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
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// See LICENSE file in the project root for full license information.
//
#include "backends/NeonWorkloadFactory.hpp"
#include "backends/NeonWorkloadUtils.hpp"
#include "backends/NeonWorkloads.hpp"
#include "backends/MemCopyWorkload.hpp"
#include "backends/NeonTensorHandle.hpp"
#include "test/CreateWorkloadClNeon.hpp"
BOOST_AUTO_TEST_SUITE(CreateWorkloadNeon)
namespace
{
bool TestNeonTensorHandleInfo(armnn::INeonTensorHandle* handle, const armnn::TensorInfo& expectedInfo)
{
using namespace armnn::armcomputetensorutils;
const arm_compute::ITensorInfo* handleInfo = handle->GetTensor().info();
const arm_compute::TensorInfo expectedAclInfo = BuildArmComputeTensorInfo(expectedInfo);
if (handleInfo->data_type() != expectedAclInfo.data_type())
{
return false;
}
if (handleInfo->num_dimensions() != expectedAclInfo.num_dimensions())
{
return false;
}
if (handleInfo->quantization_info() != expectedAclInfo.quantization_info())
{
return false;
}
for (std::size_t d = 0; d < expectedAclInfo.num_dimensions(); ++d)
{
if (handleInfo->dimension(d) != expectedAclInfo.dimension(d))
{
return false;
}
}
return true;
}
} // namespace
BOOST_AUTO_TEST_CASE(CreateActivationWorkload)
{
Graph graph;
NeonWorkloadFactory factory;
auto workload = CreateActivationWorkloadTest<NeonActivationFloat32Workload>(factory, graph);
// check that inputs/outputs are as we expect them (see definition of CreateActivationWorkloadTest)
ActivationQueueDescriptor queueDescriptor = workload->GetData();
auto inputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({1, 1}, DataType::Float32)));
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({1, 1}, DataType::Float32)));
}
BOOST_AUTO_TEST_CASE(CreateAdditionWorkload)
{
Graph graph;
NeonWorkloadFactory factory;
auto workload = CreateAdditionWorkloadTest<NeonAdditionFloat32Workload>(factory, graph);
// check that inputs/outputs are as we expect them (see definition of CreateAdditionWorkloadTest)
AdditionQueueDescriptor queueDescriptor = workload->GetData();
auto inputHandle1 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
auto inputHandle2 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[1]);
auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle1, TensorInfo({2, 3}, DataType::Float32)));
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle2, TensorInfo({2, 3}, DataType::Float32)));
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({2, 3}, DataType::Float32)));
}
BOOST_AUTO_TEST_CASE(CreateBatchNormalizationWorkload)
{
Graph graph;
NeonWorkloadFactory factory;
auto workload = CreateBatchNormalizationWorkloadTest<NeonBatchNormalizationFloat32Workload>(factory, graph);
// check that outputs and inputs are as we expect them (see definition of CreateBatchNormalizationWorkloadTest)
BatchNormalizationQueueDescriptor queueDescriptor = workload->GetData();
auto inputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({2, 3, 1, 1}, DataType::Float32)));
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({2, 3, 1, 1}, DataType::Float32)));
}
BOOST_AUTO_TEST_CASE(CreateConvolution2dWorkload)
{
Graph graph;
NeonWorkloadFactory factory;
auto workload = CreateConvolution2dWorkloadTest<NeonConvolution2dFloat32Workload>(factory, graph);
// check that outputs and inputs are as we expect them (see definition of CreateConvolution2dWorkloadTest)
Convolution2dQueueDescriptor queueDescriptor = workload->GetData();
auto inputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({2, 3, 8, 16}, DataType::Float32)));
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({2, 2, 2, 10}, DataType::Float32)));
}
BOOST_AUTO_TEST_CASE(CreateFullyConnectedWorkload)
{
Graph graph;
NeonWorkloadFactory factory;
auto workload = CreateFullyConnectedWorkloadTest<NeonFullyConnectedFloat32Workload>(factory, graph);
// check that outputs and inputs are as we expect them (see definition of CreateFullyConnectedWorkloadTest)
FullyConnectedQueueDescriptor queueDescriptor = workload->GetData();
auto inputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({3, 1, 4, 5}, DataType::Float32)));
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({3, 7}, DataType::Float32)));
}
BOOST_AUTO_TEST_CASE(CreateMultiplicationWorkload)
{
Graph graph;
NeonWorkloadFactory factory;
auto workload = CreateMultiplicationWorkloadTest<NeonMultiplicationFloat32Workload>(factory, graph);
// check that inputs/outputs are as we expect them (see definition of CreateMultiplicationWorkloadTest)
MultiplicationQueueDescriptor queueDescriptor = workload->GetData();
auto inputHandle1 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
auto inputHandle2 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[1]);
auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle1, TensorInfo({2, 3}, DataType::Float32)));
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle2, TensorInfo({2, 3}, DataType::Float32)));
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({2, 3}, DataType::Float32)));
}
BOOST_AUTO_TEST_CASE(CreateNormalizationWorkload)
{
Graph graph;
NeonWorkloadFactory factory;
auto workload = CreateNormalizationWorkloadTest<NeonNormalizationFloat32Workload>(factory, graph);
// check that outputs and inputs are as we expect them (see definition of CreateNormalizationWorkloadTest)
NormalizationQueueDescriptor queueDescriptor = workload->GetData();
auto inputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({3, 5, 5, 1}, DataType::Float32)));
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({3, 5, 5, 1}, DataType::Float32)));
}
BOOST_AUTO_TEST_CASE(CreatePooling2dWorkload)
{
Graph graph;
NeonWorkloadFactory factory;
auto workload = CreatePooling2dWorkloadTest<NeonPooling2dFloat32Workload>(factory, graph);
// check that outputs and inputs are as we expect them (see definition of CreatePooling2dWorkloadTest)
Pooling2dQueueDescriptor queueDescriptor = workload->GetData();
auto inputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({3, 2, 5, 5}, DataType::Float32)));
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({3, 2, 2, 4}, DataType::Float32)));
}
template <typename ReshapeWorkloadType>
static void NeonCreateReshapeWorkloadTest(DataType dataType)
{
Graph graph;
NeonWorkloadFactory factory;
auto workload = CreateReshapeWorkloadTest<ReshapeWorkloadType>(factory, graph);
// check that outputs and inputs are as we expect them (see definition of CreateReshapeWorkloadTest)
ReshapeQueueDescriptor queueDescriptor = workload->GetData();
auto inputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({4, 1}, dataType)));
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({1, 4}, dataType)));
}
BOOST_AUTO_TEST_CASE(CreateReshapeFloat32Workload)
{
NeonCreateReshapeWorkloadTest<NeonReshapeFloat32Workload>(DataType::Float32);
}
BOOST_AUTO_TEST_CASE(CreateReshapeUint8Workload)
{
NeonCreateReshapeWorkloadTest<NeonReshapeUint8Workload>(DataType::QuantisedAsymm8);
}
BOOST_AUTO_TEST_CASE(CreateSoftmaxWorkload)
{
Graph graph;
NeonWorkloadFactory factory;
auto workload = CreateSoftmaxWorkloadTest<NeonSoftmaxFloat32Workload>(factory, graph);
// check that outputs and inputs are as we expect them (see definition of CreateSoftmaxWorkloadTest)
SoftmaxQueueDescriptor queueDescriptor = workload->GetData();
auto inputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({4, 1}, DataType::Float32)));
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({4, 1}, DataType::Float32)));
}
BOOST_AUTO_TEST_CASE(CreateSplitterWorkload)
{
Graph graph;
NeonWorkloadFactory factory;
auto workload = CreateSplitterWorkloadTest<NeonSplitterFloat32Workload>(factory, graph);
// check that outputs are as we expect them (see definition of CreateSplitterWorkloadTest)
SplitterQueueDescriptor queueDescriptor = workload->GetData();
auto inputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({1, 7}, DataType::Float32)));
auto outputHandle0 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle0, TensorInfo({1, 4}, DataType::Float32)));
auto outputHandle1 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[1]);
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle1, TensorInfo({1, 1}, DataType::Float32)));
auto outputHandle2 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[2]);
BOOST_TEST(TestNeonTensorHandleInfo(outputHandle2, TensorInfo({1, 2}, DataType::Float32)));
}
BOOST_AUTO_TEST_CASE(CreateSplitterMerger)
{
// Test that it is possible to decide which output of the splitter layer
// should be lined to which input of the merger layer
// We test that is is possible to specify 0th output
// of the splitter to be the 1st input to the merger and the 1st output of the splitter to be 0th input
// of the merger.
Graph graph;
NeonWorkloadFactory factory;
auto workloads =
CreateSplitterMergerWorkloadTest<NeonSplitterFloat32Workload, NeonMergerFloat32Workload>(factory, graph);
auto wlSplitter = std::move(workloads.first);
auto wlMerger = std::move(workloads.second);
//check that the index of inputs/outputs matches what we declared on InputDescriptor construction.
armnn::INeonTensorHandle* sOut0 = dynamic_cast<armnn::INeonTensorHandle*>(wlSplitter->GetData().m_Outputs[0]);
armnn::INeonTensorHandle* sOut1 = dynamic_cast<armnn::INeonTensorHandle*>(wlSplitter->GetData().m_Outputs[1]);
armnn::INeonTensorHandle* mIn0 = dynamic_cast<armnn::INeonTensorHandle*>(wlMerger->GetData().m_Inputs[0]);
armnn::INeonTensorHandle* mIn1 = dynamic_cast<armnn::INeonTensorHandle*>(wlMerger->GetData().m_Inputs[1]);
BOOST_TEST(sOut0);
BOOST_TEST(sOut1);
BOOST_TEST(mIn0);
BOOST_TEST(mIn1);
bool validDataPointers = (sOut0 == mIn1) && (sOut1 == mIn0);
BOOST_TEST(validDataPointers);
}
BOOST_AUTO_TEST_CASE(CreateSingleOutputMultipleInputs)
{
// Test that it is possible to assign multiple (two) different layers to each of the outputs of a splitter layer.
// We create a splitter with two outputs. That each of those outputs is used by two different activation layers
Graph graph;
NeonWorkloadFactory factory;
std::unique_ptr<NeonSplitterFloat32Workload> wlSplitter;
std::unique_ptr<NeonActivationFloat32Workload> wlActiv0_0;
std::unique_ptr<NeonActivationFloat32Workload> wlActiv0_1;
std::unique_ptr<NeonActivationFloat32Workload> wlActiv1_0;
std::unique_ptr<NeonActivationFloat32Workload> wlActiv1_1;
CreateSplitterMultipleInputsOneOutputWorkloadTest<NeonSplitterFloat32Workload,
NeonActivationFloat32Workload>(factory, graph, wlSplitter, wlActiv0_0, wlActiv0_1, wlActiv1_0, wlActiv1_1);
armnn::INeonTensorHandle* sOut0 = dynamic_cast<armnn::INeonTensorHandle*>(wlSplitter->GetData().m_Outputs[0]);
armnn::INeonTensorHandle* sOut1 = dynamic_cast<armnn::INeonTensorHandle*>(wlSplitter->GetData().m_Outputs[1]);
armnn::INeonTensorHandle* activ0_0Im = dynamic_cast<armnn::INeonTensorHandle*>(wlActiv0_0->GetData().m_Inputs[0]);
armnn::INeonTensorHandle* activ0_1Im = dynamic_cast<armnn::INeonTensorHandle*>(wlActiv0_1->GetData().m_Inputs[0]);
armnn::INeonTensorHandle* activ1_0Im = dynamic_cast<armnn::INeonTensorHandle*>(wlActiv1_0->GetData().m_Inputs[0]);
armnn::INeonTensorHandle* activ1_1Im = dynamic_cast<armnn::INeonTensorHandle*>(wlActiv1_1->GetData().m_Inputs[0]);
BOOST_TEST(sOut0);
BOOST_TEST(sOut1);
BOOST_TEST(activ0_0Im);
BOOST_TEST(activ0_1Im);
BOOST_TEST(activ1_0Im);
BOOST_TEST(activ1_1Im);
bool validDataPointers = (sOut0 == activ0_0Im) && (sOut0 == activ0_1Im) &&
(sOut1 == activ1_0Im) && (sOut1 == activ1_1Im);
BOOST_TEST(validDataPointers);
}
BOOST_AUTO_TEST_CASE(CreateMemCopyWorkloadsNeon)
{
NeonWorkloadFactory factory;
CreateMemCopyWorkloads<CopyFromCpuToNeonWorkload,CopyFromNeonToCpuWorkload,INeonTensorHandle>(factory);
}
BOOST_AUTO_TEST_SUITE_END()
|
