1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
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_CXX11_TENSOR_TENSOR_SHUFFLING_H
11 #define EIGEN_CXX11_TENSOR_TENSOR_SHUFFLING_H
12 
13 namespace Eigen {
14 
15 /** \class TensorShuffling
16   * \ingroup CXX11_Tensor_Module
17   *
18   * \brief Tensor shuffling class.
19   *
20   *
21   */
22 namespace internal {
23 template<typename Shuffle, typename XprType>
24 struct traits<TensorShufflingOp<Shuffle, XprType> > : public traits<XprType>
25 {
26   typedef typename XprType::Scalar Scalar;
27   typedef traits<XprType> XprTraits;
28   typedef typename XprTraits::StorageKind StorageKind;
29   typedef typename XprTraits::Index Index;
30   typedef typename XprType::Nested Nested;
31   typedef typename remove_reference<Nested>::type _Nested;
32   static const int NumDimensions = XprTraits::NumDimensions;
33   static const int Layout = XprTraits::Layout;
34 };
35 
36 template<typename Shuffle, typename XprType>
37 struct eval<TensorShufflingOp<Shuffle, XprType>, Eigen::Dense>
38 {
39   typedef const TensorShufflingOp<Shuffle, XprType>& type;
40 };
41 
42 template<typename Shuffle, typename XprType>
43 struct nested<TensorShufflingOp<Shuffle, XprType>, 1, typename eval<TensorShufflingOp<Shuffle, XprType> >::type>
44 {
45   typedef TensorShufflingOp<Shuffle, XprType> type;
46 };
47 
48 }  // end namespace internal
49 
50 
51 
52 template<typename Shuffle, typename XprType>
53 class TensorShufflingOp : public TensorBase<TensorShufflingOp<Shuffle, XprType> >
54 {
55   public:
56   typedef typename Eigen::internal::traits<TensorShufflingOp>::Scalar Scalar;
57   typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
58   typedef typename XprType::CoeffReturnType CoeffReturnType;
59   typedef typename Eigen::internal::nested<TensorShufflingOp>::type Nested;
60   typedef typename Eigen::internal::traits<TensorShufflingOp>::StorageKind StorageKind;
61   typedef typename Eigen::internal::traits<TensorShufflingOp>::Index Index;
62 
63   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorShufflingOp(const XprType& expr, const Shuffle& shuffle)
64       : m_xpr(expr), m_shuffle(shuffle) {}
65 
66     EIGEN_DEVICE_FUNC
67     const Shuffle& shufflePermutation() const { return m_shuffle; }
68 
69     EIGEN_DEVICE_FUNC
70     const typename internal::remove_all<typename XprType::Nested>::type&
71     expression() const { return m_xpr; }
72 
73     EIGEN_DEVICE_FUNC
74     EIGEN_STRONG_INLINE TensorShufflingOp& operator = (const TensorShufflingOp& other)
75     {
76       typedef TensorAssignOp<TensorShufflingOp, const TensorShufflingOp> Assign;
77       Assign assign(*this, other);
78       internal::TensorExecutor<const Assign, DefaultDevice>::run(assign, DefaultDevice());
79       return *this;
80     }
81 
82     template<typename OtherDerived>
83     EIGEN_DEVICE_FUNC
84     EIGEN_STRONG_INLINE TensorShufflingOp& operator = (const OtherDerived& other)
85     {
86       typedef TensorAssignOp<TensorShufflingOp, const OtherDerived> Assign;
87       Assign assign(*this, other);
88       internal::TensorExecutor<const Assign, DefaultDevice>::run(assign, DefaultDevice());
89       return *this;
90     }
91 
92   protected:
93     typename XprType::Nested m_xpr;
94     const Shuffle m_shuffle;
95 };
96 
97 
98 // Eval as rvalue
99 template<typename Shuffle, typename ArgType, typename Device>
100 struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
101 {
102   typedef TensorShufflingOp<Shuffle, ArgType> XprType;
103   typedef typename XprType::Index Index;
104   static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
105   typedef DSizes<Index, NumDims> Dimensions;
106   typedef typename XprType::Scalar Scalar;
107   typedef typename XprType::CoeffReturnType CoeffReturnType;
108   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
109   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
110 
111   enum {
112     IsAligned = false,
113     PacketAccess = (internal::packet_traits<Scalar>::size > 1),
114     Layout = TensorEvaluator<ArgType, Device>::Layout,
115     CoordAccess = false,  // to be implemented
116     RawAccess = false
117   };
118 
119   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
120       : m_impl(op.expression(), device)
121   {
122     const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
123     const Shuffle& shuffle = op.shufflePermutation();
124     for (int i = 0; i < NumDims; ++i) {
125       m_dimensions[i] = input_dims[shuffle[i]];
126     }
127 
128     array<Index, NumDims> inputStrides;
129 
130     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
131       inputStrides[0] = 1;
132       m_outputStrides[0] = 1;
133       for (int i = 1; i < NumDims; ++i) {
134         inputStrides[i] = inputStrides[i - 1] * input_dims[i - 1];
135         m_outputStrides[i] = m_outputStrides[i - 1] * m_dimensions[i - 1];
136       }
137     } else {
138       inputStrides[NumDims - 1] = 1;
139       m_outputStrides[NumDims - 1] = 1;
140       for (int i = NumDims - 2; i >= 0; --i) {
141         inputStrides[i] = inputStrides[i + 1] * input_dims[i + 1];
142         m_outputStrides[i] = m_outputStrides[i + 1] * m_dimensions[i + 1];
143       }
144     }
145 
146     for (int i = 0; i < NumDims; ++i) {
147       m_inputStrides[i] = inputStrides[shuffle[i]];
148     }
149   }
150 
151   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
152 
153   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/) {
154     m_impl.evalSubExprsIfNeeded(NULL);
155     return true;
156   }
157   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
158     m_impl.cleanup();
159   }
160 
161   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
162   {
163     return m_impl.coeff(srcCoeff(index));
164   }
165 
166   template<int LoadMode>
167   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
168   {
169     EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
170     eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
171 
172     EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
173     for (int i = 0; i < PacketSize; ++i) {
174       values[i] = coeff(index+i);
175     }
176     PacketReturnType rslt = internal::pload<PacketReturnType>(values);
177     return rslt;
178   }
179 
180   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
181     const double compute_cost = NumDims * (2 * TensorOpCost::AddCost<Index>() +
182                                            2 * TensorOpCost::MulCost<Index>() +
183                                            TensorOpCost::DivCost<Index>());
184     return m_impl.costPerCoeff(vectorized) +
185            TensorOpCost(0, 0, compute_cost, false /* vectorized */, PacketSize);
186   }
187 
188   EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
189 
190  protected:
191   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const {
192     Index inputIndex = 0;
193     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
194       for (int i = NumDims - 1; i > 0; --i) {
195         const Index idx = index / m_outputStrides[i];
196         inputIndex += idx * m_inputStrides[i];
197         index -= idx * m_outputStrides[i];
198       }
199       return inputIndex + index * m_inputStrides[0];
200     } else {
201       for (int i = 0; i < NumDims - 1; ++i) {
202         const Index idx = index / m_outputStrides[i];
203         inputIndex += idx * m_inputStrides[i];
204         index -= idx * m_outputStrides[i];
205       }
206       return inputIndex + index * m_inputStrides[NumDims - 1];
207     }
208   }
209 
210   Dimensions m_dimensions;
211   array<Index, NumDims> m_outputStrides;
212   array<Index, NumDims> m_inputStrides;
213   TensorEvaluator<ArgType, Device> m_impl;
214 };
215 
216 
217 // Eval as lvalue
218 template<typename Shuffle, typename ArgType, typename Device>
219 struct TensorEvaluator<TensorShufflingOp<Shuffle, ArgType>, Device>
220     : public TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
221 {
222   typedef TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device> Base;
223 
224   typedef TensorShufflingOp<Shuffle, ArgType> XprType;
225   typedef typename XprType::Index Index;
226   static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
227   typedef DSizes<Index, NumDims> Dimensions;
228   typedef typename XprType::Scalar Scalar;
229   typedef typename XprType::CoeffReturnType CoeffReturnType;
230   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
231   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
232 
233   enum {
234     IsAligned = false,
235     PacketAccess = (internal::packet_traits<Scalar>::size > 1),
236     RawAccess = false
237   };
238 
239   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
240       : Base(op, device)
241   { }
242 
243   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
244   {
245     return this->m_impl.coeffRef(this->srcCoeff(index));
246   }
247 
248   template <int StoreMode> EIGEN_STRONG_INLINE
249   void writePacket(Index index, const PacketReturnType& x)
250   {
251     EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
252 
253     EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
254     internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
255     for (int i = 0; i < PacketSize; ++i) {
256       this->coeffRef(index+i) = values[i];
257     }
258   }
259 };
260 
261 
262 } // end namespace Eigen
263 
264 #endif // EIGEN_CXX11_TENSOR_TENSOR_SHUFFLING_H
265