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
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include "Optimization.hpp"
#include <armnnUtils/QuantizeHelper.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>
#include <armnnUtils/DataLayoutIndexed.hpp>
namespace armnn
{
namespace optimizations
{
namespace pad_fold
{
inline float GetZeroElement(const TensorInfo& tensorInfo)
{
return static_cast<float>(tensorInfo.IsQuantized() ? tensorInfo.GetQuantizationOffset() : 0);
}
inline float GetLowestElement(const TensorInfo& tensorInfo)
{
constexpr float negativeInfinity = -std::numeric_limits<float>::infinity();
const float scale = tensorInfo.GetQuantizationScale();
const int32_t offset = tensorInfo.GetQuantizationOffset();
switch (tensorInfo.GetDataType())
{
case DataType::Float16:
return armnnUtils::SelectiveQuantize<armnn::Half>(negativeInfinity, scale, offset);
case DataType::Float32:
return armnnUtils::SelectiveQuantize<float>(negativeInfinity, scale, offset);
case DataType::QAsymmU8:
return armnnUtils::SelectiveQuantize<uint8_t>(negativeInfinity, scale, offset);
case DataType::QSymmS16:
return armnnUtils::SelectiveQuantize<int16_t>(negativeInfinity, scale, offset);
case DataType::QSymmS8:
// Fall-through
case DataType::QAsymmS8:
return armnnUtils::SelectiveQuantize<int8_t>(negativeInfinity, scale, offset);
case DataType::BFloat16:
return armnnUtils::SelectiveQuantize<armnn::BFloat16>(negativeInfinity, scale, offset);
default:
{
ARMNN_ASSERT_MSG(false, "Unsupported DataType");
return NAN;
}
}
}
inline bool IsNeutralElement(const Convolution2dDescriptor&, const TensorInfo& tensorInfo, const float tensorValue)
{
return tensorValue == GetZeroElement(tensorInfo);
}
inline bool IsNeutralElement(const DepthwiseConvolution2dDescriptor&,
const TensorInfo& tensorInfo,
const float tensorValue)
{
return tensorValue == GetZeroElement(tensorInfo);
}
inline bool IsNeutralElement(
const Pooling2dDescriptor& descriptor, const TensorInfo& tensorInfo, const float tensorValue)
{
return (descriptor.m_PoolType == PoolingAlgorithm::Max)
? tensorValue <= GetLowestElement(tensorInfo)
: tensorValue == GetZeroElement(tensorInfo);
}
template <typename Descriptor>
bool TryFoldPadIntoLayer2d(
const PadDescriptor& padDescriptor, Descriptor& layerDescriptor, const TensorInfo& tensorInfo)
{
armnnUtils::DataLayoutIndexed layout = armnnUtils::DataLayoutIndexed(layerDescriptor.m_DataLayout);
constexpr unsigned int batchIndex = 0;
constexpr auto noPad = std::make_pair(0U, 0U);
if ((!IsNeutralElement(layerDescriptor, tensorInfo, padDescriptor.m_PadValue)) ||
(padDescriptor.m_PadList[batchIndex] != noPad) || (padDescriptor.m_PadList[layout.GetChannelsIndex()] != noPad))
{
return false;
}
const auto& padList = padDescriptor.m_PadList;
// In Convolution2dDescriptor/Pooling2dDescriptor, padLeft and padRight are defined as paddings
// on width dimension whereas padTop and padBottom - paddings on height dimension, so updating
// these according to data layout
layerDescriptor.m_PadLeft += padList[layout.GetWidthIndex()].first;
layerDescriptor.m_PadRight += padList[layout.GetWidthIndex()].second;
layerDescriptor.m_PadTop += padList[layout.GetHeightIndex()].first;
layerDescriptor.m_PadBottom += padList[layout.GetHeightIndex()].second;
return true;
}
inline bool TryFoldPadIntoLayer2d(
const PadDescriptor& padDescriptor, Pooling2dDescriptor& poolDescriptor, const TensorInfo& tensorInfo)
{
const auto poolingPadValues = std::make_tuple(poolDescriptor.m_PadLeft, poolDescriptor.m_PadRight,
poolDescriptor.m_PadTop, poolDescriptor.m_PadBottom);
bool poolHasPadding = false;
if (poolingPadValues != std::make_tuple(0U, 0U, 0U, 0U))
{
poolHasPadding = true;
}
// We cannot fold Average or L2 pooling if there's is already padding and that padding method is Exclude.
if (poolDescriptor.m_PoolType != PoolingAlgorithm::Max) // PoolingAlgorithm::Average or PoolingAlgorithm::L2
{
if ((poolHasPadding) && (poolDescriptor.m_PaddingMethod == PaddingMethod::Exclude))
{
return false;
}
}
poolDescriptor.m_PaddingMethod = PaddingMethod::IgnoreValue;
return TryFoldPadIntoLayer2d<Pooling2dDescriptor>(padDescriptor, poolDescriptor, tensorInfo);
}
template <typename Layer2dT>
Layer2dT* FoldPadIntoLayer2dImpl(Graph& graph, InputSlot& connection)
{
PadLayer& padLayer = *PolymorphicDowncast<PadLayer*>(&connection.GetConnectedOutputSlot()->GetOwningLayer());
Layer2dT& layer2d = *PolymorphicDowncast<Layer2dT*>(&connection.GetOwningLayer());
const PadDescriptor& padDescriptor = padLayer.GetParameters();
auto newLayer2dDescriptor = layer2d.GetParameters();
if (!TryFoldPadIntoLayer2d(padDescriptor, newLayer2dDescriptor, padLayer.GetOutputSlot().GetTensorInfo()))
{
return nullptr;
}
// Save original parent output slot of the pad layer
OutputSlot& parentSlot = *padLayer.GetInputSlot(0).GetConnectedOutputSlot();
// Insert new layer2d layer between the pad layer an its parent layer.
const std::string name = std::string("folded-") + padLayer.GetName() + "-into-" + layer2d.GetName();
auto& newLayer2d = *graph.InsertNewLayer<Layer2dT>(padLayer.GetInputSlot(0), newLayer2dDescriptor, name.c_str());
// Reconnect the pad layer with its original parent.
newLayer2d.GetOutputSlot().MoveAllConnections(parentSlot);
// Moves connections in old layer2d layer output to new layer.
// Old layer2d layer will be removed as it's left unconnected.
// Pad layer will be removed if left unconnected.
layer2d.GetOutputSlot().MoveAllConnections(newLayer2d.GetOutputSlot());
return &newLayer2d;
}
class FoldPadIntoConvolution2dImpl
{
public:
void Run(Graph& graph, InputSlot& connection) const
{
const auto newConv2dLayer = FoldPadIntoLayer2dImpl<Convolution2dLayer>(graph, connection);
if (newConv2dLayer != nullptr)
{
const auto conv2dLayer = PolymorphicDowncast<Convolution2dLayer*>(&connection.GetOwningLayer());
// Copy weights and bias to the new convolution layer
ARMNN_ASSERT_MSG(conv2dLayer->m_Weight != nullptr,
"FoldPadIntoConvolution2d: Weights data should not be null.");
newConv2dLayer->m_Weight = std::move(conv2dLayer->m_Weight);
if (conv2dLayer->GetParameters().m_BiasEnabled)
{
ARMNN_ASSERT_MSG(conv2dLayer->m_Bias != nullptr,
"FoldPadIntoConvolution2d: Bias data should not be null if bias is enabled.");
newConv2dLayer->m_Bias = std::move(conv2dLayer->m_Bias);
}
}
}
protected:
FoldPadIntoConvolution2dImpl() = default;
~FoldPadIntoConvolution2dImpl() = default;
};
class FoldPadIntoDepthwiseConvolution2dImpl
{
public:
void Run(Graph& graph, InputSlot& connection) const
{
const auto newConv2dLayer = FoldPadIntoLayer2dImpl<DepthwiseConvolution2dLayer>(graph, connection);
if (newConv2dLayer != nullptr)
{
const auto conv2dLayer = PolymorphicDowncast<DepthwiseConvolution2dLayer*>(&connection.GetOwningLayer());
// Copy weights and bias to the new convolution layer
ARMNN_ASSERT_MSG(conv2dLayer->m_Weight != nullptr,
"FoldPadIntoDepthwiseConvolution2d: Weights data should not be null.");
newConv2dLayer->m_Weight = std::move(conv2dLayer->m_Weight);
if (conv2dLayer->GetParameters().m_BiasEnabled)
{
ARMNN_ASSERT_MSG(conv2dLayer->m_Bias != nullptr,
"FoldPadIntoDepthwiseConvolution2d: Bias data should not be null if bias is enabled.");
newConv2dLayer->m_Bias = std::move(conv2dLayer->m_Bias);
}
}
}
protected:
FoldPadIntoDepthwiseConvolution2dImpl() = default;
~FoldPadIntoDepthwiseConvolution2dImpl() = default;
};
class FoldPadIntoPooling2dImpl
{
public:
void Run(Graph& graph, InputSlot& connection) const
{
FoldPadIntoLayer2dImpl<Pooling2dLayer>(graph, connection);
}
protected:
FoldPadIntoPooling2dImpl() = default;
~FoldPadIntoPooling2dImpl() = default;
};
} // namespace pad_fold
using FoldPadIntoConvolution2d =
OptimizeForExclusiveConnection<PadLayer, Convolution2dLayer, pad_fold::FoldPadIntoConvolution2dImpl>;
using FoldPadIntoDepthwiseConvolution2d =
OptimizeForExclusiveConnection <PadLayer,
DepthwiseConvolution2dLayer,
pad_fold::FoldPadIntoDepthwiseConvolution2dImpl>;
using FoldPadIntoPooling2d =
OptimizeForExclusiveConnection<PadLayer, Pooling2dLayer, pad_fold::FoldPadIntoPooling2dImpl>;
} // namespace optimizations
} // namespace armnn
|
