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
|
//
// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include <TestUtils.hpp>
#include <Half.hpp>
#include <Optimizer.hpp>
#include <doctest/doctest.h>
using namespace armnn;
TEST_SUITE("Optimizer")
{
using namespace armnn::optimizations;
TEST_CASE("ConvertConstantsFloatToHalfTest")
{
armnn::Graph graph;
const armnn::TensorInfo info({ 1, 1, 1, 2 }, armnn::DataType::Float16);
// Create const tensor from fp32 data
unsigned int dims[] = { 4, 1, 1, 1 };
std::vector<float> floatWeights{ 1.0f, 2.0f, 3.0f, 4.0f };
armnn::TensorInfo weightsInfo = armnn::TensorInfo(4, dims, armnn::DataType::Float32, 0.0f, 0, true);
armnn::ConstTensor weights(weightsInfo, floatWeights);
// Create simple test network
auto input = graph.AddLayer<armnn::InputLayer>(0, "input");
input->GetOutputSlot().SetTensorInfo(info);
auto fc = graph.AddLayer<armnn::FullyConnectedLayer>(armnn::FullyConnectedDescriptor(), "fc");
fc->GetOutputSlot().SetTensorInfo(info);
auto weightsLayer = graph.AddLayer<armnn::ConstantLayer>("weights");
weightsLayer->m_LayerOutput = std::make_unique<armnn::ScopedTensorHandle>(weights);
weightsLayer->GetOutputSlot().SetTensorInfo(weightsInfo);
auto output = graph.AddLayer<armnn::OutputLayer>(1, "output");
// Connect up the layers
input->GetOutputSlot().Connect(fc->GetInputSlot(0));
weightsLayer->GetOutputSlot().Connect(fc->GetInputSlot(1));
fc->GetOutputSlot().Connect(output->GetInputSlot(0));
// Check tensor data type before conversion
CHECK(weightsLayer->m_LayerOutput->GetTensorInfo().GetDataType() == armnn::DataType::Float32);
// Run the optimizer
armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(ConvertConstantsFloatToHalf()));
// Check tensor data type after conversion
CHECK(weightsLayer->m_LayerOutput->GetTensorInfo().GetDataType() == armnn::DataType::Float16);
// Check whether data matches expected fp16 data
const Half* data = weightsLayer->m_LayerOutput->GetConstTensor<Half>();
CHECK(data[0] == Half(1.0f));
CHECK(data[1] == Half(2.0f));
CHECK(data[2] == Half(3.0f));
CHECK(data[3] == Half(4.0f));
}
TEST_CASE("ConvertConstantsFloatToHalfTest_constant")
{
armnn::Graph graph;
// Create the simple test network with Weights and Biases as inputs to a FullyConnected layer.
auto input = graph.AddLayer<armnn::InputLayer>(0, "Input");
auto weights = graph.AddLayer<armnn::ConstantLayer>("Weights");
auto biases = graph.AddLayer<armnn::ConstantLayer>("Biases");
armnn::FullyConnectedDescriptor desc;
desc.m_BiasEnabled = true;
desc.m_ConstantWeights = true;
auto fcLayer = graph.AddLayer<armnn::FullyConnectedLayer>(desc, "FullyConnected");
auto output = graph.AddLayer<armnn::OutputLayer>(1, "Output");
float expectedWeightsData[] = { 1.0f, 2.0f, 3.0f, 4.0f };
float expectedBiasesData[] = { 0.0f, 2.0f };
const armnn::TensorInfo inputInfo ({ 1, 2, 2, 3 }, armnn::DataType::Float16);
const armnn::TensorInfo outputInfo ({ 1, 2, 2, 3 }, armnn::DataType::Float16);
const armnn::TensorInfo weightsInfo({ 4 }, armnn::DataType::Float32, 0.0f, 0, true);
const armnn::TensorInfo biasesInfo ({ 2 }, armnn::DataType::Float32, 0.0f, 0, true);
// Set the m_LayerOutput for the optimizer to point to.
armnn::ConstTensor weightsTensor(weightsInfo, &expectedWeightsData);
armnn::ConstTensor biasesTensor(biasesInfo, &expectedBiasesData);
weights->m_LayerOutput = std::make_unique<armnn::ScopedTensorHandle>(weightsTensor);
biases->m_LayerOutput = std::make_unique<armnn::ScopedTensorHandle>(biasesTensor);
input->GetOutputSlot().SetTensorInfo(inputInfo);
weights->GetOutputSlot().SetTensorInfo(weightsInfo);
biases->GetOutputSlot().SetTensorInfo(biasesInfo);
fcLayer->GetOutputSlot().SetTensorInfo(outputInfo);
// Connect up the layers
input->GetOutputSlot(0).Connect(fcLayer->GetInputSlot(0));
weights->GetOutputSlot(0).Connect(fcLayer->GetInputSlot(1));
biases->GetOutputSlot(0).Connect(fcLayer->GetInputSlot(2));
fcLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0));
// Check tensor data type before conversion
CHECK(5 == graph.GetNumLayers());
CHECK(weights->m_LayerOutput->GetTensorInfo().GetDataType() == armnn::DataType::Float32);
// Run the optimizer
armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(ConvertConstantsFloatToHalf()));
// Check tensor data type after conversion
CHECK(5 == graph.GetNumLayers());
CHECK(weights->m_LayerOutput->GetTensorInfo().GetDataType() == armnn::DataType::Float16);
// Check whether weights data matches expected fp16 data
const Half* data = weights->m_LayerOutput->GetConstTensor<Half>();
CHECK(data[0] == Half(1.0f));
CHECK(data[1] == Half(2.0f));
CHECK(data[2] == Half(3.0f));
CHECK(data[3] == Half(4.0f));
// Check whether bias data matches expected fp16 data
const Half* biasData = biases->m_LayerOutput->GetConstTensor<Half>();
CHECK(biasData[0] == Half(0.0f));
CHECK(biasData[1] == Half(2.0f));
}
TEST_CASE("ConvertConstantsFloatToHalfInfinityTest")
{
armnn::Graph graph;
const armnn::TensorInfo info({ 1, 1, 1, 2 }, armnn::DataType::Float16);
// Create const tensor from fp32 data
unsigned int dims[] = { 4, 1, 1, 1 };
std::vector<float> floatWeights{ std::numeric_limits<float>::infinity(),
-std::numeric_limits<float>::infinity(),
std::numeric_limits<float>::max(),
std::numeric_limits<float>::lowest() };
armnn::TensorInfo weightsInfo = armnn::TensorInfo(4, dims, armnn::DataType::Float32, 0.0f, 0, true);
armnn::ConstTensor weights(weightsInfo, floatWeights);
// Create simple test network
auto input = graph.AddLayer<armnn::InputLayer>(0, "input");
input->GetOutputSlot().SetTensorInfo(info);
auto fc = graph.AddLayer<armnn::FullyConnectedLayer>(armnn::FullyConnectedDescriptor(), "fc");
fc->GetOutputSlot().SetTensorInfo(info);
auto weightsLayer = graph.AddLayer<armnn::ConstantLayer>("weights");
weightsLayer->m_LayerOutput = std::make_unique<armnn::ScopedTensorHandle>(weights);
weightsLayer->GetOutputSlot().SetTensorInfo(weightsInfo);
auto output = graph.AddLayer<armnn::OutputLayer>(1, "output");
// Connect up the layers
input->GetOutputSlot().Connect(fc->GetInputSlot(0));
weightsLayer->GetOutputSlot().Connect(fc->GetInputSlot(1));
fc->GetOutputSlot().Connect(output->GetInputSlot(0));
// Check tensor data type before conversion
CHECK(weightsLayer->m_LayerOutput->GetTensorInfo().GetDataType() == armnn::DataType::Float32);
// Run the optimizer
armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(ConvertConstantsFloatToHalf()));
// Check tensor data type after conversion
CHECK(weightsLayer->m_LayerOutput->GetTensorInfo().GetDataType() == armnn::DataType::Float16);
// Check whether data matches expected fp16 data
const Half* data = weightsLayer->m_LayerOutput->GetConstTensor<Half>();
CHECK(data[0] == std::numeric_limits<armnn::Half>::max());
CHECK(data[1] == std::numeric_limits<armnn::Half>::lowest());
CHECK(data[2] == std::numeric_limits<armnn::Half>::max());
CHECK(data[3] == std::numeric_limits<armnn::Half>::lowest());
}
}
|
