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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include "TensorFwd.hpp"
#include "Exceptions.hpp"
#include "Optional.hpp"
#include "Types.hpp"
#include <array>
#include <initializer_list>
#include <vector>
namespace armnn
{
class TensorShape
{
public:
/// Empty (invalid) constructor.
TensorShape();
TensorShape(unsigned int numDimensions);
TensorShape(unsigned int numDimensions, const unsigned int* dimensionSizes);
TensorShape(std::initializer_list<unsigned int> dimensionSizeList);
TensorShape(const TensorShape& other);
TensorShape& operator=(const TensorShape& other);
unsigned int operator[](unsigned int i) const;
unsigned int& operator[](unsigned int i);
bool operator==(const TensorShape& other) const;
bool operator!=(const TensorShape& other) const;
unsigned int GetNumDimensions() const { return m_NumDimensions; }
unsigned int GetNumElements() const;
private:
std::array<unsigned int, MaxNumOfTensorDimensions> m_Dimensions;
unsigned int m_NumDimensions;
void CheckDimensionIndex(unsigned int i) const;
};
class TensorInfo
{
public:
/// Empty (invalid) constructor.
TensorInfo();
TensorInfo(const TensorShape& shape,
DataType dataType,
float quantizationScale = 0.0f,
int32_t quantizationOffset = 0);
TensorInfo(unsigned int numDimensions,
const unsigned int* dimensionSizes,
DataType dataType,
float quantizationScale = 0.0f,
int32_t quantizationOffset = 0);
TensorInfo(const TensorShape& shape,
DataType dataType,
const std::vector<float>& quantizationScales,
unsigned int quantizationDim);
TensorInfo(unsigned int numDimensions,
const unsigned int* dimensionSizes,
DataType dataType,
const std::vector<float>& quantizationScales,
unsigned int quantizationDim);
TensorInfo(const TensorInfo& other);
TensorInfo& operator=(const TensorInfo& other);
bool operator==(const TensorInfo& other) const;
bool operator!=(const TensorInfo& other) const;
const TensorShape& GetShape() const { return m_Shape; }
TensorShape& GetShape() { return m_Shape; }
void SetShape(const TensorShape& newShape) { m_Shape = newShape; }
unsigned int GetNumDimensions() const { return m_Shape.GetNumDimensions(); }
unsigned int GetNumElements() const { return m_Shape.GetNumElements(); }
DataType GetDataType() const { return m_DataType; }
void SetDataType(DataType type) { m_DataType = type; }
bool HasMultipleQuantizationScales() const { return m_Quantization.m_Scales.size() > 1; }
bool HasPerAxisQuantization() const;
std::vector<float> GetQuantizationScales() const;
void SetQuantizationScales(const std::vector<float>& scales);
float GetQuantizationScale() const;
void SetQuantizationScale(float scale);
int32_t GetQuantizationOffset() const;
void SetQuantizationOffset(int32_t offset);
Optional<unsigned int> GetQuantizationDim() const;
void SetQuantizationDim(const Optional<unsigned int>& quantizationDim);
bool IsQuantized() const;
/// Check that the types are the same and, if quantize, that the quantization parameters are the same.
bool IsTypeSpaceMatch(const TensorInfo& other) const;
unsigned int GetNumBytes() const;
private:
TensorShape m_Shape;
DataType m_DataType;
/// Vectors of scale and offset are used for per-axis quantization.
struct Quantization
{
Quantization()
: m_Scales{}
, m_Offset(EmptyOptional())
, m_QuantizationDim(EmptyOptional()) {}
bool operator==(const Quantization& other) const
{
return ((m_Scales == other.m_Scales) && (m_Offset == other.m_Offset) &&
(m_QuantizationDim == other.m_QuantizationDim));
}
std::vector<float> m_Scales;
Optional<int32_t> m_Offset;
Optional<unsigned int> m_QuantizationDim;
} m_Quantization;
};
using BindingPointInfo = std::pair<armnn::LayerBindingId, armnn::TensorInfo>;
template<typename MemoryType>
class BaseTensor
{
public:
/// Empty (invalid) constructor.
BaseTensor();
/// Constructor from a raw memory pointer.
/// @param memoryArea - Region of CPU-addressable memory where tensor data will be stored. Must be valid while
/// workloads are on the fly. Tensor instances do not claim ownership of referenced memory regions, that is,
/// no attempt will be made by ArmNN to free these memory regions automatically.
BaseTensor(const TensorInfo& info, MemoryType memoryArea);
/// Tensors are copyable.
BaseTensor(const BaseTensor& other);
/// Tensors are copyable.
BaseTensor& operator=(const BaseTensor&);
const TensorInfo& GetInfo() const { return m_Info; }
TensorInfo& GetInfo() { return m_Info; }
const TensorShape& GetShape() const { return m_Info.GetShape(); }
TensorShape& GetShape() { return m_Info.GetShape(); }
DataType GetDataType() const { return m_Info.GetDataType(); }
unsigned int GetNumDimensions() const { return m_Info.GetNumDimensions(); }
unsigned int GetNumBytes() const { return m_Info.GetNumBytes(); }
unsigned int GetNumElements() const { return m_Info.GetNumElements(); }
MemoryType GetMemoryArea() const { return m_MemoryArea; }
protected:
/// Protected destructor to stop users from making these
/// (could still new one on the heap and then leak it...)
~BaseTensor() {}
MemoryType m_MemoryArea;
private:
TensorInfo m_Info;
};
/// A tensor defined by a TensorInfo (shape and data type) and a mutable backing store.
class Tensor : public BaseTensor<void*>
{
public:
/// Brings in the constructors and assignment operator.
using BaseTensor<void*>::BaseTensor;
};
/// A tensor defined by a TensorInfo (shape and data type) and an immutable backing store.
class ConstTensor : public BaseTensor<const void*>
{
public:
/// Brings in the constructors and assignment operator.
using BaseTensor<const void*>::BaseTensor;
ConstTensor() : BaseTensor<const void*>() {} // This needs to be redefined explicitly??
/// Can be implicitly constructed from non-const Tensor.
ConstTensor(const Tensor& other) : BaseTensor<const void*>(other.GetInfo(), other.GetMemoryArea()) {}
/// Constructor from a backing container.
/// @param container - An stl-like container type which implements data() and size() methods.
/// Presence of data() and size() is a strong indicator of the continuous memory layout of the container,
/// which is a requirement for Tensor data. Tensor instances do not claim ownership of referenced memory regions,
/// that is, no attempt will be made by ArmNN to free these memory regions automatically.
template < template<typename, typename...> class ContainerType, typename T, typename...ContainerArgs >
ConstTensor(const TensorInfo& info, const ContainerType<T, ContainerArgs...>& container)
: BaseTensor<const void*>(info, container.data())
{
if (container.size() * sizeof(T) != info.GetNumBytes())
{
throw InvalidArgumentException("Container size is not correct");
}
}
};
using InputTensors = std::vector<std::pair<LayerBindingId, class ConstTensor>>;
using OutputTensors = std::vector<std::pair<LayerBindingId, class Tensor>>;
} // namespace armnn
|
