xref: /aosp_15_r20/external/eigen/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h (revision bf2c37156dfe67e5dfebd6d394bad8b2ab5804d4)
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2014 Benoit Steiner <[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_CXX11_TENSOR_TENSOR_IMAGE_PATCH_H
11 #define EIGEN_CXX11_TENSOR_TENSOR_IMAGE_PATCH_H
12 
13 namespace Eigen {
14 
15 /** \class TensorImagePatch
16   * \ingroup CXX11_Tensor_Module
17   *
18   * \brief Patch extraction specialized for image processing.
19   * This assumes that the input has a least 3 dimensions ordered as follow:
20   *  1st dimension: channels (of size d)
21   *  2nd dimension: rows (of size r)
22   *  3rd dimension: columns (of size c)
23   *  There can be additional dimensions such as time (for video) or batch (for
24   * bulk processing after the first 3.
25   * Calling the image patch code with patch_rows and patch_cols is equivalent
26   * to calling the regular patch extraction code with parameters d, patch_rows,
27   * patch_cols, and 1 for all the additional dimensions.
28   */
29 namespace internal {
30 
31 template<DenseIndex Rows, DenseIndex Cols, typename XprType>
32 struct traits<TensorImagePatchOp<Rows, Cols, XprType> > : public traits<XprType>
33 {
34   typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
35   typedef traits<XprType> XprTraits;
36   typedef typename XprTraits::StorageKind StorageKind;
37   typedef typename XprTraits::Index Index;
38   typedef typename XprType::Nested Nested;
39   typedef typename remove_reference<Nested>::type _Nested;
40   static const int NumDimensions = XprTraits::NumDimensions + 1;
41   static const int Layout = XprTraits::Layout;
42   typedef typename XprTraits::PointerType PointerType;
43 };
44 
45 template<DenseIndex Rows, DenseIndex Cols, typename XprType>
46 struct eval<TensorImagePatchOp<Rows, Cols, XprType>, Eigen::Dense>
47 {
48   typedef const TensorImagePatchOp<Rows, Cols, XprType>& type;
49 };
50 
51 template<DenseIndex Rows, DenseIndex Cols, typename XprType>
52 struct nested<TensorImagePatchOp<Rows, Cols, XprType>, 1, typename eval<TensorImagePatchOp<Rows, Cols, XprType> >::type>
53 {
54   typedef TensorImagePatchOp<Rows, Cols, XprType> type;
55 };
56 
57 template <typename Self, bool Vectorizable>
58 struct ImagePatchCopyOp {
59   typedef typename Self::Index Index;
60   typedef typename Self::Scalar Scalar;
61   typedef typename Self::Impl Impl;
62   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
63       const Self& self, const Index num_coeff_to_copy, const Index dst_index,
64       Scalar* dst_data, const Index src_index) {
65     const Impl& impl = self.impl();
66     for (Index i = 0; i < num_coeff_to_copy; ++i) {
67       dst_data[dst_index + i] = impl.coeff(src_index + i);
68     }
69   }
70 };
71 
72 template <typename Self>
73 struct ImagePatchCopyOp<Self, true> {
74   typedef typename Self::Index Index;
75   typedef typename Self::Scalar Scalar;
76   typedef typename Self::Impl Impl;
77   typedef typename packet_traits<Scalar>::type Packet;
78   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
79       const Self& self, const Index num_coeff_to_copy, const Index dst_index,
80       Scalar* dst_data, const Index src_index) {
81     const Impl& impl = self.impl();
82     const Index packet_size = internal::unpacket_traits<Packet>::size;
83     const Index vectorized_size =
84         (num_coeff_to_copy / packet_size) * packet_size;
85     for (Index i = 0; i < vectorized_size; i += packet_size) {
86       Packet p = impl.template packet<Unaligned>(src_index + i);
87       internal::pstoret<Scalar, Packet, Unaligned>(dst_data + dst_index + i, p);
88     }
89     for (Index i = vectorized_size; i < num_coeff_to_copy; ++i) {
90       dst_data[dst_index + i] = impl.coeff(src_index + i);
91     }
92   }
93 };
94 
95 template <typename Self>
96 struct ImagePatchPaddingOp {
97   typedef typename Self::Index Index;
98   typedef typename Self::Scalar Scalar;
99   typedef typename packet_traits<Scalar>::type Packet;
100   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
101       const Index num_coeff_to_pad, const Scalar padding_value,
102       const Index dst_index, Scalar* dst_data) {
103     const Index packet_size = internal::unpacket_traits<Packet>::size;
104     const Packet padded_packet = internal::pset1<Packet>(padding_value);
105     const Index vectorized_size =
106         (num_coeff_to_pad / packet_size) * packet_size;
107     for (Index i = 0; i < vectorized_size; i += packet_size) {
108       internal::pstoret<Scalar, Packet, Unaligned>(dst_data + dst_index + i,
109                                                    padded_packet);
110     }
111     for (Index i = vectorized_size; i < num_coeff_to_pad; ++i) {
112       dst_data[dst_index + i] = padding_value;
113     }
114   }
115 };
116 
117 }  // end namespace internal
118 
119 template<DenseIndex Rows, DenseIndex Cols, typename XprType>
120 class TensorImagePatchOp : public TensorBase<TensorImagePatchOp<Rows, Cols, XprType>, ReadOnlyAccessors>
121 {
122   public:
123   typedef typename Eigen::internal::traits<TensorImagePatchOp>::Scalar Scalar;
124   typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
125   typedef typename XprType::CoeffReturnType CoeffReturnType;
126   typedef typename Eigen::internal::nested<TensorImagePatchOp>::type Nested;
127   typedef typename Eigen::internal::traits<TensorImagePatchOp>::StorageKind StorageKind;
128   typedef typename Eigen::internal::traits<TensorImagePatchOp>::Index Index;
129 
130   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorImagePatchOp(const XprType& expr, DenseIndex patch_rows, DenseIndex patch_cols,
131                                                            DenseIndex row_strides, DenseIndex col_strides,
132                                                            DenseIndex in_row_strides, DenseIndex in_col_strides,
133                                                            DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
134                                                            PaddingType padding_type, Scalar padding_value)
135                                                            : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
136                                                            m_row_strides(row_strides), m_col_strides(col_strides),
137                                                            m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
138                                                            m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
139                                                            m_padding_explicit(false), m_padding_top(0), m_padding_bottom(0), m_padding_left(0), m_padding_right(0),
140                                                            m_padding_type(padding_type), m_padding_value(padding_value) {}
141 
142   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorImagePatchOp(const XprType& expr, DenseIndex patch_rows, DenseIndex patch_cols,
143                                                            DenseIndex row_strides, DenseIndex col_strides,
144                                                            DenseIndex in_row_strides, DenseIndex in_col_strides,
145                                                            DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
146                                                            DenseIndex padding_top, DenseIndex padding_bottom,
147                                                            DenseIndex padding_left, DenseIndex padding_right,
148                                                            Scalar padding_value)
149                                                            : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
150                                                            m_row_strides(row_strides), m_col_strides(col_strides),
151                                                            m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
152                                                            m_row_inflate_strides(row_inflate_strides), m_col_inflate_strides(col_inflate_strides),
153                                                            m_padding_explicit(true), m_padding_top(padding_top), m_padding_bottom(padding_bottom),
154                                                            m_padding_left(padding_left), m_padding_right(padding_right),
155                                                            m_padding_type(PADDING_VALID), m_padding_value(padding_value) {}
156 
157 
158     EIGEN_DEVICE_FUNC
159     DenseIndex patch_rows() const { return m_patch_rows; }
160     EIGEN_DEVICE_FUNC
161     DenseIndex patch_cols() const { return m_patch_cols; }
162     EIGEN_DEVICE_FUNC
163     DenseIndex row_strides() const { return m_row_strides; }
164     EIGEN_DEVICE_FUNC
165     DenseIndex col_strides() const { return m_col_strides; }
166     EIGEN_DEVICE_FUNC
167     DenseIndex in_row_strides() const { return m_in_row_strides; }
168     EIGEN_DEVICE_FUNC
169     DenseIndex in_col_strides() const { return m_in_col_strides; }
170     EIGEN_DEVICE_FUNC
171     DenseIndex row_inflate_strides() const { return m_row_inflate_strides; }
172     EIGEN_DEVICE_FUNC
173     DenseIndex col_inflate_strides() const { return m_col_inflate_strides; }
174     EIGEN_DEVICE_FUNC
175     bool padding_explicit() const { return m_padding_explicit; }
176     EIGEN_DEVICE_FUNC
177     DenseIndex padding_top() const { return m_padding_top; }
178     EIGEN_DEVICE_FUNC
179     DenseIndex padding_bottom() const { return m_padding_bottom; }
180     EIGEN_DEVICE_FUNC
181     DenseIndex padding_left() const { return m_padding_left; }
182     EIGEN_DEVICE_FUNC
183     DenseIndex padding_right() const { return m_padding_right; }
184     EIGEN_DEVICE_FUNC
185     PaddingType padding_type() const { return m_padding_type; }
186     EIGEN_DEVICE_FUNC
187     Scalar padding_value() const { return m_padding_value; }
188 
189     EIGEN_DEVICE_FUNC
190     const typename internal::remove_all<typename XprType::Nested>::type&
191     expression() const { return m_xpr; }
192 
193   protected:
194     typename XprType::Nested m_xpr;
195     const DenseIndex m_patch_rows;
196     const DenseIndex m_patch_cols;
197     const DenseIndex m_row_strides;
198     const DenseIndex m_col_strides;
199     const DenseIndex m_in_row_strides;
200     const DenseIndex m_in_col_strides;
201     const DenseIndex m_row_inflate_strides;
202     const DenseIndex m_col_inflate_strides;
203     const bool m_padding_explicit;
204     const DenseIndex m_padding_top;
205     const DenseIndex m_padding_bottom;
206     const DenseIndex m_padding_left;
207     const DenseIndex m_padding_right;
208     const PaddingType m_padding_type;
209     const Scalar m_padding_value;
210 };
211 
212 // Eval as rvalue
213 template<DenseIndex Rows, DenseIndex Cols, typename ArgType, typename Device>
214 struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
215 {
216   typedef TensorImagePatchOp<Rows, Cols, ArgType> XprType;
217   typedef typename XprType::Index Index;
218   static const int NumInputDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
219   static const int NumDims = NumInputDims + 1;
220   typedef DSizes<Index, NumDims> Dimensions;
221   typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
222   typedef TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>,
223                           Device> Self;
224   typedef TensorEvaluator<ArgType, Device> Impl;
225   typedef typename XprType::CoeffReturnType CoeffReturnType;
226   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
227   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
228   typedef StorageMemory<CoeffReturnType, Device> Storage;
229   typedef typename Storage::Type EvaluatorPointerType;
230 
231   enum {
232     IsAligned         = false,
233     PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
234     BlockAccess       = false,
235     PreferBlockAccess = true,
236     Layout            = TensorEvaluator<ArgType, Device>::Layout,
237     CoordAccess       = false,
238     RawAccess         = false
239   };
240 
241   //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
242   typedef internal::TensorBlockNotImplemented TensorBlock;
243   //===--------------------------------------------------------------------===//
244 
245   EIGEN_STRONG_INLINE TensorEvaluator( const XprType& op, const Device& device)
246       : m_device(device), m_impl(op.expression(), device)
247   {
248     EIGEN_STATIC_ASSERT((NumDims >= 4), YOU_MADE_A_PROGRAMMING_MISTAKE);
249 
250     m_paddingValue = op.padding_value();
251 
252     const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
253 
254     // Caches a few variables.
255     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
256       m_inputDepth = input_dims[0];
257       m_inputRows = input_dims[1];
258       m_inputCols = input_dims[2];
259     } else {
260       m_inputDepth = input_dims[NumInputDims-1];
261       m_inputRows = input_dims[NumInputDims-2];
262       m_inputCols = input_dims[NumInputDims-3];
263     }
264 
265     m_row_strides = op.row_strides();
266     m_col_strides = op.col_strides();
267 
268     // Input strides and effective input/patch size
269     m_in_row_strides = op.in_row_strides();
270     m_in_col_strides = op.in_col_strides();
271     m_row_inflate_strides = op.row_inflate_strides();
272     m_col_inflate_strides = op.col_inflate_strides();
273     // The "effective" input rows and input cols are the input rows and cols
274     // after inflating them with zeros.
275     // For examples, a 2x3 matrix with row_inflate_strides and
276     // col_inflate_strides of 2 comes from:
277     //   A B C
278     //   D E F
279     //
280     // to a matrix is 3 x 5:
281     //
282     //   A . B . C
283     //   . . . . .
284     //   D . E . F
285 
286     m_input_rows_eff = (m_inputRows - 1) * m_row_inflate_strides + 1;
287     m_input_cols_eff = (m_inputCols - 1) * m_col_inflate_strides + 1;
288     m_patch_rows_eff = op.patch_rows() + (op.patch_rows() - 1) * (m_in_row_strides - 1);
289     m_patch_cols_eff = op.patch_cols() + (op.patch_cols() - 1) * (m_in_col_strides - 1);
290 
291     if (op.padding_explicit()) {
292       m_outputRows = numext::ceil((m_input_rows_eff + op.padding_top() + op.padding_bottom() - m_patch_rows_eff + 1.f) / static_cast<float>(m_row_strides));
293       m_outputCols = numext::ceil((m_input_cols_eff + op.padding_left() + op.padding_right() - m_patch_cols_eff + 1.f) / static_cast<float>(m_col_strides));
294       m_rowPaddingTop = op.padding_top();
295       m_colPaddingLeft = op.padding_left();
296     } else {
297       // Computing padding from the type
298       switch (op.padding_type()) {
299         case PADDING_VALID:
300           m_outputRows = numext::ceil((m_input_rows_eff - m_patch_rows_eff + 1.f) / static_cast<float>(m_row_strides));
301           m_outputCols = numext::ceil((m_input_cols_eff - m_patch_cols_eff + 1.f) / static_cast<float>(m_col_strides));
302           // Calculate the padding
303           m_rowPaddingTop = numext::maxi<Index>(0, ((m_outputRows - 1) * m_row_strides + m_patch_rows_eff - m_input_rows_eff) / 2);
304           m_colPaddingLeft = numext::maxi<Index>(0, ((m_outputCols - 1) * m_col_strides + m_patch_cols_eff - m_input_cols_eff) / 2);
305           break;
306         case PADDING_SAME:
307           m_outputRows = numext::ceil(m_input_rows_eff / static_cast<float>(m_row_strides));
308           m_outputCols = numext::ceil(m_input_cols_eff / static_cast<float>(m_col_strides));
309           // Calculate the padding
310           m_rowPaddingTop = ((m_outputRows - 1) * m_row_strides + m_patch_rows_eff - m_input_rows_eff) / 2;
311           m_colPaddingLeft = ((m_outputCols - 1) * m_col_strides + m_patch_cols_eff - m_input_cols_eff) / 2;
312           // The padding size calculation for PADDING_SAME has been updated to
313           // be consistent with how TensorFlow extracts its paddings.
314           m_rowPaddingTop = numext::maxi<Index>(0, m_rowPaddingTop);
315           m_colPaddingLeft = numext::maxi<Index>(0, m_colPaddingLeft);
316           break;
317         default:
318           eigen_assert(false && "unexpected padding");
319           m_outputCols=0; // silence the uninitialised warning;
320           m_outputRows=0; //// silence the uninitialised warning;
321       }
322     }
323     eigen_assert(m_outputRows > 0);
324     eigen_assert(m_outputCols > 0);
325 
326     // Dimensions for result of extraction.
327     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
328       // ColMajor
329       // 0: depth
330       // 1: patch_rows
331       // 2: patch_cols
332       // 3: number of patches
333       // 4 and beyond: anything else (such as batch).
334       m_dimensions[0] = input_dims[0];
335       m_dimensions[1] = op.patch_rows();
336       m_dimensions[2] = op.patch_cols();
337       m_dimensions[3] = m_outputRows * m_outputCols;
338       for (int i = 4; i < NumDims; ++i) {
339         m_dimensions[i] = input_dims[i-1];
340       }
341     } else {
342       // RowMajor
343       // NumDims-1: depth
344       // NumDims-2: patch_rows
345       // NumDims-3: patch_cols
346       // NumDims-4: number of patches
347       // NumDims-5 and beyond: anything else (such as batch).
348       m_dimensions[NumDims-1] = input_dims[NumInputDims-1];
349       m_dimensions[NumDims-2] = op.patch_rows();
350       m_dimensions[NumDims-3] = op.patch_cols();
351       m_dimensions[NumDims-4] = m_outputRows * m_outputCols;
352       for (int i = NumDims-5; i >= 0; --i) {
353         m_dimensions[i] = input_dims[i];
354       }
355     }
356 
357     // Strides for moving the patch in various dimensions.
358     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
359       m_colStride = m_dimensions[1];
360       m_patchStride = m_colStride * m_dimensions[2] * m_dimensions[0];
361       m_otherStride = m_patchStride * m_dimensions[3];
362     } else {
363       m_colStride = m_dimensions[NumDims-2];
364       m_patchStride = m_colStride * m_dimensions[NumDims-3] * m_dimensions[NumDims-1];
365       m_otherStride = m_patchStride * m_dimensions[NumDims-4];
366     }
367 
368     // Strides for navigating through the input tensor.
369     m_rowInputStride = m_inputDepth;
370     m_colInputStride = m_inputDepth * m_inputRows;
371     m_patchInputStride = m_inputDepth * m_inputRows * m_inputCols;
372 
373     // Fast representations of different variables.
374     m_fastOtherStride = internal::TensorIntDivisor<Index>(m_otherStride);
375     m_fastPatchStride = internal::TensorIntDivisor<Index>(m_patchStride);
376     m_fastColStride = internal::TensorIntDivisor<Index>(m_colStride);
377     m_fastInflateRowStride = internal::TensorIntDivisor<Index>(m_row_inflate_strides);
378     m_fastInflateColStride = internal::TensorIntDivisor<Index>(m_col_inflate_strides);
379     m_fastInputColsEff = internal::TensorIntDivisor<Index>(m_input_cols_eff);
380 
381     // Number of patches in the width dimension.
382     m_fastOutputRows = internal::TensorIntDivisor<Index>(m_outputRows);
383     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
384       m_fastOutputDepth = internal::TensorIntDivisor<Index>(m_dimensions[0]);
385     } else {
386       m_fastOutputDepth = internal::TensorIntDivisor<Index>(m_dimensions[NumDims-1]);
387     }
388   }
389 
390   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
391 
392   EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
393     m_impl.evalSubExprsIfNeeded(NULL);
394     return true;
395   }
396 
397 #ifdef EIGEN_USE_THREADS
398   template <typename EvalSubExprsCallback>
399   EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
400       EvaluatorPointerType, EvalSubExprsCallback done) {
401     m_impl.evalSubExprsIfNeededAsync(nullptr, [done](bool) { done(true); });
402   }
403 #endif  // EIGEN_USE_THREADS
404 
405   EIGEN_STRONG_INLINE void cleanup() {
406     m_impl.cleanup();
407   }
408 
409   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
410   {
411     // Patch index corresponding to the passed in index.
412     const Index patchIndex = index / m_fastPatchStride;
413     // Find the offset of the element wrt the location of the first element.
414     const Index patchOffset = (index - patchIndex * m_patchStride) / m_fastOutputDepth;
415 
416     // Other ways to index this element.
417     const Index otherIndex = (NumDims == 4) ? 0 : index / m_fastOtherStride;
418     const Index patch2DIndex = (NumDims == 4) ? patchIndex : (index - otherIndex * m_otherStride) / m_fastPatchStride;
419 
420     // Calculate col index in the input original tensor.
421     const Index colIndex = patch2DIndex / m_fastOutputRows;
422     const Index colOffset = patchOffset / m_fastColStride;
423     const Index inputCol = colIndex * m_col_strides + colOffset * m_in_col_strides - m_colPaddingLeft;
424     const Index origInputCol = (m_col_inflate_strides == 1) ? inputCol : ((inputCol >= 0) ? (inputCol / m_fastInflateColStride) : 0);
425     if (inputCol < 0 || inputCol >= m_input_cols_eff ||
426         ((m_col_inflate_strides != 1) && (inputCol != origInputCol * m_col_inflate_strides))) {
427       return Scalar(m_paddingValue);
428     }
429 
430     // Calculate row index in the original input tensor.
431     const Index rowIndex = patch2DIndex - colIndex * m_outputRows;
432     const Index rowOffset = patchOffset - colOffset * m_colStride;
433     const Index inputRow = rowIndex * m_row_strides + rowOffset * m_in_row_strides - m_rowPaddingTop;
434     const Index origInputRow = (m_row_inflate_strides == 1) ? inputRow : ((inputRow >= 0) ? (inputRow / m_fastInflateRowStride) : 0);
435     if (inputRow < 0 || inputRow >= m_input_rows_eff ||
436         ((m_row_inflate_strides != 1) && (inputRow != origInputRow * m_row_inflate_strides))) {
437       return Scalar(m_paddingValue);
438     }
439 
440     const int depth_index = static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : NumDims - 1;
441     const Index depth = index - (index / m_fastOutputDepth) * m_dimensions[depth_index];
442 
443     const Index inputIndex = depth + origInputRow * m_rowInputStride + origInputCol * m_colInputStride + otherIndex * m_patchInputStride;
444     return m_impl.coeff(inputIndex);
445   }
446 
447   template<int LoadMode>
448   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
449   {
450     EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
451     eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
452 
453     if (m_in_row_strides != 1 || m_in_col_strides != 1 || m_row_inflate_strides != 1 || m_col_inflate_strides != 1) {
454       return packetWithPossibleZero(index);
455     }
456 
457     const Index indices[2] = {index, index + PacketSize - 1};
458     const Index patchIndex = indices[0] / m_fastPatchStride;
459     if (patchIndex != indices[1] / m_fastPatchStride) {
460       return packetWithPossibleZero(index);
461     }
462     const Index otherIndex = (NumDims == 4) ? 0 : indices[0] / m_fastOtherStride;
463     eigen_assert(otherIndex == indices[1] / m_fastOtherStride);
464 
465     // Find the offset of the element wrt the location of the first element.
466     const Index patchOffsets[2] = {(indices[0] - patchIndex * m_patchStride) / m_fastOutputDepth,
467                                    (indices[1] - patchIndex * m_patchStride) / m_fastOutputDepth};
468 
469     const Index patch2DIndex = (NumDims == 4) ? patchIndex : (indices[0] - otherIndex * m_otherStride) / m_fastPatchStride;
470     eigen_assert(patch2DIndex == (indices[1] - otherIndex * m_otherStride) / m_fastPatchStride);
471 
472     const Index colIndex = patch2DIndex / m_fastOutputRows;
473     const Index colOffsets[2] = {patchOffsets[0] / m_fastColStride, patchOffsets[1] / m_fastColStride};
474 
475     // Calculate col indices in the original input tensor.
476     const Index inputCols[2] = {colIndex * m_col_strides + colOffsets[0] -
477       m_colPaddingLeft, colIndex * m_col_strides + colOffsets[1] - m_colPaddingLeft};
478     if (inputCols[1] < 0 || inputCols[0] >= m_inputCols) {
479       return internal::pset1<PacketReturnType>(Scalar(m_paddingValue));
480     }
481 
482     if (inputCols[0] == inputCols[1]) {
483       const Index rowIndex = patch2DIndex - colIndex * m_outputRows;
484       const Index rowOffsets[2] = {patchOffsets[0] - colOffsets[0]*m_colStride, patchOffsets[1] - colOffsets[1]*m_colStride};
485       eigen_assert(rowOffsets[0] <= rowOffsets[1]);
486       // Calculate col indices in the original input tensor.
487       const Index inputRows[2] = {rowIndex * m_row_strides + rowOffsets[0] -
488         m_rowPaddingTop, rowIndex * m_row_strides + rowOffsets[1] - m_rowPaddingTop};
489 
490       if (inputRows[1] < 0 || inputRows[0] >= m_inputRows) {
491         return internal::pset1<PacketReturnType>(Scalar(m_paddingValue));
492       }
493 
494       if (inputRows[0] >= 0 && inputRows[1] < m_inputRows) {
495         // no padding
496         const int depth_index = static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : NumDims - 1;
497         const Index depth = index - (index / m_fastOutputDepth) * m_dimensions[depth_index];
498         const Index inputIndex = depth + inputRows[0] * m_rowInputStride + inputCols[0] * m_colInputStride + otherIndex * m_patchInputStride;
499         return m_impl.template packet<Unaligned>(inputIndex);
500       }
501     }
502 
503     return packetWithPossibleZero(index);
504   }
505 
506   EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
507 
508   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
509 
510 #ifdef EIGEN_USE_SYCL
511   // binding placeholder accessors to a command group handler for SYCL
512   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
513     m_impl.bind(cgh);
514   }
515 #endif
516 
517   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowPaddingTop() const { return m_rowPaddingTop; }
518   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colPaddingLeft() const { return m_colPaddingLeft; }
519   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index outputRows() const { return m_outputRows; }
520   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index outputCols() const { return m_outputCols; }
521   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userRowStride() const { return m_row_strides; }
522   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userColStride() const { return m_col_strides; }
523   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userInRowStride() const { return m_in_row_strides; }
524   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userInColStride() const { return m_in_col_strides; }
525   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowInflateStride() const { return m_row_inflate_strides; }
526   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colInflateStride() const { return m_col_inflate_strides; }
527 
528   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
529   costPerCoeff(bool vectorized) const {
530     // We conservatively estimate the cost for the code path where the computed
531     // index is inside the original image and
532     // TensorEvaluator<ArgType, Device>::CoordAccess is false.
533     const double compute_cost = 3 * TensorOpCost::DivCost<Index>() +
534                                 6 * TensorOpCost::MulCost<Index>() +
535                                 8 * TensorOpCost::MulCost<Index>();
536     return m_impl.costPerCoeff(vectorized) +
537            TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
538   }
539 
540  protected:
541   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const
542   {
543     EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
544     EIGEN_UNROLL_LOOP
545     for (int i = 0; i < PacketSize; ++i) {
546       values[i] = coeff(index+i);
547     }
548     PacketReturnType rslt = internal::pload<PacketReturnType>(values);
549     return rslt;
550   }
551 
552   Dimensions m_dimensions;
553 
554   Index m_otherStride;
555   Index m_patchStride;
556   Index m_colStride;
557   Index m_row_strides;
558   Index m_col_strides;
559 
560   Index m_in_row_strides;
561   Index m_in_col_strides;
562   Index m_row_inflate_strides;
563   Index m_col_inflate_strides;
564 
565   Index m_input_rows_eff;
566   Index m_input_cols_eff;
567   Index m_patch_rows_eff;
568   Index m_patch_cols_eff;
569 
570   internal::TensorIntDivisor<Index> m_fastOtherStride;
571   internal::TensorIntDivisor<Index> m_fastPatchStride;
572   internal::TensorIntDivisor<Index> m_fastColStride;
573   internal::TensorIntDivisor<Index> m_fastInflateRowStride;
574   internal::TensorIntDivisor<Index> m_fastInflateColStride;
575   internal::TensorIntDivisor<Index> m_fastInputColsEff;
576 
577   Index m_rowInputStride;
578   Index m_colInputStride;
579   Index m_patchInputStride;
580 
581   Index m_inputDepth;
582   Index m_inputRows;
583   Index m_inputCols;
584 
585   Index m_outputRows;
586   Index m_outputCols;
587 
588   Index m_rowPaddingTop;
589   Index m_colPaddingLeft;
590 
591   internal::TensorIntDivisor<Index> m_fastOutputRows;
592   internal::TensorIntDivisor<Index> m_fastOutputDepth;
593 
594   Scalar m_paddingValue;
595 
596   const Device EIGEN_DEVICE_REF m_device;
597   TensorEvaluator<ArgType, Device> m_impl;
598 };
599 
600 
601 } // end namespace Eigen
602 
603 #endif // EIGEN_CXX11_TENSOR_TENSOR_IMAGE_PATCH_H
604