• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2015 Eugene Brevdo <ebrevdo@gmail.com>
5 //                    Benoit Steiner <benoit.steiner.goog@gmail.com>
6 //
7 // This Source Code Form is subject to the terms of the Mozilla
8 // Public License v. 2.0. If a copy of the MPL was not distributed
9 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
10 
11 #ifndef EIGEN_CXX11_TENSOR_TENSOR_ARG_MAX_H
12 #define EIGEN_CXX11_TENSOR_TENSOR_ARG_MAX_H
13 
14 namespace Eigen {
15 namespace internal {
16 
17 /** \class TensorIndexTuple
18   * \ingroup CXX11_Tensor_Module
19   *
20   * \brief Tensor + Index Tuple class.
21   *
22   *
23   */
24 template<typename XprType>
25 struct traits<TensorIndexTupleOp<XprType> > : public traits<XprType>
26 {
27   typedef traits<XprType> XprTraits;
28   typedef typename XprTraits::StorageKind StorageKind;
29   typedef typename XprTraits::Index Index;
30   typedef Tuple<Index, typename XprTraits::Scalar> Scalar;
31   typedef typename XprType::Nested Nested;
32   typedef typename remove_reference<Nested>::type _Nested;
33   static const int NumDimensions = XprTraits::NumDimensions;
34   static const int Layout = XprTraits::Layout;
35 };
36 
37 template<typename XprType>
38 struct eval<TensorIndexTupleOp<XprType>, Eigen::Dense>
39 {
40   typedef const TensorIndexTupleOp<XprType>EIGEN_DEVICE_REF type;
41 };
42 
43 template<typename XprType>
44 struct nested<TensorIndexTupleOp<XprType>, 1,
45               typename eval<TensorIndexTupleOp<XprType> >::type>
46 {
47   typedef TensorIndexTupleOp<XprType> type;
48 };
49 
50 }  // end namespace internal
51 
52 template<typename XprType>
53 class TensorIndexTupleOp : public TensorBase<TensorIndexTupleOp<XprType>, ReadOnlyAccessors>
54 {
55   public:
56   typedef typename Eigen::internal::traits<TensorIndexTupleOp>::Scalar Scalar;
57   typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
58   typedef typename Eigen::internal::nested<TensorIndexTupleOp>::type Nested;
59   typedef typename Eigen::internal::traits<TensorIndexTupleOp>::StorageKind StorageKind;
60   typedef typename Eigen::internal::traits<TensorIndexTupleOp>::Index Index;
61   typedef Tuple<Index, typename XprType::CoeffReturnType> CoeffReturnType;
62 
63   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIndexTupleOp(const XprType& expr)
64       : m_xpr(expr) {}
65 
66   EIGEN_DEVICE_FUNC
67   const typename internal::remove_all<typename XprType::Nested>::type&
68   expression() const { return m_xpr; }
69 
70   protected:
71     typename XprType::Nested m_xpr;
72 };
73 
74 // Eval as rvalue
75 template<typename ArgType, typename Device>
76 struct TensorEvaluator<const TensorIndexTupleOp<ArgType>, Device>
77 {
78   typedef TensorIndexTupleOp<ArgType> XprType;
79   typedef typename XprType::Index Index;
80   typedef typename XprType::Scalar Scalar;
81   typedef typename XprType::CoeffReturnType CoeffReturnType;
82 
83   typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
84   static const int NumDims = internal::array_size<Dimensions>::value;
85   typedef StorageMemory<CoeffReturnType, Device> Storage;
86   typedef typename Storage::Type EvaluatorPointerType;
87 
88   enum {
89     IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/ false,
90     PacketAccess = /*TensorEvaluator<ArgType, Device>::PacketAccess*/ false,
91     BlockAccess = false,
92     PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
93     Layout = TensorEvaluator<ArgType, Device>::Layout,
94     CoordAccess = false,  // to be implemented
95     RawAccess = false
96   };
97 
98   //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
99   typedef internal::TensorBlockNotImplemented TensorBlock;
100   //===--------------------------------------------------------------------===//
101 
102   EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
103       : m_impl(op.expression(), device) { }
104 
105   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const {
106     return m_impl.dimensions();
107   }
108 
109   EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
110     m_impl.evalSubExprsIfNeeded(NULL);
111     return true;
112   }
113   EIGEN_STRONG_INLINE void cleanup() {
114     m_impl.cleanup();
115   }
116 
117   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
118   {
119     return CoeffReturnType(index, m_impl.coeff(index));
120   }
121 
122   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
123   costPerCoeff(bool vectorized) const {
124     return m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, 1);
125   }
126 
127   EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
128 
129 #ifdef EIGEN_USE_SYCL
130   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
131     m_impl.bind(cgh);
132   }
133 #endif
134 
135  protected:
136   TensorEvaluator<ArgType, Device> m_impl;
137 };
138 
139 namespace internal {
140 
141 /** \class TensorTupleIndex
142   * \ingroup CXX11_Tensor_Module
143   *
144   * \brief Converts to Tensor<Tuple<Index, Scalar> > and reduces to Tensor<Index>.
145   *
146   */
147 template<typename ReduceOp, typename Dims, typename XprType>
148 struct traits<TensorTupleReducerOp<ReduceOp, Dims, XprType> > : public traits<XprType>
149 {
150   typedef traits<XprType> XprTraits;
151   typedef typename XprTraits::StorageKind StorageKind;
152   typedef typename XprTraits::Index Index;
153   typedef Index Scalar;
154   typedef typename XprType::Nested Nested;
155   typedef typename remove_reference<Nested>::type _Nested;
156   static const int NumDimensions = XprTraits::NumDimensions - array_size<Dims>::value;
157   static const int Layout = XprTraits::Layout;
158 };
159 
160 template<typename ReduceOp, typename Dims, typename XprType>
161 struct eval<TensorTupleReducerOp<ReduceOp, Dims, XprType>, Eigen::Dense>
162 {
163   typedef const TensorTupleReducerOp<ReduceOp, Dims, XprType>EIGEN_DEVICE_REF type;
164 };
165 
166 template<typename ReduceOp, typename Dims, typename XprType>
167 struct nested<TensorTupleReducerOp<ReduceOp, Dims, XprType>, 1,
168               typename eval<TensorTupleReducerOp<ReduceOp, Dims, XprType> >::type>
169 {
170   typedef TensorTupleReducerOp<ReduceOp, Dims, XprType> type;
171 };
172 
173 }  // end namespace internal
174 
175 template<typename ReduceOp, typename Dims, typename XprType>
176 class TensorTupleReducerOp : public TensorBase<TensorTupleReducerOp<ReduceOp, Dims, XprType>, ReadOnlyAccessors>
177 {
178   public:
179   typedef typename Eigen::internal::traits<TensorTupleReducerOp>::Scalar Scalar;
180   typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
181   typedef typename Eigen::internal::nested<TensorTupleReducerOp>::type Nested;
182   typedef typename Eigen::internal::traits<TensorTupleReducerOp>::StorageKind StorageKind;
183   typedef typename Eigen::internal::traits<TensorTupleReducerOp>::Index Index;
184   typedef Index CoeffReturnType;
185 
186   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorTupleReducerOp(const XprType& expr,
187                                                           const ReduceOp& reduce_op,
188                                                           const Index return_dim,
189                                                           const Dims& reduce_dims)
190       : m_xpr(expr), m_reduce_op(reduce_op), m_return_dim(return_dim), m_reduce_dims(reduce_dims) {}
191 
192   EIGEN_DEVICE_FUNC
193   const typename internal::remove_all<typename XprType::Nested>::type&
194   expression() const { return m_xpr; }
195 
196   EIGEN_DEVICE_FUNC
197   const ReduceOp& reduce_op() const { return m_reduce_op; }
198 
199   EIGEN_DEVICE_FUNC
200   const Dims& reduce_dims() const { return m_reduce_dims; }
201 
202   EIGEN_DEVICE_FUNC
203   Index return_dim() const { return m_return_dim; }
204 
205   protected:
206     typename XprType::Nested m_xpr;
207     const ReduceOp m_reduce_op;
208     const Index m_return_dim;
209     const Dims m_reduce_dims;
210 };
211 
212 // Eval as rvalue
213 template<typename ReduceOp, typename Dims, typename ArgType, typename Device>
214 struct TensorEvaluator<const TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Device>
215 {
216   typedef TensorTupleReducerOp<ReduceOp, Dims, ArgType> XprType;
217   typedef typename XprType::Index Index;
218   typedef typename XprType::Scalar Scalar;
219   typedef typename XprType::CoeffReturnType CoeffReturnType;
220   typedef typename TensorIndexTupleOp<ArgType>::CoeffReturnType TupleType;
221   typedef typename TensorEvaluator<const TensorReductionOp<ReduceOp, Dims, const TensorIndexTupleOp<ArgType> >, Device>::Dimensions Dimensions;
222   typedef typename TensorEvaluator<const TensorIndexTupleOp<ArgType> , Device>::Dimensions InputDimensions;
223   static const int NumDims = internal::array_size<InputDimensions>::value;
224   typedef array<Index, NumDims> StrideDims;
225   typedef StorageMemory<CoeffReturnType, Device> Storage;
226   typedef typename Storage::Type EvaluatorPointerType;
227   typedef StorageMemory<TupleType, Device> TupleStorageMem;
228 
229   enum {
230     IsAligned         = /*TensorEvaluator<ArgType, Device>::IsAligned*/ false,
231     PacketAccess      = /*TensorEvaluator<ArgType, Device>::PacketAccess*/ false,
232     BlockAccess       = false,
233     PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
234     Layout            = TensorEvaluator<const TensorReductionOp<ReduceOp, Dims, const TensorIndexTupleOp<ArgType> >, Device>::Layout,
235     CoordAccess       = false,  // to be implemented
236     RawAccess         = false
237   };
238 
239   //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
240   typedef internal::TensorBlockNotImplemented TensorBlock;
241   //===--------------------------------------------------------------------===//
242 
243   EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
244       : m_orig_impl(op.expression(), device),
245         m_impl(op.expression().index_tuples().reduce(op.reduce_dims(), op.reduce_op()), device),
246         m_return_dim(op.return_dim())
247   {
248     gen_strides(m_orig_impl.dimensions(), m_strides);
249     if (Layout == static_cast<int>(ColMajor)) {
250       const Index total_size = internal::array_prod(m_orig_impl.dimensions());
251       m_stride_mod = (m_return_dim < NumDims - 1) ? m_strides[m_return_dim + 1] : total_size;
252     } else {
253       const Index total_size = internal::array_prod(m_orig_impl.dimensions());
254       m_stride_mod = (m_return_dim > 0) ? m_strides[m_return_dim - 1] : total_size;
255     }
256     // If m_return_dim is not a valid index, returns 1 or this can crash on Windows.
257     m_stride_div = ((m_return_dim >= 0) &&
258                     (m_return_dim < static_cast<Index>(m_strides.size())))
259                    ? m_strides[m_return_dim] : 1;
260   }
261 
262   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const {
263     return m_impl.dimensions();
264   }
265 
266   EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
267     m_impl.evalSubExprsIfNeeded(NULL);
268     return true;
269   }
270   EIGEN_STRONG_INLINE void cleanup() {
271     m_impl.cleanup();
272   }
273 
274   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
275     const TupleType v = m_impl.coeff(index);
276     return (m_return_dim < 0) ? v.first : (v.first % m_stride_mod) / m_stride_div;
277   }
278 
279   EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
280 #ifdef EIGEN_USE_SYCL
281   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
282     m_impl.bind(cgh);
283     m_orig_impl.bind(cgh);
284   }
285 #endif
286 
287   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
288   costPerCoeff(bool vectorized) const {
289     const double compute_cost = 1.0 +
290         (m_return_dim < 0 ? 0.0 : (TensorOpCost::ModCost<Index>() + TensorOpCost::DivCost<Index>()));
291     return m_orig_impl.costPerCoeff(vectorized) +
292            m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, compute_cost);
293   }
294 
295  private:
296   EIGEN_DEVICE_FUNC void gen_strides(const InputDimensions& dims, StrideDims& strides) {
297     if (m_return_dim < 0) {
298       return;  // Won't be using the strides.
299     }
300     eigen_assert(m_return_dim < NumDims &&
301                  "Asking to convert index to a dimension outside of the rank");
302 
303     // Calculate m_stride_div and m_stride_mod, which are used to
304     // calculate the value of an index w.r.t. the m_return_dim.
305     if (Layout == static_cast<int>(ColMajor)) {
306       strides[0] = 1;
307       for (int i = 1; i < NumDims; ++i) {
308         strides[i] = strides[i-1] * dims[i-1];
309       }
310     } else {
311       strides[NumDims-1] = 1;
312       for (int i = NumDims - 2; i >= 0; --i) {
313         strides[i] = strides[i+1] * dims[i+1];
314       }
315     }
316   }
317 
318  protected:
319   TensorEvaluator<const TensorIndexTupleOp<ArgType>, Device> m_orig_impl;
320   TensorEvaluator<const TensorReductionOp<ReduceOp, Dims, const TensorIndexTupleOp<ArgType> >, Device> m_impl;
321   const Index m_return_dim;
322   StrideDims m_strides;
323   Index m_stride_mod;
324   Index m_stride_div;
325 };
326 
327 } // end namespace Eigen
328 
329 #endif // EIGEN_CXX11_TENSOR_TENSOR_ARG_MAX_H
330