// -*- C++ -*- //===-- parallel_backend_tbb.h --------------------------------------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// #ifndef _PSTL_PARALLEL_BACKEND_TBB_H #define _PSTL_PARALLEL_BACKEND_TBB_H #include #include #include "parallel_backend_utils.h" // Bring in minimal required subset of Intel TBB #include #include #include #include #include #include #include #if TBB_INTERFACE_VERSION < 10000 # error Intel(R) Threading Building Blocks 2018 is required; older versions are not supported. #endif namespace __pstl { namespace __par_backend { //! Raw memory buffer with automatic freeing and no exceptions. /** Some of our algorithms need to start with raw memory buffer, not an initialize array, because initialization/destruction would make the span be at least O(N). */ // tbb::allocator can improve performance in some cases. template class __buffer { tbb::tbb_allocator<_Tp> _M_allocator; _Tp* _M_ptr; const std::size_t _M_buf_size; __buffer(const __buffer&) = delete; void operator=(const __buffer&) = delete; public: //! Try to obtain buffer of given size to store objects of _Tp type __buffer(std::size_t n) : _M_allocator(), _M_ptr(_M_allocator.allocate(n)), _M_buf_size(n) {} //! True if buffer was successfully obtained, zero otherwise. operator bool() const { return _M_ptr != NULL; } //! Return pointer to buffer, or NULL if buffer could not be obtained. _Tp* get() const { return _M_ptr; } //! Destroy buffer ~__buffer() { _M_allocator.deallocate(_M_ptr, _M_buf_size); } }; // Wrapper for tbb::task inline void __cancel_execution() { tbb::task::self().group()->cancel_group_execution(); } //------------------------------------------------------------------------ // parallel_for //------------------------------------------------------------------------ template class __parallel_for_body { public: __parallel_for_body(const _RealBody& __body) : _M_body(__body) {} __parallel_for_body(const __parallel_for_body& __body) : _M_body(__body._M_body) {} void operator()(const tbb::blocked_range<_Index>& __range) const { _M_body(__range.begin(), __range.end()); } private: _RealBody _M_body; }; //! Evaluation of brick f[i,j) for each subrange [i,j) of [first,last) // wrapper over tbb::parallel_for template void __parallel_for(_ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f) { tbb::this_task_arena::isolate([=]() { tbb::parallel_for(tbb::blocked_range<_Index>(__first, __last), __parallel_for_body<_Index, _Fp>(__f)); }); } //! Evaluation of brick f[i,j) for each subrange [i,j) of [first,last) // wrapper over tbb::parallel_reduce template _Value __parallel_reduce(_ExecutionPolicy&&, _Index __first, _Index __last, const _Value& __identity, const _RealBody& __real_body, const _Reduction& __reduction) { return tbb::this_task_arena::isolate([__first, __last, &__identity, &__real_body, &__reduction]() -> _Value { return tbb::parallel_reduce( tbb::blocked_range<_Index>(__first, __last), __identity, [__real_body](const tbb::blocked_range<_Index>& __r, const _Value& __value) -> _Value { return __real_body(__r.begin(), __r.end(), __value); }, __reduction); }); } //------------------------------------------------------------------------ // parallel_transform_reduce // // Notation: // r(i,j,init) returns reduction of init with reduction over [i,j) // u(i) returns f(i,i+1,identity) for a hypothetical left identity element of r // c(x,y) combines values x and y that were the result of r or u //------------------------------------------------------------------------ template struct __par_trans_red_body { alignas(_Tp) char _M_sum_storage[sizeof(_Tp)]; // Holds generalized non-commutative sum when has_sum==true _Rp _M_brick_reduce; // Most likely to have non-empty layout _Up _M_u; _Cp _M_combine; bool _M_has_sum; // Put last to minimize size of class _Tp& sum() { _PSTL_ASSERT_MSG(_M_has_sum, "sum expected"); return *(_Tp*)_M_sum_storage; } __par_trans_red_body(_Up __u, _Tp __init, _Cp __c, _Rp __r) : _M_brick_reduce(__r), _M_u(__u), _M_combine(__c), _M_has_sum(true) { new (_M_sum_storage) _Tp(__init); } __par_trans_red_body(__par_trans_red_body& __left, tbb::split) : _M_brick_reduce(__left._M_brick_reduce), _M_u(__left._M_u), _M_combine(__left._M_combine), _M_has_sum(false) { } ~__par_trans_red_body() { // 17.6.5.12 tells us to not worry about catching exceptions from destructors. if (_M_has_sum) sum().~_Tp(); } void join(__par_trans_red_body& __rhs) { sum() = _M_combine(sum(), __rhs.sum()); } void operator()(const tbb::blocked_range<_Index>& __range) { _Index __i = __range.begin(); _Index __j = __range.end(); if (!_M_has_sum) { _PSTL_ASSERT_MSG(__range.size() > 1, "there should be at least 2 elements"); new (&_M_sum_storage) _Tp(_M_combine(_M_u(__i), _M_u(__i + 1))); // The condition i+1 < j is provided by the grain size of 3 _M_has_sum = true; std::advance(__i, 2); if (__i == __j) return; } sum() = _M_brick_reduce(__i, __j, sum()); } }; template _Tp __parallel_transform_reduce(_ExecutionPolicy&&, _Index __first, _Index __last, _Up __u, _Tp __init, _Cp __combine, _Rp __brick_reduce) { __par_backend::__par_trans_red_body<_Index, _Up, _Tp, _Cp, _Rp> __body(__u, __init, __combine, __brick_reduce); // The grain size of 3 is used in order to provide mininum 2 elements for each body tbb::this_task_arena::isolate( [__first, __last, &__body]() { tbb::parallel_reduce(tbb::blocked_range<_Index>(__first, __last, 3), __body); }); return __body.sum(); } //------------------------------------------------------------------------ // parallel_scan //------------------------------------------------------------------------ template class __trans_scan_body { alignas(_Tp) char _M_sum_storage[sizeof(_Tp)]; // Holds generalized non-commutative sum when has_sum==true _Rp _M_brick_reduce; // Most likely to have non-empty layout _Up _M_u; _Cp _M_combine; _Sp _M_scan; bool _M_has_sum; // Put last to minimize size of class public: __trans_scan_body(_Up __u, _Tp __init, _Cp __combine, _Rp __reduce, _Sp __scan) : _M_brick_reduce(__reduce), _M_u(__u), _M_combine(__combine), _M_scan(__scan), _M_has_sum(true) { new (_M_sum_storage) _Tp(__init); } __trans_scan_body(__trans_scan_body& __b, tbb::split) : _M_brick_reduce(__b._M_brick_reduce), _M_u(__b._M_u), _M_combine(__b._M_combine), _M_scan(__b._M_scan), _M_has_sum(false) { } ~__trans_scan_body() { // 17.6.5.12 tells us to not worry about catching exceptions from destructors. if (_M_has_sum) sum().~_Tp(); } _Tp& sum() const { _PSTL_ASSERT_MSG(_M_has_sum, "sum expected"); return *const_cast<_Tp*>(reinterpret_cast<_Tp const*>(_M_sum_storage)); } void operator()(const tbb::blocked_range<_Index>& __range, tbb::pre_scan_tag) { _Index __i = __range.begin(); _Index __j = __range.end(); if (!_M_has_sum) { new (&_M_sum_storage) _Tp(_M_u(__i)); _M_has_sum = true; ++__i; if (__i == __j) return; } sum() = _M_brick_reduce(__i, __j, sum()); } void operator()(const tbb::blocked_range<_Index>& __range, tbb::final_scan_tag) { sum() = _M_scan(__range.begin(), __range.end(), sum()); } void reverse_join(__trans_scan_body& __a) { if (_M_has_sum) { sum() = _M_combine(__a.sum(), sum()); } else { new (&_M_sum_storage) _Tp(__a.sum()); _M_has_sum = true; } } void assign(__trans_scan_body& __b) { sum() = __b.sum(); } }; template _Index __split(_Index __m) { _Index __k = 1; while (2 * __k < __m) __k *= 2; return __k; } //------------------------------------------------------------------------ // __parallel_strict_scan //------------------------------------------------------------------------ template void __upsweep(_Index __i, _Index __m, _Index __tilesize, _Tp* __r, _Index __lastsize, _Rp __reduce, _Cp __combine) { if (__m == 1) __r[0] = __reduce(__i * __tilesize, __lastsize); else { _Index __k = __split(__m); tbb::parallel_invoke( [=] { __par_backend::__upsweep(__i, __k, __tilesize, __r, __tilesize, __reduce, __combine); }, [=] { __par_backend::__upsweep(__i + __k, __m - __k, __tilesize, __r + __k, __lastsize, __reduce, __combine); }); if (__m == 2 * __k) __r[__m - 1] = __combine(__r[__k - 1], __r[__m - 1]); } } template void __downsweep(_Index __i, _Index __m, _Index __tilesize, _Tp* __r, _Index __lastsize, _Tp __initial, _Cp __combine, _Sp __scan) { if (__m == 1) __scan(__i * __tilesize, __lastsize, __initial); else { const _Index __k = __split(__m); tbb::parallel_invoke( [=] { __par_backend::__downsweep(__i, __k, __tilesize, __r, __tilesize, __initial, __combine, __scan); }, // Assumes that __combine never throws. //TODO: Consider adding a requirement for user functors to be constant. [=, &__combine] { __par_backend::__downsweep(__i + __k, __m - __k, __tilesize, __r + __k, __lastsize, __combine(__initial, __r[__k - 1]), __combine, __scan); }); } } // Adapted from Intel(R) Cilk(TM) version from cilkpub. // Let i:len denote a counted interval of length n starting at i. s denotes a generalized-sum value. // Expected actions of the functors are: // reduce(i,len) -> s -- return reduction value of i:len. // combine(s1,s2) -> s -- return merged sum // apex(s) -- do any processing necessary between reduce and scan. // scan(i,len,initial) -- perform scan over i:len starting with initial. // The initial range 0:n is partitioned into consecutive subranges. // reduce and scan are each called exactly once per subrange. // Thus callers can rely upon side effects in reduce. // combine must not throw an exception. // apex is called exactly once, after all calls to reduce and before all calls to scan. // For example, it's useful for allocating a __buffer used by scan but whose size is the sum of all reduction values. // T must have a trivial constructor and destructor. template void __parallel_strict_scan(_ExecutionPolicy&&, _Index __n, _Tp __initial, _Rp __reduce, _Cp __combine, _Sp __scan, _Ap __apex) { tbb::this_task_arena::isolate([=, &__combine]() { if (__n > 1) { _Index __p = tbb::this_task_arena::max_concurrency(); const _Index __slack = 4; _Index __tilesize = (__n - 1) / (__slack * __p) + 1; _Index __m = (__n - 1) / __tilesize; __buffer<_Tp> __buf(__m + 1); _Tp* __r = __buf.get(); __par_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce, __combine); // When __apex is a no-op and __combine has no side effects, a good optimizer // should be able to eliminate all code between here and __apex. // Alternatively, provide a default value for __apex that can be // recognized by metaprogramming that conditionlly executes the following. size_t __k = __m + 1; _Tp __t = __r[__k - 1]; while ((__k &= __k - 1)) __t = __combine(__r[__k - 1], __t); __apex(__combine(__initial, __t)); __par_backend::__downsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __initial, __combine, __scan); return; } // Fewer than 2 elements in sequence, or out of memory. Handle has single block. _Tp __sum = __initial; if (__n) __sum = __combine(__sum, __reduce(_Index(0), __n)); __apex(__sum); if (__n) __scan(_Index(0), __n, __initial); }); } template _Tp __parallel_transform_scan(_ExecutionPolicy&&, _Index __n, _Up __u, _Tp __init, _Cp __combine, _Rp __brick_reduce, _Sp __scan) { __trans_scan_body<_Index, _Up, _Tp, _Cp, _Rp, _Sp> __body(__u, __init, __combine, __brick_reduce, __scan); auto __range = tbb::blocked_range<_Index>(0, __n); tbb::this_task_arena::isolate([__range, &__body]() { tbb::parallel_scan(__range, __body); }); return __body.sum(); } //------------------------------------------------------------------------ // parallel_stable_sort //------------------------------------------------------------------------ //------------------------------------------------------------------------ // stable_sort utilities // // These are used by parallel implementations but do not depend on them. //------------------------------------------------------------------------ template class __merge_task : public tbb::task { /*override*/ tbb::task* execute(); _RandomAccessIterator1 _M_xs, _M_xe; _RandomAccessIterator2 _M_ys, _M_ye; _RandomAccessIterator3 _M_zs; _Compare _M_comp; _Cleanup _M_cleanup; _LeafMerge _M_leaf_merge; public: __merge_task(_RandomAccessIterator1 __xs, _RandomAccessIterator1 __xe, _RandomAccessIterator2 __ys, _RandomAccessIterator2 __ye, _RandomAccessIterator3 __zs, _Compare __comp, _Cleanup __cleanup, _LeafMerge __leaf_merge) : _M_xs(__xs), _M_xe(__xe), _M_ys(__ys), _M_ye(__ye), _M_zs(__zs), _M_comp(__comp), _M_cleanup(__cleanup), _M_leaf_merge(__leaf_merge) { } }; #define _PSTL_MERGE_CUT_OFF 2000 template tbb::task* __merge_task<_RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator3, __M_Compare, _Cleanup, _LeafMerge>::execute() { typedef typename std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType1; typedef typename std::iterator_traits<_RandomAccessIterator2>::difference_type _DifferenceType2; typedef typename std::common_type<_DifferenceType1, _DifferenceType2>::type _SizeType; const _SizeType __n = (_M_xe - _M_xs) + (_M_ye - _M_ys); const _SizeType __merge_cut_off = _PSTL_MERGE_CUT_OFF; if (__n <= __merge_cut_off) { _M_leaf_merge(_M_xs, _M_xe, _M_ys, _M_ye, _M_zs, _M_comp); //we clean the buffer one time on last step of the sort _M_cleanup(_M_xs, _M_xe); _M_cleanup(_M_ys, _M_ye); return nullptr; } else { _RandomAccessIterator1 __xm; _RandomAccessIterator2 __ym; if (_M_xe - _M_xs < _M_ye - _M_ys) { __ym = _M_ys + (_M_ye - _M_ys) / 2; __xm = std::upper_bound(_M_xs, _M_xe, *__ym, _M_comp); } else { __xm = _M_xs + (_M_xe - _M_xs) / 2; __ym = std::lower_bound(_M_ys, _M_ye, *__xm, _M_comp); } const _RandomAccessIterator3 __zm = _M_zs + ((__xm - _M_xs) + (__ym - _M_ys)); tbb::task* __right = new (tbb::task::allocate_additional_child_of(*parent())) __merge_task(__xm, _M_xe, __ym, _M_ye, __zm, _M_comp, _M_cleanup, _M_leaf_merge); tbb::task::spawn(*__right); tbb::task::recycle_as_continuation(); _M_xe = __xm; _M_ye = __ym; } return this; } template class __stable_sort_task : public tbb::task { public: typedef typename std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType1; typedef typename std::iterator_traits<_RandomAccessIterator2>::difference_type _DifferenceType2; typedef typename std::common_type<_DifferenceType1, _DifferenceType2>::type _SizeType; private: /*override*/ tbb::task* execute(); _RandomAccessIterator1 _M_xs, _M_xe; _RandomAccessIterator2 _M_zs; _Compare _M_comp; _LeafSort _M_leaf_sort; int32_t _M_inplace; _SizeType _M_nsort; public: __stable_sort_task(_RandomAccessIterator1 __xs, _RandomAccessIterator1 __xe, _RandomAccessIterator2 __zs, int32_t __inplace, _Compare __comp, _LeafSort __leaf_sort, _SizeType __n) : _M_xs(__xs), _M_xe(__xe), _M_zs(__zs), _M_comp(__comp), _M_leaf_sort(__leaf_sort), _M_inplace(__inplace), _M_nsort(__n) { } }; //! Binary operator that does nothing struct __binary_no_op { template void operator()(_Tp, _Tp) { } }; #define _PSTL_STABLE_SORT_CUT_OFF 500 template tbb::task* __stable_sort_task<_RandomAccessIterator1, _RandomAccessIterator2, _Compare, _LeafSort>::execute() { const _SizeType __n = _M_xe - _M_xs; const _SizeType __nmerge = _M_nsort > 0 ? _M_nsort : __n; const _SizeType __sort_cut_off = _PSTL_STABLE_SORT_CUT_OFF; if (__n <= __sort_cut_off) { _M_leaf_sort(_M_xs, _M_xe, _M_comp); if (_M_inplace != 2) __par_backend::__init_buf(_M_xs, _M_xe, _M_zs, _M_inplace == 0); return NULL; } else { const _RandomAccessIterator1 __xm = _M_xs + __n / 2; const _RandomAccessIterator2 __zm = _M_zs + (__xm - _M_xs); const _RandomAccessIterator2 __ze = _M_zs + __n; task* __m; auto __move_values = [](_RandomAccessIterator2 __x, _RandomAccessIterator1 __z) { *__z = std::move(*__x); }; auto __move_sequences = [](_RandomAccessIterator2 __first1, _RandomAccessIterator2 __last1, _RandomAccessIterator1 __first2) { return std::move(__first1, __last1, __first2); }; if (_M_inplace == 2) __m = new (tbb::task::allocate_continuation()) __merge_task<_RandomAccessIterator2, _RandomAccessIterator2, _RandomAccessIterator1, _Compare, __serial_destroy, __par_backend::__serial_move_merge>( _M_zs, __zm, __zm, __ze, _M_xs, _M_comp, __serial_destroy(), __par_backend::__serial_move_merge( __nmerge, __move_values, __move_sequences)); else if (_M_inplace) __m = new (tbb::task::allocate_continuation()) __merge_task<_RandomAccessIterator2, _RandomAccessIterator2, _RandomAccessIterator1, _Compare, __par_backend::__binary_no_op, __par_backend::__serial_move_merge>( _M_zs, __zm, __zm, __ze, _M_xs, _M_comp, __par_backend::__binary_no_op(), __par_backend::__serial_move_merge( __nmerge, __move_values, __move_sequences)); else { auto __move_values = [](_RandomAccessIterator1 __x, _RandomAccessIterator2 __z) { *__z = std::move(*__x); }; auto __move_sequences = [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2) { return std::move(__first1, __last1, __first2); }; __m = new (tbb::task::allocate_continuation()) __merge_task<_RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2, _Compare, __par_backend::__binary_no_op, __par_backend::__serial_move_merge>( _M_xs, __xm, __xm, _M_xe, _M_zs, _M_comp, __par_backend::__binary_no_op(), __par_backend::__serial_move_merge( __nmerge, __move_values, __move_sequences)); } __m->set_ref_count(2); task* __right = new (__m->allocate_child()) __stable_sort_task(__xm, _M_xe, __zm, !_M_inplace, _M_comp, _M_leaf_sort, __nmerge); tbb::task::spawn(*__right); tbb::task::recycle_as_child_of(*__m); _M_xe = __xm; _M_inplace = !_M_inplace; } return this; } template void __parallel_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator __xs, _RandomAccessIterator __xe, _Compare __comp, _LeafSort __leaf_sort, std::size_t __nsort = 0) { tbb::this_task_arena::isolate([=, &__nsort]() { //sorting based on task tree and parallel merge typedef typename std::iterator_traits<_RandomAccessIterator>::value_type _ValueType; typedef typename std::iterator_traits<_RandomAccessIterator>::difference_type _DifferenceType; const _DifferenceType __n = __xe - __xs; if (__nsort == 0) __nsort = __n; const _DifferenceType __sort_cut_off = _PSTL_STABLE_SORT_CUT_OFF; if (__n > __sort_cut_off) { _PSTL_ASSERT(__nsort > 0 && __nsort <= __n); __buffer<_ValueType> __buf(__n); using tbb::task; task::spawn_root_and_wait(*new (task::allocate_root()) __stable_sort_task<_RandomAccessIterator, _ValueType*, _Compare, _LeafSort>( __xs, __xe, (_ValueType*)__buf.get(), 2, __comp, __leaf_sort, __nsort)); return; } //serial sort __leaf_sort(__xs, __xe, __comp); }); } //------------------------------------------------------------------------ // parallel_merge //------------------------------------------------------------------------ template void __parallel_merge(_ExecutionPolicy&&, _RandomAccessIterator1 __xs, _RandomAccessIterator1 __xe, _RandomAccessIterator2 __ys, _RandomAccessIterator2 __ye, _RandomAccessIterator3 __zs, _Compare __comp, _LeafMerge __leaf_merge) { typedef typename std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType1; typedef typename std::iterator_traits<_RandomAccessIterator2>::difference_type _DifferenceType2; typedef typename std::common_type<_DifferenceType1, _DifferenceType2>::type _SizeType; const _SizeType __n = (__xe - __xs) + (__ye - __ys); const _SizeType __merge_cut_off = _PSTL_MERGE_CUT_OFF; if (__n <= __merge_cut_off) { // Fall back on serial merge __leaf_merge(__xs, __xe, __ys, __ye, __zs, __comp); } else { tbb::this_task_arena::isolate([=]() { typedef __merge_task<_RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator3, _Compare, __par_backend::__binary_no_op, _LeafMerge> _TaskType; tbb::task::spawn_root_and_wait(*new (tbb::task::allocate_root()) _TaskType( __xs, __xe, __ys, __ye, __zs, __comp, __par_backend::__binary_no_op(), __leaf_merge)); }); } } //------------------------------------------------------------------------ // parallel_invoke //------------------------------------------------------------------------ template void __parallel_invoke(_ExecutionPolicy&&, _F1&& __f1, _F2&& __f2) { //TODO: a version of tbb::this_task_arena::isolate with variadic arguments pack should be added in the future tbb::this_task_arena::isolate([&]() { tbb::parallel_invoke(std::forward<_F1>(__f1), std::forward<_F2>(__f2)); }); } } // namespace __par_backend } // namespace __pstl #endif /* _PSTL_PARALLEL_BACKEND_TBB_H */