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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include "Optimization.hpp"
#include <armnnUtils/FloatingPointConverter.hpp>
#include <backendsCommon/TensorHandle.hpp>
#include <armnn/utility/IgnoreUnused.hpp>
#include <BFloat16.hpp>
#include <Half.hpp>
namespace armnn
{
namespace optimizations
{
struct BFloat16ToFloat32
{
static void Func(std::shared_ptr<ConstTensorHandle>& handle)
{
const TensorInfo& info = handle->GetTensorInfo();
if (info.GetDataType() == DataType::BFloat16)
{
std::vector<float> newValues(info.GetNumElements());
armnnUtils::FloatingPointConverter::ConvertBFloat16ToFloat32(handle->GetConstTensor<BFloat16>(),
info.GetNumElements(),
newValues.data());
TensorInfo newInfo(info.GetShape(), DataType::Float32);
ConstTensor newInput(newInfo, newValues);
handle.reset(new ScopedTensorHandle(newInput));
}
}
};
struct Float16ToFloat32
{
static void Func(std::shared_ptr<ConstTensorHandle>& handle)
{
const TensorInfo& info = handle->GetTensorInfo();
if (info.GetDataType() == DataType::Float16)
{
std::vector<float> newValues(info.GetNumElements());
armnnUtils::FloatingPointConverter::ConvertFloat16To32(handle->GetConstTensor<Half>(),
info.GetNumElements(),
newValues.data());
TensorInfo newInfo(info.GetShape(), DataType::Float32);
ConstTensor newInput(newInfo, newValues);
handle.reset(new ScopedTensorHandle(newInput));
}
}
};
struct Float32ToBFloat16
{
static void Func(std::shared_ptr<ConstTensorHandle>& handle)
{
const TensorInfo& info = handle->GetTensorInfo();
if (info.GetDataType() == DataType::Float32)
{
std::vector<BFloat16> newValues(info.GetNumElements());
armnnUtils::FloatingPointConverter::ConvertFloat32ToBFloat16(handle->GetConstTensor<float>(),
info.GetNumElements(),
newValues.data());
TensorInfo newInfo(info.GetShape(), DataType::BFloat16);
ConstTensor newInput(newInfo, newValues);
handle.reset(new ScopedTensorHandle(newInput));
}
}
};
struct Float32ToFloat16
{
static void Func(std::shared_ptr<ConstTensorHandle>& handle)
{
const TensorInfo& info = handle->GetTensorInfo();
if (info.GetDataType() == DataType::Float32)
{
std::vector<Half> newValues(info.GetNumElements());
armnnUtils::FloatingPointConverter::ConvertFloat32To16(handle->GetConstTensor<float>(),
info.GetNumElements(),
newValues.data());
TensorInfo newInfo(info.GetShape(), DataType::Float16);
ConstTensor newInput(newInfo, newValues);
handle.reset(new ScopedTensorHandle(newInput));
}
}
};
template<typename Converter, typename Predicate>
class ConvertConstants : public Optimization
{
public:
ConvertConstants() = default;
ConvertConstants(const ConvertConstants&) = default;
virtual ~ConvertConstants() = default;
void Run(Graph& graph, Layer& layer) const override
{
IgnoreUnused(graph);
if (Predicate::Test(layer))
{
layer.OperateOnConstantTensors(Converter::Func);
}
}
protected:
};
struct IsFloat32Layer
{
static bool Test(const Layer& layer)
{
return layer.GetDataType() == DataType::Float32;
}
};
struct IsFloat16Layer
{
static bool Test(const Layer& layer)
{
return layer.GetDataType() == DataType::Float16;
}
};
struct IsBFloat16Layer
{
static bool Test(const Layer& layer)
{
return layer.GetDataType() == DataType::BFloat16;
}
};
using ConvertConstantsBFloatToFloat = ConvertConstants<BFloat16ToFloat32, IsFloat32Layer>;
using ConvertConstantsFloatToBFloat = ConvertConstants<Float32ToBFloat16, IsBFloat16Layer>;
using ConvertConstantsHalfToFloat = ConvertConstants<Float16ToFloat32, IsFloat32Layer>;
using ConvertConstantsFloatToHalf = ConvertConstants<Float32ToFloat16, IsFloat16Layer>;
} //namespace optimizations
} //namespace armnn
|
