(__num_bins_cache))); + // Power of 2 and at least one element per bin, at most the TLB size. #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB - // 2 TLB entries needed per bin. - _M_num_bins = std::min( - static_cast<_DifferenceType>(__s.TLB_size / 2), _M_num_bins); + // 2 TLB entries needed per bin. + __num_bins = std::min(static_cast<_DifferenceType>(__s.TLB_size / 2), + __num_bins); #endif - _M_num_bins = __round_up_to_pow2(_M_num_bins); + __num_bins = __round_up_to_pow2(__num_bins); #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1 - } + } #endif - __num_threads = std::min<_BinIndex>(__num_threads, _M_num_bins); + __num_threads = std::min<_BinIndex>(__num_threads, __num_bins); - if (__num_threads <= 1) - return __sequential_random_shuffle(__begin, __end, __rng); + if (__num_threads <= 1) + return __sequential_random_shuffle(__begin, __end, __rng); - _DRandomShufflingGlobalData<_RAIter> _M_sd(__begin); - _DRSSorterPU<_RAIter, _RandomNumber >* __pus; - _DifferenceType* _M_starts; + _DRandomShufflingGlobalData<_RAIter> __sd(__begin); + _DRSSorterPU<_RAIter, _RandomNumber >* __pus; + _DifferenceType* __starts; -# pragma omp parallel num_threads(__num_threads) +# pragma omp parallel num_threads(__num_threads) { - _ThreadIndex __num_threads = omp_get_num_threads(); + _ThreadIndex __num_threads = omp_get_num_threads(); # pragma omp single - { - __pus = new _DRSSorterPU<_RAIter, _RandomNumber> - [__num_threads]; - - _M_sd._M_temporaries = new _ValueType*[__num_threads]; - _M_sd._M_dist = new _DifferenceType*[_M_num_bins + 1]; - _M_sd._M_bin_proc = new _ThreadIndex[_M_num_bins]; - for (_BinIndex __b = 0; __b < _M_num_bins + 1; ++__b) - _M_sd._M_dist[__b] = new _DifferenceType[__num_threads + 1]; - for (_BinIndex __b = 0; __b < (_M_num_bins + 1); ++__b) - { - _M_sd._M_dist[0][0] = 0; - _M_sd._M_dist[__b][0] = 0; - } - _M_starts = _M_sd._M_starts - = new _DifferenceType[__num_threads + 1]; - int bin_cursor = 0; - _M_sd._M_num_bins = _M_num_bins; - _M_sd._M_num_bits = __rd_log2(_M_num_bins); - - _DifferenceType __chunk_length = __n / __num_threads, - __split = __n % __num_threads, __start = 0; - _DifferenceType bin_chunk_length = _M_num_bins / __num_threads, - bin_split = _M_num_bins % __num_threads; - for (_ThreadIndex __i = 0; __i < __num_threads; ++__i) - { - _M_starts[__i] = __start; - __start += (__i < __split) - ? (__chunk_length + 1) : __chunk_length; - int __j = __pus[__i]._M_bins_begin = bin_cursor; - - // Range of bins for this processor. - bin_cursor += (__i < bin_split) ? - (bin_chunk_length + 1) : bin_chunk_length; - __pus[__i].__bins_end = bin_cursor; - for (; __j < bin_cursor; ++__j) - _M_sd._M_bin_proc[__j] = __i; - __pus[__i]._M_num_threads = __num_threads; - __pus[__i]._M_seed = - __rng(std::numeric_limits::max()); - __pus[__i]._M_sd = &_M_sd; - } - _M_starts[__num_threads] = __start; - } //single - // Now shuffle in parallel. - __parallel_random_shuffle_drs_pu(__pus); + { + __pus = new _DRSSorterPU<_RAIter, _RandomNumber>[__num_threads]; + + __sd._M_temporaries = new _ValueType*[__num_threads]; + __sd._M_dist = new _DifferenceType*[__num_bins + 1]; + __sd._M_bin_proc = new _ThreadIndex[__num_bins]; + for (_BinIndex __b = 0; __b < __num_bins + 1; ++__b) + __sd._M_dist[__b] = new _DifferenceType[__num_threads + 1]; + for (_BinIndex __b = 0; __b < (__num_bins + 1); ++__b) + { + __sd._M_dist[0][0] = 0; + __sd._M_dist[__b][0] = 0; + } + __starts = __sd._M_starts = new _DifferenceType[__num_threads + 1]; + int __bin_cursor = 0; + __sd._M_num_bins = __num_bins; + __sd._M_num_bits = __rd_log2(__num_bins); + + _DifferenceType __chunk_length = __n / __num_threads, + __split = __n % __num_threads, + __start = 0; + _DifferenceType __bin_chunk_length = __num_bins / __num_threads, + __bin_split = __num_bins % __num_threads; + for (_ThreadIndex __i = 0; __i < __num_threads; ++__i) + { + __starts[__i] = __start; + __start += (__i < __split + ? (__chunk_length + 1) : __chunk_length); + int __j = __pus[__i]._M_bins_begin = __bin_cursor; + + // Range of bins for this processor. + __bin_cursor += (__i < __bin_split + ? (__bin_chunk_length + 1) + : __bin_chunk_length); + __pus[__i].__bins_end = __bin_cursor; + for (; __j < __bin_cursor; ++__j) + __sd._M_bin_proc[__j] = __i; + __pus[__i]._M_num_threads = __num_threads; + __pus[__i]._M_seed = __rng(std::numeric_limits::max()); + __pus[__i]._M_sd = &__sd; + } + __starts[__num_threads] = __start; + } //single + // Now shuffle in parallel. + __parallel_random_shuffle_drs_pu(__pus); } // parallel - delete[] _M_starts; - delete[] _M_sd._M_bin_proc; - for (int __s = 0; __s < (_M_num_bins + 1); ++__s) - delete[] _M_sd._M_dist[__s]; - delete[] _M_sd._M_dist; - delete[] _M_sd._M_temporaries; + delete[] __starts; + delete[] __sd._M_bin_proc; + for (int __s = 0; __s < (__num_bins + 1); ++__s) + delete[] __sd._M_dist[__s]; + delete[] __sd._M_dist; + delete[] __sd._M_temporaries; - delete[] __pus; - } + delete[] __pus; + } -/** @brief Sequential cache-efficient random shuffle. - * @param __begin Begin iterator of sequence. - * @param __end End iterator of sequence. - * @param __rng Random number generator to use. - */ -template - void - __sequential_random_shuffle(_RAIter __begin, - _RAIter __end, - _RandomNumberGenerator& __rng) - { - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::value_type _ValueType; - typedef typename _TraitsType::difference_type _DifferenceType; + /** @brief Sequential cache-efficient random shuffle. + * @param __begin Begin iterator of sequence. + * @param __end End iterator of sequence. + * @param __rng Random number generator to use. + */ + template + void + __sequential_random_shuffle(_RAIter __begin, _RAIter __end, + _RandomNumberGenerator& __rng) + { + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::value_type _ValueType; + typedef typename _TraitsType::difference_type _DifferenceType; - _DifferenceType __n = __end - __begin; - const _Settings& __s = _Settings::get(); + _DifferenceType __n = __end - __begin; + const _Settings& __s = _Settings::get(); - _BinIndex _M_num_bins, __num_bins_cache; + _BinIndex __num_bins, __num_bins_cache; #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1 - // Try the L1 cache first, must fit into L1. - __num_bins_cache = - std::max<_DifferenceType> - (1, __n / (__s.L1_cache_size_lb / sizeof(_ValueType))); - __num_bins_cache = __round_up_to_pow2(__num_bins_cache); - - // No more buckets than TLB entries, power of 2 - // Power of 2 and at least one element per bin, at most the TLB size - _M_num_bins = std::min(__n, (_DifferenceType)__num_bins_cache); + // Try the L1 cache first, must fit into L1. + __num_bins_cache = std::max<_DifferenceType> + (1, __n / (__s.L1_cache_size_lb / sizeof(_ValueType))); + __num_bins_cache = __round_up_to_pow2(__num_bins_cache); + + // No more buckets than TLB entries, power of 2 + // Power of 2 and at least one element per bin, at most the TLB size + __num_bins = std::min(__n, (_DifferenceType)__num_bins_cache); #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB - // 2 TLB entries needed per bin - _M_num_bins = std::min((_DifferenceType)__s.TLB_size / 2, _M_num_bins); + // 2 TLB entries needed per bin + __num_bins = std::min((_DifferenceType)__s.TLB_size / 2, __num_bins); #endif - _M_num_bins = __round_up_to_pow2(_M_num_bins); + __num_bins = __round_up_to_pow2(__num_bins); - if (_M_num_bins < __num_bins_cache) - { + if (__num_bins < __num_bins_cache) + { #endif - // Now try the L2 cache, must fit into L2. - __num_bins_cache = - static_cast<_BinIndex>(std::max<_DifferenceType>( - 1, __n / (__s.L2_cache_size / sizeof(_ValueType)))); - __num_bins_cache = __round_up_to_pow2(__num_bins_cache); + // Now try the L2 cache, must fit into L2. + __num_bins_cache = static_cast<_BinIndex> + (std::max<_DifferenceType>(1, __n / (__s.L2_cache_size + / sizeof(_ValueType)))); + __num_bins_cache = __round_up_to_pow2(__num_bins_cache); - // No more buckets than TLB entries, power of 2 - // Power of 2 and at least one element per bin, at most the TLB size. - _M_num_bins = static_cast<_BinIndex> - (std::min(__n, static_cast<_DifferenceType>(__num_bins_cache))); + // No more buckets than TLB entries, power of 2 + // Power of 2 and at least one element per bin, at most the TLB size. + __num_bins = static_cast<_BinIndex> + (std::min(__n, static_cast<_DifferenceType>(__num_bins_cache))); #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB - // 2 TLB entries needed per bin - _M_num_bins = - std::min<_DifferenceType>(__s.TLB_size / 2, _M_num_bins); + // 2 TLB entries needed per bin + __num_bins = std::min<_DifferenceType>(__s.TLB_size / 2, __num_bins); #endif - _M_num_bins = __round_up_to_pow2(_M_num_bins); + __num_bins = __round_up_to_pow2(__num_bins); #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1 - } + } #endif - int _M_num_bits = __rd_log2(_M_num_bins); + int __num_bits = __rd_log2(__num_bins); - if (_M_num_bins > 1) - { - _ValueType* __target = static_cast<_ValueType*>( - ::operator new(sizeof(_ValueType) * __n)); - _BinIndex* __oracles = new _BinIndex[__n]; - _DifferenceType* __dist0 = new _DifferenceType[_M_num_bins + 1], - * __dist1 = new _DifferenceType[_M_num_bins + 1]; - - for (int __b = 0; __b < _M_num_bins + 1; ++__b) - __dist0[__b] = 0; - - _RandomNumber bitrng(__rng(0xFFFFFFFF)); - - for (_DifferenceType __i = 0; __i < __n; ++__i) - { - _BinIndex __oracle = __random_number_pow2(_M_num_bits, bitrng); - __oracles[__i] = __oracle; - - // To allow prefix (partial) sum. - ++(__dist0[__oracle + 1]); - } - - // Sum up bins. - __gnu_sequential:: - partial_sum(__dist0, __dist0 + _M_num_bins + 1, __dist0); - - for (int __b = 0; __b < _M_num_bins + 1; ++__b) - __dist1[__b] = __dist0[__b]; - - // Distribute according to oracles. - for (_DifferenceType __i = 0; __i < __n; ++__i) - ::new(&(__target[(__dist0[__oracles[__i]])++])) - _ValueType(*(__begin + __i)); - - for (int __b = 0; __b < _M_num_bins; ++__b) - { - __sequential_random_shuffle(__target + __dist1[__b], - __target + __dist1[__b + 1], - __rng); - } - - // Copy elements back. - std::copy(__target, __target + __n, __begin); - - delete[] __dist0; - delete[] __dist1; - delete[] __oracles; - ::operator delete(__target); - } - else - __gnu_sequential::random_shuffle(__begin, __end, __rng); - } - -/** @brief Parallel random public call. - * @param __begin Begin iterator of sequence. - * @param __end End iterator of sequence. - * @param __rng Random number generator to use. - */ -template - inline void - __parallel_random_shuffle(_RAIter __begin, - _RAIter __end, - _RandomNumberGenerator __rng = _RandomNumber()) - { - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::difference_type _DifferenceType; - _DifferenceType __n = __end - __begin; - __parallel_random_shuffle_drs( - __begin, __end, __n, __get_max_threads(), __rng) ; - } + if (__num_bins > 1) + { + _ValueType* __target = + static_cast<_ValueType*>(::operator new(sizeof(_ValueType) * __n)); + _BinIndex* __oracles = new _BinIndex[__n]; + _DifferenceType* __dist0 = new _DifferenceType[__num_bins + 1], + * __dist1 = new _DifferenceType[__num_bins + 1]; + + for (int __b = 0; __b < __num_bins + 1; ++__b) + __dist0[__b] = 0; + + _RandomNumber __bitrng(__rng(0xFFFFFFFF)); + for (_DifferenceType __i = 0; __i < __n; ++__i) + { + _BinIndex __oracle = __random_number_pow2(__num_bits, __bitrng); + __oracles[__i] = __oracle; + + // To allow prefix (partial) sum. + ++(__dist0[__oracle + 1]); + } + + // Sum up bins. + __gnu_sequential::partial_sum(__dist0, __dist0 + __num_bins + 1, + __dist0); + + for (int __b = 0; __b < __num_bins + 1; ++__b) + __dist1[__b] = __dist0[__b]; + + // Distribute according to oracles. + for (_DifferenceType __i = 0; __i < __n; ++__i) + ::new(&(__target[(__dist0[__oracles[__i]])++])) + _ValueType(*(__begin + __i)); + + for (int __b = 0; __b < __num_bins; ++__b) + __sequential_random_shuffle(__target + __dist1[__b], + __target + __dist1[__b + 1], __rng); + + // Copy elements back. + std::copy(__target, __target + __n, __begin); + + delete[] __dist0; + delete[] __dist1; + delete[] __oracles; + ::operator delete(__target); + } + else + __gnu_sequential::random_shuffle(__begin, __end, __rng); + } + + /** @brief Parallel random public call. + * @param __begin Begin iterator of sequence. + * @param __end End iterator of sequence. + * @param __rng Random number generator to use. + */ + template + inline void + __parallel_random_shuffle(_RAIter __begin, _RAIter __end, + _RandomNumberGenerator __rng = _RandomNumber()) + { + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::difference_type _DifferenceType; + _DifferenceType __n = __end - __begin; + __parallel_random_shuffle_drs(__begin, __end, __n, + __get_max_threads(), __rng); + } } #endif /* _GLIBCXX_PARALLEL_RANDOM_SHUFFLE_H */ diff --git a/libstdc++-v3/include/parallel/search.h b/libstdc++-v3/include/parallel/search.h index 1fc8ceaa5c8..9709925618d 100644 --- a/libstdc++-v3/include/parallel/search.h +++ b/libstdc++-v3/include/parallel/search.h @@ -38,7 +38,6 @@ #include #include - namespace __gnu_parallel { /** @@ -47,24 +46,24 @@ namespace __gnu_parallel * @param __length Length of sequence to search for. * @param __advances Returned __offsets. */ -template - void - __calc_borders(_RAIter __elements, _DifferenceTp __length, - _DifferenceTp* __off) - { - typedef _DifferenceTp _DifferenceType; - - __off[0] = -1; - if (__length > 1) - __off[1] = 0; - _DifferenceType __k = 0; - for (_DifferenceType __j = 2; __j <= __length; __j++) - { - while ((__k >= 0) && !(__elements[__k] == __elements[__j-1])) - __k = __off[__k]; - __off[__j] = ++__k; - } - } + template + void + __calc_borders(_RAIter __elements, _DifferenceTp __length, + _DifferenceTp* __off) + { + typedef _DifferenceTp _DifferenceType; + + __off[0] = -1; + if (__length > 1) + __off[1] = 0; + _DifferenceType __k = 0; + for (_DifferenceType __j = 2; __j <= __length; __j++) + { + while ((__k >= 0) && !(__elements[__k] == __elements[__j-1])) + __k = __off[__k]; + __off[__j] = ++__k; + } + } // Generic parallel find algorithm (requires random access iterator). @@ -75,100 +74,99 @@ template * @param __end2 End iterator of second sequence. * @param __pred Find predicate. * @return Place of finding in first sequences. */ -template - __RAIter1 - __search_template(__RAIter1 __begin1, __RAIter1 __end1, - __RAIter2 __begin2, __RAIter2 __end2, - _Pred __pred) - { - typedef std::iterator_traits<__RAIter1> _TraitsType; - typedef typename _TraitsType::difference_type _DifferenceType; + template + __RAIter1 + __search_template(__RAIter1 __begin1, __RAIter1 __end1, + __RAIter2 __begin2, __RAIter2 __end2, + _Pred __pred) + { + typedef std::iterator_traits<__RAIter1> _TraitsType; + typedef typename _TraitsType::difference_type _DifferenceType; - _GLIBCXX_CALL((__end1 - __begin1) + (__end2 - __begin2)); + _GLIBCXX_CALL((__end1 - __begin1) + (__end2 - __begin2)); - _DifferenceType __pattern_length = __end2 - __begin2; + _DifferenceType __pattern_length = __end2 - __begin2; - // Pattern too short. - if(__pattern_length <= 0) - return __end1; + // Pattern too short. + if(__pattern_length <= 0) + return __end1; - // Last point to start search. - _DifferenceType __input_length = (__end1 - __begin1) - __pattern_length; + // Last point to start search. + _DifferenceType __input_length = (__end1 - __begin1) - __pattern_length; - // Where is first occurrence of pattern? defaults to end. - _DifferenceType __result = (__end1 - __begin1); - _DifferenceType *__splitters; + // Where is first occurrence of pattern? defaults to end. + _DifferenceType __result = (__end1 - __begin1); + _DifferenceType *__splitters; - // Pattern too long. - if (__input_length < 0) - return __end1; + // Pattern too long. + if (__input_length < 0) + return __end1; - omp_lock_t __result_lock; - omp_init_lock(&__result_lock); + omp_lock_t __result_lock; + omp_init_lock(&__result_lock); - _ThreadIndex __num_threads = - std::max<_DifferenceType>(1, - std::min<_DifferenceType>(__input_length, __get_max_threads())); + _ThreadIndex __num_threads = std::max<_DifferenceType> + (1, std::min<_DifferenceType>(__input_length, + __get_max_threads())); - _DifferenceType __advances[__pattern_length]; - __calc_borders(__begin2, __pattern_length, __advances); + _DifferenceType __advances[__pattern_length]; + __calc_borders(__begin2, __pattern_length, __advances); -# pragma omp parallel num_threads(__num_threads) +# pragma omp parallel num_threads(__num_threads) { # pragma omp single - { - __num_threads = omp_get_num_threads(); - __splitters = new _DifferenceType[__num_threads + 1]; - equally_split(__input_length, __num_threads, __splitters); - } - - _ThreadIndex __iam = omp_get_thread_num(); - - _DifferenceType __start = __splitters[__iam], - __stop = __splitters[__iam + 1]; - - _DifferenceType __pos_in_pattern = 0; - bool __found_pattern = false; - - while (__start <= __stop && !__found_pattern) - { - // Get new value of result. - #pragma omp flush(__result) - // No chance for this thread to find first occurrence. - if (__result < __start) - break; - while (__pred(__begin1[__start + __pos_in_pattern], - __begin2[__pos_in_pattern])) - { - ++__pos_in_pattern; - if (__pos_in_pattern == __pattern_length) - { - // Found new candidate for result. - omp_set_lock(&__result_lock); - __result = std::min(__result, __start); - omp_unset_lock(&__result_lock); - - __found_pattern = true; - break; - } - } - // Make safe jump. - __start += (__pos_in_pattern - __advances[__pos_in_pattern]); - __pos_in_pattern = - (__advances[__pos_in_pattern] < 0) ? - 0 : __advances[__pos_in_pattern]; - } + { + __num_threads = omp_get_num_threads(); + __splitters = new _DifferenceType[__num_threads + 1]; + equally_split(__input_length, __num_threads, __splitters); + } + + _ThreadIndex __iam = omp_get_thread_num(); + + _DifferenceType __start = __splitters[__iam], + __stop = __splitters[__iam + 1]; + + _DifferenceType __pos_in_pattern = 0; + bool __found_pattern = false; + + while (__start <= __stop && !__found_pattern) + { + // Get new value of result. +#pragma omp flush(__result) + // No chance for this thread to find first occurrence. + if (__result < __start) + break; + while (__pred(__begin1[__start + __pos_in_pattern], + __begin2[__pos_in_pattern])) + { + ++__pos_in_pattern; + if (__pos_in_pattern == __pattern_length) + { + // Found new candidate for result. + omp_set_lock(&__result_lock); + __result = std::min(__result, __start); + omp_unset_lock(&__result_lock); + + __found_pattern = true; + break; + } + } + // Make safe jump. + __start += (__pos_in_pattern - __advances[__pos_in_pattern]); + __pos_in_pattern = (__advances[__pos_in_pattern] < 0 + ? 0 : __advances[__pos_in_pattern]); + } } //parallel - omp_destroy_lock(&__result_lock); - - delete[] __splitters; + omp_destroy_lock(&__result_lock); - // Return iterator on found element. - return (__begin1 + __result); - } + delete[] __splitters; + + // Return iterator on found element. + return (__begin1 + __result); + } } // end namespace #endif /* _GLIBCXX_PARALLEL_SEARCH_H */ diff --git a/libstdc++-v3/include/parallel/set_operations.h b/libstdc++-v3/include/parallel/set_operations.h index ac669c55d5d..6dd63c9b128 100644 --- a/libstdc++-v3/include/parallel/set_operations.h +++ b/libstdc++-v3/include/parallel/set_operations.h @@ -41,490 +41,489 @@ namespace __gnu_parallel { -template - _OutputIterator - copy_tail(std::pair<_IIter, _IIter> __b, - std::pair<_IIter, _IIter> __e, _OutputIterator __r) - { - if (__b.first != __e.first) + template + _OutputIterator + __copy_tail(std::pair<_IIter, _IIter> __b, + std::pair<_IIter, _IIter> __e, _OutputIterator __r) + { + if (__b.first != __e.first) + { + do + { + *__r++ = *__b.first++; + } + while (__b.first != __e.first); + } + else + { + while (__b.second != __e.second) + *__r++ = *__b.second++; + } + return __r; + } + + template + struct __symmetric_difference_func + { + typedef std::iterator_traits<_IIter> _TraitsType; + typedef typename _TraitsType::difference_type _DifferenceType; + typedef typename std::pair<_IIter, _IIter> _IteratorPair; + + __symmetric_difference_func(_Compare __comp) : _M_comp(__comp) {} + + _Compare _M_comp; + + _OutputIterator + _M_invoke(_IIter __a, _IIter __b, _IIter __c, _IIter __d, + _OutputIterator __r) const { - do + while (__a != __b && __c != __d) { - *__r++ = *__b.first++; + if (_M_comp(*__a, *__c)) + { + *__r = *__a; + ++__a; + ++__r; + } + else if (_M_comp(*__c, *__a)) + { + *__r = *__c; + ++__c; + ++__r; + } + else + { + ++__a; + ++__c; + } } - while (__b.first != __e.first); + return std::copy(__c, __d, std::copy(__a, __b, __r)); } - else + + _DifferenceType + __count(_IIter __a, _IIter __b, _IIter __c, _IIter d) const { - while (__b.second != __e.second) - *__r++ = *__b.second++; + _DifferenceType __counter = 0; + + while (__a != __b && __c != d) + { + if (_M_comp(*__a, *__c)) + { + ++__a; + ++__counter; + } + else if (_M_comp(*__c, *__a)) + { + ++__c; + ++__counter; + } + else + { + ++__a; + ++__c; + } + } + + return __counter + (__b - __a) + (d - __c); } - return __r; - } -template - struct symmetric_difference_func - { - typedef std::iterator_traits<_IIter> _TraitsType; - typedef typename _TraitsType::difference_type _DifferenceType; - typedef typename std::pair<_IIter, _IIter> _IteratorPair; + _OutputIterator + __first_empty(_IIter __c, _IIter d, _OutputIterator __out) const + { return std::copy(__c, d, __out); } - symmetric_difference_func(_Compare __comp) : _M_comp(__comp) {} + _OutputIterator + __second_empty(_IIter __a, _IIter __b, _OutputIterator __out) const + { return std::copy(__a, __b, __out); } + }; - _Compare _M_comp; - _OutputIterator - _M_invoke(_IIter __a, _IIter __b, - _IIter __c, _IIter d, - _OutputIterator __r) const + template + struct __difference_func { - while (__a != __b && __c != d) - { - if (_M_comp(*__a, *__c)) - { - *__r = *__a; - ++__a; - ++__r; - } - else if (_M_comp(*__c, *__a)) - { - *__r = *__c; - ++__c; - ++__r; - } - else - { - ++__a; - ++__c; - } - } - return std::copy(__c, d, std::copy(__a, __b, __r)); - } + typedef std::iterator_traits<_IIter> _TraitsType; + typedef typename _TraitsType::difference_type _DifferenceType; + typedef typename std::pair<_IIter, _IIter> _IteratorPair; - _DifferenceType - __count(_IIter __a, _IIter __b, - _IIter __c, _IIter d) const - { - _DifferenceType __counter = 0; + __difference_func(_Compare __comp) : _M_comp(__comp) {} - while (__a != __b && __c != d) - { - if (_M_comp(*__a, *__c)) - { - ++__a; - ++__counter; - } - else if (_M_comp(*__c, *__a)) - { - ++__c; - ++__counter; - } - else - { - ++__a; - ++__c; - } - } + _Compare _M_comp; - return __counter + (__b - __a) + (d - __c); - } + _OutputIterator + _M_invoke(_IIter __a, _IIter __b, _IIter __c, _IIter d, + _OutputIterator __r) const + { + while (__a != __b && __c != d) + { + if (_M_comp(*__a, *__c)) + { + *__r = *__a; + ++__a; + ++__r; + } + else if (_M_comp(*__c, *__a)) + { ++__c; } + else + { + ++__a; + ++__c; + } + } + return std::copy(__a, __b, __r); + } - _OutputIterator - __first_empty(_IIter __c, _IIter d, _OutputIterator __out) const - { return std::copy(__c, d, __out); } + _DifferenceType + __count(_IIter __a, _IIter __b, + _IIter __c, _IIter d) const + { + _DifferenceType __counter = 0; - _OutputIterator - __second_empty(_IIter __a, _IIter __b, _OutputIterator __out) const - { return std::copy(__a, __b, __out); } - }; + while (__a != __b && __c != d) + { + if (_M_comp(*__a, *__c)) + { + ++__a; + ++__counter; + } + else if (_M_comp(*__c, *__a)) + { ++__c; } + else + { ++__a; ++__c; } + } + return __counter + (__b - __a); + } -template - struct __difference_func - { - typedef std::iterator_traits<_IIter> _TraitsType; - typedef typename _TraitsType::difference_type _DifferenceType; - typedef typename std::pair<_IIter, _IIter> _IteratorPair; + _OutputIterator + __first_empty(_IIter, _IIter, _OutputIterator __out) const + { return __out; } - __difference_func(_Compare __comp) : _M_comp(__comp) {} + _OutputIterator + __second_empty(_IIter __a, _IIter __b, _OutputIterator __out) const + { return std::copy(__a, __b, __out); } + }; - _Compare _M_comp; - _OutputIterator - _M_invoke(_IIter __a, _IIter __b, _IIter __c, _IIter d, - _OutputIterator __r) const + template + struct __intersection_func { - while (__a != __b && __c != d) - { - if (_M_comp(*__a, *__c)) - { - *__r = *__a; - ++__a; - ++__r; - } - else if (_M_comp(*__c, *__a)) - { ++__c; } - else - { - ++__a; - ++__c; - } - } - return std::copy(__a, __b, __r); - } + typedef std::iterator_traits<_IIter> _TraitsType; + typedef typename _TraitsType::difference_type _DifferenceType; + typedef typename std::pair<_IIter, _IIter> _IteratorPair; - _DifferenceType - __count(_IIter __a, _IIter __b, - _IIter __c, _IIter d) const - { - _DifferenceType __counter = 0; + __intersection_func(_Compare __comp) : _M_comp(__comp) {} - while (__a != __b && __c != d) - { - if (_M_comp(*__a, *__c)) - { - ++__a; - ++__counter; - } - else if (_M_comp(*__c, *__a)) - { ++__c; } - else - { ++__a; ++__c; } - } + _Compare _M_comp; - return __counter + (__b - __a); - } + _OutputIterator + _M_invoke(_IIter __a, _IIter __b, _IIter __c, _IIter __d, + _OutputIterator __r) const + { + while (__a != __b && __c != __d) + { + if (_M_comp(*__a, *__c)) + { ++__a; } + else if (_M_comp(*__c, *__a)) + { ++__c; } + else + { + *__r = *__a; + ++__a; + ++__c; + ++__r; + } + } - inline _OutputIterator - __first_empty(_IIter __c, _IIter d, _OutputIterator __out) const - { return __out; } + return __r; + } - inline _OutputIterator - __second_empty(_IIter __a, _IIter __b, _OutputIterator __out) const - { return std::copy(__a, __b, __out); } - }; + _DifferenceType + __count(_IIter __a, _IIter __b, _IIter __c, _IIter __d) const + { + _DifferenceType __counter = 0; + while (__a != __b && __c != __d) + { + if (_M_comp(*__a, *__c)) + { ++__a; } + else if (_M_comp(*__c, *__a)) + { ++__c; } + else + { + ++__a; + ++__c; + ++__counter; + } + } -template - struct __intersection_func - { - typedef std::iterator_traits<_IIter> _TraitsType; - typedef typename _TraitsType::difference_type _DifferenceType; - typedef typename std::pair<_IIter, _IIter> _IteratorPair; + return __counter; + } - __intersection_func(_Compare __comp) : _M_comp(__comp) {} + _OutputIterator + __first_empty(_IIter, _IIter, _OutputIterator __out) const + { return __out; } - _Compare _M_comp; + _OutputIterator + __second_empty(_IIter, _IIter, _OutputIterator __out) const + { return __out; } + }; - _OutputIterator - _M_invoke(_IIter __a, _IIter __b, _IIter __c, _IIter d, - _OutputIterator __r) const + template + struct __union_func { - while (__a != __b && __c != d) - { - if (_M_comp(*__a, *__c)) - { ++__a; } - else if (_M_comp(*__c, *__a)) - { ++__c; } - else - { - *__r = *__a; - ++__a; - ++__c; - ++__r; - } - } + typedef typename std::iterator_traits<_IIter>::difference_type + _DifferenceType; - return __r; - } + __union_func(_Compare __comp) : _M_comp(__comp) {} - _DifferenceType - __count(_IIter __a, _IIter __b, - _IIter __c, _IIter d) const - { - _DifferenceType __counter = 0; - - while (__a != __b && __c != d) - { - if (_M_comp(*__a, *__c)) - { ++__a; } - else if (_M_comp(*__c, *__a)) - { ++__c; } - else - { - ++__a; - ++__c; - ++__counter; - } - } + _Compare _M_comp; - return __counter; - } + _OutputIterator + _M_invoke(_IIter __a, const _IIter __b, _IIter __c, + const _IIter __d, _OutputIterator __r) const + { + while (__a != __b && __c != __d) + { + if (_M_comp(*__a, *__c)) + { + *__r = *__a; + ++__a; + } + else if (_M_comp(*__c, *__a)) + { + *__r = *__c; + ++__c; + } + else + { + *__r = *__a; + ++__a; + ++__c; + } + ++__r; + } + return std::copy(__c, __d, std::copy(__a, __b, __r)); + } - inline _OutputIterator - __first_empty(_IIter __c, _IIter d, _OutputIterator __out) const - { return __out; } + _DifferenceType + __count(_IIter __a, _IIter __b, _IIter __c, _IIter __d) const + { + _DifferenceType __counter = 0; - inline _OutputIterator - __second_empty(_IIter __a, _IIter __b, _OutputIterator __out) const - { return __out; } - }; + while (__a != __b && __c != __d) + { + if (_M_comp(*__a, *__c)) + { ++__a; } + else if (_M_comp(*__c, *__a)) + { ++__c; } + else + { + ++__a; + ++__c; + } + ++__counter; + } -template - struct __union_func - { - typedef typename std::iterator_traits<_IIter>::difference_type - _DifferenceType; + __counter += (__b - __a); + __counter += (__d - __c); + return __counter; + } - __union_func(_Compare __comp) : _M_comp(__comp) {} + _OutputIterator + __first_empty(_IIter __c, _IIter __d, _OutputIterator __out) const + { return std::copy(__c, __d, __out); } - _Compare _M_comp; + _OutputIterator + __second_empty(_IIter __a, _IIter __b, _OutputIterator __out) const + { return std::copy(__a, __b, __out); } + }; + template _OutputIterator - _M_invoke(_IIter __a, const _IIter __b, _IIter __c, - const _IIter d, _OutputIterator __r) const + __parallel_set_operation(_IIter __begin1, _IIter __end1, + _IIter __begin2, _IIter __end2, + _OutputIterator __result, Operation __op) { - while (__a != __b && __c != d) - { - if (_M_comp(*__a, *__c)) - { - *__r = *__a; - ++__a; - } - else if (_M_comp(*__c, *__a)) - { - *__r = *__c; - ++__c; - } - else - { - *__r = *__a; - ++__a; - ++__c; - } - ++__r; - } - return std::copy(__c, d, std::copy(__a, __b, __r)); - } + _GLIBCXX_CALL((__end1 - __begin1) + (__end2 - __begin2)) - _DifferenceType - __count(_IIter __a, _IIter __b, - _IIter __c, _IIter d) const - { - _DifferenceType __counter = 0; - - while (__a != __b && __c != d) - { - if (_M_comp(*__a, *__c)) - { ++__a; } - else if (_M_comp(*__c, *__a)) - { ++__c; } - else - { - ++__a; - ++__c; - } - ++__counter; - } + typedef std::iterator_traits<_IIter> _TraitsType; + typedef typename _TraitsType::difference_type _DifferenceType; + typedef typename std::pair<_IIter, _IIter> _IteratorPair; - __counter += (__b - __a); - __counter += (d - __c); - return __counter; - } + if (__begin1 == __end1) + return __op.__first_empty(__begin2, __end2, __result); - inline _OutputIterator - __first_empty(_IIter __c, _IIter d, _OutputIterator __out) const - { return std::copy(__c, d, __out); } + if (__begin2 == __end2) + return __op.__second_empty(__begin1, __end1, __result); - inline _OutputIterator - __second_empty(_IIter __a, _IIter __b, _OutputIterator __out) const - { return std::copy(__a, __b, __out); } - }; - -template - _OutputIterator - __parallel_set_operation(_IIter __begin1, _IIter __end1, - _IIter __begin2, _IIter __end2, - _OutputIterator __result, Operation __op) - { - _GLIBCXX_CALL((__end1 - __begin1) + (__end2 - __begin2)) - - typedef std::iterator_traits<_IIter> _TraitsType; - typedef typename _TraitsType::difference_type _DifferenceType; - typedef typename std::pair<_IIter, _IIter> _IteratorPair; - - if (__begin1 == __end1) - return __op.__first_empty(__begin2, __end2, __result); - - if (__begin2 == __end2) - return __op.__second_empty(__begin1, __end1, __result); - - const _DifferenceType size = (__end1 - __begin1) + (__end2 - __begin2); - - const _IteratorPair __sequence[ 2 ] = - { std::make_pair(__begin1, __end1), std::make_pair(__begin2, __end2) }; - _OutputIterator return_value = __result; - _DifferenceType *__borders; - _IteratorPair *__block_begins; - _DifferenceType* __lengths; - - _ThreadIndex __num_threads = - std::min<_DifferenceType>(__get_max_threads(), - std::min(__end1 - __begin1, __end2 - __begin2)); - -# pragma omp parallel num_threads(__num_threads) + const _DifferenceType __size = (__end1 - __begin1) + (__end2 - __begin2); + + const _IteratorPair __sequence[2] = { std::make_pair(__begin1, __end1), + std::make_pair(__begin2, __end2) }; + _OutputIterator __return_value = __result; + _DifferenceType *__borders; + _IteratorPair *__block_begins; + _DifferenceType* __lengths; + + _ThreadIndex __num_threads = + std::min<_DifferenceType>(__get_max_threads(), + std::min(__end1 - __begin1, __end2 - __begin2)); + +# pragma omp parallel num_threads(__num_threads) { # pragma omp single - { - __num_threads = omp_get_num_threads(); - - __borders = new _DifferenceType[__num_threads + 2]; - equally_split(size, __num_threads + 1, __borders); - __block_begins = new _IteratorPair[__num_threads + 1]; - // Very __start. - __block_begins[0] = std::make_pair(__begin1, __begin2); - __lengths = new _DifferenceType[__num_threads]; - } //single - - _ThreadIndex __iam = omp_get_thread_num(); - - // _Result from multiseq_partition. - _IIter __offset[2]; - const _DifferenceType __rank = __borders[__iam + 1]; - - multiseq_partition(__sequence, __sequence + 2, - __rank, __offset, __op._M_comp); - - // allowed to read? - // together - // *(__offset[ 0 ] - 1) == *__offset[ 1 ] - if (__offset[ 0 ] != __begin1 && __offset[ 1 ] != __end2 - && !__op._M_comp(*(__offset[ 0 ] - 1), *__offset[ 1 ]) - && !__op._M_comp(*__offset[ 1 ], *(__offset[ 0 ] - 1))) - { - // Avoid split between globally equal elements: move one to - // front in first sequence. - --__offset[ 0 ]; - } - - _IteratorPair block_end = __block_begins[ __iam + 1 ] = - _IteratorPair(__offset[ 0 ], __offset[ 1 ]); + { + __num_threads = omp_get_num_threads(); + + __borders = new _DifferenceType[__num_threads + 2]; + equally_split(__size, __num_threads + 1, __borders); + __block_begins = new _IteratorPair[__num_threads + 1]; + // Very __start. + __block_begins[0] = std::make_pair(__begin1, __begin2); + __lengths = new _DifferenceType[__num_threads]; + } //single + + _ThreadIndex __iam = omp_get_thread_num(); + + // _Result from multiseq_partition. + _IIter __offset[2]; + const _DifferenceType __rank = __borders[__iam + 1]; + + multiseq_partition(__sequence, __sequence + 2, + __rank, __offset, __op._M_comp); + + // allowed to read? + // together + // *(__offset[ 0 ] - 1) == *__offset[ 1 ] + if (__offset[ 0 ] != __begin1 && __offset[1] != __end2 + && !__op._M_comp(*(__offset[0] - 1), *__offset[1]) + && !__op._M_comp(*__offset[1], *(__offset[0] - 1))) + { + // Avoid split between globally equal elements: move one to + // front in first sequence. + --__offset[0]; + } + + _IteratorPair __block_end = __block_begins[__iam + 1] = + _IteratorPair(__offset[0], __offset[1]); + + // Make sure all threads have their block_begin result written out. +# pragma omp barrier - // Make sure all threads have their block_begin result written out. + _IteratorPair __block_begin = __block_begins[__iam]; + + // Begin working for the first block, while the others except + // the last start to count. + if (__iam == 0) + { + // The first thread can copy already. + __lengths[ __iam ] = + __op._M_invoke(__block_begin.first, __block_end.first, + __block_begin.second, __block_end.second, + __result) - __result; + } + else + { + __lengths[ __iam ] = + __op.__count(__block_begin.first, __block_end.first, + __block_begin.second, __block_end.second); + } + + // Make sure everyone wrote their lengths. # pragma omp barrier - _IteratorPair __block_begin = __block_begins[ __iam ]; + _OutputIterator __r = __result; - // Begin working for the first block, while the others except - // the last start to count. - if (__iam == 0) - { - // The first thread can copy already. - __lengths[ __iam ] = - __op._M_invoke(__block_begin.first, block_end.first, - __block_begin.second, block_end.second, __result) - - __result; - } - else - { - __lengths[ __iam ] = - __op.__count(__block_begin.first, block_end.first, - __block_begin.second, block_end.second); - } + if (__iam == 0) + { + // Do the last block. + for (int __i = 0; __i < __num_threads; ++__i) + __r += __lengths[__i]; - // Make sure everyone wrote their lengths. -# pragma omp barrier + __block_begin = __block_begins[__num_threads]; - _OutputIterator __r = __result; + // Return the result iterator of the last block. + __return_value = + __op._M_invoke(__block_begin.first, __end1, + __block_begin.second, __end2, __r); - if (__iam == 0) - { - // Do the last block. - for (int __i = 0; __i < __num_threads; ++__i) - __r += __lengths[__i]; + } + else + { + for (int __i = 0; __i < __iam; ++__i) + __r += __lengths[ __i ]; - __block_begin = __block_begins[__num_threads]; + // Reset begins for copy pass. + __op._M_invoke(__block_begin.first, __block_end.first, + __block_begin.second, __block_end.second, __r); + } + } + return __return_value; + } - // Return the result iterator of the last block. - return_value = __op._M_invoke( - __block_begin.first, __end1, __block_begin.second, __end2, __r); + template + inline _OutputIterator + __parallel_set_union(_IIter __begin1, _IIter __end1, + _IIter __begin2, _IIter __end2, + _OutputIterator __result, _Compare __comp) + { + return __parallel_set_operation(__begin1, __end1, __begin2, __end2, + __result, + __union_func< _IIter, _OutputIterator, + _Compare>(__comp)); + } - } - else - { - for (int __i = 0; __i < __iam; ++__i) - __r += __lengths[ __i ]; + template + inline _OutputIterator + __parallel_set_intersection(_IIter __begin1, _IIter __end1, + _IIter __begin2, _IIter __end2, + _OutputIterator __result, _Compare __comp) + { + return __parallel_set_operation(__begin1, __end1, __begin2, __end2, + __result, + __intersection_func<_IIter, + _OutputIterator, _Compare>(__comp)); + } - // Reset begins for copy pass. - __op._M_invoke(__block_begin.first, block_end.first, - __block_begin.second, block_end.second, __r); - } - } - return return_value; - } - - -template - inline _OutputIterator - __parallel_set_union(_IIter __begin1, _IIter __end1, - _IIter __begin2, _IIter __end2, - _OutputIterator __result, _Compare _M_comp) - { - return __parallel_set_operation(__begin1, __end1, __begin2, __end2, - __result, __union_func< _IIter, _OutputIterator, _Compare>(_M_comp)); - } - -template - inline _OutputIterator - __parallel_set_intersection(_IIter __begin1, _IIter __end1, + template + inline _OutputIterator + __parallel_set_difference(_IIter __begin1, _IIter __end1, _IIter __begin2, _IIter __end2, - _OutputIterator __result, _Compare _M_comp) - { - return __parallel_set_operation( - __begin1, __end1, __begin2, __end2, __result, - __intersection_func<_IIter, _OutputIterator, _Compare>(_M_comp)); - } - -template - inline _OutputIterator - __parallel_set_difference(_IIter __begin1, _IIter __end1, - _IIter __begin2, _IIter __end2, - _OutputIterator __result, _Compare _M_comp) - { - return __parallel_set_operation( - __begin1, __end1, __begin2, __end2, __result, - __difference_func<_IIter, _OutputIterator, _Compare>(_M_comp)); - } - -template - inline _OutputIterator - __parallel_set_symmetric_difference(_IIter __begin1, _IIter __end1, - _IIter __begin2, _IIter __end2, - _OutputIterator __result, - _Compare _M_comp) - { - return __parallel_set_operation( - __begin1, __end1, __begin2, __end2, __result, - symmetric_difference_func<_IIter, _OutputIterator, _Compare> - (_M_comp)); - } + _OutputIterator __result, _Compare __comp) + { + return __parallel_set_operation(__begin1, __end1, __begin2, __end2, + __result, + __difference_func<_IIter, + _OutputIterator, _Compare>(__comp)); + } + template + inline _OutputIterator + __parallel_set_symmetric_difference(_IIter __begin1, _IIter __end1, + _IIter __begin2, _IIter __end2, + _OutputIterator __result, + _Compare __comp) + { + return __parallel_set_operation(__begin1, __end1, __begin2, __end2, + __result, + __symmetric_difference_func<_IIter, + _OutputIterator, _Compare>(__comp)); + } } #endif /* _GLIBCXX_PARALLEL_SET_OPERATIONS_H */ diff --git a/libstdc++-v3/include/parallel/sort.h b/libstdc++-v3/include/parallel/sort.h index 2d38cad4d96..f1a163c63b6 100644 --- a/libstdc++-v3/include/parallel/sort.h +++ b/libstdc++-v3/include/parallel/sort.h @@ -54,12 +54,12 @@ namespace __gnu_parallel { - //prototype + //prototype template - void - __parallel_sort(_RAIter __begin, _RAIter __end, - _Compare __comp, _Parallelism __parallelism); + void + __parallel_sort(_RAIter __begin, _RAIter __end, + _Compare __comp, _Parallelism __parallelism); /** * @brief Choose multiway mergesort, splitting variant at run-time, @@ -70,19 +70,19 @@ namespace __gnu_parallel * @callgraph */ template - inline void - __parallel_sort(_RAIter __begin, _RAIter __end, - _Compare __comp, multiway_mergesort_tag __parallelism) - { - _GLIBCXX_CALL(__end - __begin) + inline void + __parallel_sort(_RAIter __begin, _RAIter __end, + _Compare __comp, multiway_mergesort_tag __parallelism) + { + _GLIBCXX_CALL(__end - __begin) - if(_Settings::get().sort_splitting == EXACT) - parallel_sort_mwms<__stable, true> - (__begin, __end, __comp, __parallelism.__get_num_threads()); - else - parallel_sort_mwms<__stable, false> - (__begin, __end, __comp, __parallelism.__get_num_threads()); - } + if(_Settings::get().sort_splitting == EXACT) + parallel_sort_mwms<__stable, true> + (__begin, __end, __comp, __parallelism.__get_num_threads()); + else + parallel_sort_mwms<__stable, false> + (__begin, __end, __comp, __parallelism.__get_num_threads()); + } /** * @brief Choose multiway mergesort with exact splitting, @@ -93,15 +93,16 @@ namespace __gnu_parallel * @callgraph */ template - inline void - __parallel_sort(_RAIter __begin, _RAIter __end, - _Compare __comp, multiway_mergesort_exact_tag __parallelism) - { - _GLIBCXX_CALL(__end - __begin) + inline void + __parallel_sort(_RAIter __begin, _RAIter __end, + _Compare __comp, + multiway_mergesort_exact_tag __parallelism) + { + _GLIBCXX_CALL(__end - __begin) parallel_sort_mwms<__stable, true> (__begin, __end, __comp, __parallelism.__get_num_threads()); - } + } /** * @brief Choose multiway mergesort with splitting by sampling, @@ -112,15 +113,16 @@ namespace __gnu_parallel * @callgraph */ template - inline void - __parallel_sort(_RAIter __begin, _RAIter __end, - _Compare __comp, multiway_mergesort_sampling_tag __parallelism) - { - _GLIBCXX_CALL(__end - __begin) + inline void + __parallel_sort(_RAIter __begin, _RAIter __end, + _Compare __comp, + multiway_mergesort_sampling_tag __parallelism) + { + _GLIBCXX_CALL(__end - __begin) - parallel_sort_mwms<__stable, false> + parallel_sort_mwms<__stable, false> (__begin, __end, __comp, __parallelism.__get_num_threads()); - } + } /** * @brief Choose quicksort for parallel sorting. @@ -130,17 +132,17 @@ namespace __gnu_parallel * @callgraph */ template - inline void - __parallel_sort(_RAIter __begin, _RAIter __end, - _Compare __comp, quicksort_tag __parallelism) - { - _GLIBCXX_CALL(__end - __begin) + inline void + __parallel_sort(_RAIter __begin, _RAIter __end, + _Compare __comp, quicksort_tag __parallelism) + { + _GLIBCXX_CALL(__end - __begin) - _GLIBCXX_PARALLEL_ASSERT(__stable == false); + _GLIBCXX_PARALLEL_ASSERT(__stable == false); - __parallel_sort_qs(__begin, __end, __comp, - __parallelism.__get_num_threads()); - } + __parallel_sort_qs(__begin, __end, __comp, + __parallelism.__get_num_threads()); + } /** * @brief Choose balanced quicksort for parallel sorting. @@ -150,19 +152,18 @@ namespace __gnu_parallel * @param __stable Sort __stable. * @callgraph */ - template - inline void - __parallel_sort(_RAIter __begin, _RAIter __end, - _Compare __comp, balanced_quicksort_tag __parallelism) - { - _GLIBCXX_CALL(__end - __begin) - - _GLIBCXX_PARALLEL_ASSERT(__stable == false); + template + inline void + __parallel_sort(_RAIter __begin, _RAIter __end, + _Compare __comp, balanced_quicksort_tag __parallelism) + { + _GLIBCXX_CALL(__end - __begin) - __parallel_sort_qsb(__begin, __end, __comp, - __parallelism.__get_num_threads()); - } + _GLIBCXX_PARALLEL_ASSERT(__stable == false); + __parallel_sort_qsb(__begin, __end, __comp, + __parallelism.__get_num_threads()); + } /** * @brief Choose multiway mergesort with exact splitting, @@ -173,17 +174,16 @@ namespace __gnu_parallel * @callgraph */ template - inline void - __parallel_sort(_RAIter __begin, _RAIter __end, - _Compare __comp, default_parallel_tag __parallelism) - { - _GLIBCXX_CALL(__end - __begin) - - __parallel_sort<__stable> - (__begin, __end, __comp, - multiway_mergesort_exact_tag(__parallelism.__get_num_threads())); - } + inline void + __parallel_sort(_RAIter __begin, _RAIter __end, + _Compare __comp, default_parallel_tag __parallelism) + { + _GLIBCXX_CALL(__end - __begin) + __parallel_sort<__stable> + (__begin, __end, __comp, + multiway_mergesort_exact_tag(__parallelism.__get_num_threads())); + } /** * @brief Choose a parallel sorting algorithm. @@ -196,7 +196,7 @@ namespace __gnu_parallel template inline void __parallel_sort(_RAIter __begin, _RAIter __end, - _Compare __comp, parallel_tag __parallelism) + _Compare __comp, parallel_tag __parallelism) { _GLIBCXX_CALL(__end - __begin) typedef std::iterator_traits<_RAIter> _TraitsType; diff --git a/libstdc++-v3/include/parallel/tags.h b/libstdc++-v3/include/parallel/tags.h index 43561d28d5c..bc47b2699b4 100644 --- a/libstdc++-v3/include/parallel/tags.h +++ b/libstdc++-v3/include/parallel/tags.h @@ -51,20 +51,16 @@ namespace __gnu_parallel public: /** @brief Default constructor. Use default number of threads. */ parallel_tag() - { - this->_M_num_threads = 0; - } + { _M_num_threads = 0; } /** @brief Default constructor. Recommend number of threads to use. * @param __num_threads Desired number of threads. */ parallel_tag(_ThreadIndex __num_threads) - { - this->_M_num_threads = __num_threads; - } + { _M_num_threads = __num_threads; } /** @brief Find out desired number of threads. * @return Desired number of threads. */ - inline _ThreadIndex __get_num_threads() + _ThreadIndex __get_num_threads() { if(_M_num_threads == 0) return omp_get_max_threads(); @@ -74,19 +70,17 @@ namespace __gnu_parallel /** @brief Set the desired number of threads. * @param __num_threads Desired number of threads. */ - inline void set_num_threads(_ThreadIndex __num_threads) - { - this->_M_num_threads = __num_threads; - } + void set_num_threads(_ThreadIndex __num_threads) + { _M_num_threads = __num_threads; } }; /** @brief Recommends parallel execution using the default parallel algorithm. */ struct default_parallel_tag : public parallel_tag { - default_parallel_tag() { } - default_parallel_tag(_ThreadIndex __num_threads) - : parallel_tag(__num_threads) { } + default_parallel_tag() { } + default_parallel_tag(_ThreadIndex __num_threads) + : parallel_tag(__num_threads) { } }; /** @brief Recommends parallel execution using dynamic @@ -114,18 +108,18 @@ namespace __gnu_parallel * with exact splitting, at compile time. */ struct exact_tag : public parallel_tag { - exact_tag() { } - exact_tag(_ThreadIndex __num_threads) - : parallel_tag(__num_threads) { } + exact_tag() { } + exact_tag(_ThreadIndex __num_threads) + : parallel_tag(__num_threads) { } }; /** @brief Forces parallel merging * with exact splitting, at compile time. */ struct sampling_tag : public parallel_tag { - sampling_tag() { } - sampling_tag(_ThreadIndex __num_threads) - : parallel_tag(__num_threads) { } + sampling_tag() { } + sampling_tag(_ThreadIndex __num_threads) + : parallel_tag(__num_threads) { } }; @@ -133,45 +127,45 @@ namespace __gnu_parallel * at compile time. */ struct multiway_mergesort_tag : public parallel_tag { - multiway_mergesort_tag() { } - multiway_mergesort_tag(_ThreadIndex __num_threads) - : parallel_tag(__num_threads) { } + multiway_mergesort_tag() { } + multiway_mergesort_tag(_ThreadIndex __num_threads) + : parallel_tag(__num_threads) { } }; /** @brief Forces parallel sorting using multiway mergesort * with exact splitting at compile time. */ struct multiway_mergesort_exact_tag : public parallel_tag { - multiway_mergesort_exact_tag() { } - multiway_mergesort_exact_tag(_ThreadIndex __num_threads) - : parallel_tag(__num_threads) { } + multiway_mergesort_exact_tag() { } + multiway_mergesort_exact_tag(_ThreadIndex __num_threads) + : parallel_tag(__num_threads) { } }; /** @brief Forces parallel sorting using multiway mergesort * with splitting by sampling at compile time. */ struct multiway_mergesort_sampling_tag : public parallel_tag { - multiway_mergesort_sampling_tag() { } - multiway_mergesort_sampling_tag(_ThreadIndex __num_threads) - : parallel_tag(__num_threads) { } + multiway_mergesort_sampling_tag() { } + multiway_mergesort_sampling_tag(_ThreadIndex __num_threads) + : parallel_tag(__num_threads) { } }; /** @brief Forces parallel sorting using unbalanced quicksort * at compile time. */ struct quicksort_tag : public parallel_tag { - quicksort_tag() { } - quicksort_tag(_ThreadIndex __num_threads) - : parallel_tag(__num_threads) { } + quicksort_tag() { } + quicksort_tag(_ThreadIndex __num_threads) + : parallel_tag(__num_threads) { } }; /** @brief Forces parallel sorting using balanced quicksort * at compile time. */ struct balanced_quicksort_tag : public parallel_tag { - balanced_quicksort_tag() { } - balanced_quicksort_tag(_ThreadIndex __num_threads) - : parallel_tag(__num_threads) { } + balanced_quicksort_tag() { } + balanced_quicksort_tag(_ThreadIndex __num_threads) + : parallel_tag(__num_threads) { } }; diff --git a/libstdc++-v3/include/parallel/unique_copy.h b/libstdc++-v3/include/parallel/unique_copy.h index 327870e7d86..1ae5a12316d 100644 --- a/libstdc++-v3/include/parallel/unique_copy.h +++ b/libstdc++-v3/include/parallel/unique_copy.h @@ -37,155 +37,160 @@ namespace __gnu_parallel { - -/** @brief Parallel std::unique_copy(), w/__o explicit equality predicate. - * @param __first Begin iterator of input sequence. - * @param __last End iterator of input sequence. - * @param __result Begin iterator of result __sequence. - * @param __binary_pred Equality predicate. - * @return End iterator of result __sequence. */ -template - _OutputIterator - __parallel_unique_copy(_IIter __first, _IIter __last, - _OutputIterator __result, _BinaryPredicate __binary_pred) - { - _GLIBCXX_CALL(__last - __first) - - typedef std::iterator_traits<_IIter> _TraitsType; - typedef typename _TraitsType::value_type _ValueType; - typedef typename _TraitsType::difference_type _DifferenceType; - - _DifferenceType size = __last - __first; - - if (size == 0) - return __result; - - // Let the first thread process two parts. - _DifferenceType *__counter; - _DifferenceType *__borders; - - _ThreadIndex __num_threads = __get_max_threads(); - // First part contains at least one element. -# pragma omp parallel num_threads(__num_threads) + /** @brief Parallel std::unique_copy(), w/__o explicit equality predicate. + * @param __first Begin iterator of input sequence. + * @param __last End iterator of input sequence. + * @param __result Begin iterator of result __sequence. + * @param __binary_pred Equality predicate. + * @return End iterator of result __sequence. */ + template + _OutputIterator + __parallel_unique_copy(_IIter __first, _IIter __last, + _OutputIterator __result, + _BinaryPredicate __binary_pred) + { + _GLIBCXX_CALL(__last - __first) + + typedef std::iterator_traits<_IIter> _TraitsType; + typedef typename _TraitsType::value_type _ValueType; + typedef typename _TraitsType::difference_type _DifferenceType; + + _DifferenceType __size = __last - __first; + + if (__size == 0) + return __result; + + // Let the first thread process two parts. + _DifferenceType *__counter; + _DifferenceType *__borders; + + _ThreadIndex __num_threads = __get_max_threads(); + // First part contains at least one element. +# pragma omp parallel num_threads(__num_threads) { # pragma omp single + { + __num_threads = omp_get_num_threads(); + __borders = new _DifferenceType[__num_threads + 2]; + equally_split(__size, __num_threads + 1, __borders); + __counter = new _DifferenceType[__num_threads + 1]; + } + + _ThreadIndex __iam = omp_get_thread_num(); + + _DifferenceType __begin, __end; + + // Check for length without duplicates + // Needed for position in output + _DifferenceType __i = 0; + _OutputIterator __out = __result; + + if (__iam == 0) { - __num_threads = omp_get_num_threads(); - __borders = new _DifferenceType[__num_threads + 2]; - equally_split(size, __num_threads + 1, __borders); - __counter = new _DifferenceType[__num_threads + 1]; + __begin = __borders[0] + 1; // == 1 + __end = __borders[__iam + 1]; + + ++__i; + *__out++ = *__first; + + for (_IIter __iter = __first + __begin; __iter < __first + __end; + ++__iter) + { + if (!__binary_pred(*__iter, *(__iter - 1))) + { + ++__i; + *__out++ = *__iter; + } + } } + else + { + __begin = __borders[__iam]; //one part + __end = __borders[__iam + 1]; - _ThreadIndex __iam = omp_get_thread_num(); + for (_IIter __iter = __first + __begin; __iter < __first + __end; + ++__iter) + { + if (!__binary_pred(*__iter, *(__iter - 1))) + ++__i; + } + } + __counter[__iam] = __i; - _DifferenceType __begin, __end; + // Last part still untouched. + _DifferenceType __begin_output; - // Check for length without duplicates - // Needed for position in output - _DifferenceType __i = 0; - _OutputIterator __out = __result; +# pragma omp barrier - if (__iam == 0) - { - __begin = __borders[0] + 1; // == 1 - __end = __borders[__iam + 1]; + // Store result in output on calculated positions. + __begin_output = 0; - ++__i; - *__out++ = *__first; + if (__iam == 0) + { + for (int __t = 0; __t < __num_threads; ++__t) + __begin_output += __counter[__t]; - for (_IIter iter = __first + __begin; iter < __first + __end; ++iter) - { - if (!__binary_pred(*iter, *(iter-1))) - { - ++__i; - *__out++ = *iter; - } - } - } - else - { - __begin = __borders[__iam]; //one part - __end = __borders[__iam + 1]; - - for (_IIter iter = __first + __begin; iter < __first + __end; ++iter) - { - if (!__binary_pred(*iter, *(iter - 1))) - ++__i; - } - } - __counter[__iam] = __i; - - // Last part still untouched. - _DifferenceType __begin_output; - -# pragma omp barrier - - // Store result in output on calculated positions. - __begin_output = 0; - - if (__iam == 0) - { - for (int __t = 0; __t < __num_threads; ++__t) - __begin_output += __counter[__t]; - - __i = 0; - - _OutputIterator __iter_out = __result + __begin_output; - - __begin = __borders[__num_threads]; - __end = size; - - for (_IIter iter = __first + __begin; iter < __first + __end; ++iter) - { - if (iter == __first || !__binary_pred(*iter, *(iter - 1))) - { - ++__i; - *__iter_out++ = *iter; - } - } - - __counter[__num_threads] = __i; - } - else - { - for (int __t = 0; __t < __iam; __t++) - __begin_output += __counter[__t]; - - _OutputIterator __iter_out = __result + __begin_output; - for (_IIter iter = __first + __begin; iter < __first + __end; ++iter) - { - if (!__binary_pred(*iter, *(iter-1))) - *__iter_out++ = *iter; - } - } + __i = 0; + + _OutputIterator __iter_out = __result + __begin_output; + + __begin = __borders[__num_threads]; + __end = __size; + + for (_IIter __iter = __first + __begin; __iter < __first + __end; + ++__iter) + { + if (__iter == __first + || !__binary_pred(*__iter, *(__iter - 1))) + { + ++__i; + *__iter_out++ = *__iter; + } + } + + __counter[__num_threads] = __i; + } + else + { + for (int __t = 0; __t < __iam; __t++) + __begin_output += __counter[__t]; + + _OutputIterator __iter_out = __result + __begin_output; + for (_IIter __iter = __first + __begin; __iter < __first + __end; + ++__iter) + { + if (!__binary_pred(*__iter, *(__iter - 1))) + *__iter_out++ = *__iter; + } + } + } + + _DifferenceType __end_output = 0; + for (int __t = 0; __t < __num_threads + 1; __t++) + __end_output += __counter[__t]; + + delete[] __borders; + + return __result + __end_output; } - _DifferenceType __end_output = 0; - for (int __t = 0; __t < __num_threads + 1; __t++) - __end_output += __counter[__t]; - - delete[] __borders; - - return __result + __end_output; - } - -/** @brief Parallel std::unique_copy(), without explicit equality predicate - * @param __first Begin iterator of input sequence. - * @param __last End iterator of input sequence. - * @param __result Begin iterator of result __sequence. - * @return End iterator of result __sequence. */ -template - inline _OutputIterator - __parallel_unique_copy(_IIter __first, _IIter __last, - _OutputIterator __result) - { - typedef typename std::iterator_traits<_IIter>::value_type - _ValueType; - return __parallel_unique_copy(__first, __last, __result, - std::equal_to<_ValueType>()); - } + /** @brief Parallel std::unique_copy(), without explicit equality predicate + * @param __first Begin iterator of input sequence. + * @param __last End iterator of input sequence. + * @param __result Begin iterator of result __sequence. + * @return End iterator of result __sequence. */ + template + inline _OutputIterator + __parallel_unique_copy(_IIter __first, _IIter __last, + _OutputIterator __result) + { + typedef typename std::iterator_traits<_IIter>::value_type + _ValueType; + return __parallel_unique_copy(__first, __last, __result, + std::equal_to<_ValueType>()); + } }//namespace __gnu_parallel diff --git a/libstdc++-v3/include/parallel/workstealing.h b/libstdc++-v3/include/parallel/workstealing.h index 638057ca740..4ac155d923f 100644 --- a/libstdc++-v3/include/parallel/workstealing.h +++ b/libstdc++-v3/include/parallel/workstealing.h @@ -49,261 +49,264 @@ namespace __gnu_parallel #define _GLIBCXX_JOB_VOLATILE volatile -/** @brief One __job for a certain thread. */ -template - struct _Job - { - typedef _DifferenceTp _DifferenceType; - - /** @brief First element. - * - * Changed by owning and stealing thread. By stealing thread, - * always incremented. */ - _GLIBCXX_JOB_VOLATILE _DifferenceType _M_first; - - /** @brief Last element. - * - * Changed by owning thread only. */ - _GLIBCXX_JOB_VOLATILE _DifferenceType _M_last; - - /** @brief Number of elements, i.e. @__c _M_last-_M_first+1. - * - * Changed by owning thread only. */ - _GLIBCXX_JOB_VOLATILE _DifferenceType _M_load; - }; - -/** @brief Work stealing algorithm for random access iterators. - * - * Uses O(1) additional memory. Synchronization at job lists is - * done with atomic operations. - * @param __begin Begin iterator of element sequence. - * @param __end End iterator of element sequence. - * @param __op User-supplied functor (comparator, predicate, adding - * functor, ...). - * @param __f Functor to "process" an element with __op (depends on - * desired functionality, e. g. for std::for_each(), ...). - * @param __r Functor to "add" a single __result to the already - * processed elements (depends on functionality). - * @param __base Base value for reduction. - * @param __output Pointer to position where final result is written to - * @param __bound Maximum number of elements processed (e. g. for - * std::count_n()). - * @return User-supplied functor (that may contain a part of the result). - */ -template - _Op - __for_each_template_random_access_workstealing( - _RAIter __begin, _RAIter __end, _Op __op, _Fu& __f, _Red __r, - _Result __base, _Result& __output, - typename std::iterator_traits<_RAIter>::difference_type __bound) - { - _GLIBCXX_CALL(__end - __begin) - - typedef std::iterator_traits<_RAIter> _TraitsType; - typedef typename _TraitsType::difference_type _DifferenceType; - - const _Settings& __s = _Settings::get(); - - _DifferenceType __chunk_size = - static_cast<_DifferenceType>(__s.workstealing_chunk_size); - - // How many jobs? - _DifferenceType __length = (__bound < 0) ? (__end - __begin) : __bound; - - // To avoid false sharing in a cache line. - const int __stride = - __s.cache_line_size * 10 / sizeof(_Job<_DifferenceType>) + 1; - - // Total number of threads currently working. - _ThreadIndex __busy = 0; - - _Job<_DifferenceType> *__job; - - omp_lock_t __output_lock; - omp_init_lock(&__output_lock); - - // Write base value to output. - __output = __base; - - // No more threads than jobs, at least one thread. - _ThreadIndex __num_threads = - __gnu_parallel::max<_ThreadIndex>(1, - __gnu_parallel::min<_DifferenceType>(__length, __get_max_threads())); - -# pragma omp parallel shared(__busy) num_threads(__num_threads) + /** @brief One __job for a certain thread. */ + template + struct _Job + { + typedef _DifferenceTp _DifferenceType; + + /** @brief First element. + * + * Changed by owning and stealing thread. By stealing thread, + * always incremented. */ + _GLIBCXX_JOB_VOLATILE _DifferenceType _M_first; + + /** @brief Last element. + * + * Changed by owning thread only. */ + _GLIBCXX_JOB_VOLATILE _DifferenceType _M_last; + + /** @brief Number of elements, i.e. @__c _M_last-_M_first+1. + * + * Changed by owning thread only. */ + _GLIBCXX_JOB_VOLATILE _DifferenceType _M_load; + }; + + /** @brief Work stealing algorithm for random access iterators. + * + * Uses O(1) additional memory. Synchronization at job lists is + * done with atomic operations. + * @param __begin Begin iterator of element sequence. + * @param __end End iterator of element sequence. + * @param __op User-supplied functor (comparator, predicate, adding + * functor, ...). + * @param __f Functor to "process" an element with __op (depends on + * desired functionality, e. g. for std::for_each(), ...). + * @param __r Functor to "add" a single __result to the already + * processed elements (depends on functionality). + * @param __base Base value for reduction. + * @param __output Pointer to position where final result is written to + * @param __bound Maximum number of elements processed (e. g. for + * std::count_n()). + * @return User-supplied functor (that may contain a part of the result). + */ + template + _Op + __for_each_template_random_access_workstealing(_RAIter __begin, + _RAIter __end, _Op __op, + _Fu& __f, _Red __r, + _Result __base, + _Result& __output, + typename std::iterator_traits<_RAIter>::difference_type __bound) + { + _GLIBCXX_CALL(__end - __begin) + + typedef std::iterator_traits<_RAIter> _TraitsType; + typedef typename _TraitsType::difference_type _DifferenceType; + + const _Settings& __s = _Settings::get(); + + _DifferenceType __chunk_size = + static_cast<_DifferenceType>(__s.workstealing_chunk_size); + + // How many jobs? + _DifferenceType __length = (__bound < 0) ? (__end - __begin) : __bound; + + // To avoid false sharing in a cache line. + const int __stride = (__s.cache_line_size * 10 + / sizeof(_Job<_DifferenceType>) + 1); + + // Total number of threads currently working. + _ThreadIndex __busy = 0; + + _Job<_DifferenceType> *__job; + + omp_lock_t __output_lock; + omp_init_lock(&__output_lock); + + // Write base value to output. + __output = __base; + + // No more threads than jobs, at least one thread. + _ThreadIndex __num_threads = __gnu_parallel::max<_ThreadIndex> + (1, __gnu_parallel::min<_DifferenceType>(__length, + __get_max_threads())); + +# pragma omp parallel shared(__busy) num_threads(__num_threads) { - # pragma omp single - { - __num_threads = omp_get_num_threads(); + { + __num_threads = omp_get_num_threads(); - // Create job description array. - __job = new _Job<_DifferenceType>[__num_threads * __stride]; - } + // Create job description array. + __job = new _Job<_DifferenceType>[__num_threads * __stride]; + } - // Initialization phase. + // Initialization phase. - // Flags for every thread if it is doing productive work. - bool __iam_working = false; + // Flags for every thread if it is doing productive work. + bool __iam_working = false; - // Thread id. - _ThreadIndex __iam = omp_get_thread_num(); + // Thread id. + _ThreadIndex __iam = omp_get_thread_num(); - // This job. - _Job<_DifferenceType>& __my_job = __job[__iam * __stride]; + // This job. + _Job<_DifferenceType>& __my_job = __job[__iam * __stride]; - // Random number (for work stealing). - _ThreadIndex __victim; + // Random number (for work stealing). + _ThreadIndex __victim; - // Local value for reduction. - _Result __result = _Result(); + // Local value for reduction. + _Result __result = _Result(); - // Number of elements to steal in one attempt. - _DifferenceType __steal; + // Number of elements to steal in one attempt. + _DifferenceType __steal; - // Every thread has its own random number generator - // (modulo __num_threads). - _RandomNumber rand_gen(__iam, __num_threads); + // Every thread has its own random number generator + // (modulo __num_threads). + _RandomNumber __rand_gen(__iam, __num_threads); - // This thread is currently working. + // This thread is currently working. # pragma omp atomic - ++__busy; + ++__busy; - __iam_working = true; + __iam_working = true; - // How many jobs per thread? last thread gets the rest. - __my_job._M_first = - static_cast<_DifferenceType>(__iam * (__length / __num_threads)); + // How many jobs per thread? last thread gets the rest. + __my_job._M_first = static_cast<_DifferenceType> + (__iam * (__length / __num_threads)); - __my_job._M_last = (__iam == (__num_threads - 1)) ? - (__length - 1) : ((__iam + 1) * (__length / __num_threads) - 1); - __my_job._M_load = __my_job._M_last - __my_job._M_first + 1; + __my_job._M_last = (__iam == (__num_threads - 1) + ? (__length - 1) + : ((__iam + 1) * (__length / __num_threads) - 1)); + __my_job._M_load = __my_job._M_last - __my_job._M_first + 1; - // Init result with _M_first value (to have a base value for reduction) - if (__my_job._M_first <= __my_job._M_last) - { - // Cannot use volatile variable directly. - _DifferenceType __my_first = __my_job._M_first; - __result = __f(__op, __begin + __my_first); - ++__my_job._M_first; - --__my_job._M_load; - } + // Init result with _M_first value (to have a base value for reduction) + if (__my_job._M_first <= __my_job._M_last) + { + // Cannot use volatile variable directly. + _DifferenceType __my_first = __my_job._M_first; + __result = __f(__op, __begin + __my_first); + ++__my_job._M_first; + --__my_job._M_load; + } - _RAIter __current; + _RAIter __current; # pragma omp barrier - // Actual work phase - // Work on own or stolen current start - while (__busy > 0) - { - // Work until no productive thread left. + // Actual work phase + // Work on own or stolen current start + while (__busy > 0) + { + // Work until no productive thread left. # pragma omp flush(__busy) - // Thread has own work to do - while (__my_job._M_first <= __my_job._M_last) - { - // fetch-and-add call - // Reserve current job block (size __chunk_size) in my queue. - _DifferenceType __current_job = - __fetch_and_add<_DifferenceType>( - &(__my_job._M_first), __chunk_size); - - // Update _M_load, to make the three values consistent, - // _M_first might have been changed in the meantime - __my_job._M_load = __my_job._M_last - __my_job._M_first + 1; - for (_DifferenceType __job_counter = 0; - __job_counter < __chunk_size - && __current_job <= __my_job._M_last; - ++__job_counter) - { - // Yes: process it! - __current = __begin + __current_job; - ++__current_job; - - // Do actual work. - __result = __r(__result, __f(__op, __current)); - } + // Thread has own work to do + while (__my_job._M_first <= __my_job._M_last) + { + // fetch-and-add call + // Reserve current job block (size __chunk_size) in my queue. + _DifferenceType __current_job = + __fetch_and_add<_DifferenceType>(&(__my_job._M_first), + __chunk_size); + + // Update _M_load, to make the three values consistent, + // _M_first might have been changed in the meantime + __my_job._M_load = __my_job._M_last - __my_job._M_first + 1; + for (_DifferenceType __job_counter = 0; + __job_counter < __chunk_size + && __current_job <= __my_job._M_last; + ++__job_counter) + { + // Yes: process it! + __current = __begin + __current_job; + ++__current_job; + + // Do actual work. + __result = __r(__result, __f(__op, __current)); + } # pragma omp flush(__busy) - } + } - // After reaching this point, a thread's __job list is empty. - if (__iam_working) - { - // This thread no longer has work. + // After reaching this point, a thread's __job list is empty. + if (__iam_working) + { + // This thread no longer has work. # pragma omp atomic - --__busy; + --__busy; - __iam_working = false; - } + __iam_working = false; + } - _DifferenceType __supposed_first, __supposed_last, __supposed_load; - do - { - // Find random nonempty deque (not own), do consistency check. - __yield(); + _DifferenceType __supposed_first, __supposed_last, + __supposed_load; + do + { + // Find random nonempty deque (not own), do consistency check. + __yield(); # pragma omp flush(__busy) - __victim = rand_gen(); - __supposed_first = __job[__victim * __stride]._M_first; - __supposed_last = __job[__victim * __stride]._M_last; - __supposed_load = __job[__victim * __stride]._M_load; - } - while (__busy > 0 - && ((__supposed_load <= 0) - || ((__supposed_first + __supposed_load - 1) - != __supposed_last))); - - if (__busy == 0) - break; - - if (__supposed_load > 0) - { - // Has work and work to do. - // Number of elements to steal (at least one). - __steal = (__supposed_load < 2) ? 1 : __supposed_load / 2; - - // Push __victim's current start forward. - _DifferenceType __stolen_first = - __fetch_and_add<_DifferenceType>( - &(__job[__victim * __stride]._M_first), __steal); - _DifferenceType __stolen_try = - __stolen_first + __steal - _DifferenceType(1); - - __my_job._M_first = __stolen_first; - __my_job._M_last = - __gnu_parallel::min(__stolen_try, __supposed_last); - __my_job._M_load = __my_job._M_last - __my_job._M_first + 1; - - // Has potential work again. + __victim = __rand_gen(); + __supposed_first = __job[__victim * __stride]._M_first; + __supposed_last = __job[__victim * __stride]._M_last; + __supposed_load = __job[__victim * __stride]._M_load; + } + while (__busy > 0 + && ((__supposed_load <= 0) + || ((__supposed_first + __supposed_load - 1) + != __supposed_last))); + + if (__busy == 0) + break; + + if (__supposed_load > 0) + { + // Has work and work to do. + // Number of elements to steal (at least one). + __steal = (__supposed_load < 2) ? 1 : __supposed_load / 2; + + // Push __victim's current start forward. + _DifferenceType __stolen_first = + __fetch_and_add<_DifferenceType> + (&(__job[__victim * __stride]._M_first), __steal); + _DifferenceType __stolen_try = (__stolen_first + __steal + - _DifferenceType(1)); + + __my_job._M_first = __stolen_first; + __my_job._M_last = __gnu_parallel::min(__stolen_try, + __supposed_last); + __my_job._M_load = __my_job._M_last - __my_job._M_first + 1; + + // Has potential work again. # pragma omp atomic - ++__busy; - __iam_working = true; + ++__busy; + __iam_working = true; # pragma omp flush(__busy) - } + } # pragma omp flush(__busy) - } // end while __busy > 0 - // Add accumulated result to output. - omp_set_lock(&__output_lock); - __output = __r(__output, __result); - omp_unset_lock(&__output_lock); + } // end while __busy > 0 + // Add accumulated result to output. + omp_set_lock(&__output_lock); + __output = __r(__output, __result); + omp_unset_lock(&__output_lock); } - delete[] __job; + delete[] __job; - // Points to last element processed (needed as return value for - // some algorithms like transform) - __f._M_finish_iterator = __begin + __length; + // Points to last element processed (needed as return value for + // some algorithms like transform) + __f._M_finish_iterator = __begin + __length; - omp_destroy_lock(&__output_lock); + omp_destroy_lock(&__output_lock); - return __op; - } + return __op; + } } // end namespace #endif /* _GLIBCXX_PARALLEL_WORKSTEALING_H */ -- cgit v1.2.3