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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include <armnn/backends/TensorHandle.hpp>
#include <armnn/Tensor.hpp>
#include <armnn/Types.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>
#include <reference/RefTensorHandle.hpp>
#include <BFloat16.hpp>
#include <Half.hpp>
namespace armnn
{
////////////////////////////////////////////
/// float32 helpers
////////////////////////////////////////////
inline const TensorInfo& GetTensorInfo(const ITensorHandle* tensorHandle)
{
// We know that reference workloads use RefTensorHandles for inputs and outputs
const RefTensorHandle* refTensorHandle =
PolymorphicDowncast<const RefTensorHandle*>(tensorHandle);
return refTensorHandle->GetTensorInfo();
}
template <typename DataType, typename PayloadType>
const DataType* GetInputTensorData(unsigned int idx, const PayloadType& data)
{
const ITensorHandle* tensorHandle = data.m_Inputs[idx];
return reinterpret_cast<const DataType*>(tensorHandle->Map());
}
template <typename DataType, typename PayloadType>
DataType* GetOutputTensorData(unsigned int idx, const PayloadType& data)
{
ITensorHandle* tensorHandle = data.m_Outputs[idx];
return reinterpret_cast<DataType*>(tensorHandle->Map());
}
template <typename DataType>
DataType* GetOutputTensorData(ITensorHandle* tensorHandle)
{
return reinterpret_cast<DataType*>(tensorHandle->Map());
}
template <typename PayloadType>
const float* GetInputTensorDataFloat(unsigned int idx, const PayloadType& data)
{
return GetInputTensorData<float>(idx, data);
}
template <typename PayloadType>
float* GetOutputTensorDataFloat(unsigned int idx, const PayloadType& data)
{
return GetOutputTensorData<float>(idx, data);
}
template <typename PayloadType>
const Half* GetInputTensorDataHalf(unsigned int idx, const PayloadType& data)
{
return GetInputTensorData<Half>(idx, data);
}
template <typename PayloadType>
Half* GetOutputTensorDataHalf(unsigned int idx, const PayloadType& data)
{
return GetOutputTensorData<Half>(idx, data);
}
template <typename PayloadType>
const BFloat16* GetInputTensorDataBFloat16(unsigned int idx, const PayloadType& data)
{
return GetInputTensorData<BFloat16>(idx, data);
}
template <typename PayloadType>
BFloat16* GetOutputTensorDataBFloat16(unsigned int idx, const PayloadType& data)
{
return GetOutputTensorData<BFloat16>(idx, data);
}
////////////////////////////////////////////
/// u8 helpers
////////////////////////////////////////////
template<typename T>
std::vector<float> Dequantize(const T* quant, const TensorInfo& info)
{
std::vector<float> ret(info.GetNumElements());
for (size_t i = 0; i < info.GetNumElements(); i++)
{
ret[i] = armnn::Dequantize(quant[i], info.GetQuantizationScale(), info.GetQuantizationOffset());
}
return ret;
}
template<typename T>
inline void Dequantize(const T* inputData, float* outputData, const TensorInfo& info)
{
for (unsigned int i = 0; i < info.GetNumElements(); i++)
{
outputData[i] = Dequantize<T>(inputData[i], info.GetQuantizationScale(), info.GetQuantizationOffset());
}
}
inline void Quantize(uint8_t* quant, const float* dequant, const TensorInfo& info)
{
for (size_t i = 0; i < info.GetNumElements(); i++)
{
quant[i] = armnn::Quantize<uint8_t>(dequant[i], info.GetQuantizationScale(), info.GetQuantizationOffset());
}
}
} //namespace armnn
|
