aboutsummaryrefslogtreecommitdiffhomepage
path: root/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
blob: 0999815d754b01023fd44d92dbd751624bf5b2e2 (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
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
// 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_SHUFFLING_H
#define EIGEN_CXX11_TENSOR_TENSOR_SHUFFLING_H

namespace Eigen {

/** \class TensorShuffling
  * \ingroup CXX11_Tensor_Module
  *
  * \brief Tensor shuffling class.
  *
  *
  */
namespace internal {
template<typename Shuffle, typename XprType>
struct traits<TensorShufflingOp<Shuffle, XprType> > : public traits<XprType>
{
  typedef typename XprType::Scalar Scalar;
  typedef traits<XprType> XprTraits;
  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;
  typedef typename XprTraits::PointerType PointerType;
};

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

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

}  // end namespace internal



template<typename Shuffle, typename XprType>
class TensorShufflingOp : public TensorBase<TensorShufflingOp<Shuffle, XprType> >
{
  public:
    typedef TensorBase<TensorShufflingOp<Shuffle, XprType> > Base;
    typedef typename Eigen::internal::traits<TensorShufflingOp>::Scalar Scalar;
    typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
    typedef typename XprType::CoeffReturnType CoeffReturnType;
    typedef typename Eigen::internal::nested<TensorShufflingOp>::type Nested;
    typedef typename Eigen::internal::traits<TensorShufflingOp>::StorageKind StorageKind;
    typedef typename Eigen::internal::traits<TensorShufflingOp>::Index Index;

    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorShufflingOp(const XprType& expr, const Shuffle& shfl)
      : m_xpr(expr), m_shuffle(shfl) {}

    EIGEN_DEVICE_FUNC
    const Shuffle& shufflePermutation() const { return m_shuffle; }

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

    EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(TensorShufflingOp)


  protected:
    typename XprType::Nested m_xpr;
    const Shuffle m_shuffle;
};


// Eval as rvalue
template<typename Shuffle, typename ArgType, typename Device>
struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
{
  typedef TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device> Self;
  typedef TensorShufflingOp<Shuffle, ArgType> XprType;
  typedef typename XprType::Index Index;
  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
  typedef DSizes<Index, NumDims> Dimensions;
  typedef typename XprType::Scalar Scalar;
  typedef typename XprType::CoeffReturnType CoeffReturnType;
  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
  typedef StorageMemory<CoeffReturnType, Device> Storage;
  typedef typename Storage::Type EvaluatorPointerType;

  enum {
    IsAligned         = false,
    PacketAccess      = (PacketType<CoeffReturnType, Device>::size > 1),
    BlockAccess       = TensorEvaluator<ArgType, Device>::RawAccess,
    PreferBlockAccess = true,
    Layout            = TensorEvaluator<ArgType, Device>::Layout,
    CoordAccess       = false,  // to be implemented
    RawAccess         = false
  };

  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;

  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;

  typedef typename internal::TensorMaterializedBlock<ScalarNoConst, NumDims,
                                                     Layout, Index>
      TensorBlock;
  //===--------------------------------------------------------------------===//

  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
      : m_device(device),
        m_impl(op.expression(), device)
  {
    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
    const Shuffle& shuffle = op.shufflePermutation();
    m_is_identity = true;
    for (int i = 0; i < NumDims; ++i) {
      m_shuffle[i] = static_cast<int>(shuffle[i]);
      m_dimensions[i] = input_dims[shuffle[i]];
      m_inverseShuffle[shuffle[i]] = i;
      if (m_is_identity && shuffle[i] != i) {
        m_is_identity = false;
      }
    }

    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
      m_unshuffledInputStrides[0] = 1;
      m_outputStrides[0] = 1;

      for (int i = 1; i < NumDims; ++i) {
        m_unshuffledInputStrides[i] =
            m_unshuffledInputStrides[i - 1] * input_dims[i - 1];
        m_outputStrides[i] = m_outputStrides[i - 1] * m_dimensions[i - 1];
        m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i]);
      }
    } else {
      m_unshuffledInputStrides[NumDims - 1] = 1;
      m_outputStrides[NumDims - 1] = 1;
      for (int i = NumDims - 2; i >= 0; --i) {
        m_unshuffledInputStrides[i] =
            m_unshuffledInputStrides[i + 1] * input_dims[i + 1];
        m_outputStrides[i] = m_outputStrides[i + 1] * m_dimensions[i + 1];
        m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(m_outputStrides[i]);
      }
    }

    for (int i = 0; i < NumDims; ++i) {
      m_inputStrides[i] = m_unshuffledInputStrides[shuffle[i]];
    }
  }

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

  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
    m_impl.evalSubExprsIfNeeded(NULL);
    return true;
  }

#ifdef EIGEN_USE_THREADS
  template <typename EvalSubExprsCallback>
  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
      EvaluatorPointerType, EvalSubExprsCallback done) {
    m_impl.evalSubExprsIfNeededAsync(nullptr, [done](bool) { done(true); });
  }
#endif  // EIGEN_USE_THREADS

  EIGEN_STRONG_INLINE void cleanup() {
    m_impl.cleanup();
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
  {
    if (m_is_identity) {
      return m_impl.coeff(index);
    } else {
      return m_impl.coeff(srcCoeff(index));
    }
  }

  template <int LoadMode, typename Self, bool ImplPacketAccess>
  struct PacketLoader {
    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
    static PacketReturnType Run(const Self& self, Index index) {
      EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
      EIGEN_UNROLL_LOOP
      for (int i = 0; i < PacketSize; ++i) {
        values[i] = self.coeff(index + i);
      }
      PacketReturnType rslt = internal::pload<PacketReturnType>(values);
      return rslt;
    }
  };

  template<int LoadMode, typename Self>
  struct PacketLoader<LoadMode, Self, true> {
    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
    static PacketReturnType Run(const Self& self, Index index) {
      if (self.m_is_identity) {
        return self.m_impl.template packet<LoadMode>(index);
      } else {
        EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
        EIGEN_UNROLL_LOOP
        for (int i = 0; i < PacketSize; ++i) {
          values[i] = self.coeff(index + i);
        }
        PacketReturnType rslt = internal::pload<PacketReturnType>(values);
        return rslt;
      }
    }
  };

  template<int LoadMode>
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
  {
    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
        eigen_assert(index + PacketSize - 1 < dimensions().TotalSize());
    return PacketLoader<LoadMode, Self, TensorEvaluator<ArgType, Device>::PacketAccess>::Run(*this, index);
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
  internal::TensorBlockResourceRequirements getResourceRequirements() const {
    static const int inner_dim =
        Layout == static_cast<int>(ColMajor) ? 0 : NumDims - 1;

    const size_t target_size = m_device.firstLevelCacheSize();
    const bool inner_dim_shuffled = m_shuffle[inner_dim] != inner_dim;

    // Shuffled inner dimensions leads to a random memory access, which is not
    // captured by default cost model bytes loaded/stored. We add this cost
    // explicitly. The number of cycles picked based on the benchmarks.
    // TODO(ezhulenev): This number was picked based on a very questionable
    // benchmarks, add benchmarks that are representative of real workloads.
    using BlockRequirements = internal::TensorBlockResourceRequirements;
    if (inner_dim_shuffled) {
      return BlockRequirements::uniform<Scalar>(target_size)
          .addCostPerCoeff({0, 0, NumDims * 28});
    } else {
      return BlockRequirements::skewed<Scalar>(target_size);
    }
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
  block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
          bool root_of_expr_ast = false) const {
    assert(m_impl.data() != NULL);

    typedef internal::TensorBlockIO<ScalarNoConst, Index, NumDims, Layout>
        TensorBlockIO;
    typedef typename TensorBlockIO::Dst TensorBlockIODst;
    typedef typename TensorBlockIO::Src TensorBlockIOSrc;

    const typename TensorBlock::Storage block_storage =
        TensorBlock::prepareStorage(
            desc, scratch, /*allow_strided_storage=*/root_of_expr_ast);

    typename TensorBlockIO::Dimensions input_strides(m_unshuffledInputStrides);
    TensorBlockIOSrc src(input_strides, m_impl.data(), srcCoeff(desc.offset()));

    TensorBlockIODst dst(block_storage.dimensions(), block_storage.strides(),
                         block_storage.data());

    typename TensorBlockIO::DimensionsMap dst_to_src_dim_map(m_shuffle);
    TensorBlockIO::Copy(dst, src, dst_to_src_dim_map);

    return block_storage.AsTensorMaterializedBlock();
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
    const double compute_cost = m_is_identity ? TensorOpCost::AddCost<Index>() :
                                NumDims * (2 * TensorOpCost::AddCost<Index>() +
                                           2 * TensorOpCost::MulCost<Index>() +
                                           TensorOpCost::DivCost<Index>());
    return m_impl.costPerCoeff(vectorized) +
           TensorOpCost(0, 0, compute_cost, m_is_identity /* vectorized */, PacketSize);
  }

  EIGEN_DEVICE_FUNC typename Storage::Type data() const { return NULL; }

#ifdef EIGEN_USE_SYCL
   // binding placeholder accessors to a command group handler for SYCL
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
    m_impl.bind(cgh);
  }
#endif
 protected:
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index GetBlockOutputIndex(
      Index input_index,
      const DSizes<Index, NumDims>& input_block_strides,
      const DSizes<Index, NumDims>& output_block_strides,
      const DSizes<internal::TensorIntDivisor<Index>, NumDims>& fast_input_block_strides) const {
    Index output_index = 0;
    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
      for (int i = NumDims - 1; i > 0; --i) {
        const Index idx = input_index / fast_input_block_strides[i];
        output_index += idx * output_block_strides[m_inverseShuffle[i]];
        input_index -= idx * input_block_strides[i];
      }
      return output_index + input_index *
          output_block_strides[m_inverseShuffle[0]];
    } else {
      for (int i = 0; i < NumDims - 1; ++i) {
        const Index idx = input_index / fast_input_block_strides[i];
        output_index += idx * output_block_strides[m_inverseShuffle[i]];
        input_index -= idx * input_block_strides[i];
      }
      return output_index + input_index *
          output_block_strides[m_inverseShuffle[NumDims - 1]];
    }
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const {
    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_fastOutputStrides[i];
        inputIndex += idx * m_inputStrides[i];
        index -= idx * m_outputStrides[i];
      }
      return inputIndex + index * m_inputStrides[0];
    } else {
      for (int i = 0; i < NumDims - 1; ++i) {
        const Index idx = index / m_fastOutputStrides[i];
        inputIndex += idx * m_inputStrides[i];
        index -= idx * m_outputStrides[i];
      }
      return inputIndex + index * m_inputStrides[NumDims - 1];
    }
  }

  Dimensions m_dimensions;
  bool m_is_identity;
  array<int, NumDims> m_shuffle;
  array<Index, NumDims> m_inverseShuffle;  // TODO(ezhulenev): Make it int type.
  array<Index, NumDims> m_outputStrides;
  array<internal::TensorIntDivisor<Index>, NumDims> m_fastOutputStrides;
  array<Index, NumDims> m_inputStrides;
  array<Index, NumDims> m_unshuffledInputStrides;

  const Device EIGEN_DEVICE_REF m_device;
  TensorEvaluator<ArgType, Device> m_impl;
};


// Eval as lvalue
template<typename Shuffle, typename ArgType, typename Device>
struct TensorEvaluator<TensorShufflingOp<Shuffle, ArgType>, Device>
    : public TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
{
  typedef TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device> Base;

  typedef TensorShufflingOp<Shuffle, ArgType> XprType;
  typedef typename XprType::Index Index;
  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
  typedef DSizes<Index, NumDims> Dimensions;
  typedef typename XprType::Scalar Scalar;
  typedef typename XprType::CoeffReturnType CoeffReturnType;
  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;

  enum {
    IsAligned         = false,
    PacketAccess      = (PacketType<CoeffReturnType, Device>::size > 1),
    BlockAccess       = TensorEvaluator<ArgType, Device>::RawAccess,
    PreferBlockAccess = true,
    Layout            = TensorEvaluator<ArgType, Device>::Layout,
    RawAccess         = false
  };

  typedef typename internal::remove_const<Scalar>::type ScalarNoConst;

  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
  //===--------------------------------------------------------------------===//

  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
      : Base(op, device)
  { }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
  {
    return this->m_impl.coeffRef(this->srcCoeff(index));
  }

  template <int StoreMode> EIGEN_STRONG_INLINE
  void writePacket(Index index, const PacketReturnType& x)
  {
    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)

    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
    internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
    EIGEN_UNROLL_LOOP
    for (int i = 0; i < PacketSize; ++i) {
      this->coeffRef(index+i) = values[i];
    }
  }

  template <typename TensorBlock>
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writeBlock(
      const TensorBlockDesc& desc, const TensorBlock& block) {
    eigen_assert(this->m_impl.data() != NULL);

    typedef internal::TensorBlockIO<ScalarNoConst, Index, NumDims, Layout>
        TensorBlockIO;
    typedef typename TensorBlockIO::Dst TensorBlockIODst;
    typedef typename TensorBlockIO::Src TensorBlockIOSrc;

    const Scalar* block_buffer = block.data();

    // TODO(ezhulenev): TensorBlockIO should be able to read from any Eigen
    // expression with coefficient and packet access as `src`.
    void* mem = NULL;
    if (block_buffer == NULL) {
      mem = this->m_device.allocate(desc.size() * sizeof(Scalar));
      ScalarNoConst* buf = static_cast<ScalarNoConst*>(mem);

      typedef internal::TensorBlockAssignment<
          ScalarNoConst, NumDims, typename TensorBlock::XprType, Index>
          TensorBlockAssignment;

      TensorBlockAssignment::Run(
          TensorBlockAssignment::target(
              desc.dimensions(), internal::strides<Layout>(desc.dimensions()),
              buf),
          block.expr());

      block_buffer = buf;
    }

    // Read from block.
    TensorBlockIOSrc src(internal::strides<Layout>(desc.dimensions()),
                         block_buffer);

    // Write to the output buffer.
    typename TensorBlockIO::Dimensions output_strides(
        this->m_unshuffledInputStrides);
    typename TensorBlockIO::Dimensions output_dimensions;
    for (int i = 0; i < NumDims; ++i) {
      output_dimensions[this->m_shuffle[i]] = desc.dimension(i);
    }
    TensorBlockIODst dst(output_dimensions, output_strides, this->m_impl.data(),
                         this->srcCoeff(desc.offset()));

    // Reorder dimensions according to the shuffle.
    typename TensorBlockIO::DimensionsMap dst_to_src_dim_map;
    for (int i = 0; i < NumDims; ++i) {
      dst_to_src_dim_map[i] = static_cast<int>(this->m_inverseShuffle[i]);
    }
    TensorBlockIO::Copy(dst, src, dst_to_src_dim_map);

    // Deallocate temporary buffer used for the block materialization.
    if (mem != NULL) this->m_device.deallocate(mem);
  }
};


} // end namespace Eigen

#endif // EIGEN_CXX11_TENSOR_TENSOR_SHUFFLING_H