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 © 2020 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include <TestUtils.hpp>
#include <Optimizer.hpp>
#include <doctest/doctest.h>
TEST_SUITE("Optimizer")
{
using namespace armnn::optimizations;
TEST_CASE("Fp32NetworkToBf16OptimizationNoConversionTest")
{
armnn::Graph graph;
const armnn::TensorInfo infoFP32({ 2, 2, 1, 3 }, armnn::DataType::Float32);
// Create the simple test network without Conv2D/FullyConnected.
auto input = graph.AddLayer<armnn::InputLayer>(0, "input");
input->GetOutputSlot().SetTensorInfo(infoFP32);
auto floor = graph.AddLayer<armnn::FloorLayer>("floor");
floor->GetOutputSlot().SetTensorInfo(infoFP32);
auto output = graph.AddLayer<armnn::OutputLayer>(1, "output");
// Connect up the layers
input->GetOutputSlot().Connect(floor->GetInputSlot(0));
floor->GetOutputSlot().Connect(output->GetInputSlot(0));
CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>,
&IsLayerOfType<armnn::FloorLayer>, &IsLayerOfType<armnn::OutputLayer>));
// Run the optimizer
armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(Fp32NetworkToBf16Converter()));
CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>,
&IsLayerOfType<armnn::FloorLayer>,
&IsLayerOfType<armnn::OutputLayer>));
}
TEST_CASE("Fp32NetworkToBf16OptimizationConv2DTest")
{
armnn::Graph graph;
const armnn::TensorInfo infoFP32({ 2, 3, 8, 1 }, armnn::DataType::Float32);
// Create const tensor fp32 data
unsigned int dims[] = { 4, 2, 1, 1 };
std::vector<float> floatWeights{ 0.0f, -1.0f,
3.8f, // 0x40733333 Round down
3.1055E+29f, // 0x707ADC3C Round up
9.149516E-10f, // 0x307B7FFF Round down
-3.8f, // 0xC0733333 Round down
-3.1055E+29f, // 0xF07ADC3C Round up
-9.149516E-10f // 0xB07B7FFF Round down
};
armnn::ConstTensor weights(armnn::TensorInfo(4, dims, armnn::DataType::Float32, 0.0f, 0, true), floatWeights);
// Create const bias fp32 data
unsigned int biasDims[] {4};
std::vector<float> floatBias{ 1.0f, 2.0f, 3.0f, 4.0f };
armnn::ConstTensor bias(armnn::TensorInfo(1, biasDims, armnn::DataType::Float32, 0.0f, 0, true), floatBias);
// A network with Convolution2d layer
auto input = graph.AddLayer<armnn::InputLayer>(0, "input");
input->GetOutputSlot().SetTensorInfo(infoFP32);
armnn::Convolution2dDescriptor descriptor;
descriptor.m_BiasEnabled = true;
auto conv = graph.AddLayer<armnn::Convolution2dLayer>(descriptor, "conv2d");
conv->GetOutputSlot().SetTensorInfo(infoFP32);
auto weightsLayer = graph.AddLayer<armnn::ConstantLayer>("Weights");
weightsLayer->m_LayerOutput = std::make_shared<armnn::ScopedTensorHandle>(weights);
weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsLayer->m_LayerOutput->GetTensorInfo());
auto biasLayer = graph.AddLayer<armnn::ConstantLayer>("Bias");
biasLayer->m_LayerOutput = std::make_shared<armnn::ScopedTensorHandle>(bias);
biasLayer->GetOutputSlot(0).SetTensorInfo(biasLayer->m_LayerOutput->GetTensorInfo());
auto output = graph.AddLayer<armnn::OutputLayer>(1, "output");
// Connect up the layers
input->GetOutputSlot().Connect(conv->GetInputSlot(0));
weightsLayer->GetOutputSlot(0).Connect(conv->GetInputSlot(1));
biasLayer->GetOutputSlot(0).Connect(conv->GetInputSlot(2));
conv->GetOutputSlot().Connect(output->GetInputSlot(0));
CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>,
&IsLayerOfType<armnn::Convolution2dLayer>, &IsLayerOfType<armnn::ConstantLayer>,
&IsLayerOfType<armnn::ConstantLayer>, &IsLayerOfType<armnn::OutputLayer>));
// Run the optimizer
armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(RedirectMembersToConstantInputs(),
Fp32NetworkToBf16Converter()));
CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>,
&IsLayerOfType<armnn::ConvertFp32ToBf16Layer>,
&IsLayerOfType<armnn::ConstantLayer>, &IsLayerOfType<armnn::ConvertFp32ToBf16Layer>,
&IsLayerOfType<armnn::ConstantLayer>, &IsLayerOfType<armnn::Convolution2dLayer>,
&IsLayerOfType<armnn::OutputLayer>));
armnn::TensorInfo inputTensor = conv->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo();
armnn::TensorInfo weightTensor = conv->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo();
armnn::TensorInfo biasTensor = conv->GetInputSlot(2).GetConnectedOutputSlot()->GetTensorInfo();
armnn::TensorInfo outputTensor = conv->GetOutputSlot(0).GetTensorInfo();
CHECK((conv->GetDataType() == armnn::DataType::BFloat16));
CHECK((conv->m_Weight->GetTensorInfo().GetDataType() == armnn::DataType::BFloat16));
CHECK((conv->m_Bias->GetTensorInfo().GetDataType() == armnn::DataType::Float32));
CHECK((inputTensor.GetDataType() == armnn::DataType::BFloat16));
CHECK((weightTensor.GetDataType() == armnn::DataType::BFloat16));
CHECK((biasTensor.GetDataType() == armnn::DataType::Float32));
CHECK((outputTensor.GetDataType() == armnn::DataType::Float32));
// Check whether data matches expected Bf16 data
const armnn::BFloat16* data = conv->m_Weight->GetConstTensor<armnn::BFloat16>();
CHECK(data[0] == armnn::BFloat16(0.0f));
CHECK(data[1] == armnn::BFloat16(-1.0f));
CHECK(data[2] == armnn::BFloat16(3.796875f)); // 0x4073
CHECK(data[3] == armnn::BFloat16(3.1072295E29f)); // 0x707B
CHECK(data[4] == armnn::BFloat16(9.131327E-10f)); // 0x307B
CHECK(data[5] == armnn::BFloat16(-3.796875f)); // 0xC073
CHECK(data[6] == armnn::BFloat16(-3.1072295E29f)); // 0xF07B
CHECK(data[7] == armnn::BFloat16(-9.131327E-10f)); // 0xB07B
}
TEST_CASE("Fp32NetworkToBf16OptimizationFullyConnectedTest")
{
armnn::Graph graph;
const armnn::TensorInfo infoFP32({ 2, 3, 8, 1 }, armnn::DataType::Float32);
// Create const tensor fp32 data
unsigned int dims[] = { 4, 2, 1, 1 };
std::vector<float> floatWeights{ 0.0f, -1.0f,
3.8f, // 0x40733333 Round down
3.1055E+29f, // 0x707ADC3C Round up
9.149516E-10f, // 0x307B7FFF Round down
-3.8f, // 0xC0733333 Round down
-3.1055E+29f, // 0xF07ADC3C Round up
-9.149516E-10f // 0xB07B7FFF Round down
};
armnn::ConstTensor weights(armnn::TensorInfo(4, dims, armnn::DataType::Float32, 0.0f, 0, true), floatWeights);
// Create const bias fp32 data
unsigned int biasDims[] {4};
std::vector<float> floatBias{ 1.0f, 2.0f, 3.0f, 4.0f };
armnn::ConstTensor bias(armnn::TensorInfo(1, biasDims, armnn::DataType::Float32, 0.0f, 0, true), floatBias);
// A network with FullyConnected layer
auto input = graph.AddLayer<armnn::InputLayer>(0, "input");
input->GetOutputSlot().SetTensorInfo(infoFP32);
armnn::FullyConnectedDescriptor descriptor;
descriptor.m_BiasEnabled = true;
auto fc = graph.AddLayer<armnn::FullyConnectedLayer>(descriptor, "fully");
fc->GetOutputSlot().SetTensorInfo(infoFP32);
auto weightsLayer = graph.AddLayer<armnn::ConstantLayer>("Weights");
weightsLayer->m_LayerOutput = std::make_shared<armnn::ScopedTensorHandle>(weights);
weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsLayer->m_LayerOutput->GetTensorInfo());
auto biasLayer = graph.AddLayer<armnn::ConstantLayer>("Bias");
biasLayer->m_LayerOutput = std::make_shared<armnn::ScopedTensorHandle>(bias);
biasLayer->GetOutputSlot(0).SetTensorInfo(biasLayer->m_LayerOutput->GetTensorInfo());
auto output = graph.AddLayer<armnn::OutputLayer>(1, "output");
// Connect up the layers
input->GetOutputSlot().Connect(fc->GetInputSlot(0));
weightsLayer->GetOutputSlot(0).Connect(fc->GetInputSlot(1));
biasLayer->GetOutputSlot(0).Connect(fc->GetInputSlot(2));
fc->GetOutputSlot().Connect(output->GetInputSlot(0));
CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>,
&IsLayerOfType<armnn::FullyConnectedLayer>, &IsLayerOfType<armnn::ConstantLayer>,
&IsLayerOfType<armnn::ConstantLayer>, &IsLayerOfType<armnn::OutputLayer>));
// Run the optimizer
armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(RedirectMembersToConstantInputs(),
Fp32NetworkToBf16Converter()));
CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>,
&IsLayerOfType<armnn::ConvertFp32ToBf16Layer>, &IsLayerOfType<armnn::ConstantLayer>,
&IsLayerOfType<armnn::ConvertFp32ToBf16Layer>, &IsLayerOfType<armnn::ConstantLayer>,
&IsLayerOfType<armnn::FullyConnectedLayer>, &IsLayerOfType<armnn::OutputLayer>));
armnn::TensorInfo inputTensor = fc->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo();
armnn::TensorInfo weightTensor = fc->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo();
armnn::TensorInfo biasTensor = fc->GetInputSlot(2).GetConnectedOutputSlot()->GetTensorInfo();
armnn::TensorInfo outputTensor = fc->GetOutputSlot(0).GetTensorInfo();
CHECK((fc->GetDataType() == armnn::DataType::BFloat16));
CHECK((fc->m_Weight->GetTensorInfo().GetDataType() == armnn::DataType::BFloat16));
CHECK((fc->m_Bias->GetTensorInfo().GetDataType() == armnn::DataType::Float32));
CHECK((inputTensor.GetDataType() == armnn::DataType::BFloat16));
CHECK((weightTensor.GetDataType() == armnn::DataType::BFloat16));
CHECK((biasTensor.GetDataType() == armnn::DataType::Float32));
CHECK((outputTensor.GetDataType() == armnn::DataType::Float32));
// Check whether data matches expected Bf16 data
const armnn::BFloat16* data = fc->m_Weight->GetConstTensor<armnn::BFloat16>();
CHECK(data[0] == armnn::BFloat16(0.0f));
CHECK(data[1] == armnn::BFloat16(-1.0f));
CHECK(data[2] == armnn::BFloat16(3.796875f)); // 0x4073
CHECK(data[3] == armnn::BFloat16(3.1072295E29f)); // 0x707B
CHECK(data[4] == armnn::BFloat16(9.131327E-10f)); // 0x307B
CHECK(data[5] == armnn::BFloat16(-3.796875f)); // 0xC073
CHECK(data[6] == armnn::BFloat16(-3.1072295E29f)); // 0xF07B
CHECK(data[7] == armnn::BFloat16(-9.131327E-10f)); // 0xB07B
}
}
|