aboutsummaryrefslogtreecommitdiffhomepage
path: root/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h
blob: 91e32d200cef1d1e6e6086875bcb12e5a05530e3 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.

#ifndef EIGEN_CXX11_TENSOR_TENSOR_PADDING_H
#define EIGEN_CXX11_TENSOR_TENSOR_PADDING_H

namespace Eigen {

/** \class TensorPadding
  * \ingroup CXX11_Tensor_Module
  *
  * \brief Tensor padding class.
  * At the moment only 0-padding is supported.
  *
  */
namespace internal {
template<typename PaddingDimensions, typename XprType>
struct traits<TensorPaddingOp<PaddingDimensions, XprType> > : public traits<XprType>
{
  typedef typename XprType::Scalar Scalar;
  typedef traits<XprType> XprTraits;
  typedef typename packet_traits<Scalar>::type Packet;
  typedef typename XprTraits::StorageKind StorageKind;
  typedef typename XprTraits::Index Index;
  typedef typename XprType::Nested Nested;
  typedef typename remove_reference<Nested>::type _Nested;
  static const int NumDimensions = XprTraits::NumDimensions;
  static const int Layout = XprTraits::Layout;
};

template<typename PaddingDimensions, typename XprType>
struct eval<TensorPaddingOp<PaddingDimensions, XprType>, Eigen::Dense>
{
  typedef const TensorPaddingOp<PaddingDimensions, XprType>& type;
};

template<typename PaddingDimensions, typename XprType>
struct nested<TensorPaddingOp<PaddingDimensions, XprType>, 1, typename eval<TensorPaddingOp<PaddingDimensions, XprType> >::type>
{
  typedef TensorPaddingOp<PaddingDimensions, XprType> type;
};

}  // end namespace internal



template<typename PaddingDimensions, typename XprType>
class TensorPaddingOp : public TensorBase<TensorPaddingOp<PaddingDimensions, XprType>, ReadOnlyAccessors>
{
  public:
  typedef typename Eigen::internal::traits<TensorPaddingOp>::Scalar Scalar;
  typedef typename Eigen::internal::traits<TensorPaddingOp>::Packet Packet;
  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
  typedef typename XprType::CoeffReturnType CoeffReturnType;
  typedef typename XprType::PacketReturnType PacketReturnType;
  typedef typename Eigen::internal::nested<TensorPaddingOp>::type Nested;
  typedef typename Eigen::internal::traits<TensorPaddingOp>::StorageKind StorageKind;
  typedef typename Eigen::internal::traits<TensorPaddingOp>::Index Index;

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorPaddingOp(const XprType& expr, const PaddingDimensions& padding_dims)
      : m_xpr(expr), m_padding_dims(padding_dims) {}

    EIGEN_DEVICE_FUNC
    const PaddingDimensions& padding() const { return m_padding_dims; }

    EIGEN_DEVICE_FUNC
    const typename internal::remove_all<typename XprType::Nested>::type&
    expression() const { return m_xpr; }

  protected:
    typename XprType::Nested m_xpr;
    const PaddingDimensions m_padding_dims;
};


// Eval as rvalue
template<typename PaddingDimensions, typename ArgType, typename Device>
struct TensorEvaluator<const TensorPaddingOp<PaddingDimensions, ArgType>, Device>
{
  typedef TensorPaddingOp<PaddingDimensions, ArgType> XprType;
  typedef typename XprType::Index Index;
  static const int NumDims = internal::array_size<PaddingDimensions>::value;
  typedef DSizes<Index, NumDims> Dimensions;

  enum {
    IsAligned = false,
    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
    Layout = TensorEvaluator<ArgType, Device>::Layout,
    CoordAccess = true,
  };

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
      : m_impl(op.expression(), device), m_padding(op.padding())
  {
    // The padding op doesn't change the rank of the tensor. Directly padding a scalar would lead
    // to a vector, which doesn't make sense. Instead one should reshape the scalar into a vector
    // of 1 element first and then pad.
    EIGEN_STATIC_ASSERT(NumDims > 0, YOU_MADE_A_PROGRAMMING_MISTAKE);

    // Compute dimensions
    m_dimensions = m_impl.dimensions();
    for (int i = 0; i < NumDims; ++i) {
      m_dimensions[i] += m_padding[i].first + m_padding[i].second;
    }
    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
      m_inputStrides[0] = 1;
      m_outputStrides[0] = 1;
      for (int i = 1; i < NumDims; ++i) {
        m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
        m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1];
      }
      m_outputStrides[NumDims] = m_outputStrides[NumDims-1] * m_dimensions[NumDims-1];
    } else {
      m_inputStrides[NumDims - 1] = 1;
      m_outputStrides[NumDims] = 1;
      for (int i = NumDims - 2; i >= 0; --i) {
        m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1];
        m_outputStrides[i+1] = m_outputStrides[i+2] * m_dimensions[i+1];
      }
      m_outputStrides[0] = m_outputStrides[1] * m_dimensions[0];
    }
  }

  typedef typename XprType::Scalar Scalar;
  typedef typename XprType::CoeffReturnType CoeffReturnType;
  typedef typename XprType::PacketReturnType PacketReturnType;

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) {
    m_impl.evalSubExprsIfNeeded(NULL);
    return true;
  }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
    m_impl.cleanup();
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
  {
    eigen_assert(index < dimensions().TotalSize());
    Index inputIndex = 0;
    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
      for (int i = NumDims - 1; i > 0; --i) {
        const Index idx = index / m_outputStrides[i];
        if (idx < m_padding[i].first || idx >= m_dimensions[i] - m_padding[i].second) {
          return Scalar(0);
        }
        inputIndex += (idx - m_padding[i].first) * m_inputStrides[i];
        index -= idx * m_outputStrides[i];
      }
      if (index < m_padding[0].first || index >= m_dimensions[0] - m_padding[0].second) {
        return Scalar(0);
      }
      inputIndex += (index - m_padding[0].first);
    } else {
      for (int i = 0; i < NumDims - 1; ++i) {
        const Index idx = index / m_outputStrides[i+1];
        if (idx < m_padding[i].first || idx >= m_dimensions[i] - m_padding[i].second) {
          return Scalar(0);
        }
        inputIndex += (idx - m_padding[i].first) * m_inputStrides[i];
        index -= idx * m_outputStrides[i+1];
      }
      if (index < m_padding[NumDims-1].first ||
          index >= m_dimensions[NumDims-1] - m_padding[NumDims-1].second) {
        return Scalar(0);
      }
      inputIndex += (index - m_padding[NumDims-1].first);
    }
    return m_impl.coeff(inputIndex);
  }

  template<int LoadMode>
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
  {
    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
      return packetColMajor(index);
    }
    return packetRowMajor(index);
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(const array<Index, NumDims>& coords) const
  {
    Index inputIndex;
    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
      {
        const Index idx = coords[0];
        if (idx < m_padding[0].first || idx >= m_dimensions[0] - m_padding[0].second) {
          return Scalar(0);
        }
        inputIndex = idx - m_padding[0].first;
      }
      for (int i = 1; i < NumDims; ++i) {
        const Index idx = coords[i];
        if (idx < m_padding[i].first || idx >= m_dimensions[i] - m_padding[i].second) {
          return Scalar(0);
        }
        inputIndex += (idx - m_padding[i].first) * m_inputStrides[i];
      }
    } else {
      {
        const Index idx = coords[NumDims-1];
        if (idx < m_padding[NumDims-1].first || idx >= m_dimensions[NumDims-1] - m_padding[NumDims-1].second) {
          return Scalar(0);
        }
        inputIndex = idx - m_padding[NumDims-1].first;
      }
      for (int i = NumDims - 2; i >= 0; --i) {
        const Index idx = coords[i];
        if (idx < m_padding[i].first || idx >= m_dimensions[i] - m_padding[i].second) {
          return Scalar(0);
        }
        inputIndex += (idx - m_padding[i].first) * m_inputStrides[i];
      }
    }
    return m_impl.coeff(inputIndex);
  }

  EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }

 protected:

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetColMajor(Index index) const
  {
    const int packetSize = internal::unpacket_traits<PacketReturnType>::size;
    EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE)
    eigen_assert(index+packetSize-1 < dimensions().TotalSize());

    const Index initialIndex = index;
    Index inputIndex = 0;
    for (int i = NumDims - 1; i > 0; --i) {
      const Index first = index;
      const Index last = index + packetSize - 1;
      const Index lastPaddedLeft = m_padding[i].first * m_outputStrides[i];
      const Index firstPaddedRight = (m_dimensions[i] - m_padding[i].second) * m_outputStrides[i];
      const Index lastPaddedRight = m_outputStrides[i+1];

      if (last < lastPaddedLeft) {
        // all the coefficient are in the padding zone.
        return internal::pset1<PacketReturnType>(Scalar(0));
      }
      else if (first >= firstPaddedRight && last < lastPaddedRight) {
        // all the coefficient are in the padding zone.
        return internal::pset1<PacketReturnType>(Scalar(0));
      }
      else if (first >= lastPaddedLeft && last < firstPaddedRight) {
        // all the coefficient are between the 2 padding zones.
        const Index idx = index / m_outputStrides[i];
        inputIndex += (idx - m_padding[i].first) * m_inputStrides[i];
        index -= idx * m_outputStrides[i];
      }
      else {
        // Every other case
        return packetWithPossibleZero(initialIndex);
      }
    }

    const Index last = index + packetSize - 1;
    const Index first = index;
    const Index lastPaddedLeft = m_padding[0].first;
    const Index firstPaddedRight = (m_dimensions[0] - m_padding[0].second);
    const Index lastPaddedRight = m_outputStrides[1];

    if (last < lastPaddedLeft) {
      // all the coefficient are in the padding zone.
      return internal::pset1<PacketReturnType>(Scalar(0));
    }
    else if (first >= firstPaddedRight && last < lastPaddedRight) {
      // all the coefficient are in the padding zone.
      return internal::pset1<PacketReturnType>(Scalar(0));
    }
    else if (first >= lastPaddedLeft && last < firstPaddedRight) {
      // all the coefficient are between the 2 padding zones.
      inputIndex += (index - m_padding[0].first);
      return m_impl.template packet<Unaligned>(inputIndex);
    }
    // Every other case
    return packetWithPossibleZero(initialIndex);
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetRowMajor(Index index) const
  {
    const int packetSize = internal::unpacket_traits<PacketReturnType>::size;
    EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE)
    eigen_assert(index+packetSize-1 < dimensions().TotalSize());

    const Index initialIndex = index;
    Index inputIndex = 0;

    for (int i = 0; i < NumDims - 1; ++i) {
      const Index first = index;
      const Index last = index + packetSize - 1;
      const Index lastPaddedLeft = m_padding[i].first * m_outputStrides[i+1];
      const Index firstPaddedRight = (m_dimensions[i] - m_padding[i].second) * m_outputStrides[i+1];
      const Index lastPaddedRight = m_outputStrides[i];

      if (last < lastPaddedLeft) {
        // all the coefficient are in the padding zone.
        return internal::pset1<PacketReturnType>(Scalar(0));
      }
      else if (first >= firstPaddedRight && last < lastPaddedRight) {
        // all the coefficient are in the padding zone.
        return internal::pset1<PacketReturnType>(Scalar(0));
      }
      else if (first >= lastPaddedLeft && last < firstPaddedRight) {
        // all the coefficient are between the 2 padding zones.
        const Index idx = index / m_outputStrides[i+1];
        inputIndex += (idx - m_padding[i].first) * m_inputStrides[i];
        index -= idx * m_outputStrides[i+1];
      }
      else {
        // Every other case
        return packetWithPossibleZero(initialIndex);
      }
    }

    const Index last = index + packetSize - 1;
    const Index first = index;
    const Index lastPaddedLeft = m_padding[NumDims-1].first;
    const Index firstPaddedRight = (m_dimensions[NumDims-1] - m_padding[NumDims-1].second);
    const Index lastPaddedRight = m_outputStrides[NumDims-1];

    if (last < lastPaddedLeft) {
      // all the coefficient are in the padding zone.
      return internal::pset1<PacketReturnType>(Scalar(0));
    }
    else if (first >= firstPaddedRight && last < lastPaddedRight) {
      // all the coefficient are in the padding zone.
      return internal::pset1<PacketReturnType>(Scalar(0));
    }
    else if (first >= lastPaddedLeft && last < firstPaddedRight) {
      // all the coefficient are between the 2 padding zones.
      inputIndex += (index - m_padding[NumDims-1].first);
      return m_impl.template packet<Unaligned>(inputIndex);
    }
    // Every other case
    return packetWithPossibleZero(initialIndex);
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const
  {
    const int packetSize = internal::unpacket_traits<PacketReturnType>::size;
    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[packetSize];
    for (int i = 0; i < packetSize; ++i) {
      values[i] = coeff(index+i);
    }
    PacketReturnType rslt = internal::pload<PacketReturnType>(values);
    return rslt;
  }

  Dimensions m_dimensions;
  array<Index, NumDims+1> m_outputStrides;
  array<Index, NumDims> m_inputStrides;
  TensorEvaluator<ArgType, Device> m_impl;
  PaddingDimensions m_padding;
};




} // end namespace Eigen

#endif // EIGEN_CXX11_TENSOR_TENSOR_PADDING_H