xref: /aosp_15_r20/external/eigen/Eigen/src/Core/PartialReduxEvaluator.h (revision bf2c37156dfe67e5dfebd6d394bad8b2ab5804d4)
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2011-2018 Gael Guennebaud <[email protected]>
5 //
6 // This Source Code Form is subject to the terms of the Mozilla
7 // Public License v. 2.0. If a copy of the MPL was not distributed
8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9 
10 #ifndef EIGEN_PARTIALREDUX_H
11 #define EIGEN_PARTIALREDUX_H
12 
13 namespace Eigen {
14 
15 namespace internal {
16 
17 
18 /***************************************************************************
19 *
20 * This file provides evaluators for partial reductions.
21 * There are two modes:
22 *
23 *  - scalar path: simply calls the respective function on the column or row.
24 *    -> nothing special here, all the tricky part is handled by the return
25 *       types of VectorwiseOp's members. They embed the functor calling the
26 *       respective DenseBase's member function.
27 *
28 *  - vectorized path: implements a packet-wise reductions followed by
29 *    some (optional) processing of the outcome, e.g., division by n for mean.
30 *
31 * For the vectorized path let's observe that the packet-size and outer-unrolling
32 * are both decided by the assignement logic. So all we have to do is to decide
33 * on the inner unrolling.
34 *
35 * For the unrolling, we can reuse "internal::redux_vec_unroller" from Redux.h,
36 * but be need to be careful to specify correct increment.
37 *
38 ***************************************************************************/
39 
40 
41 /* logic deciding a strategy for unrolling of vectorized paths */
42 template<typename Func, typename Evaluator>
43 struct packetwise_redux_traits
44 {
45   enum {
46     OuterSize = int(Evaluator::IsRowMajor) ? Evaluator::RowsAtCompileTime : Evaluator::ColsAtCompileTime,
47     Cost = OuterSize == Dynamic ? HugeCost
48          : OuterSize * Evaluator::CoeffReadCost + (OuterSize-1) * functor_traits<Func>::Cost,
49     Unrolling = Cost <= EIGEN_UNROLLING_LIMIT ? CompleteUnrolling : NoUnrolling
50   };
51 
52 };
53 
54 /* Value to be returned when size==0 , by default let's return 0 */
55 template<typename PacketType,typename Func>
56 EIGEN_DEVICE_FUNC
packetwise_redux_empty_value(const Func &)57 PacketType packetwise_redux_empty_value(const Func& ) { return pset1<PacketType>(0); }
58 
59 /* For products the default is 1 */
60 template<typename PacketType,typename Scalar>
61 EIGEN_DEVICE_FUNC
packetwise_redux_empty_value(const scalar_product_op<Scalar,Scalar> &)62 PacketType packetwise_redux_empty_value(const scalar_product_op<Scalar,Scalar>& ) { return pset1<PacketType>(1); }
63 
64 /* Perform the actual reduction */
65 template<typename Func, typename Evaluator,
66          int Unrolling = packetwise_redux_traits<Func, Evaluator>::Unrolling
67 >
68 struct packetwise_redux_impl;
69 
70 /* Perform the actual reduction with unrolling */
71 template<typename Func, typename Evaluator>
72 struct packetwise_redux_impl<Func, Evaluator, CompleteUnrolling>
73 {
74   typedef redux_novec_unroller<Func,Evaluator, 0, Evaluator::SizeAtCompileTime> Base;
75   typedef typename Evaluator::Scalar Scalar;
76 
77   template<typename PacketType>
78   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE
79   PacketType run(const Evaluator &eval, const Func& func, Index /*size*/)
80   {
81     return redux_vec_unroller<Func, Evaluator, 0, packetwise_redux_traits<Func, Evaluator>::OuterSize>::template run<PacketType>(eval,func);
82   }
83 };
84 
85 /* Add a specialization of redux_vec_unroller for size==0 at compiletime.
86  * This specialization is not required for general reductions, which is
87  * why it is defined here.
88  */
89 template<typename Func, typename Evaluator, int Start>
90 struct redux_vec_unroller<Func, Evaluator, Start, 0>
91 {
92   template<typename PacketType>
93   EIGEN_DEVICE_FUNC
94   static EIGEN_STRONG_INLINE PacketType run(const Evaluator &, const Func& f)
95   {
96     return packetwise_redux_empty_value<PacketType>(f);
97   }
98 };
99 
100 /* Perform the actual reduction for dynamic sizes */
101 template<typename Func, typename Evaluator>
102 struct packetwise_redux_impl<Func, Evaluator, NoUnrolling>
103 {
104   typedef typename Evaluator::Scalar Scalar;
105   typedef typename redux_traits<Func, Evaluator>::PacketType PacketScalar;
106 
107   template<typename PacketType>
108   EIGEN_DEVICE_FUNC
109   static PacketType run(const Evaluator &eval, const Func& func, Index size)
110   {
111     if(size==0)
112       return packetwise_redux_empty_value<PacketType>(func);
113 
114     const Index size4 = (size-1)&(~3);
115     PacketType p = eval.template packetByOuterInner<Unaligned,PacketType>(0,0);
116     Index i = 1;
117     // This loop is optimized for instruction pipelining:
118     // - each iteration generates two independent instructions
119     // - thanks to branch prediction and out-of-order execution we have independent instructions across loops
120     for(; i<size4; i+=4)
121       p = func.packetOp(p,
122             func.packetOp(
123               func.packetOp(eval.template packetByOuterInner<Unaligned,PacketType>(i+0,0),eval.template packetByOuterInner<Unaligned,PacketType>(i+1,0)),
124               func.packetOp(eval.template packetByOuterInner<Unaligned,PacketType>(i+2,0),eval.template packetByOuterInner<Unaligned,PacketType>(i+3,0))));
125     for(; i<size; ++i)
126       p = func.packetOp(p, eval.template packetByOuterInner<Unaligned,PacketType>(i,0));
127     return p;
128   }
129 };
130 
131 template< typename ArgType, typename MemberOp, int Direction>
132 struct evaluator<PartialReduxExpr<ArgType, MemberOp, Direction> >
133   : evaluator_base<PartialReduxExpr<ArgType, MemberOp, Direction> >
134 {
135   typedef PartialReduxExpr<ArgType, MemberOp, Direction> XprType;
136   typedef typename internal::nested_eval<ArgType,1>::type ArgTypeNested;
137   typedef typename internal::add_const_on_value_type<ArgTypeNested>::type ConstArgTypeNested;
138   typedef typename internal::remove_all<ArgTypeNested>::type ArgTypeNestedCleaned;
139   typedef typename ArgType::Scalar InputScalar;
140   typedef typename XprType::Scalar Scalar;
141   enum {
142     TraversalSize = Direction==int(Vertical) ? int(ArgType::RowsAtCompileTime) :  int(ArgType::ColsAtCompileTime)
143   };
144   typedef typename MemberOp::template Cost<int(TraversalSize)> CostOpType;
145   enum {
146     CoeffReadCost = TraversalSize==Dynamic ? HugeCost
147                   : TraversalSize==0 ? 1
148                   : int(TraversalSize) * int(evaluator<ArgType>::CoeffReadCost) + int(CostOpType::value),
149 
150     _ArgFlags = evaluator<ArgType>::Flags,
151 
152     _Vectorizable =  bool(int(_ArgFlags)&PacketAccessBit)
153                   && bool(MemberOp::Vectorizable)
154                   && (Direction==int(Vertical) ? bool(_ArgFlags&RowMajorBit) : (_ArgFlags&RowMajorBit)==0)
155                   && (TraversalSize!=0),
156 
157     Flags = (traits<XprType>::Flags&RowMajorBit)
158           | (evaluator<ArgType>::Flags&(HereditaryBits&(~RowMajorBit)))
159           | (_Vectorizable ? PacketAccessBit : 0)
160           | LinearAccessBit,
161 
162     Alignment = 0 // FIXME this will need to be improved once PartialReduxExpr is vectorized
163   };
164 
165   EIGEN_DEVICE_FUNC explicit evaluator(const XprType xpr)
166     : m_arg(xpr.nestedExpression()), m_functor(xpr.functor())
167   {
168     EIGEN_INTERNAL_CHECK_COST_VALUE(TraversalSize==Dynamic ? HugeCost : (TraversalSize==0 ? 1 : int(CostOpType::value)));
169     EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
170   }
171 
172   typedef typename XprType::CoeffReturnType CoeffReturnType;
173 
174   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
175   const Scalar coeff(Index i, Index j) const
176   {
177     return coeff(Direction==Vertical ? j : i);
178   }
179 
180   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
181   const Scalar coeff(Index index) const
182   {
183     return m_functor(m_arg.template subVector<DirectionType(Direction)>(index));
184   }
185 
186   template<int LoadMode,typename PacketType>
187   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
188   PacketType packet(Index i, Index j) const
189   {
190     return packet<LoadMode,PacketType>(Direction==Vertical ? j : i);
191   }
192 
193   template<int LoadMode,typename PacketType>
194   EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
195   PacketType packet(Index idx) const
196   {
197     enum { PacketSize = internal::unpacket_traits<PacketType>::size };
198     typedef Block<const ArgTypeNestedCleaned,
199                   Direction==Vertical ? int(ArgType::RowsAtCompileTime) : int(PacketSize),
200                   Direction==Vertical ? int(PacketSize) : int(ArgType::ColsAtCompileTime),
201                   true /* InnerPanel */> PanelType;
202 
203     PanelType panel(m_arg,
204                     Direction==Vertical ? 0 : idx,
205                     Direction==Vertical ? idx : 0,
206                     Direction==Vertical ? m_arg.rows() : Index(PacketSize),
207                     Direction==Vertical ? Index(PacketSize) : m_arg.cols());
208 
209     // FIXME
210     // See bug 1612, currently if PacketSize==1 (i.e. complex<double> with 128bits registers) then the storage-order of panel get reversed
211     // and methods like packetByOuterInner do not make sense anymore in this context.
212     // So let's just by pass "vectorization" in this case:
213     if(PacketSize==1)
214       return internal::pset1<PacketType>(coeff(idx));
215 
216     typedef typename internal::redux_evaluator<PanelType> PanelEvaluator;
217     PanelEvaluator panel_eval(panel);
218     typedef typename MemberOp::BinaryOp BinaryOp;
219     PacketType p = internal::packetwise_redux_impl<BinaryOp,PanelEvaluator>::template run<PacketType>(panel_eval,m_functor.binaryFunc(),m_arg.outerSize());
220     return p;
221   }
222 
223 protected:
224   ConstArgTypeNested m_arg;
225   const MemberOp m_functor;
226 };
227 
228 } // end namespace internal
229 
230 } // end namespace Eigen
231 
232 #endif // EIGEN_PARTIALREDUX_H
233