1 //
2 // Copyright © 2017 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6
7 #include <Graph.hpp>
8 #include <Network.hpp>
9
10 #include <reference/RefWorkloadFactory.hpp>
11
12 #include <boost/test/unit_test.hpp>
13
14 BOOST_AUTO_TEST_SUITE(OptimizedNetwork)
15
BOOST_AUTO_TEST_CASE(SerializeToDot)16 BOOST_AUTO_TEST_CASE(SerializeToDot)
17 {
18 armnn::Network net;
19
20 //Defines layers.
21 auto input = net.AddInputLayer(0);
22 auto add = net.AddAdditionLayer();
23 auto output = net.AddOutputLayer(0);
24
25 // Connects layers.
26 input->GetOutputSlot(0).Connect(add->GetInputSlot(0));
27 input->GetOutputSlot(0).Connect(add->GetInputSlot(1));
28 add->GetOutputSlot(0).Connect(output->GetInputSlot(0));
29
30 armnn::TensorShape shape({4});
31 armnn::TensorInfo info(shape, armnn::DataType::Float32);
32 input->GetOutputSlot(0).SetTensorInfo(info);
33 add->GetOutputSlot(0).SetTensorInfo(info);
34
35 armnn::IRuntime::CreationOptions options;
36 armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
37
38 std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
39 armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec());
40
41 std::ostringstream ss;
42 optimizedNet->SerializeToDot(ss);
43
44 auto inputId = input->GetGuid();
45 auto addId = add->GetGuid();
46 auto outputId = output->GetGuid();
47
48 std::stringstream expected;
49 expected <<
50 "digraph Optimized {\n"
51 " node [shape=\"record\"];\n"
52 " edge [fontsize=8 fontcolor=\"blue\" fontname=\"arial-bold\"];\n"
53 " " << inputId << " [label=\"{Input|Guid : " << inputId << "\\lLayerType : Input\\l"
54 "BackendID : CpuRef\\l}\"];\n"
55 " " << addId << " [label=\"{Addition|Guid : " << addId << "\\lLayerType : Addition\\l"
56 "BackendID : CpuRef\\l}\"];\n"
57 " " << outputId << " [label=\"{Output|Guid : " << outputId << "\\lLayerType : Output\\l"
58 "BackendID : CpuRef\\l}\"];\n"
59 " " << inputId << " -> " << addId << " [label=< [4] >];\n"
60 " " << inputId << " -> " << addId << " [label=< [4] >];\n"
61 " " << addId << " -> " << outputId << " [label=< [4] >];\n"
62 "}\n";
63
64 BOOST_TEST(ss.str() == expected.str());
65 }
66
BOOST_AUTO_TEST_CASE(OptimizeValidateDeviceNonSupportLayerNoFallback)67 BOOST_AUTO_TEST_CASE(OptimizeValidateDeviceNonSupportLayerNoFallback)
68 {
69 // build up the structure of the network
70 armnn::INetworkPtr net(armnn::INetwork::Create());
71
72 armnn::IConnectableLayer* input = net->AddInputLayer(0);
73
74 // This layer configuration isn't supported by CpuAcc and isn't allowed to fall back, so Optimize will return null.
75 armnn::NormalizationDescriptor descriptor;
76 armnn::IConnectableLayer* normalize = net->AddNormalizationLayer(descriptor);
77
78 armnn::IConnectableLayer* output = net->AddOutputLayer(0);
79
80 input->GetOutputSlot(0).Connect(normalize->GetInputSlot(0));
81 normalize->GetOutputSlot(0).Connect(output->GetInputSlot(0));
82
83 input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
84 normalize->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
85
86 armnn::IRuntime::CreationOptions options;
87 armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
88
89 std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
90 std::vector<std::string> errMessages;
91
92 try
93 {
94 Optimize(*net, backends, runtime->GetDeviceSpec(), armnn::OptimizerOptions(), errMessages);
95 BOOST_FAIL("Should have thrown an exception.");
96 }
97 catch (const armnn::InvalidArgumentException& e)
98 {
99 // Different exceptions are thrown on different backends
100 }
101 BOOST_CHECK(errMessages.size() > 0);
102 }
103
BOOST_AUTO_TEST_CASE(OptimizeValidateDeviceNonSupportLayerWithFallback)104 BOOST_AUTO_TEST_CASE(OptimizeValidateDeviceNonSupportLayerWithFallback)
105 {
106 // build up the structure of the network
107 armnn::INetworkPtr net(armnn::INetwork::Create());
108
109 armnn::IConnectableLayer* input = net->AddInputLayer(0);
110
111 // This layer configuration isn't supported by CpuAcc but it allows to fallback to CpuRef.
112 armnn::NormalizationDescriptor descriptor;
113 armnn::IConnectableLayer* normalize = net->AddNormalizationLayer(descriptor);
114
115 armnn::IConnectableLayer* output = net->AddOutputLayer(0);
116
117 input->GetOutputSlot(0).Connect(normalize->GetInputSlot(0));
118 normalize->GetOutputSlot(0).Connect(output->GetInputSlot(0));
119
120 input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
121 normalize->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
122
123 armnn::IRuntime::CreationOptions options;
124 armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
125
126 std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc, armnn::Compute::CpuRef };
127 armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
128 BOOST_REQUIRE(optNet);
129
130 for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
131 {
132 // If NEON is enabled, Input and Output layers are supported by CpuAcc,
133 // the other layers are supported by CpuRef.
134 // If NEON is not enabled, all layers are supported by CpuRef.
135 #if defined(ARMCOMPUTENEON_ENABLED)
136 if (layer->GetType() == armnn::LayerType::Input || layer->GetType() == armnn::LayerType::Output)
137 {
138 BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuAcc);
139 }
140 else if (layer->GetType() == armnn::LayerType::Normalization)
141 {
142 BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
143 }
144 #else
145 BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
146 #endif
147 }
148 }
149
BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsUndefinedComputeDevice)150 BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsUndefinedComputeDevice)
151 {
152 const armnn::TensorInfo desc({3, 5}, armnn::DataType::Float32);
153
154 armnn::Network net;
155
156 armnn::NormalizationDescriptor nmDesc;
157 armnn::ActivationDescriptor acDesc;
158
159 // in
160 // |
161 // nm
162 // / |
163 // ac |
164 // \ |
165 // ml
166 // |
167 // sm
168 // |
169 // ot
170 armnn::IConnectableLayer* layer = net.AddInputLayer(0, "in");
171 layer->GetOutputSlot(0).SetTensorInfo(desc);
172
173 armnn::IConnectableLayer* const normLayer = net.AddNormalizationLayer(nmDesc, "nm");
174
175 layer->GetOutputSlot(0).Connect(normLayer->GetInputSlot(0));
176 normLayer->GetOutputSlot(0).SetTensorInfo(desc);
177
178 layer = net.AddActivationLayer(acDesc, "ac");
179
180 normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
181 layer->GetOutputSlot(0).SetTensorInfo(desc);
182
183 armnn::IConnectableLayer* prevLayer = layer;
184 layer = net.AddMultiplicationLayer("ml");
185
186 prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
187 normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1));
188 layer->GetOutputSlot(0).SetTensorInfo(desc);
189
190 prevLayer = layer;
191 armnn::SoftmaxDescriptor softmaxDescriptor;
192 layer = net.AddSoftmaxLayer(softmaxDescriptor, "sm");
193
194 prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
195 layer->GetOutputSlot(0).SetTensorInfo(desc);
196
197 prevLayer = layer;
198 layer = net.AddOutputLayer(0, "ot");
199
200 prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
201
202 armnn::IRuntime::CreationOptions options;
203 armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
204
205 std::vector<armnn::BackendId> backends = { armnn::Compute::Undefined };
206 std::vector<std::string> errMessages;
207
208 try
209 {
210 Optimize(net, backends, runtime->GetDeviceSpec(), armnn::OptimizerOptions(), errMessages);
211 BOOST_FAIL("Should have thrown an exception.");
212 }
213 catch (const armnn::InvalidArgumentException& e)
214 {
215 // Different exceptions are thrown on different backends
216 }
217 BOOST_CHECK(errMessages.size() > 0);
218 }
219
BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsUndefinedComputeDeviceWithFallback)220 BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsUndefinedComputeDeviceWithFallback)
221 {
222 const armnn::TensorInfo desc({3, 5}, armnn::DataType::Float32);
223
224 armnn::Network net;
225
226 armnn::NormalizationDescriptor nmDesc;
227 armnn::ActivationDescriptor acDesc;
228
229 // in
230 // |
231 // nm
232 // / |
233 // ac |
234 // \ |
235 // ml
236 // |
237 // sm
238 // |
239 // ot
240 armnn::IConnectableLayer* layer = net.AddInputLayer(0, "in");
241 layer->GetOutputSlot(0).SetTensorInfo(desc);
242
243 armnn::IConnectableLayer* const normLayer = net.AddNormalizationLayer(nmDesc, "nm");
244
245 layer->GetOutputSlot(0).Connect(normLayer->GetInputSlot(0));
246 normLayer->GetOutputSlot(0).SetTensorInfo(desc);
247
248 layer = net.AddActivationLayer(acDesc, "ac");
249
250 normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
251 layer->GetOutputSlot(0).SetTensorInfo(desc);
252
253 armnn::IConnectableLayer* prevLayer = layer;
254 layer = net.AddMultiplicationLayer("ml");
255
256 prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
257 normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1));
258 layer->GetOutputSlot(0).SetTensorInfo(desc);
259
260 prevLayer = layer;
261 armnn::SoftmaxDescriptor softmaxDescriptor;
262 layer = net.AddSoftmaxLayer(softmaxDescriptor, "sm");
263
264 prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
265 layer->GetOutputSlot(0).SetTensorInfo(desc);
266
267 prevLayer = layer;
268 layer = net.AddOutputLayer(0, "ot");
269
270 prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
271
272 armnn::IRuntime::CreationOptions options;
273 armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
274
275 std::vector<armnn::BackendId> backends = { armnn::Compute::Undefined, armnn::Compute::CpuRef };
276
277 armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec());
278 BOOST_CHECK(optNet);
279
280 // validate workloads
281 armnn::RefWorkloadFactory fact;
282 for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
283 {
284 BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
285 BOOST_CHECK_NO_THROW(
286 layer->CreateWorkload(fact));
287 }
288 }
289
BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsDuplicateComputeDeviceWithFallback)290 BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsDuplicateComputeDeviceWithFallback)
291 {
292 // build up the structure of the network
293 armnn::INetworkPtr net(armnn::INetwork::Create());
294
295 armnn::IConnectableLayer* input = net->AddInputLayer(0);
296
297 // This layer configuration isn't supported by CpuAcc but it allows to fallback to CpuRef.
298 armnn::NormalizationDescriptor descriptor;
299 armnn::IConnectableLayer* normalize = net->AddNormalizationLayer(descriptor);
300
301 armnn::IConnectableLayer* output = net->AddOutputLayer(0);
302
303 input->GetOutputSlot(0).Connect(normalize->GetInputSlot(0));
304 normalize->GetOutputSlot(0).Connect(output->GetInputSlot(0));
305
306 input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
307 normalize->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
308
309 armnn::IRuntime::CreationOptions options;
310 armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
311
312 std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc,
313 armnn::Compute::GpuAcc,
314 armnn::Compute::CpuRef };
315
316 armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
317 BOOST_REQUIRE(optNet);
318
319 for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
320 {
321 // If NEON is enabled, Input and Output layers are supported by CpuAcc,
322 // the other layers are supported by CpuRef.
323 // If only CL is enabled, Input and Output layers are supported by GpuAcc,
324 // the other layers are supported by CpuRef.
325 // If neither NEON, nor CL is enabled, all layers are supported by CpuRef.
326 #if defined(ARMCOMPUTENEON_ENABLED)
327 if (layer->GetType() == armnn::LayerType::Input || layer->GetType() == armnn::LayerType::Output)
328 {
329 BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuAcc);
330 }
331 else if (layer->GetType() == armnn::LayerType::Normalization)
332 {
333 BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
334 }
335 #elif defined(ARMCOMPUTECL_ENABLED)
336 if (layer->GetType() == armnn::LayerType::Input || layer->GetType() == armnn::LayerType::Output)
337 {
338 BOOST_CHECK(layer->GetBackendId() == armnn::Compute::GpuAcc);
339 }
340 else if (layer->GetType() == armnn::LayerType::Normalization)
341 {
342 BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
343 }
344 #else
345 BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
346 #endif
347 }
348 }
349
350 BOOST_AUTO_TEST_SUITE_END()
351