Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include <armnn/ArmNN.hpp> |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 7 | |
Aron Virginas-Tar | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 8 | #include <Graph.hpp> |
| 9 | #include <Network.hpp> |
| 10 | |
| 11 | #include <reference/RefWorkloadFactory.hpp> |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 12 | |
| 13 | #include <boost/test/unit_test.hpp> |
| 14 | |
| 15 | BOOST_AUTO_TEST_SUITE(OptimizedNetwork) |
| 16 | |
| 17 | BOOST_AUTO_TEST_CASE(SerializeToDot) |
| 18 | { |
| 19 | armnn::Network net; |
| 20 | |
| 21 | //Defines layers. |
| 22 | auto input = net.AddInputLayer(0); |
| 23 | auto add = net.AddAdditionLayer(); |
| 24 | auto output = net.AddOutputLayer(0); |
| 25 | |
| 26 | // Connects layers. |
| 27 | input->GetOutputSlot(0).Connect(add->GetInputSlot(0)); |
| 28 | input->GetOutputSlot(0).Connect(add->GetInputSlot(1)); |
| 29 | add->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 30 | |
| 31 | armnn::TensorShape shape({4}); |
| 32 | armnn::TensorInfo info(shape, armnn::DataType::Float32); |
| 33 | input->GetOutputSlot(0).SetTensorInfo(info); |
| 34 | add->GetOutputSlot(0).SetTensorInfo(info); |
| 35 | |
| 36 | armnn::IRuntime::CreationOptions options; |
| 37 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 38 | |
| 39 | std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef}; |
| 40 | armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec()); |
| 41 | |
| 42 | std::ostringstream ss; |
| 43 | optimizedNet->SerializeToDot(ss); |
| 44 | |
| 45 | auto inputId = input->GetGuid(); |
| 46 | auto addId = add->GetGuid(); |
| 47 | auto outputId = output->GetGuid(); |
| 48 | |
| 49 | std::stringstream expected; |
| 50 | expected << |
| 51 | "digraph Optimized {\n" |
| 52 | " node [shape=\"record\"];\n" |
| 53 | " edge [fontsize=8 fontcolor=\"blue\" fontname=\"arial-bold\"];\n" |
| 54 | " " << inputId << " [label=\"{Input}\"];\n" |
| 55 | " " << addId << " [label=\"{Addition}\"];\n" |
| 56 | " " << outputId << " [label=\"{Output}\"];\n" |
| 57 | " " << inputId << " -> " << addId << " [label=< [4] >];\n" |
| 58 | " " << inputId << " -> " << addId << " [label=< [4] >];\n" |
| 59 | " " << addId << " -> " << outputId << " [label=< [4] >];\n" |
| 60 | "}\n"; |
| 61 | |
| 62 | BOOST_TEST(ss.str() == expected.str()); |
| 63 | } |
| 64 | |
| 65 | BOOST_AUTO_TEST_CASE(OptimizeValidateDeviceNonSupportLayerNoFallback) |
| 66 | { |
| 67 | // build up the structure of the network |
| 68 | armnn::INetworkPtr net(armnn::INetwork::Create()); |
| 69 | |
| 70 | armnn::IConnectableLayer* input = net->AddInputLayer(0); |
| 71 | |
| 72 | // This layer configuration isn't supported by CpuAcc and isn't allowed to fall back, so Optimize will return null. |
| 73 | armnn::NormalizationDescriptor descriptor; |
| 74 | armnn::IConnectableLayer* normalize = net->AddNormalizationLayer(descriptor); |
| 75 | |
| 76 | armnn::IConnectableLayer* output = net->AddOutputLayer(0); |
| 77 | |
| 78 | input->GetOutputSlot(0).Connect(normalize->GetInputSlot(0)); |
| 79 | normalize->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 80 | |
| 81 | input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32)); |
| 82 | normalize->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32)); |
| 83 | |
| 84 | armnn::IRuntime::CreationOptions options; |
| 85 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 86 | |
| 87 | std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc }; |
| 88 | armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec()); |
| 89 | BOOST_CHECK(!optNet); |
| 90 | } |
| 91 | |
| 92 | BOOST_AUTO_TEST_CASE(OptimizeValidateDeviceNonSupportLayerWithFallback) |
| 93 | { |
| 94 | // build up the structure of the network |
| 95 | armnn::INetworkPtr net(armnn::INetwork::Create()); |
| 96 | |
| 97 | armnn::IConnectableLayer* input = net->AddInputLayer(0); |
| 98 | |
| 99 | // This layer configuration isn't supported by CpuAcc but it allows to fallback to CpuRef. |
| 100 | armnn::NormalizationDescriptor descriptor; |
| 101 | armnn::IConnectableLayer* normalize = net->AddNormalizationLayer(descriptor); |
| 102 | |
| 103 | armnn::IConnectableLayer* output = net->AddOutputLayer(0); |
| 104 | |
| 105 | input->GetOutputSlot(0).Connect(normalize->GetInputSlot(0)); |
| 106 | normalize->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 107 | |
| 108 | input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32)); |
| 109 | normalize->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32)); |
| 110 | |
| 111 | armnn::IRuntime::CreationOptions options; |
| 112 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 113 | |
| 114 | std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc, armnn::Compute::CpuRef }; |
| 115 | armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec()); |
| 116 | BOOST_REQUIRE(optNet); |
| 117 | |
| 118 | for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph()) |
| 119 | { |
| 120 | // If NEON is enabled, Input and Output layers are supported by CpuAcc, |
| 121 | // the other layers are supported by CpuRef. |
| 122 | // If NEON is not enabled, all layers are supported by CpuRef. |
Matteo Martincigh | d95e906 | 2019-01-31 15:35:59 +0000 | [diff] [blame] | 123 | #if defined(ARMCOMPUTENEON_ENABLED) |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 124 | if (layer->GetType() == armnn::LayerType::Input || layer->GetType() == armnn::LayerType::Output) |
| 125 | { |
| 126 | BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuAcc); |
| 127 | } |
| 128 | else if (layer->GetType() == armnn::LayerType::Normalization) |
| 129 | { |
| 130 | BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef); |
| 131 | } |
| 132 | #else |
| 133 | BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef); |
| 134 | #endif |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsUndefinedComputeDevice) |
| 139 | { |
| 140 | const armnn::TensorInfo desc({3, 5}, armnn::DataType::Float32); |
| 141 | |
| 142 | armnn::Network net; |
| 143 | |
| 144 | armnn::NormalizationDescriptor nmDesc; |
| 145 | armnn::ActivationDescriptor acDesc; |
| 146 | |
| 147 | // in |
| 148 | // | |
| 149 | // nm |
| 150 | // / | |
| 151 | // ac | |
| 152 | // \ | |
| 153 | // ml |
| 154 | // | |
| 155 | // sm |
| 156 | // | |
| 157 | // ot |
| 158 | armnn::IConnectableLayer* layer = net.AddInputLayer(0, "in"); |
| 159 | layer->GetOutputSlot(0).SetTensorInfo(desc); |
| 160 | |
| 161 | armnn::IConnectableLayer* const normLayer = net.AddNormalizationLayer(nmDesc, "nm"); |
| 162 | |
| 163 | layer->GetOutputSlot(0).Connect(normLayer->GetInputSlot(0)); |
| 164 | normLayer->GetOutputSlot(0).SetTensorInfo(desc); |
| 165 | |
| 166 | layer = net.AddActivationLayer(acDesc, "ac"); |
| 167 | |
| 168 | normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); |
| 169 | layer->GetOutputSlot(0).SetTensorInfo(desc); |
| 170 | |
| 171 | armnn::IConnectableLayer* prevLayer = layer; |
| 172 | layer = net.AddMultiplicationLayer("ml"); |
| 173 | |
| 174 | prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); |
| 175 | normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1)); |
| 176 | layer->GetOutputSlot(0).SetTensorInfo(desc); |
| 177 | |
| 178 | prevLayer = layer; |
| 179 | armnn::SoftmaxDescriptor softmaxDescriptor; |
| 180 | layer = net.AddSoftmaxLayer(softmaxDescriptor, "sm"); |
| 181 | |
| 182 | prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); |
| 183 | layer->GetOutputSlot(0).SetTensorInfo(desc); |
| 184 | |
| 185 | prevLayer = layer; |
| 186 | layer = net.AddOutputLayer(0, "ot"); |
| 187 | |
| 188 | prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); |
| 189 | |
| 190 | armnn::IRuntime::CreationOptions options; |
| 191 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 192 | |
| 193 | std::vector<armnn::BackendId> backends = { armnn::Compute::Undefined }; |
| 194 | |
| 195 | armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec()); |
| 196 | BOOST_CHECK(!optNet); |
| 197 | |
| 198 | } |
| 199 | |
| 200 | BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsUndefinedComputeDeviceWithFallback) |
| 201 | { |
| 202 | const armnn::TensorInfo desc({3, 5}, armnn::DataType::Float32); |
| 203 | |
| 204 | armnn::Network net; |
| 205 | |
| 206 | armnn::NormalizationDescriptor nmDesc; |
| 207 | armnn::ActivationDescriptor acDesc; |
| 208 | |
| 209 | // in |
| 210 | // | |
| 211 | // nm |
| 212 | // / | |
| 213 | // ac | |
| 214 | // \ | |
| 215 | // ml |
| 216 | // | |
| 217 | // sm |
| 218 | // | |
| 219 | // ot |
| 220 | armnn::IConnectableLayer* layer = net.AddInputLayer(0, "in"); |
| 221 | layer->GetOutputSlot(0).SetTensorInfo(desc); |
| 222 | |
| 223 | armnn::IConnectableLayer* const normLayer = net.AddNormalizationLayer(nmDesc, "nm"); |
| 224 | |
| 225 | layer->GetOutputSlot(0).Connect(normLayer->GetInputSlot(0)); |
| 226 | normLayer->GetOutputSlot(0).SetTensorInfo(desc); |
| 227 | |
| 228 | layer = net.AddActivationLayer(acDesc, "ac"); |
| 229 | |
| 230 | normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); |
| 231 | layer->GetOutputSlot(0).SetTensorInfo(desc); |
| 232 | |
| 233 | armnn::IConnectableLayer* prevLayer = layer; |
| 234 | layer = net.AddMultiplicationLayer("ml"); |
| 235 | |
| 236 | prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); |
| 237 | normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1)); |
| 238 | layer->GetOutputSlot(0).SetTensorInfo(desc); |
| 239 | |
| 240 | prevLayer = layer; |
| 241 | armnn::SoftmaxDescriptor softmaxDescriptor; |
| 242 | layer = net.AddSoftmaxLayer(softmaxDescriptor, "sm"); |
| 243 | |
| 244 | prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); |
| 245 | layer->GetOutputSlot(0).SetTensorInfo(desc); |
| 246 | |
| 247 | prevLayer = layer; |
| 248 | layer = net.AddOutputLayer(0, "ot"); |
| 249 | |
| 250 | prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); |
| 251 | |
| 252 | armnn::IRuntime::CreationOptions options; |
| 253 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 254 | |
| 255 | std::vector<armnn::BackendId> backends = { armnn::Compute::Undefined, armnn::Compute::CpuRef }; |
| 256 | |
| 257 | armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec()); |
| 258 | BOOST_CHECK(optNet); |
| 259 | |
| 260 | // validate workloads |
| 261 | armnn::RefWorkloadFactory fact; |
| 262 | for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph()) |
| 263 | { |
| 264 | BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef); |
| 265 | BOOST_CHECK_NO_THROW( |
| 266 | layer->CreateWorkload(static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph(), fact)); |
| 267 | } |
| 268 | } |
| 269 | |
| 270 | BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsDuplicateComputeDeviceWithFallback) |
| 271 | { |
| 272 | // build up the structure of the network |
| 273 | armnn::INetworkPtr net(armnn::INetwork::Create()); |
| 274 | |
| 275 | armnn::IConnectableLayer* input = net->AddInputLayer(0); |
| 276 | |
| 277 | // This layer configuration isn't supported by CpuAcc but it allows to fallback to CpuRef. |
| 278 | armnn::NormalizationDescriptor descriptor; |
| 279 | armnn::IConnectableLayer* normalize = net->AddNormalizationLayer(descriptor); |
| 280 | |
| 281 | armnn::IConnectableLayer* output = net->AddOutputLayer(0); |
| 282 | |
| 283 | input->GetOutputSlot(0).Connect(normalize->GetInputSlot(0)); |
| 284 | normalize->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 285 | |
| 286 | input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32)); |
| 287 | normalize->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32)); |
| 288 | |
| 289 | armnn::IRuntime::CreationOptions options; |
| 290 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 291 | |
| 292 | std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc, |
| 293 | armnn::Compute::GpuAcc, |
| 294 | armnn::Compute::CpuRef }; |
| 295 | |
| 296 | armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec()); |
| 297 | BOOST_REQUIRE(optNet); |
| 298 | |
| 299 | for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph()) |
| 300 | { |
| 301 | // If NEON is enabled, Input and Output layers are supported by CpuAcc, |
| 302 | // the other layers are supported by CpuRef. |
| 303 | // If only CL is enabled, Input and Output layers are supported by GpuAcc, |
| 304 | // the other layers are supported by CpuRef. |
| 305 | // If neither NEON, nor CL is enabled, all layers are supported by CpuRef. |
Matteo Martincigh | d95e906 | 2019-01-31 15:35:59 +0000 | [diff] [blame] | 306 | #if defined(ARMCOMPUTENEON_ENABLED) |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 307 | if (layer->GetType() == armnn::LayerType::Input || layer->GetType() == armnn::LayerType::Output) |
| 308 | { |
| 309 | BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuAcc); |
| 310 | } |
| 311 | else if (layer->GetType() == armnn::LayerType::Normalization) |
| 312 | { |
| 313 | BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef); |
| 314 | } |
Matteo Martincigh | d95e906 | 2019-01-31 15:35:59 +0000 | [diff] [blame] | 315 | #elif defined(ARMCOMPUTECL_ENABLED) |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 316 | if (layer->GetType() == armnn::LayerType::Input || layer->GetType() == armnn::LayerType::Output) |
| 317 | { |
| 318 | BOOST_CHECK(layer->GetBackendId() == armnn::Compute::GpuAcc); |
| 319 | } |
| 320 | else if (layer->GetType() == armnn::LayerType::Normalization) |
| 321 | { |
| 322 | BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef); |
| 323 | } |
| 324 | #else |
| 325 | BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef); |
| 326 | #endif |
| 327 | } |
| 328 | } |
| 329 | |
| 330 | BOOST_AUTO_TEST_SUITE_END() |