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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include "BaseIterator.hpp"
#include <armnnUtils/FloatingPointConverter.hpp>
#include <armnnUtils/TensorUtils.hpp>
#include <boost/assert.hpp>
namespace armnn
{
namespace
{
inline std::unique_ptr<Decoder<float>> MakeSigned32PerAxisDecoder(const TensorInfo& info, const void* data)
{
auto params = armnnUtils::GetPerAxisParams(info);
return std::make_unique<ScaledInt32PerAxisDecoder>(
static_cast<const int32_t*>(data),
params.second,
params.first);
}
inline std::unique_ptr<Decoder<float>> MakeSigned32Decoder(const TensorInfo& info, const void* data)
{
if(info.HasMultipleQuantizationScales())
{
// NOTE: If we have multiple quantization scales, we create a ScaledInt32PerAxisDecoder.
// This will be used to decode per-axis quantized convolution biases.
return MakeSigned32PerAxisDecoder(info, data);
}
else
{
if (info.GetQuantizationDim().has_value())
{
// NOTE: Even though we only have a single quantization scale, if the quantization
// dimension is set, the tensor has per-axis quantization and we need to create a
// ScaledInt32PerAxisDecoder
return MakeSigned32PerAxisDecoder(info, data);
}
const float scale = info.GetQuantizationScale();
if (scale == 0.f)
{
// NOTE:: If no quantization scale is set, we create an Int32Decoder, which simply
// casts the int value to float. This will be used for any INT32 data other than
// convolution biases.
return std::make_unique<Int32Decoder>(static_cast<const int32_t*>(data));
}
// NOTE: If we only have a single (non-zero) quantization scale and no quantization
// dimension is specified, we need to create a ScaledInt32Decoder. This will be used
// to decode per-tensor quantized convolution biases.
return std::make_unique<ScaledInt32Decoder>(static_cast<const int32_t*>(data), scale);
}
}
} // anonymous namespace
template<typename T>
inline std::unique_ptr<Decoder<T>> MakeDecoder(const TensorInfo& info, const void* data = nullptr);
template<>
inline std::unique_ptr<Decoder<float>> MakeDecoder(const TensorInfo& info, const void* data)
{
switch(info.GetDataType())
{
ARMNN_NO_DEPRECATE_WARN_BEGIN
case armnn::DataType::QuantizedSymm8PerAxis:
{
std::pair<unsigned int, std::vector<float>> params = armnnUtils::GetPerAxisParams(info);
return std::make_unique<QSymm8PerAxisDecoder>(
static_cast<const int8_t*>(data),
params.second,
params.first);
}
ARMNN_NO_DEPRECATE_WARN_END
case DataType::QAsymmU8:
{
return std::make_unique<QASymm8Decoder>(
static_cast<const uint8_t*>(data),
info.GetQuantizationScale(),
info.GetQuantizationOffset());
}
case DataType::QSymmS16:
{
return std::make_unique<QSymm16Decoder>(
static_cast<const int16_t*>(data),
info.GetQuantizationScale(),
info.GetQuantizationOffset());
}
case DataType::Float16:
{
return std::make_unique<Float16Decoder>(static_cast<const Half*>(data));
}
case DataType::Float32:
{
return std::make_unique<Float32Decoder>(static_cast<const float*>(data));
}
case DataType::Signed32:
{
return MakeSigned32Decoder(info, data);
}
case DataType::QSymmS8:
{
if (info.HasPerAxisQuantization())
{
std::pair<unsigned int, std::vector<float>> params = armnnUtils::GetPerAxisParams(info);
return std::make_unique<QSymm8PerAxisDecoder>(
static_cast<const int8_t*>(data),
params.second,
params.first);
}
else
{
return std::make_unique<QSymmS8Decoder>(
static_cast<const int8_t*>(data),
info.GetQuantizationScale(),
info.GetQuantizationOffset());
}
}
default:
{
BOOST_ASSERT_MSG(false, "Unsupported Data Type!");
break;
}
}
return nullptr;
}
} //namespace armnn
|
