aboutsummaryrefslogtreecommitdiffhomepage
path: root/unsupported/Eigen/CXX11/src/Tensor/TensorScan.h
blob: 1f545ef1aa41609883460918e9126976ec58e95d (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
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2016 Igor Babuschkin <igor@babuschk.in>
//
// 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_SCAN_H
#define EIGEN_CXX11_TENSOR_TENSOR_SCAN_H

namespace Eigen {

namespace internal {

template <typename Op, typename XprType>
struct traits<TensorScanOp<Op, XprType> >
    : public traits<XprType> {
  typedef typename XprType::Scalar Scalar;
  typedef traits<XprType> XprTraits;
  typedef typename XprTraits::StorageKind StorageKind;
  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 Op, typename XprType>
struct eval<TensorScanOp<Op, XprType>, Eigen::Dense>
{
  typedef const TensorScanOp<Op, XprType>& type;
};

template<typename Op, typename XprType>
struct nested<TensorScanOp<Op, XprType>, 1,
            typename eval<TensorScanOp<Op, XprType> >::type>
{
  typedef TensorScanOp<Op, XprType> type;
};
} // end namespace internal

/** \class TensorScan
  * \ingroup CXX11_Tensor_Module
  *
  * \brief Tensor scan class.
  */
template <typename Op, typename XprType>
class TensorScanOp
    : public TensorBase<TensorScanOp<Op, XprType>, ReadOnlyAccessors> {
public:
  typedef typename Eigen::internal::traits<TensorScanOp>::Scalar Scalar;
  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
  typedef typename XprType::CoeffReturnType CoeffReturnType;
  typedef typename Eigen::internal::nested<TensorScanOp>::type Nested;
  typedef typename Eigen::internal::traits<TensorScanOp>::StorageKind StorageKind;
  typedef typename Eigen::internal::traits<TensorScanOp>::Index Index;

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorScanOp(
      const XprType& expr, const Index& axis, bool exclusive = false, const Op& op = Op())
      : m_expr(expr), m_axis(axis), m_accumulator(op), m_exclusive(exclusive) {}

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
  const Index axis() const { return m_axis; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
  const XprType& expression() const { return m_expr; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
  const Op accumulator() const { return m_accumulator; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
  bool exclusive() const { return m_exclusive; }

protected:
  typename XprType::Nested m_expr;
  const Index m_axis;
  const Op m_accumulator;
  const bool m_exclusive;
};

template <typename Self, typename Reducer, typename Device>
struct ScanLauncher;

// Eval as rvalue
template <typename Op, typename ArgType, typename Device>
struct TensorEvaluator<const TensorScanOp<Op, ArgType>, Device> {

  typedef TensorScanOp<Op, 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 internal::remove_const<typename XprType::Scalar>::type Scalar;
  typedef typename XprType::CoeffReturnType CoeffReturnType;
  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
  typedef TensorEvaluator<const TensorScanOp<Op, ArgType>, Device> Self;

  enum {
    IsAligned = false,
    PacketAccess = (internal::unpacket_traits<PacketReturnType>::size > 1),
    BlockAccess = false,
    Layout = TensorEvaluator<ArgType, Device>::Layout,
    CoordAccess = false,
    RawAccess = true
  };

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op,
                                                        const Device& device)
      : m_impl(op.expression(), device),
        m_device(device),
        m_exclusive(op.exclusive()),
        m_accumulator(op.accumulator()),
        m_size(m_impl.dimensions()[op.axis()]),
        m_stride(1),
        m_output(NULL) {

    // Accumulating a scalar isn't supported.
    EIGEN_STATIC_ASSERT((NumDims > 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
    eigen_assert(op.axis() >= 0 && op.axis() < NumDims);

    // Compute stride of scan axis
    const Dimensions& dims = m_impl.dimensions();
    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
      for (int i = 0; i < op.axis(); ++i) {
        m_stride = m_stride * dims[i];
      }
    } else {
      for (int i = NumDims - 1; i > op.axis(); --i) {
        m_stride = m_stride * dims[i];
      }
    }
  }

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

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Index& stride() const {
    return m_stride;
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Index& size() const {
    return m_size;
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Op& accumulator() const {
    return m_accumulator;
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool exclusive() const {
    return m_exclusive;
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& inner() const {
    return m_impl;
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const {
    return m_device;
  }

  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) {
    m_impl.evalSubExprsIfNeeded(NULL);
    ScanLauncher<Self, Op, Device> launcher;
    if (data) {
      launcher(*this, data);
      return false;
    }

    const Index total_size = internal::array_prod(dimensions());
    m_output = static_cast<CoeffReturnType*>(m_device.allocate(total_size * sizeof(Scalar)));
    launcher(*this, m_output);
    return true;
  }

  template<int LoadMode>
  EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const {
    return internal::ploadt<PacketReturnType, LoadMode>(m_output + index);
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const
  {
    return m_output;
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
  {
    return m_output[index];
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool) const {
    return TensorOpCost(sizeof(CoeffReturnType), 0, 0);
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
    if (m_output != NULL) {
      m_device.deallocate(m_output);
      m_output = NULL;
    }
    m_impl.cleanup();
  }

protected:
  TensorEvaluator<ArgType, Device> m_impl;
  const Device& m_device;
  const bool m_exclusive;
  Op m_accumulator;
  const Index m_size;
  Index m_stride;
  CoeffReturnType* m_output;
};

// CPU implementation of scan
// TODO(ibab) This single-threaded implementation should be parallelized,
// at least by running multiple scans at the same time.
template <typename Self, typename Reducer, typename Device>
struct ScanLauncher {
  void operator()(Self& self, typename Self::CoeffReturnType *data) {
    Index total_size = internal::array_prod(self.dimensions());

    // We fix the index along the scan axis to 0 and perform a
    // scan per remaining entry. The iteration is split into two nested
    // loops to avoid an integer division by keeping track of each idx1 and idx2.
    for (Index idx1 = 0; idx1 < total_size; idx1 += self.stride() * self.size()) {
      for (Index idx2 = 0; idx2 < self.stride(); idx2++) {
        // Calculate the starting offset for the scan
        Index offset = idx1 + idx2;

        // Compute the scan along the axis, starting at the calculated offset
        typename Self::CoeffReturnType accum = self.accumulator().initialize();
        for (Index idx3 = 0; idx3 < self.size(); idx3++) {
          Index curr = offset + idx3 * self.stride();

          if (self.exclusive()) {
            data[curr] = self.accumulator().finalize(accum);
            self.accumulator().reduce(self.inner().coeff(curr), &accum);
          } else {
            self.accumulator().reduce(self.inner().coeff(curr), &accum);
            data[curr] = self.accumulator().finalize(accum);
          }
        }
      }
    }
  }
};

#if defined(EIGEN_USE_GPU) && defined(EIGEN_CUDACC)

// GPU implementation of scan
// TODO(ibab) This placeholder implementation performs multiple scans in
// parallel, but it would be better to use a parallel scan algorithm and
// optimize memory access.
template <typename Self, typename Reducer>
__global__ void ScanKernel(Self self, Index total_size, typename Self::CoeffReturnType* data) {
  // Compute offset as in the CPU version
  Index val = threadIdx.x + blockIdx.x * blockDim.x;
  Index offset = (val / self.stride()) * self.stride() * self.size() + val % self.stride();

  if (offset + (self.size() - 1) * self.stride() < total_size) {
    // Compute the scan along the axis, starting at the calculated offset
    typename Self::CoeffReturnType accum = self.accumulator().initialize();
    for (Index idx = 0; idx < self.size(); idx++) {
      Index curr = offset + idx * self.stride();
      if (self.exclusive()) {
        data[curr] = self.accumulator().finalize(accum);
        self.accumulator().reduce(self.inner().coeff(curr), &accum);
      } else {
        self.accumulator().reduce(self.inner().coeff(curr), &accum);
        data[curr] = self.accumulator().finalize(accum);
      }
    }
  }
  __syncthreads();

}

template <typename Self, typename Reducer>
struct ScanLauncher<Self, Reducer, GpuDevice> {
  void operator()(const Self& self, typename Self::CoeffReturnType* data) {
     Index total_size = internal::array_prod(self.dimensions());
     Index num_blocks = (total_size / self.size() + 63) / 64;
     Index block_size = 64;
     LAUNCH_CUDA_KERNEL((ScanKernel<Self, Reducer>), num_blocks, block_size, 0, self.device(), self, total_size, data);
  }
};
#endif  // EIGEN_USE_GPU && EIGEN_CUDACC

}  // end namespace Eigen

#endif  // EIGEN_CXX11_TENSOR_TENSOR_SCAN_H