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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "TransposeConvolution2d.hpp"
#include <DataLayoutIndexed.hpp>
namespace armnn
{
using namespace armnnUtils;
void TransposeConvolution2dImpl(const TransposeConvolution2dDescriptor& descriptor,
const TensorShape& inputShape,
Decoder<float>& inputDecoder,
const TensorShape& outputShape,
Encoder<float>& outputEncoder,
const TensorShape& weightsShape,
Decoder<float>& weightsDecoder,
Decoder<float>* biasesDecoder)
{
if (descriptor.m_BiasEnabled && !biasesDecoder)
{
throw InvalidArgumentException("Biases enabled but no bias data provided");
}
const DataLayoutIndexed dataLayoutIndexed(descriptor.m_DataLayout);
const unsigned int channelsIndex = dataLayoutIndexed.GetChannelsIndex();
const unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex();
const unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex();
unsigned int numBatches = inputShape[0];
unsigned int inputWidth = inputShape[widthIndex];
unsigned int inputHeight = inputShape[heightIndex];
unsigned int inputDepth = inputShape[channelsIndex];
unsigned int weightsHeight = weightsShape[heightIndex];
unsigned int weightsWidth = weightsShape[widthIndex];
unsigned int outputHeight = outputShape[heightIndex];
unsigned int outputWidth = outputShape[widthIndex];
unsigned int outputDepth = outputShape[channelsIndex];
unsigned int paddingLeft = descriptor.m_PadLeft;
unsigned int paddingTop = descriptor.m_PadTop;
unsigned int strideX = descriptor.m_StrideX;
unsigned int strideY = descriptor.m_StrideY;
// Set the initial output values to be logically 0 otherwise the algorithm doesn't work.
for (unsigned int i = 0u; i < outputShape.GetNumElements(); ++i)
{
outputEncoder.Set(0.f);
++outputEncoder;
}
for (unsigned int batch = 0u; batch < numBatches; ++batch)
{
for (unsigned int yInput = 0u; yInput < inputHeight; ++yInput)
{
for (unsigned int xInput = 0u; xInput < inputWidth; ++xInput)
{
unsigned int xOutputOrigin = xInput * strideX - paddingLeft;
unsigned int yOutputOrigin = yInput * strideY - paddingTop;
for (unsigned int dOutput = 0u; dOutput < outputDepth; ++dOutput)
{
for (unsigned int yWeights = 0u; yWeights < weightsHeight; ++yWeights)
{
for (unsigned int xWeights = 0u; xWeights < weightsWidth; ++xWeights)
{
unsigned int yOutput = yOutputOrigin + yWeights;
unsigned int xOutput = xOutputOrigin + xWeights;
if (yOutput < outputHeight && xOutput< outputWidth)
{
for (unsigned int dInput = 0u; dInput < inputDepth; dInput++)
{
const unsigned int inputIndex =
dataLayoutIndexed.GetIndex(inputShape, batch, dInput, yInput, xInput);
inputDecoder[inputIndex];
const unsigned int weightsIndex =
dataLayoutIndexed.GetIndex(weightsShape, dOutput, dInput, yWeights, xWeights);
weightsDecoder[weightsIndex];
const unsigned int outputIndex =
dataLayoutIndexed.GetIndex(outputShape, batch, dOutput, yOutput, xOutput);
outputEncoder[outputIndex];
float output = outputEncoder.Get();
output += inputDecoder.Get() * weightsDecoder.Get();
outputEncoder.Set(output);
}
}
}
}
}
}
}
}
// Apply bias (if enabled)
if (descriptor.m_BiasEnabled)
{
outputEncoder[0];
Decoder<float>& rBiasesDecoder = *biasesDecoder;
for (unsigned int batch = 0u; batch < numBatches; ++batch)
{
for (unsigned int dOutput = 0u; dOutput < outputDepth; ++dOutput)
{
rBiasesDecoder[dOutput];
for (unsigned int yOutput = 0u; yOutput < outputHeight; ++yOutput)
{
for (unsigned int xOutput = 0u; xOutput < outputWidth; ++xOutput)
{
const unsigned int outputIndex =
dataLayoutIndexed.GetIndex(outputShape, batch, dOutput, yOutput, xOutput);
outputEncoder[outputIndex];
outputEncoder.Set(outputEncoder.Get() + rBiasesDecoder.Get());
}
}
}
}
}
}
} // namespace armnn
|
