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
|
//
// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include <Graph.hpp>
#include <Network.hpp>
#include <neon/NeonTensorHandle.hpp>
#include <neon/NeonTensorHandleFactory.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>
#include <test/GraphUtils.hpp>
#include <arm_compute/runtime/Allocator.h>
#include <boost/test/unit_test.hpp>
BOOST_AUTO_TEST_SUITE(NeonTensorHandleTests)
using namespace armnn;
BOOST_AUTO_TEST_CASE(NeonTensorHandleGetCapabilitiesNoPadding)
{
std::shared_ptr<NeonMemoryManager> memoryManager = std::make_shared<NeonMemoryManager>();
NeonTensorHandleFactory handleFactory(memoryManager);
INetworkPtr network(INetwork::Create());
// Add the layers
IConnectableLayer* input = network->AddInputLayer(0);
SoftmaxDescriptor descriptor;
descriptor.m_Beta = 1.0f;
IConnectableLayer* softmax = network->AddSoftmaxLayer(descriptor);
IConnectableLayer* output = network->AddOutputLayer(2);
// Establish connections
input->GetOutputSlot(0).Connect(softmax->GetInputSlot(0));
softmax->GetOutputSlot(0).Connect(output->GetInputSlot(0));
// No padding required for input
std::vector<Capability> capabilities = handleFactory.GetCapabilities(input,
softmax,
CapabilityClass::PaddingRequired);
BOOST_TEST(capabilities.empty());
// No padding required for Softmax
capabilities = handleFactory.GetCapabilities(softmax, output, CapabilityClass::PaddingRequired);
BOOST_TEST(capabilities.empty());
// No padding required for output
capabilities = handleFactory.GetCapabilities(output, nullptr, CapabilityClass::PaddingRequired);
BOOST_TEST(capabilities.empty());
}
BOOST_AUTO_TEST_CASE(NeonTensorHandleGetCapabilitiesPadding)
{
std::shared_ptr<NeonMemoryManager> memoryManager = std::make_shared<NeonMemoryManager>();
NeonTensorHandleFactory handleFactory(memoryManager);
INetworkPtr network(INetwork::Create());
// Add the layers
IConnectableLayer* input = network->AddInputLayer(0);
Pooling2dDescriptor descriptor;
IConnectableLayer* pooling = network->AddPooling2dLayer(descriptor);
IConnectableLayer* output = network->AddOutputLayer(2);
// Establish connections
input->GetOutputSlot(0).Connect(pooling->GetInputSlot(0));
pooling->GetOutputSlot(0).Connect(output->GetInputSlot(0));
// No padding required for input
std::vector<Capability> capabilities = handleFactory.GetCapabilities(input,
pooling,
CapabilityClass::PaddingRequired);
BOOST_TEST(capabilities.empty());
// No padding required for output
capabilities = handleFactory.GetCapabilities(output, nullptr, CapabilityClass::PaddingRequired);
BOOST_TEST(capabilities.empty());
// Padding required for Pooling2d
capabilities = handleFactory.GetCapabilities(pooling, output, CapabilityClass::PaddingRequired);
BOOST_TEST(capabilities.size() == 1);
BOOST_TEST((capabilities[0].m_CapabilityClass == CapabilityClass::PaddingRequired));
BOOST_TEST(capabilities[0].m_Value);
}
BOOST_AUTO_TEST_CASE(ConcatOnXorYSubTensorsNoPaddinRequiredTest)
{
armnn::INetworkPtr net(armnn::INetwork::Create());
// Set up tensor infos
const armnn::TensorInfo inputInfo = armnn::TensorInfo({2, 3, 2, 2}, armnn::DataType::Float32);
const armnn::TensorInfo intermediateInfo = armnn::TensorInfo({2, 3, 2, 2}, armnn::DataType::Float32);
const armnn::TensorInfo outputInfo = armnn::TensorInfo({2, 3, 4, 2}, armnn::DataType::Float32);
armnn::ElementwiseUnaryDescriptor descriptor(armnn::UnaryOperation::Abs);
// Create the network
armnn::IConnectableLayer* const input0Layer = net->AddInputLayer(0, "input_0");
input0Layer->GetOutputSlot(0).SetTensorInfo(inputInfo);
armnn::IConnectableLayer* elementwiseUnaryLayer0 = net->AddElementwiseUnaryLayer(descriptor, "elementwiseUnary_0");
elementwiseUnaryLayer0->GetOutputSlot(0).SetTensorInfo(intermediateInfo);
input0Layer->GetOutputSlot(0).Connect(elementwiseUnaryLayer0->GetInputSlot(0));
armnn::IConnectableLayer* const input1Layer = net->AddInputLayer(1, "input_1");
input1Layer->GetOutputSlot(0).SetTensorInfo(inputInfo);
armnn::IConnectableLayer* elementwiseUnaryLayer1 = net->AddElementwiseUnaryLayer(descriptor, "elementwiseUnary_1");
elementwiseUnaryLayer1->GetOutputSlot(0).SetTensorInfo(intermediateInfo);
input1Layer->GetOutputSlot(0).Connect(elementwiseUnaryLayer1->GetInputSlot(0));
std::array<armnn::TensorShape, 2> concatInputShapes = { intermediateInfo.GetShape(), intermediateInfo.GetShape() };
armnn::IConnectableLayer* const concatLayer = net->AddConcatLayer(armnn::CreateDescriptorForConcatenation(
concatInputShapes.begin(), concatInputShapes.end(), 2), "concatenation");
concatLayer->GetOutputSlot(0).SetTensorInfo(outputInfo);
elementwiseUnaryLayer0->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(0));
elementwiseUnaryLayer1->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(1));
armnn::IConnectableLayer* const outputLayer = net->AddOutputLayer(0, "output");
concatLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
const armnn::Graph& theGraph = static_cast<armnn::OptimizedNetwork*>(optimizedNet.get())->GetGraph();
// Load graph into runtime
armnn::NetworkId networkIdentifier;
runtime->LoadNetwork(networkIdentifier, std::move(optimizedNet));
// now check the concat how many sub-tensors it is using..
auto TraceSubTensorHandleAncestry = [](armnn::ITensorHandle* const subTensorHandle)
{
if (subTensorHandle && subTensorHandle->GetParent())
{
return true;
}
return false;
};
for (auto&& layer : theGraph)
{
if(layer->GetType() == armnn::LayerType::Concat)
{
unsigned int numberOfSubTensors = 0;
for (unsigned int i = 0; i < layer->GetNumInputSlots(); ++i)
{
const armnn::OutputSlot* slot = layer->GetInputSlot(i).GetConnectedOutputSlot();
if (TraceSubTensorHandleAncestry(slot->GetOutputHandler().GetData()))
{
++numberOfSubTensors;
}
}
// sub-tensors should be supported in this configuration
BOOST_CHECK(numberOfSubTensors > 0);
}
}
}
BOOST_AUTO_TEST_CASE(NeonTensorHandleFactoryMemoryManaged)
{
std::shared_ptr<NeonMemoryManager> memoryManager = std::make_shared<NeonMemoryManager>(
std::make_unique<arm_compute::Allocator>(),
BaseMemoryManager::MemoryAffinity::Offset);
NeonTensorHandleFactory handleFactory(memoryManager);
TensorInfo info({ 1, 1, 2, 1 }, DataType::Float32);
// create TensorHandle with memory managed
auto handle = handleFactory.CreateTensorHandle(info, true);
handle->Manage();
handle->Allocate();
memoryManager->Acquire();
{
float* buffer = reinterpret_cast<float*>(handle->Map());
BOOST_CHECK(buffer != nullptr); // Yields a valid pointer
buffer[0] = 1.5f;
buffer[1] = 2.5f;
BOOST_CHECK(buffer[0] == 1.5f); // Memory is writable and readable
BOOST_CHECK(buffer[1] == 2.5f); // Memory is writable and readable
}
memoryManager->Release();
memoryManager->Acquire();
{
float* buffer = reinterpret_cast<float*>(handle->Map());
BOOST_CHECK(buffer != nullptr); // Yields a valid pointer
buffer[0] = 3.5f;
buffer[1] = 4.5f;
BOOST_CHECK(buffer[0] == 3.5f); // Memory is writable and readable
BOOST_CHECK(buffer[1] == 4.5f); // Memory is writable and readable
}
memoryManager->Release();
float testPtr[2] = { 2.5f, 5.5f };
// Cannot import as import is disabled
BOOST_CHECK(!handle->Import(static_cast<void*>(testPtr), MemorySource::Malloc));
}
BOOST_AUTO_TEST_CASE(NeonTensorHandleFactoryImport)
{
std::shared_ptr<NeonMemoryManager> memoryManager = std::make_shared<NeonMemoryManager>(
std::make_unique<arm_compute::Allocator>(),
BaseMemoryManager::MemoryAffinity::Offset);
NeonTensorHandleFactory handleFactory(memoryManager);
TensorInfo info({ 1, 1, 2, 1 }, DataType::Float32);
// create TensorHandle without memory managed
auto handle = handleFactory.CreateTensorHandle(info, false);
handle->Manage();
handle->Allocate();
memoryManager->Acquire();
// No buffer allocated when import is enabled
BOOST_CHECK((PolymorphicDowncast<NeonTensorHandle*>(handle.get()))->GetTensor().buffer() == nullptr);
float testPtr[2] = { 2.5f, 5.5f };
// Correctly import
BOOST_CHECK(handle->Import(static_cast<void*>(testPtr), MemorySource::Malloc));
float* buffer = reinterpret_cast<float*>(handle->Map());
BOOST_CHECK(buffer != nullptr); // Yields a valid pointer after import
BOOST_CHECK(buffer == testPtr); // buffer is pointing to testPtr
// Memory is writable and readable with correct value
BOOST_CHECK(buffer[0] == 2.5f);
BOOST_CHECK(buffer[1] == 5.5f);
buffer[0] = 3.5f;
buffer[1] = 10.0f;
BOOST_CHECK(buffer[0] == 3.5f);
BOOST_CHECK(buffer[1] == 10.0f);
memoryManager->Release();
}
BOOST_AUTO_TEST_SUITE_END()
|