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
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "RefNormalizationWorkload.hpp"
#include <armnn/Logging.hpp>
#include <armnn/Tensor.hpp>
#include <armnnUtils/DataLayoutIndexed.hpp>
#include <armnn/utility/NumericCast.hpp>
#include <Profiling.hpp>
#include "RefWorkloadUtils.hpp"
#include "Decoders.hpp"
#include "Encoders.hpp"
using namespace armnn;
using namespace armnnUtils;
namespace
{
// Helper function to compute "Within" normalization using Krichevsky 2012: Local Brightness Normalization.
void NormalizeWithinUingLbr(Decoder<float>& inputData,
Encoder<float>& outputData,
const TensorShape& tensorShape,
uint32_t norm_size,
float alpha,
float beta,
float kappa)
{
const unsigned int batchSize = tensorShape[0];
const unsigned int depth = tensorShape[1];
const unsigned int rows = tensorShape[2];
const unsigned int cols = tensorShape[3];
int radius = armnn::numeric_cast<int>(norm_size / 2u); /* Strong Assumption on rounding Mode */
for (unsigned int n = 0; n < batchSize; n++)
{
for (unsigned int c = 0; c < depth; c++)
{
for (unsigned int h = 0; h < rows; h++)
{
for (unsigned int w = 0; w < cols; w++)
{
float accumulated_scale = 0.0;
for (int y = -radius; y <= radius; y++)
{
for (int x = -radius; x <= radius; x++)
{
int i = armnn::numeric_cast<int>(w) + x;
int j = armnn::numeric_cast<int>(h) + y;
if ((i < 0) || (i >= armnn::numeric_cast<int>(cols)))
{
continue;
}
if ((j < 0) || (j >= armnn::numeric_cast<int>(rows)))
{
continue;
}
unsigned int inputIndex = n * cols * rows * depth +
c * cols * rows +
armnn::numeric_cast<unsigned int>(j) * cols +
armnn::numeric_cast<unsigned int>(i);
inputData[inputIndex];
float inval = inputData.Get();
accumulated_scale += inval*inval;
}
}
unsigned int index = n * cols * rows * depth +
c * cols * rows +
h * cols +
w;
inputData[index];
outputData[index];
outputData.Set(inputData.Get() / (powf((kappa + (accumulated_scale * alpha)), beta)));
}
}
}
}
}
// Helper function to compute "Across" normalization using Krichevsky 2012: Local Brightness Normalization.
void NormalizeAcrossUingLbr(Decoder<float>& inputData,
Encoder<float>& outputData,
const TensorShape& tensorShape,
uint32_t norm_size,
float alpha,
float beta,
float kappa,
DataLayout dataLayout)
{
DataLayoutIndexed dataLayoutIndexed(dataLayout);
const unsigned int batchSize = tensorShape[0];
const unsigned int depth = tensorShape[dataLayoutIndexed.GetChannelsIndex()];
const unsigned int rows = tensorShape[dataLayoutIndexed.GetHeightIndex()];
const unsigned int cols = tensorShape[dataLayoutIndexed.GetWidthIndex()];
int radius = armnn::numeric_cast<int>(norm_size / 2u); /* Strong Assumption on rounding Mode */
for (unsigned int n = 0; n < batchSize; n++)
{
for (unsigned int c = 0; c < depth; c++)
{
for (unsigned int h = 0; h < rows; h++)
{
for (unsigned int w = 0; w < cols; w++)
{
float accumulated_scale = 0.0;
for (int z = -radius; z <= radius; z++)
{
int k = armnn::numeric_cast<int>(c) + z;
if ((k < 0) || (k >= armnn::numeric_cast<int>(depth)))
{
continue;
}
unsigned inputIndex = dataLayoutIndexed.GetIndex(tensorShape,
n,
armnn::numeric_cast<unsigned int>(k),
h,
w);
inputData[inputIndex];
float inval = inputData.Get();
accumulated_scale += inval * inval;
}
float scale = kappa + (accumulated_scale * alpha);
scale = powf(scale, -beta);
unsigned index = dataLayoutIndexed.GetIndex(tensorShape, n, c, h, w);
inputData[index];
outputData[index];
outputData.Set(scale * inputData.Get());
}
}
}
}
}
} // Anonymous namespace
namespace armnn
{
RefNormalizationWorkload::RefNormalizationWorkload(const NormalizationQueueDescriptor& descriptor,
const WorkloadInfo& info)
: RefBaseWorkload(descriptor, info)
{}
void RefNormalizationWorkload::Execute() const
{
Execute(m_Data.m_Inputs, m_Data.m_Outputs);
}
void RefNormalizationWorkload::ExecuteAsync(WorkingMemDescriptor &workingMemDescriptor)
{
Execute(workingMemDescriptor.m_Inputs, workingMemDescriptor.m_Outputs);
}
void RefNormalizationWorkload::Execute(std::vector<ITensorHandle*> inputs, std::vector<ITensorHandle*> outputs) const
{
ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefNormalizationWorkload_Execute");
const TensorInfo& inputInfo = GetTensorInfo(inputs[0]);
auto inputDecoder = MakeDecoder<float>(inputInfo, inputs[0]->Map());
auto outputEncoder = MakeEncoder<float>(inputInfo, outputs[0]->Map());
if (NormalizationAlgorithmMethod::LocalBrightness == m_Data.m_Parameters.m_NormMethodType)
{
if (NormalizationAlgorithmChannel::Within == m_Data.m_Parameters.m_NormChannelType)
{
NormalizeWithinUingLbr(*inputDecoder,
*outputEncoder,
inputInfo.GetShape(),
m_Data.m_Parameters.m_NormSize,
m_Data.m_Parameters.m_Alpha,
m_Data.m_Parameters.m_Beta,
m_Data.m_Parameters.m_K);
}
else if (NormalizationAlgorithmChannel::Across == m_Data.m_Parameters.m_NormChannelType)
{
NormalizeAcrossUingLbr(*inputDecoder,
*outputEncoder,
inputInfo.GetShape(),
m_Data.m_Parameters.m_NormSize,
m_Data.m_Parameters.m_Alpha,
m_Data.m_Parameters.m_Beta,
m_Data.m_Parameters.m_K,
m_Data.m_Parameters.m_DataLayout);
}
else
{
ARMNN_LOG(warning) << "Illegal NORMALIZATION mode in normalization_f32";
return;
}
}
else
{
ARMNN_LOG(warning) << "Lcr method (Jarret 2009: Local Contrast Normalization) not supported yet.";
return;
}
}
} // namespace armnn
|
