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
|
// 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_CONCATENATION_H
#define EIGEN_CXX11_TENSOR_TENSOR_CONCATENATION_H
namespace Eigen {
/** \class TensorConcatenationOp
* \ingroup CXX11_Tensor_Module
*
* \brief Tensor concatenation class.
*
*
*/
namespace internal {
template<typename Axis, typename LhsXprType, typename RhsXprType>
struct traits<TensorConcatenationOp<Axis, LhsXprType, RhsXprType> >
{
// Type promotion to handle the case where the types of the lhs and the rhs are different.
typedef typename promote_storage_type<typename LhsXprType::Scalar,
typename RhsXprType::Scalar>::ret Scalar;
typedef typename packet_traits<Scalar>::type Packet;
typedef typename promote_storage_type<typename traits<LhsXprType>::StorageKind,
typename traits<RhsXprType>::StorageKind>::ret StorageKind;
typedef typename promote_index_type<typename traits<LhsXprType>::Index,
typename traits<RhsXprType>::Index>::type Index;
typedef typename LhsXprType::Nested LhsNested;
typedef typename RhsXprType::Nested RhsNested;
typedef typename remove_reference<LhsNested>::type _LhsNested;
typedef typename remove_reference<RhsNested>::type _RhsNested;
enum { Flags = 0 };
};
template<typename Axis, typename LhsXprType, typename RhsXprType>
struct eval<TensorConcatenationOp<Axis, LhsXprType, RhsXprType>, Eigen::Dense>
{
typedef const TensorConcatenationOp<Axis, LhsXprType, RhsXprType>& type;
};
template<typename Axis, typename LhsXprType, typename RhsXprType>
struct nested<TensorConcatenationOp<Axis, LhsXprType, RhsXprType>, 1, typename eval<TensorConcatenationOp<Axis, LhsXprType, RhsXprType> >::type>
{
typedef TensorConcatenationOp<Axis, LhsXprType, RhsXprType> type;
};
} // end namespace internal
template<typename Axis, typename LhsXprType, typename RhsXprType>
class TensorConcatenationOp : public TensorBase<TensorConcatenationOp<Axis, LhsXprType, RhsXprType>, WriteAccessors>
{
public:
typedef typename internal::traits<TensorConcatenationOp>::Scalar Scalar;
typedef typename internal::traits<TensorConcatenationOp>::Packet Packet;
typedef typename internal::traits<TensorConcatenationOp>::StorageKind StorageKind;
typedef typename internal::traits<TensorConcatenationOp>::Index Index;
typedef typename internal::nested<TensorConcatenationOp>::type Nested;
typedef typename internal::promote_storage_type<typename LhsXprType::CoeffReturnType,
typename RhsXprType::CoeffReturnType>::ret CoeffReturnType;
typedef typename internal::promote_storage_type<typename LhsXprType::PacketReturnType,
typename RhsXprType::PacketReturnType>::ret PacketReturnType;
typedef typename NumTraits<Scalar>::Real RealScalar;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorConcatenationOp(const LhsXprType& lhs, const RhsXprType& rhs, Axis axis)
: m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_axis(axis) {}
EIGEN_DEVICE_FUNC
const typename internal::remove_all<typename LhsXprType::Nested>::type&
lhsExpression() const { return m_lhs_xpr; }
EIGEN_DEVICE_FUNC
const typename internal::remove_all<typename RhsXprType::Nested>::type&
rhsExpression() const { return m_rhs_xpr; }
EIGEN_DEVICE_FUNC Axis axis() const { return m_axis; }
protected:
typename LhsXprType::Nested m_lhs_xpr;
typename RhsXprType::Nested m_rhs_xpr;
const Axis m_axis;
};
// Eval as rvalue
template<typename Axis, typename LeftArgType, typename RightArgType, typename Device>
struct TensorEvaluator<const TensorConcatenationOp<Axis, LeftArgType, RightArgType>, Device>
{
typedef TensorConcatenationOp<Axis, LeftArgType, RightArgType> XprType;
typedef typename XprType::Index Index;
static const int NumDims = internal::array_size<typename TensorEvaluator<LeftArgType, Device>::Dimensions>::value;
static const int RightNumDims = internal::array_size<typename TensorEvaluator<RightArgType, Device>::Dimensions>::value;
typedef DSizes<Index, NumDims> Dimensions;
typedef typename XprType::Scalar Scalar;
typedef typename XprType::CoeffReturnType CoeffReturnType;
typedef typename XprType::PacketReturnType PacketReturnType;
enum {
IsAligned = false,
PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess & TensorEvaluator<RightArgType, Device>::PacketAccess,
};
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
: m_leftImpl(op.lhsExpression(), device), m_rightImpl(op.rhsExpression(), device), m_axis(op.axis())
{
EIGEN_STATIC_ASSERT(NumDims == RightNumDims, YOU_MADE_A_PROGRAMMING_MISTAKE)
eigen_assert(0 <= m_axis && m_axis < NumDims);
const Dimensions& lhs_dims = m_leftImpl.dimensions();
const Dimensions& rhs_dims = m_rightImpl.dimensions();
int i = 0;
for (; i < m_axis; ++i) {
eigen_assert(lhs_dims[i] > 0);
eigen_assert(lhs_dims[i] == rhs_dims[i]);
m_dimensions[i] = lhs_dims[i];
}
eigen_assert(lhs_dims[i] > 0); // Now i == m_axis.
eigen_assert(rhs_dims[i] > 0);
m_dimensions[i] = lhs_dims[i] + rhs_dims[i];
for (++i; i < NumDims; ++i) {
eigen_assert(lhs_dims[i] > 0);
eigen_assert(lhs_dims[i] == rhs_dims[i]);
m_dimensions[i] = lhs_dims[i];
}
m_leftStrides[0] = 1;
m_rightStrides[0] = 1;
m_outputStrides[0] = 1;
for (int i = 1; i < NumDims; ++i) {
m_leftStrides[i] = m_leftStrides[i-1] * lhs_dims[i-1];
m_rightStrides[i] = m_rightStrides[i-1] * rhs_dims[i-1];
m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1];
}
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
// TODO(phli): Add short-circuit memcpy evaluation if underlying data are linear?
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/)
{
m_leftImpl.evalSubExprsIfNeeded(NULL);
m_rightImpl.evalSubExprsIfNeeded(NULL);
return true;
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup()
{
m_leftImpl.cleanup();
m_rightImpl.cleanup();
}
// TODO(phli): attempt to speed this up. The integer divisions and modulo are slow.
// See CL/76180724 comments for more ideas.
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
{
// Collect dimension-wise indices (subs).
array<Index, NumDims> subs;
for (int i = NumDims - 1; i > 0; --i) {
subs[i] = index / m_outputStrides[i];
index -= subs[i] * m_outputStrides[i];
}
subs[0] = index;
const Dimensions& left_dims = m_leftImpl.dimensions();
if (subs[m_axis] < left_dims[m_axis]) {
Index left_index = subs[0];
for (int i = 1; i < NumDims; ++i) {
left_index += (subs[i] % left_dims[i]) * m_leftStrides[i];
}
return m_leftImpl.coeff(left_index);
} else {
subs[m_axis] -= left_dims[m_axis];
const Dimensions& right_dims = m_rightImpl.dimensions();
Index right_index = subs[0];
for (int i = 1; i < NumDims; ++i) {
right_index += (subs[i] % right_dims[i]) * m_rightStrides[i];
}
return m_rightImpl.coeff(right_index);
}
}
// TODO(phli): Add a real vectorization.
template<int LoadMode>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
{
static 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());
EIGEN_ALIGN_DEFAULT CoeffReturnType values[packetSize];
for (int i = 0; i < packetSize; ++i) {
values[i] = coeff(index+i);
}
PacketReturnType rslt = internal::pload<PacketReturnType>(values);
return rslt;
}
Scalar* data() const { return NULL; }
protected:
Dimensions m_dimensions;
array<Index, NumDims> m_outputStrides;
array<Index, NumDims> m_leftStrides;
array<Index, NumDims> m_rightStrides;
TensorEvaluator<LeftArgType, Device> m_leftImpl;
TensorEvaluator<RightArgType, Device> m_rightImpl;
const Axis m_axis;
};
} // end namespace Eigen
#endif // EIGEN_CXX11_TENSOR_TENSOR_CONCATENATION_H
|
