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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include <armnn/Types.hpp>
#include <cmath>
#include <algorithm>
namespace armnn
{
using OffsetScalePair = std::pair<float, int>;
struct IQuantizationScheme
{
virtual OffsetScalePair ComputeScheme(double min, double max) const = 0;
virtual int NumBits() const = 0;
virtual DataType GetDataType() const = 0;
virtual ~IQuantizationScheme() {}
};
struct QAsymmU8QuantizationScheme : IQuantizationScheme
{
OffsetScalePair ComputeScheme(double min, double max) const override
{
if (min > max)
{
throw InvalidArgumentException("min > max will result in invalid quantization.");
}
double highest = (1 << NumBits()) - 1;
min = std::min(0.0, min); // min <= 0.0
max = std::max(0.0, max); // max >= 0.0
// To avoid dividing by zero when quantizing a zero filled tensor
if (min == 0.0 && max == 0.0)
{
max = 1.0;
}
// Assumes quantization range [0-highest]
double scale = (max-min) / highest;
double offset = -min / scale;
// Clamp offset [0-highest]
offset = std::max(0.0, std::min(highest, offset));
return std::make_pair(static_cast<float>(scale), static_cast<int>(std::round(offset)));
}
int NumBits() const override { return 8; }
DataType GetDataType() const override { return DataType::QAsymmU8; }
};
struct QAsymmS8QuantizationScheme : IQuantizationScheme
{
OffsetScalePair ComputeScheme(double min, double max) const override
{
if (min > max)
{
throw InvalidArgumentException("min > max will result in invalid quantization.");
}
double highest = (1 << NumBits()) - 1;
min = std::min(0.0, min); // min <= 0.0
max = std::max(0.0, max); // max >= 0.0
// To avoid dividing by zero when quantizing a zero filled tensor
if (min == 0.0 && max == 0.0)
{
max = 1.0;
}
// Assumes quantization range [0-255]
double scale = (max-min) / highest ;
double offset = - min / scale;
//Clamp 0 to Highest
offset = std::max(0.0, std::min(highest, offset));
//-128 on offset to cast to signed range
return std::make_pair(static_cast<float>(scale), static_cast<int>(std::round(offset)-128));
}
int NumBits() const override { return 8; }
DataType GetDataType() const override { return DataType::QAsymmS8; }
};
struct QSymmS8QuantizationScheme : IQuantizationScheme
{
OffsetScalePair ComputeScheme(double min, double max) const override
{
if (min > max)
{
throw InvalidArgumentException("min > max will result in invalid quantization.");
}
// To avoid dividing by zero when quantizing a zero filled tensor
if (min == 0.0 && max == 0.0)
{
max = 1.0;
}
double highest = (1 << (NumBits()-1)) - 1; // (numbits-1) accounts for the sign bit
double extent = std::max(std::abs(min), std::abs(max));
double scale = extent / highest;
return std::make_pair(static_cast<float>(scale), 0);
}
int NumBits() const override { return 8; }
DataType GetDataType() const override { return DataType::QSymmS8; }
};
struct QSymm16QuantizationScheme : IQuantizationScheme
{
OffsetScalePair ComputeScheme(double min, double max) const override
{
if (min > max)
{
throw InvalidArgumentException("min > max will result in invalid quantization.");
}
// To avoid dividing by zero when quantizing a zero filled tensor
if (min == 0.0 && max == 0.0)
{
max = 1.0;
}
double highest = (1 << (NumBits()-1)) - 1; // (numbits-1) accounts for the sign bit
double extent = std::max(std::abs(min), std::abs(max));
double scale = extent / highest;
return std::make_pair(static_cast<float>(scale), 0);
}
int NumBits() const override { return 16; }
DataType GetDataType() const override { return DataType::QSymmS16; }
};
} // namespace armnn
|
