telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // See LICENSE file in the project root for full license information. |
| 4 | // |
| 5 | #include "Network.hpp" |
| 6 | #include "Graph.hpp" |
| 7 | #include "Layer.hpp" |
| 8 | #include "backends/CpuTensorHandle.hpp" |
| 9 | #include "backends/WorkloadFactory.hpp" |
| 10 | #include "Layers.hpp" |
| 11 | #include "Optimizer.hpp" |
| 12 | |
| 13 | #include <armnn/Utils.hpp> |
| 14 | |
| 15 | #include <fcntl.h> |
| 16 | #include <algorithm> |
| 17 | #include <fstream> |
| 18 | #include <memory> |
| 19 | |
| 20 | #include <boost/assert.hpp> |
| 21 | #include <boost/format.hpp> |
| 22 | #include <boost/log/trivial.hpp> |
| 23 | #include <boost/numeric/conversion/converter_policies.hpp> |
| 24 | #include <boost/cast.hpp> |
| 25 | |
| 26 | namespace armnn |
| 27 | { |
| 28 | |
| 29 | armnn::INetwork* INetwork::CreateRaw() |
| 30 | { |
| 31 | return new Network(); |
| 32 | } |
| 33 | |
| 34 | armnn::INetworkPtr INetwork::Create() |
| 35 | { |
| 36 | return INetworkPtr(CreateRaw(), &INetwork::Destroy); |
| 37 | } |
| 38 | |
| 39 | void INetwork::Destroy(INetwork* network) |
| 40 | { |
| 41 | delete boost::polymorphic_downcast<Network*>(network); |
| 42 | } |
| 43 | |
| 44 | Status Network::PrintGraph() |
| 45 | { |
| 46 | m_Graph->Print(); |
| 47 | return Status::Success; |
| 48 | } |
| 49 | |
| 50 | void IOptimizedNetwork::Destroy(IOptimizedNetwork* network) |
| 51 | { |
| 52 | delete boost::polymorphic_downcast<OptimizedNetwork*>(network); |
| 53 | } |
| 54 | |
| 55 | Status OptimizedNetwork::PrintGraph() |
| 56 | { |
| 57 | m_Graph->Print(); |
| 58 | return Status::Success; |
| 59 | } |
| 60 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame^] | 61 | Status OptimizedNetwork::SerializeToDot(std::ostream& stream) const |
| 62 | { |
| 63 | return m_Graph->SerializeToDot(stream); |
| 64 | } |
| 65 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 66 | IOptimizedNetworkPtr Optimize(const INetwork& inNetwork, const DeviceSpec& deviceSpec) |
| 67 | { |
| 68 | const Network& network = *boost::polymorphic_downcast<const Network*>(&inNetwork); |
| 69 | std::unique_ptr<Graph> graph = std::make_unique<Graph>(network.GetGraph()); |
| 70 | |
| 71 | OptimizedNetwork* optNet = new OptimizedNetwork(std::move(graph)); |
| 72 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame^] | 73 | Optimizer::Optimize(optNet->GetGraph()); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 74 | |
| 75 | // Infer the tensor infos for all output slots. Throws an exception on failure. |
| 76 | optNet->GetGraph().InferTensorInfos(); |
| 77 | |
| 78 | // Assign a compute device for all nodes |
| 79 | for (auto&& layer : optNet->GetGraph()) |
| 80 | { |
| 81 | DataType dataType = layer->GetDataType(); |
| 82 | |
| 83 | // Default to the user-requested compute device from the Runtime |
| 84 | layer->SetComputeDevice(deviceSpec.DefaultComputeDevice); |
| 85 | |
| 86 | // If the layer is unsupported by this device, fall back to reference |
| 87 | std::string reasonIfUnsupported; |
| 88 | if (!IWorkloadFactory::IsLayerSupported(*layer, dataType, reasonIfUnsupported)) |
| 89 | { |
| 90 | BOOST_LOG_TRIVIAL(warning) << "Layer of type " << GetLayerTypeAsCString(layer->GetType()) << |
| 91 | " is not supported on requested backend " << layer->GetComputeDevice() << " (reason: " << |
| 92 | reasonIfUnsupported << "), falling back to CpuRef backend."; |
| 93 | layer->SetComputeDevice(Compute::CpuRef); |
| 94 | } |
| 95 | |
| 96 | BOOST_ASSERT_MSG(IWorkloadFactory::IsLayerSupported(*layer, dataType, reasonIfUnsupported), |
| 97 | "Layer has no valid compute device"); |
| 98 | } |
| 99 | |
| 100 | optNet->GetGraph().AddCopyLayers(); |
| 101 | |
| 102 | return {optNet, &IOptimizedNetwork::Destroy}; |
| 103 | } |
| 104 | |
| 105 | Network::Network() |
| 106 | : m_Graph(std::make_unique<Graph>()) |
| 107 | { |
| 108 | } |
| 109 | |
| 110 | Network::~Network() |
| 111 | { |
| 112 | } |
| 113 | |
| 114 | IConnectableLayer* Network::AddInputLayer(LayerBindingId id, const char* name) |
| 115 | { |
| 116 | return m_Graph->AddLayer<InputLayer>(id, name); |
| 117 | } |
| 118 | |
| 119 | IConnectableLayer* Network::AddFullyConnectedLayerImpl(const FullyConnectedDescriptor& fullyConnectedDescriptor, |
| 120 | const ConstTensor& weights, |
| 121 | const ConstTensor* biases, |
| 122 | const char* name) |
| 123 | { |
| 124 | if (fullyConnectedDescriptor.m_BiasEnabled && (biases == nullptr)) |
| 125 | { |
| 126 | throw InvalidArgumentException("AddFullyConnectedLayer: biases cannot be NULL"); |
| 127 | } |
| 128 | |
| 129 | const auto layer = m_Graph->AddLayer<FullyConnectedLayer>(fullyConnectedDescriptor, name); |
| 130 | |
| 131 | layer->m_Weight = std::make_unique<ScopedCpuTensorHandle>(weights); |
| 132 | |
| 133 | if (fullyConnectedDescriptor.m_BiasEnabled) |
| 134 | { |
| 135 | layer->m_Bias = std::make_unique<ScopedCpuTensorHandle>(*biases); |
| 136 | } |
| 137 | |
| 138 | return layer; |
| 139 | } |
| 140 | |
| 141 | IConnectableLayer* Network::AddFullyConnectedLayer(const FullyConnectedDescriptor& fullyConnectedDescriptor, |
| 142 | const ConstTensor& weights, |
| 143 | const char* name) |
| 144 | { |
| 145 | return AddFullyConnectedLayerImpl(fullyConnectedDescriptor, weights, nullptr, name); |
| 146 | } |
| 147 | |
| 148 | IConnectableLayer* Network::AddFullyConnectedLayer(const FullyConnectedDescriptor& fullyConnectedDescriptor, |
| 149 | const ConstTensor& weights, |
| 150 | const ConstTensor& biases, |
| 151 | const char* name) |
| 152 | { |
| 153 | return AddFullyConnectedLayerImpl(fullyConnectedDescriptor, weights, &biases, name); |
| 154 | } |
| 155 | |
| 156 | IConnectableLayer* Network::AddConvolution2dLayerImpl(const Convolution2dDescriptor& convolution2dDescriptor, |
| 157 | const ConstTensor& weights, |
| 158 | const ConstTensor* biases, |
| 159 | const char* name) |
| 160 | { |
| 161 | if (convolution2dDescriptor.m_BiasEnabled && (biases == nullptr)) |
| 162 | { |
| 163 | throw InvalidArgumentException("AddConvolution2dLayer: biases cannot be NULL"); |
| 164 | } |
| 165 | |
| 166 | const auto layer = m_Graph->AddLayer<Convolution2dLayer>(convolution2dDescriptor, name); |
| 167 | |
| 168 | layer->m_Weight = std::make_unique<ScopedCpuTensorHandle>(weights); |
| 169 | |
| 170 | if (convolution2dDescriptor.m_BiasEnabled) |
| 171 | { |
| 172 | layer->m_Bias = std::make_unique<ScopedCpuTensorHandle>(*biases); |
| 173 | } |
| 174 | |
| 175 | return layer; |
| 176 | } |
| 177 | |
| 178 | IConnectableLayer* Network::AddConvolution2dLayer(const Convolution2dDescriptor& convolution2dDescriptor, |
| 179 | const ConstTensor& weights, |
| 180 | const char* name) |
| 181 | { |
| 182 | return AddConvolution2dLayerImpl(convolution2dDescriptor, weights, nullptr, name); |
| 183 | } |
| 184 | IConnectableLayer* Network::AddConvolution2dLayer(const Convolution2dDescriptor& convolution2dDescriptor, |
| 185 | const ConstTensor& weights, |
| 186 | const ConstTensor& biases, |
| 187 | const char* name) |
| 188 | { |
| 189 | return AddConvolution2dLayerImpl(convolution2dDescriptor, weights, &biases, name); |
| 190 | } |
| 191 | |
| 192 | IConnectableLayer* Network::AddDepthwiseConvolution2dLayerImpl( |
| 193 | const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, |
| 194 | const ConstTensor& weights, |
| 195 | const ConstTensor* biases, |
| 196 | const char* name) |
| 197 | { |
| 198 | if (convolution2dDescriptor.m_BiasEnabled && (biases == nullptr)) |
| 199 | { |
| 200 | throw InvalidArgumentException("AddDepthwiseConvolution2dLayer: biases cannot be NULL"); |
| 201 | } |
| 202 | |
| 203 | const auto layer = m_Graph->AddLayer<DepthwiseConvolution2dLayer>(convolution2dDescriptor, name); |
| 204 | |
| 205 | layer->m_Weight = std::make_unique<ScopedCpuTensorHandle>(weights); |
| 206 | |
| 207 | if (convolution2dDescriptor.m_BiasEnabled) |
| 208 | { |
| 209 | layer->m_Bias = std::make_unique<ScopedCpuTensorHandle>(*biases); |
| 210 | } |
| 211 | |
| 212 | return layer; |
| 213 | } |
| 214 | |
| 215 | IConnectableLayer* Network::AddDepthwiseConvolution2dLayer( |
| 216 | const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, |
| 217 | const ConstTensor& weights, |
| 218 | const char* name) |
| 219 | { |
| 220 | return AddDepthwiseConvolution2dLayerImpl(convolution2dDescriptor, weights, nullptr, name); |
| 221 | } |
| 222 | IConnectableLayer* Network::AddDepthwiseConvolution2dLayer( |
| 223 | const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, |
| 224 | const ConstTensor& weights, |
| 225 | const ConstTensor& biases, |
| 226 | const char* name) |
| 227 | { |
| 228 | return AddDepthwiseConvolution2dLayerImpl(convolution2dDescriptor, weights, &biases, name); |
| 229 | } |
| 230 | |
| 231 | IConnectableLayer* Network::AddPermuteLayer(const PermuteDescriptor& permuteDescriptor, |
| 232 | const char* name) |
| 233 | { |
| 234 | return m_Graph->AddLayer<PermuteLayer>(permuteDescriptor, name); |
| 235 | } |
| 236 | |
| 237 | IConnectableLayer* Network::AddPooling2dLayer(const Pooling2dDescriptor& pooling2dDescriptor, |
| 238 | const char* name) |
| 239 | { |
| 240 | return m_Graph->AddLayer<Pooling2dLayer>(pooling2dDescriptor, name); |
| 241 | } |
| 242 | |
| 243 | IConnectableLayer* Network::AddActivationLayer(const ActivationDescriptor& activationDescriptor, |
| 244 | const char* name) |
| 245 | { |
| 246 | return m_Graph->AddLayer<ActivationLayer>(activationDescriptor, name); |
| 247 | } |
| 248 | |
| 249 | IConnectableLayer* Network::AddNormalizationLayer(const NormalizationDescriptor& normalizationDescriptor, |
| 250 | const char* name) |
| 251 | { |
| 252 | return m_Graph->AddLayer<NormalizationLayer>(normalizationDescriptor, name); |
| 253 | } |
| 254 | |
| 255 | IConnectableLayer* Network::AddSoftmaxLayer(const SoftmaxDescriptor& softmaxDescriptor, |
| 256 | const char* name) |
| 257 | { |
| 258 | return m_Graph->AddLayer<SoftmaxLayer>(softmaxDescriptor, name); |
| 259 | } |
| 260 | |
| 261 | IConnectableLayer* Network::AddSplitterLayer(const ViewsDescriptor& splitterDescriptor, |
| 262 | const char* name) |
| 263 | { |
| 264 | return m_Graph->AddLayer<SplitterLayer>(splitterDescriptor, name); |
| 265 | } |
| 266 | |
| 267 | IConnectableLayer* Network::AddMergerLayer(const OriginsDescriptor& mergerDescriptor, |
| 268 | const char* name) |
| 269 | { |
| 270 | return m_Graph->AddLayer<MergerLayer>(mergerDescriptor, name); |
| 271 | } |
| 272 | |
| 273 | IConnectableLayer* Network::AddAdditionLayer(const char* name) |
| 274 | { |
| 275 | return m_Graph->AddLayer<AdditionLayer>(name); |
| 276 | } |
| 277 | |
| 278 | IConnectableLayer* Network::AddMultiplicationLayer(const char* name) |
| 279 | { |
| 280 | return m_Graph->AddLayer<MultiplicationLayer>(name); |
| 281 | } |
| 282 | |
| 283 | IConnectableLayer* Network::AddOutputLayer(LayerBindingId id, const char* name) |
| 284 | { |
| 285 | return m_Graph->AddLayer<OutputLayer>(id, name); |
| 286 | } |
| 287 | |
| 288 | IConnectableLayer* Network::AddBatchNormalizationLayer(const BatchNormalizationDescriptor& desc, |
| 289 | const ConstTensor& mean, |
| 290 | const ConstTensor& variance, |
| 291 | const ConstTensor& beta, |
| 292 | const ConstTensor& gamma, |
| 293 | const char* name) |
| 294 | { |
| 295 | const auto layer = m_Graph->AddLayer<BatchNormalizationLayer>(desc, name); |
| 296 | |
| 297 | layer->m_Mean = std::make_unique<ScopedCpuTensorHandle>(mean); |
| 298 | layer->m_Variance = std::make_unique<ScopedCpuTensorHandle>(variance); |
| 299 | layer->m_Beta = std::make_unique<ScopedCpuTensorHandle>(beta); |
| 300 | layer->m_Gamma = std::make_unique<ScopedCpuTensorHandle>(gamma); |
| 301 | |
| 302 | return layer; |
| 303 | } |
| 304 | |
| 305 | IConnectableLayer* Network::AddResizeBilinearLayer(const ResizeBilinearDescriptor& resizeDescriptor, const char* name) |
| 306 | { |
| 307 | return m_Graph->AddLayer<ResizeBilinearLayer>(resizeDescriptor,name); |
| 308 | } |
| 309 | |
| 310 | IConnectableLayer* Network::AddL2NormalizationLayer(const char* name) |
| 311 | { |
| 312 | return m_Graph->AddLayer<L2NormalizationLayer>(name); |
| 313 | } |
| 314 | |
| 315 | IConnectableLayer* Network::AddConstantLayer(const ConstTensor& input, const char* name) |
| 316 | { |
| 317 | return m_Graph->AddLayer<ConstantLayer>(std::make_shared<ScopedCpuTensorHandle>(input), name); |
| 318 | } |
| 319 | |
| 320 | IConnectableLayer* Network::AddReshapeLayer(const ReshapeDescriptor& reshapeDescriptor, const char* name) |
| 321 | { |
| 322 | return m_Graph->AddLayer<ReshapeLayer>(reshapeDescriptor, name); |
| 323 | } |
| 324 | |
| 325 | IConnectableLayer* Network::AddFloorLayer(const char* name) |
| 326 | { |
| 327 | return m_Graph->AddLayer<FloorLayer>(name); |
| 328 | } |
| 329 | |
| 330 | OptimizedNetwork::OptimizedNetwork(std::unique_ptr<Graph> graph) |
| 331 | : m_Graph(std::move(graph)) |
| 332 | { |
| 333 | } |
| 334 | |
| 335 | OptimizedNetwork::~OptimizedNetwork() |
| 336 | { |
| 337 | } |
| 338 | |
| 339 | } // namespace armnn |
| 340 | |