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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include <Graph.hpp>
#include <Network.hpp>
#include <reference/RefWorkloadFactory.hpp>
#include <boost/test/unit_test.hpp>
#include <test/GraphUtils.hpp>
BOOST_AUTO_TEST_SUITE(RefOptimizedNetwork)
BOOST_AUTO_TEST_CASE(OptimizeValidateCpuRefWorkloads)
{
const armnn::TensorInfo desc({3, 5}, armnn::DataType::Float32);
// build up the structure of the network
armnn::INetworkPtr net(armnn::INetwork::Create());
armnn::NormalizationDescriptor nmDesc;
armnn::ActivationDescriptor acDesc;
// in
// |
// nm
// / |
// ac |
// \ |
// ml
// |
// sm
// |
// ot
armnn::IConnectableLayer* layer = net->AddInputLayer(0, "in");
layer->GetOutputSlot(0).SetTensorInfo(desc);
armnn::IConnectableLayer* const normLayer = net->AddNormalizationLayer(nmDesc, "nm");
layer->GetOutputSlot(0).Connect(normLayer->GetInputSlot(0));
normLayer->GetOutputSlot(0).SetTensorInfo(desc);
layer = net->AddActivationLayer(acDesc, "ac");
normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
layer->GetOutputSlot(0).SetTensorInfo(desc);
armnn::IConnectableLayer* prevLayer = layer;
layer = net->AddMultiplicationLayer("ml");
prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1));
layer->GetOutputSlot(0).SetTensorInfo(desc);
prevLayer = layer;
armnn::SoftmaxDescriptor softmaxDescriptor;
layer = net->AddSoftmaxLayer(softmaxDescriptor, "sm");
prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
layer->GetOutputSlot(0).SetTensorInfo(desc);
prevLayer = layer;
layer = net->AddOutputLayer(0, "ot");
prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
armnn::Graph& graph = GetGraphForTesting(optNet.get());
graph.AllocateDynamicBuffers();
BOOST_CHECK(optNet);
// Validates workloads.
armnn::RefWorkloadFactory fact;
for (auto&& layer : graph)
{
BOOST_CHECK_NO_THROW(layer->CreateWorkload(fact));
}
}
BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsCpuRefPermuteLayer)
{
// Create runtime in which test will run
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
// build up the structure of the network
armnn::INetworkPtr net(armnn::INetwork::Create());
armnn::IConnectableLayer* input = net->AddInputLayer(0);
armnn::PermuteDescriptor descriptor({0, 2, 3, 1});
armnn::IConnectableLayer* permute = net->AddPermuteLayer(descriptor);
armnn::IConnectableLayer* output = net->AddOutputLayer(0);
input->GetOutputSlot(0).Connect(permute->GetInputSlot(0));
permute->GetOutputSlot(0).Connect(output->GetInputSlot(0));
input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
permute->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 4, 1, 4 }, armnn::DataType::Float32));
// optimize the network
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
armnn::Graph& graph = GetGraphForTesting(optNet.get());
graph.AllocateDynamicBuffers();
for (auto&& layer : graph)
{
BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
}
}
BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsCpuRefMeanLayer)
{
// Create runtime in which test will run
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
// build up the structure of the network
armnn::INetworkPtr net(armnn::INetwork::Create());
armnn::IConnectableLayer* input = net->AddInputLayer(0);
armnn::MeanDescriptor descriptor({ 0, 1 }, false);
armnn::IConnectableLayer* meanLayer = net->AddMeanLayer(descriptor);
armnn::IConnectableLayer* output = net->AddOutputLayer(0);
input->GetOutputSlot(0).Connect(meanLayer->GetInputSlot(0));
meanLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0));
input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 4, 3, 2 }, armnn::DataType::Float32));
meanLayer->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 2 }, armnn::DataType::Float32));
// optimize the network
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
armnn::Graph& graph = GetGraphForTesting(optNet.get());
graph.AllocateDynamicBuffers();
for (auto&& layer : graph)
{
BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
}
}
BOOST_AUTO_TEST_CASE(DebugTestOnCpuRef)
{
// build up the structure of the network
armnn::INetworkPtr net(armnn::INetwork::Create());
armnn::ActivationDescriptor activation1Descriptor;
activation1Descriptor.m_Function = armnn::ActivationFunction::BoundedReLu;
activation1Descriptor.m_A = 1.f;
activation1Descriptor.m_B = -1.f;
// Defines layers.
auto input = net->AddInputLayer(0, "InputLayer");
auto activation = net->AddActivationLayer(activation1Descriptor, "ActivationLayer");
auto output = net->AddOutputLayer(0, "OutputLayer");
// Connects layers.
input->GetOutputSlot(0).Connect(activation->GetInputSlot(0));
activation->GetOutputSlot(0).Connect(output->GetInputSlot(0));
armnn::TensorShape shape({4});
armnn::TensorInfo info(shape, armnn::DataType::Float32);
input->GetOutputSlot(0).SetTensorInfo(info);
activation->GetOutputSlot(0).SetTensorInfo(info);
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
armnn::OptimizerOptions optimizerOptions;
optimizerOptions.m_Debug = true;
armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec(),
optimizerOptions);
armnn::Graph& graph = GetGraphForTesting(optimizedNet.get());
graph.AllocateDynamicBuffers();
// Tests that all layers are present in the graph.
BOOST_TEST(graph.GetNumLayers() == 5);
// Tests that the vertices exist and have correct names.
BOOST_TEST(GraphHasNamedLayer(graph, "InputLayer"));
BOOST_TEST(GraphHasNamedLayer(graph, "DebugLayerAfterInputLayer_0"));
BOOST_TEST(GraphHasNamedLayer(graph, "ActivationLayer"));
BOOST_TEST(GraphHasNamedLayer(graph, "DebugLayerAfterActivationLayer_0"));
BOOST_TEST(GraphHasNamedLayer(graph, "OutputLayer"));
}
BOOST_AUTO_TEST_SUITE_END()
|
