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
|
//
// Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "LayersFwd.hpp"
#include <Network.hpp>
#include <TestUtils.hpp>
#include <doctest/doctest.h>
#include <armnn/backends/TensorHandle.hpp>
#include <Optimizer.hpp>
TEST_SUITE("Optimizer")
{
using namespace armnn;
using namespace armnn::optimizations;
// Helpers for testing
auto checkConstantFloat32 = [](const armnn::Layer *const layer)
{
return IsLayerOfType<ConstantLayer>(layer) && (layer->GetDataType() == DataType::Float32);
};
auto checkConstantFloat16 = [](const armnn::Layer *const layer)
{
return IsLayerOfType<ConstantLayer>(layer) && (layer->GetDataType() == DataType::Float16);
};
auto checkConstantQAsymmS8 = [](const armnn::Layer *const layer)
{
return IsLayerOfType<ConstantLayer>(layer) && (layer->GetDataType() == DataType::QAsymmS8);
};
auto checkPadFoldedIntoConv2d = [](const Layer* const layer)
{
const auto conv2dLayer = static_cast<const Convolution2dLayer*>(layer);
const auto conv2dLayerParams = conv2dLayer->GetParameters();
return IsLayerOfType<Convolution2dLayer>(layer) &&
(layer->GetNameStr() == "folded-pad-into-conv2d") &&
(conv2dLayerParams.m_PadLeft == 2) &&
(conv2dLayerParams.m_PadRight == 2) &&
(conv2dLayerParams.m_PadTop == 2) &&
(conv2dLayerParams.m_PadBottom == 2) &&
(conv2dLayerParams.m_StrideX == 1) &&
(conv2dLayerParams.m_StrideY == 1) &&
(conv2dLayerParams.m_BiasEnabled == false) &&
(conv2dLayerParams.m_DataLayout == DataLayout::NHWC);
};
TEST_CASE("ConvertConstFloat16DequantizeToConstFloat32")
{
Graph graph;
const unsigned int shape[] = {1, 2, 2, 3};
const TensorInfo constTensorInfo(4, shape, DataType::Float16, 1.0, 0, true);
const TensorInfo outputDequantizeInfo(4, shape, DataType::Float32, 1.0, 0, true);
auto constantLayer = graph.AddLayer<ConstantLayer>("constant");
std::vector<float> constantValues(constTensorInfo.GetNumElements(), 4.5f);
ConstTensor constTensor(constTensorInfo, constantValues.data());
constantLayer->m_LayerOutput = std::make_shared<ScopedTensorHandle>(constTensor);
constantLayer->GetOutputSlot().SetTensorInfo(constTensorInfo);
auto dequantizeLayer = graph.AddLayer<DequantizeLayer>("dequantize");
dequantizeLayer->GetOutputSlot().SetTensorInfo(outputDequantizeInfo);
auto output = graph.AddLayer<OutputLayer>(0, "output");
// Connect up constant -> dequantize -> output
constantLayer->GetOutputSlot().Connect(dequantizeLayer->GetInputSlot(0));
dequantizeLayer->GetOutputSlot().Connect(output->GetInputSlot(0));
CHECK(CheckSequence(graph.cbegin(), graph.cend(),
checkConstantFloat16,
&IsLayerOfType<DequantizeLayer>,
&IsLayerOfType<OutputLayer>));
armnn::Optimizer::Pass(graph, MakeOptimizations(ConvertConstDequantisationLayersToConstLayers()));
CHECK(CheckSequence(graph.cbegin(), graph.cend(),
checkConstantFloat32,
&IsLayerOfType<OutputLayer>));
}
TEST_CASE("ConvertConstFloat16DequantizeToConstFloat32PlusFusePadWithConv2d")
{
Graph graph;
const unsigned int shape[] = {1, 2, 2, 3};
const TensorInfo constTensorInfo(4, shape, DataType::Float16, 1.0, 0, true);
const TensorInfo outputDequantizeInfo(4, shape, DataType::Float32, 1.0, 0, true);
auto constantLayer = graph.AddLayer<ConstantLayer>("constant");
std::vector<float> constantValues(constTensorInfo.GetNumElements(), 4.5f);
ConstTensor constTensor(constTensorInfo, constantValues.data());
constantLayer->m_LayerOutput = std::make_shared<ScopedTensorHandle>(constTensor);
constantLayer->GetOutputSlot().SetTensorInfo(constTensorInfo);
auto dequantizeLayer = graph.AddLayer<DequantizeLayer>("dequantize");
dequantizeLayer->GetOutputSlot().SetTensorInfo(outputDequantizeInfo);
auto output = graph.AddLayer<OutputLayer>(0, "output");
Convolution2dDescriptor convolution2dDescriptor;
convolution2dDescriptor.m_BiasEnabled = false;
convolution2dDescriptor.m_StrideX = 1;
convolution2dDescriptor.m_StrideY = 1;
convolution2dDescriptor.m_DataLayout = DataLayout::NHWC;
auto conv2d = graph.AddLayer<Convolution2dLayer>(convolution2dDescriptor, "conv2d");
auto inputLayer = graph.AddLayer<InputLayer>(0, "input");
PadDescriptor padDescriptor({{0, 0},
{2, 2},
{2, 2},
{0, 0}});
const unsigned int paddedShape[] = {1, 6, 6, 3};
TensorInfo paddedInfo(4, paddedShape, DataType::Float32);
auto padLayer = graph.AddLayer<PadLayer>(padDescriptor, "pad");
padLayer->GetOutputSlot().SetTensorInfo(paddedInfo);
// Connect up:
// input -> pad -> conv2d -> output
// constant -> dequantize ->
constantLayer->GetOutputSlot().Connect(dequantizeLayer->GetInputSlot(0));
dequantizeLayer->GetOutputSlot().Connect(conv2d->GetInputSlot(1));
inputLayer->GetOutputSlot().Connect(padLayer->GetInputSlot(0));
padLayer->GetOutputSlot().Connect(conv2d->GetInputSlot(0));
conv2d->GetOutputSlot().Connect(output->GetInputSlot(0));
CHECK(CheckSequence(graph.cbegin(), graph.cend(),
&IsLayerOfType<InputLayer>,
checkConstantFloat16,
&IsLayerOfType<DequantizeLayer>,
&IsLayerOfType<Convolution2dLayer>,
&IsLayerOfType<PadLayer>,
&IsLayerOfType<OutputLayer>));
armnn::Optimizer::Pass(graph, MakeOptimizations(ConvertConstDequantisationLayersToConstLayers()));
armnn::Optimizer::Pass(graph, MakeOptimizations(FoldPadIntoConvolution2d()));
// Ensure that the const and dequantize are now constant of type fp32
// Ensure pad and conv2d are now just convolution
CHECK(CheckSequence(graph.cbegin(), graph.cend(),
&IsLayerOfType<InputLayer>,
checkConstantFloat32,
checkPadFoldedIntoConv2d,
&IsLayerOfType<OutputLayer>));
}
TEST_CASE("ConvertConstInt8DequantizeToConstFloat32")
{
Graph graph;
const unsigned int shape[] = {1, 2, 2, 3};
const TensorInfo constTensorInfo(4, shape, DataType::QAsymmS8, 1.0, 0, true);
const TensorInfo outputDequantizeInfo(4, shape, DataType::Float32, 1.0, 0, true);
auto constantLayer = graph.AddLayer<ConstantLayer>("constant");
std::vector<int8_t> constantValues(constTensorInfo.GetNumElements(), 5);
ConstTensor constTensor(constTensorInfo, constantValues.data());
constantLayer->m_LayerOutput = std::make_shared<ScopedTensorHandle>(constTensor);
constantLayer->GetOutputSlot().SetTensorInfo(constTensorInfo);
auto dequantizeLayer = graph.AddLayer<DequantizeLayer>("dequantize");
dequantizeLayer->GetOutputSlot().SetTensorInfo(outputDequantizeInfo);
auto output = graph.AddLayer<OutputLayer>(0, "output");
// Connect up constant -> dequantize -> output
constantLayer->GetOutputSlot().Connect(dequantizeLayer->GetInputSlot(0));
dequantizeLayer->GetOutputSlot().Connect(output->GetInputSlot(0));
CHECK(CheckSequence(graph.cbegin(), graph.cend(),
checkConstantQAsymmS8,
&IsLayerOfType<DequantizeLayer>,
&IsLayerOfType<OutputLayer>));
armnn::Optimizer::Pass(graph, MakeOptimizations(ConvertConstDequantisationLayersToConstLayers()));
CHECK(CheckSequence(graph.cbegin(), graph.cend(),
checkConstantFloat32,
&IsLayerOfType<OutputLayer>));
}
}
|
