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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "ConvImpl.hpp"
#include <boost/assert.hpp>
#include <cmath>
#include <limits>
namespace armnn
{
QuantizedMultiplierSmallerThanOne::QuantizedMultiplierSmallerThanOne(float multiplier)
{
BOOST_ASSERT(multiplier >= 0.0f && multiplier < 1.0f);
if (multiplier == 0.0f)
{
m_Multiplier = 0;
m_RightShift = 0;
}
else
{
const double q = std::frexp(multiplier, &m_RightShift);
m_RightShift = -m_RightShift;
int64_t qFixed = static_cast<int64_t>(std::round(q * (1ll << 31)));
BOOST_ASSERT(qFixed <= (1ll << 31));
if (qFixed == (1ll << 31))
{
qFixed /= 2;
--m_RightShift;
}
BOOST_ASSERT(m_RightShift >= 0);
BOOST_ASSERT(qFixed <= std::numeric_limits<int32_t>::max());
m_Multiplier = static_cast<int32_t>(qFixed);
}
}
int32_t QuantizedMultiplierSmallerThanOne::operator*(int32_t rhs) const
{
int32_t x = SaturatingRoundingDoublingHighMul(rhs, m_Multiplier);
return RoundingDivideByPOT(x, m_RightShift);
}
int32_t QuantizedMultiplierSmallerThanOne::SaturatingRoundingDoublingHighMul(int32_t a, int32_t b)
{
// Check for overflow.
if (a == b && a == std::numeric_limits<int32_t>::min())
{
return std::numeric_limits<int32_t>::max();
}
int64_t a_64(a);
int64_t b_64(b);
int64_t ab_64 = a_64 * b_64;
int32_t nudge = ab_64 >= 0 ? (1 << 30) : (1 - (1 << 30));
int32_t ab_x2_high32 = static_cast<std::int32_t>((ab_64 + nudge) / (1ll << 31));
return ab_x2_high32;
}
int32_t QuantizedMultiplierSmallerThanOne::RoundingDivideByPOT(int32_t x, int exponent)
{
BOOST_ASSERT(exponent >= 0 && exponent <= 31);
int32_t mask = (1 << exponent) - 1;
int32_t remainder = x & mask;
int32_t threshold = (mask >> 1) + (x < 0 ? 1 : 0);
return (x >> exponent) + (remainder > threshold ? 1 : 0);
}
void Convolve(const TensorShape& rInputShape,
Decoder<float>& rInputDecoder,
const TensorShape& rOutputShape,
Encoder<float>& rOutputEncoder,
const TensorShape& rFilterShape,
Decoder<float>& rFilterDecoder,
bool biasEnabled,
Decoder<float>* pBiasDecoder,
DataLayout dataLayout,
unsigned int paddingTop,
unsigned int paddingLeft,
unsigned int xStride,
unsigned int yStride,
unsigned int xDilation,
unsigned int yDilation,
bool depthwise)
{
if (biasEnabled && !pBiasDecoder)
{
throw InvalidArgumentException("Bias is enabled but the bias data is invalid");
}
const armnnUtils::DataLayoutIndexed dataLayoutIndexed(dataLayout);
const unsigned int channelsIndex = dataLayoutIndexed.GetChannelsIndex();
const unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex();
const unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex();
unsigned int depthMultiplier = depthwise ? rFilterShape[0] : 1;
unsigned int inputChannels = depthwise ? rFilterShape[1] : rFilterShape[channelsIndex];
unsigned int outputChannels = depthwise ? inputChannels * depthMultiplier : rFilterShape[0];
unsigned int batchSize = rOutputShape[0];
unsigned int outputHeight = rOutputShape[heightIndex];
unsigned int outputWidth = rOutputShape[widthIndex];
unsigned int inputHeight = rInputShape[heightIndex];
unsigned int inputWidth = rInputShape[widthIndex];
unsigned int filterHeight = depthwise ? rFilterShape[2] : rFilterShape[heightIndex];
unsigned int filterWidth = depthwise ? rFilterShape[3] : rFilterShape[widthIndex];
for (unsigned int batchIdx = 0; batchIdx < batchSize; batchIdx++)
{
for (unsigned int cOutput = 0; cOutput < outputChannels; cOutput++)
{
for (unsigned int yOutput = 0; yOutput < outputHeight; yOutput++)
{
for (unsigned int xOutput = 0; xOutput < outputWidth; xOutput++)
{
// This loop goes over each output element.
float sum = 0.0f;
// For depthwise, each output channel corresponds to exactly one input channel.
// For normal, must loop over each input channel.
for (unsigned int cInput = 0; cInput < (depthwise ? 1 : inputChannels); cInput++)
{
unsigned int depthwiseMultiplierIdx = 0;
if (depthwise)
{
cInput = cOutput / depthMultiplier;
depthwiseMultiplierIdx = cOutput % depthMultiplier;
}
for (unsigned int yFilter = 0; yFilter < filterHeight; yFilter++)
{
for (unsigned int xFilter = 0; xFilter < filterWidth; xFilter++)
{
// This loop goes over each input element for each output element.
unsigned int filterIndex = 0;
// Since dimensionality of kernel depends on depthwiseness, so does index.
if (depthwise)
{
filterIndex = depthwiseMultiplierIdx * filterWidth * filterHeight * inputChannels +
cInput * filterWidth * filterHeight +
yFilter * filterWidth +
xFilter;
}
else
{
filterIndex = dataLayoutIndexed.GetIndex(rFilterShape,
cOutput,
cInput,
yFilter,
xFilter);
}
rFilterDecoder[filterIndex];
float filterValue = rFilterDecoder.Get();
unsigned int yInput = yOutput * yStride + yFilter * yDilation;
unsigned int xInput = xOutput * xStride + xFilter * xDilation;
float inputValue;
// Check if we're in the padding.
if (yInput < paddingTop || yInput >= inputHeight + paddingTop ||
xInput < paddingLeft || xInput >= inputWidth + paddingLeft )
{
inputValue = 0.0f;
}
else
{
unsigned int inputIndex = dataLayoutIndexed.GetIndex(rInputShape,
batchIdx,
cInput,
yInput - paddingTop,
xInput - paddingLeft);
rInputDecoder[inputIndex];
inputValue = rInputDecoder.Get();
}
sum += filterValue * inputValue;
}
}
}
if (biasEnabled)
{
(*pBiasDecoder)[cOutput];
sum += pBiasDecoder->Get();
}
unsigned int outIdx = dataLayoutIndexed.GetIndex(rOutputShape, batchIdx, cOutput, yOutput, xOutput);
rOutputEncoder[outIdx];
rOutputEncoder.Set(sum);
}
}
}
}
}
} //namespace armnn
|
