Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame^] | 1 | // |
| 2 | // Copyright © 2022 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "ArmnnDriverImpl.hpp" |
| 7 | #include "ArmnnPreparedModel.hpp" |
| 8 | #include "CacheDataHandler.hpp" |
| 9 | #include "ModelToINetworkTransformer.hpp" |
| 10 | #include "SystemPropertiesUtils.hpp" |
| 11 | |
| 12 | #include <armnnDeserializer/IDeserializer.hpp> |
| 13 | |
| 14 | #include <log/log.h> |
| 15 | #include <sys/stat.h> |
| 16 | |
| 17 | namespace |
| 18 | { |
| 19 | |
| 20 | Capabilities GenerateCapabilities() |
| 21 | { |
| 22 | VLOG(DRIVER) << "ArmnnDriverImpl::GenerateCapabilities()"; |
| 23 | |
| 24 | float defaultPerfValue = .1f; |
| 25 | const Capabilities::PerformanceInfo defaultPerfInfo = { /* execTime */ defaultPerfValue, |
| 26 | /* powerUsage */ defaultPerfValue |
| 27 | }; |
| 28 | std::vector<OperandType> operandsTypes({ |
| 29 | OperandType::FLOAT32, |
| 30 | OperandType::INT32, |
| 31 | OperandType::UINT32, |
| 32 | OperandType::TENSOR_FLOAT32, |
| 33 | OperandType::TENSOR_INT32, |
| 34 | OperandType::TENSOR_QUANT8_ASYMM, |
| 35 | OperandType::BOOL, |
| 36 | OperandType::TENSOR_QUANT16_SYMM, |
| 37 | OperandType::TENSOR_FLOAT16, |
| 38 | OperandType::TENSOR_BOOL8, |
| 39 | OperandType::FLOAT16, |
| 40 | OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL, |
| 41 | OperandType::TENSOR_QUANT16_ASYMM, |
| 42 | OperandType::TENSOR_QUANT8_SYMM, |
| 43 | OperandType::TENSOR_QUANT8_ASYMM_SIGNED, |
| 44 | }); |
| 45 | |
| 46 | std::vector<Capabilities::OperandPerformance> operandPerformances; |
| 47 | operandPerformances.reserve(operandsTypes.size()); |
| 48 | |
| 49 | for (auto opType : operandsTypes) |
| 50 | { |
| 51 | operandPerformances.push_back( |
| 52 | Capabilities::OperandPerformance{ /* type */ opType, /* info */ defaultPerfInfo }); |
| 53 | } |
| 54 | |
| 55 | auto operandPerformanceTable = |
| 56 | Capabilities::OperandPerformanceTable::create(std::move(operandPerformances)).value(); |
| 57 | |
| 58 | return { /* relaxedFloat32toFloat16PerformanceScalar */ defaultPerfInfo, |
| 59 | /* relaxedFloat32toFloat16PerformanceTensor */ defaultPerfInfo, |
| 60 | /* operandPerformance */ std::move(operandPerformanceTable), |
| 61 | /* ifPerformance */ defaultPerfInfo, |
| 62 | /* whilePerformance */ defaultPerfInfo }; |
| 63 | } |
| 64 | |
| 65 | } // anonymous namespace |
| 66 | |
| 67 | using namespace android::nn; |
| 68 | |
| 69 | namespace armnn_driver |
| 70 | { |
| 71 | |
| 72 | bool ArmnnDriverImpl::ValidateSharedHandle(const SharedHandle& sharedHandle) |
| 73 | { |
| 74 | bool valid = true; |
| 75 | |
| 76 | if (*sharedHandle < 0) |
| 77 | { |
| 78 | return !valid; |
| 79 | } |
| 80 | |
| 81 | int dataCacheFileAccessMode = fcntl(*sharedHandle, F_GETFL) & O_ACCMODE; |
| 82 | if (dataCacheFileAccessMode != O_RDWR) |
| 83 | { |
| 84 | return !valid; |
| 85 | } |
| 86 | |
| 87 | return valid; |
| 88 | } |
| 89 | |
| 90 | bool ArmnnDriverImpl::ValidateDataCacheHandle(const std::vector<SharedHandle>& dataCacheHandle, const size_t dataSize) |
| 91 | { |
| 92 | bool valid = true; |
| 93 | // DataCacheHandle size should always be 1 for ArmNN model |
| 94 | if (dataCacheHandle.size() != 1) |
| 95 | { |
| 96 | return !valid; |
| 97 | } |
| 98 | |
| 99 | if (dataSize == 0) |
| 100 | { |
| 101 | return !valid; |
| 102 | } |
| 103 | |
| 104 | struct stat statBuffer; |
| 105 | if (fstat(*dataCacheHandle[0], &statBuffer) == 0) |
| 106 | { |
| 107 | unsigned long bufferSize = statBuffer.st_size; |
| 108 | if (bufferSize != dataSize) |
| 109 | { |
| 110 | return !valid; |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | return ValidateSharedHandle(dataCacheHandle[0]); |
| 115 | } |
| 116 | |
| 117 | std::vector<armnn::NetworkId>& ArmnnDriverImpl::GetLoadedNetworks() |
| 118 | { |
| 119 | return m_NetworkIDs; |
| 120 | } |
| 121 | |
| 122 | GeneralResult<SharedPreparedModel> ArmnnDriverImpl::PrepareArmnnModel( |
| 123 | const armnn::IRuntimePtr& runtime, |
| 124 | const armnn::IGpuAccTunedParametersPtr& clTunedParameters, |
| 125 | const DriverOptions& options, |
| 126 | const Model& model, |
| 127 | const std::vector<SharedHandle>& modelCacheHandle, |
| 128 | const std::vector<SharedHandle>& dataCacheHandle, |
| 129 | const CacheToken& token, |
| 130 | bool float32ToFloat16, |
| 131 | Priority priority) |
| 132 | { |
| 133 | VLOG(DRIVER) << "ArmnnDriverImpl::PrepareArmnnModel()"; |
| 134 | |
| 135 | if (!runtime) |
| 136 | { |
| 137 | return NN_ERROR(ErrorStatus::DEVICE_UNAVAILABLE) << "Device unavailable"; |
| 138 | } |
| 139 | |
| 140 | if (const auto result = validate(model); !result.ok()) |
| 141 | { |
| 142 | return NN_ERROR(ErrorStatus::INVALID_ARGUMENT) << "Invalid model passed as input"; |
| 143 | } |
| 144 | |
| 145 | // Deliberately ignore any unsupported operations requested by the options - |
| 146 | // at this point we're being asked to prepare a model that we've already declared support for |
| 147 | // and the operation indices may be different to those in getSupportedOperations anyway. |
| 148 | std::set<unsigned int> unsupportedOperations; |
| 149 | ModelToINetworkTransformer modelConverter(options.GetBackends(), |
| 150 | model, |
| 151 | unsupportedOperations); |
| 152 | |
| 153 | if (modelConverter.GetConversionResult() != ConversionResult::Success) |
| 154 | { |
| 155 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) << "ModelToINetworkConverter failed"; |
| 156 | } |
| 157 | |
| 158 | // Serialize the network graph to a .armnn file if an output directory |
| 159 | // has been specified in the drivers' arguments. |
| 160 | std::vector<uint8_t> dataCacheData; |
| 161 | bool serializeToFile = dataCacheHandle.size() < 1 ? false : true; |
| 162 | auto serializedNetworkFileName = |
| 163 | SerializeNetwork(*modelConverter.GetINetwork(), |
| 164 | options.GetRequestInputsAndOutputsDumpDir(), |
| 165 | dataCacheData, |
| 166 | serializeToFile); |
| 167 | |
| 168 | // Optimize the network |
| 169 | armnn::IOptimizedNetworkPtr optNet(nullptr, nullptr); |
| 170 | armnn::OptimizerOptions OptOptions; |
| 171 | OptOptions.m_ReduceFp32ToFp16 = float32ToFloat16; |
| 172 | OptOptions.m_ProfilingEnabled = options.IsGpuProfilingEnabled(); |
| 173 | |
| 174 | int cachedFd = -1; |
| 175 | bool saveCachedNetwork = options.SaveCachedNetwork(); |
| 176 | |
| 177 | unsigned int numberOfCachedModelFiles = 0; |
| 178 | if (modelCacheHandle.size() > 0) |
| 179 | { |
| 180 | unsigned int index = 0; |
| 181 | for (auto& backend : options.GetBackends()) |
| 182 | { |
| 183 | // modelCacheHandle size should be equal to numberOfCachedModelFiles |
| 184 | // modelCacheHandle vector should be in same order as backends |
| 185 | auto numberOfCacheFiles = GetNumberOfCacheFiles(backend); |
| 186 | if (numberOfCacheFiles > 0) |
| 187 | { |
| 188 | numberOfCachedModelFiles += numberOfCacheFiles; |
| 189 | // For GpuAcc numberOfCachedFiles is 1 |
| 190 | if (backend == armnn::Compute::GpuAcc) |
| 191 | { |
| 192 | cachedFd = *modelCacheHandle[index]; |
| 193 | saveCachedNetwork = true; |
| 194 | } |
| 195 | index += numberOfCachedModelFiles; |
| 196 | } |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | armnn::BackendOptions gpuAcc("GpuAcc", |
| 201 | { |
| 202 | { "FastMathEnabled", options.IsFastMathEnabled() }, |
| 203 | { "SaveCachedNetwork", saveCachedNetwork }, |
| 204 | { "CachedNetworkFilePath", options.GetCachedNetworkFilePath() }, |
| 205 | { "MLGOTuningFilePath", options.GetClMLGOTunedParametersFile() }, |
| 206 | { "CachedFileDescriptor", cachedFd } |
| 207 | }); |
| 208 | |
| 209 | armnn::BackendOptions cpuAcc("CpuAcc", |
| 210 | { |
| 211 | { "FastMathEnabled", options.IsFastMathEnabled() }, |
| 212 | { "NumberOfThreads", options.GetNumberOfThreads() } |
| 213 | }); |
| 214 | OptOptions.m_ModelOptions.push_back(gpuAcc); |
| 215 | OptOptions.m_ModelOptions.push_back(cpuAcc); |
| 216 | |
| 217 | std::vector<std::string> errMessages; |
| 218 | try |
| 219 | { |
| 220 | optNet = armnn::Optimize(*modelConverter.GetINetwork(), |
| 221 | options.GetBackends(), |
| 222 | runtime->GetDeviceSpec(), |
| 223 | OptOptions, |
| 224 | errMessages); |
| 225 | } |
| 226 | catch (std::exception& e) |
| 227 | { |
| 228 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) << e.what(); |
| 229 | } |
| 230 | |
| 231 | // Check that the optimized network is valid. |
| 232 | if (!optNet) |
| 233 | { |
| 234 | std::stringstream message; |
| 235 | message << "Invalid optimized network"; |
| 236 | for (const std::string& msg : errMessages) |
| 237 | { |
| 238 | message << "\n" << msg; |
| 239 | } |
| 240 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) << message.str(); |
| 241 | } |
| 242 | |
| 243 | // Export the optimized network graph to a dot file if an output dump directory |
| 244 | // has been specified in the drivers' arguments. |
| 245 | std::string dotGraphFileName = ExportNetworkGraphToDotFile(*optNet, |
| 246 | options.GetRequestInputsAndOutputsDumpDir()); |
| 247 | |
| 248 | // Load it into the runtime. |
| 249 | armnn::NetworkId netId = 0; |
| 250 | std::string msg; |
| 251 | armnn::INetworkProperties networkProperties(options.isAsyncModelExecutionEnabled(), |
| 252 | MemorySource::Undefined, |
| 253 | MemorySource::Undefined, |
| 254 | options.IsGpuProfilingEnabled()); |
| 255 | auto numInputs = getMainModel(model).inputIndexes.size(); |
| 256 | auto numOutputs = getMainModel(model).outputIndexes.size(); |
| 257 | try |
| 258 | { |
| 259 | if (runtime->LoadNetwork(netId, move(optNet), msg, networkProperties) != armnn::Status::Success) |
| 260 | { |
| 261 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) << "Network could not be loaded"; |
| 262 | } |
| 263 | } |
| 264 | catch (std::exception& e) |
| 265 | { |
| 266 | std::stringstream message; |
| 267 | message << "Exception (" << e.what()<< ") caught from LoadNetwork."; |
| 268 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) << message.str(); |
| 269 | } |
| 270 | |
| 271 | // Now that we have a networkId for the graph rename the exported files to use it |
| 272 | // so that we can associate the graph file and the input/output tensor exported files |
| 273 | RenameExportedFiles(serializedNetworkFileName, |
| 274 | dotGraphFileName, |
| 275 | options.GetRequestInputsAndOutputsDumpDir(), |
| 276 | netId); |
| 277 | |
| 278 | // Cache the model |
| 279 | size_t hashValue = 0; |
| 280 | if (dataCacheHandle.size() == 1 ) |
| 281 | { |
| 282 | write(*dataCacheHandle[0], dataCacheData.data(), dataCacheData.size()); |
| 283 | hashValue = CacheDataHandlerInstance().Hash(dataCacheData); |
| 284 | } |
| 285 | |
| 286 | // Cache the model data |
| 287 | if (modelCacheHandle.size() > 0) |
| 288 | { |
| 289 | if (modelCacheHandle.size() == numberOfCachedModelFiles) |
| 290 | { |
| 291 | for (uint32_t i = 0; i < modelCacheHandle.size(); ++i) |
| 292 | { |
| 293 | int modelCacheFileAccessMode = fcntl(*modelCacheHandle[i], F_GETFL) & O_ACCMODE; |
| 294 | if (modelCacheFileAccessMode != O_RDONLY) |
| 295 | { |
| 296 | struct stat statBuffer; |
| 297 | if (fstat(*modelCacheHandle[i], &statBuffer) == 0) |
| 298 | { |
| 299 | long modelDataSize = statBuffer.st_size; |
| 300 | if (modelDataSize > 0) |
| 301 | { |
| 302 | std::vector<uint8_t> modelData(modelDataSize); |
| 303 | pread(*modelCacheHandle[i], modelData.data(), modelData.size(), 0); |
| 304 | hashValue ^= CacheDataHandlerInstance().Hash(modelData); |
| 305 | } |
| 306 | } |
| 307 | } |
| 308 | } |
| 309 | } |
| 310 | } |
| 311 | if (hashValue != 0) |
| 312 | { |
| 313 | CacheDataHandlerInstance().Register(token, hashValue, dataCacheData.size()); |
| 314 | } |
| 315 | |
| 316 | bool executeWithDummyInputs = (std::find(options.GetBackends().begin(), |
| 317 | options.GetBackends().end(), |
| 318 | armnn::Compute::GpuAcc) != options.GetBackends().end()); |
| 319 | |
| 320 | m_NetworkIDs.push_back(netId); |
| 321 | auto preparedModel = std::make_shared<const ArmnnPreparedModel>(netId, |
| 322 | runtime.get(), |
| 323 | model, |
| 324 | options.GetRequestInputsAndOutputsDumpDir(), |
| 325 | options.IsGpuProfilingEnabled(), |
| 326 | priority); |
| 327 | |
| 328 | // Run a single 'dummy' inference of the model. This means that CL kernels will get compiled (and tuned if |
| 329 | // this is enabled) before the first 'real' inference which removes the overhead of the first inference. |
| 330 | // Only run this if the GpuAcc backend has been added to options |
| 331 | if (std::find(options.GetBackends().begin(), |
| 332 | options.GetBackends().end(), |
| 333 | armnn::Compute::GpuAcc) != options.GetBackends().end()) |
| 334 | { |
| 335 | if (!preparedModel->ExecuteWithDummyInputs(numInputs, numOutputs)) |
| 336 | { |
| 337 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) << "Network could not be executed"; |
| 338 | } |
| 339 | |
| 340 | if (clTunedParameters && |
| 341 | options.GetClTunedParametersMode() == armnn::IGpuAccTunedParameters::Mode::UpdateTunedParameters) |
| 342 | { |
| 343 | // Now that we've done one inference the CL kernel parameters will have been tuned, |
| 344 | // so save the updated file. |
| 345 | try |
| 346 | { |
| 347 | clTunedParameters->Save(options.GetClTunedParametersFile().c_str()); |
| 348 | } |
| 349 | catch (std::exception& error) |
| 350 | { |
| 351 | VLOG(DRIVER) << "ArmnnDriverImpl::prepareModel: Failed to save CL tuned parameters file" |
| 352 | << options.GetClTunedParametersFile().c_str() << error.what(); |
| 353 | } |
| 354 | } |
| 355 | } |
| 356 | return std::move(preparedModel); |
| 357 | } |
| 358 | |
| 359 | std::vector<armnn::NetworkId> ArmnnDriverImpl::m_NetworkIDs = {}; |
| 360 | |
| 361 | GeneralResult<SharedPreparedModel> ArmnnDriverImpl::PrepareArmnnModelFromCache( |
| 362 | const armnn::IRuntimePtr& runtime, |
| 363 | const armnn::IGpuAccTunedParametersPtr& clTunedParameters, |
| 364 | const DriverOptions& options, |
| 365 | const std::vector<SharedHandle>& modelCacheHandle, |
| 366 | const std::vector<SharedHandle>& dataCacheHandle, |
| 367 | const CacheToken& token, |
| 368 | bool float32ToFloat16) |
| 369 | { |
| 370 | VLOG(DRIVER) << "ArmnnDriverImpl::PrepareArmnnModelFromCache()"; |
| 371 | |
| 372 | if (!runtime) |
| 373 | { |
| 374 | return NN_ERROR(ErrorStatus::DEVICE_UNAVAILABLE) |
| 375 | << "ArmnnDriverImpl::prepareModelFromCache(): Device unavailable"; |
| 376 | } |
| 377 | |
| 378 | if (token.size() != ANEURALNETWORKS_BYTE_SIZE_OF_CACHE_TOKEN) |
| 379 | { |
| 380 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) |
| 381 | << "ArmnnDriverImpl::prepareModelFromCache(): Token size does not match!"; |
| 382 | } |
| 383 | |
| 384 | // Validate dataCacheHandle |
| 385 | auto dataSize = CacheDataHandlerInstance().GetCacheSize(token); |
| 386 | if (!ValidateDataCacheHandle(dataCacheHandle, dataSize)) |
| 387 | { |
| 388 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) |
| 389 | << "ArmnnDriverImpl::prepareModelFromCache(): Not valid data cache handle!"; |
| 390 | } |
| 391 | |
| 392 | // Check if model files cached they match the expected value |
| 393 | unsigned int numberOfCachedModelFiles = 0; |
| 394 | for (auto& backend : options.GetBackends()) |
| 395 | { |
| 396 | numberOfCachedModelFiles += GetNumberOfCacheFiles(backend); |
| 397 | } |
| 398 | if (modelCacheHandle.size() != numberOfCachedModelFiles) |
| 399 | { |
| 400 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) |
| 401 | << "ArmnnDriverImpl::prepareModelFromCache(): Model cache handle size does not match."; |
| 402 | } |
| 403 | |
| 404 | // Read the model |
| 405 | std::vector<uint8_t> dataCacheData(dataSize); |
| 406 | pread(*dataCacheHandle[0], dataCacheData.data(), dataCacheData.size(), 0); |
| 407 | auto hashValue = CacheDataHandlerInstance().Hash(dataCacheData); |
| 408 | |
| 409 | int gpuAccCachedFd = -1; |
| 410 | if (modelCacheHandle.size() > 0) |
| 411 | { |
| 412 | unsigned int index = 0; |
| 413 | for (auto& backend : options.GetBackends()) |
| 414 | { |
| 415 | // modelCacheHandle size should be equal to numberOfCachedModelFiles |
| 416 | // modelCacheHandle vector should be in same order as backends |
| 417 | auto numberOfCacheFiles = GetNumberOfCacheFiles(backend); |
| 418 | if (numberOfCacheFiles > 0) |
| 419 | { |
| 420 | if (!ValidateSharedHandle(modelCacheHandle[index])) |
| 421 | { |
| 422 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) |
| 423 | << "ArmnnDriverImpl::prepareModelFromCache(): Invalid model cache handle!"; |
| 424 | } |
| 425 | int cachedFd = *modelCacheHandle[index]; |
| 426 | struct stat statBuffer; |
| 427 | if (fstat(cachedFd, &statBuffer) == 0) |
| 428 | { |
| 429 | long modelDataSize = statBuffer.st_size; |
| 430 | if (modelDataSize > 0) |
| 431 | { |
| 432 | std::vector<uint8_t> modelData(modelDataSize); |
| 433 | pread(cachedFd, modelData.data(), modelData.size(), 0); |
| 434 | hashValue ^= CacheDataHandlerInstance().Hash(modelData); |
| 435 | |
| 436 | if (backend == armnn::Compute::GpuAcc) |
| 437 | { |
| 438 | gpuAccCachedFd = cachedFd; |
| 439 | } |
| 440 | } |
| 441 | } |
| 442 | index += numberOfCacheFiles; |
| 443 | } |
| 444 | } |
| 445 | } |
| 446 | |
| 447 | if (!CacheDataHandlerInstance().Validate(token, hashValue, dataCacheData.size())) |
| 448 | { |
| 449 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) |
| 450 | << "ArmnnDriverImpl::prepareModelFromCache(): ValidateHash() failed!"; |
| 451 | } |
| 452 | |
| 453 | // Deserialize the network.. |
| 454 | armnn::INetworkPtr network = armnn::INetworkPtr(nullptr, [](armnn::INetwork*){}); |
| 455 | try |
| 456 | { |
| 457 | network = armnnDeserializer::IDeserializer::Create()->CreateNetworkFromBinary(dataCacheData); |
| 458 | } |
| 459 | catch (std::exception&) |
| 460 | { |
| 461 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) |
| 462 | << "ArmnnDriverImpl::prepareModelFromCache(): Exception caught from Deserializer!"; |
| 463 | } |
| 464 | |
| 465 | // Optimize the network |
| 466 | armnn::IOptimizedNetworkPtr optNet(nullptr, nullptr); |
| 467 | armnn::OptimizerOptions OptOptions; |
| 468 | OptOptions.m_ReduceFp32ToFp16 = float32ToFloat16; |
| 469 | OptOptions.m_ProfilingEnabled = options.IsGpuProfilingEnabled(); |
| 470 | |
| 471 | armnn::BackendOptions gpuAcc("GpuAcc", |
| 472 | { |
| 473 | { "FastMathEnabled", options.IsFastMathEnabled() }, |
| 474 | { "SaveCachedNetwork", false }, |
| 475 | { "CachedNetworkFilePath", options.GetCachedNetworkFilePath() }, |
| 476 | { "MLGOTuningFilePath", options.GetClMLGOTunedParametersFile() }, |
| 477 | { "CachedFileDescriptor", gpuAccCachedFd } |
| 478 | }); |
| 479 | |
| 480 | armnn::BackendOptions cpuAcc("CpuAcc", |
| 481 | { |
| 482 | { "FastMathEnabled", options.IsFastMathEnabled() }, |
| 483 | { "NumberOfThreads", options.GetNumberOfThreads() } |
| 484 | }); |
| 485 | OptOptions.m_ModelOptions.push_back(gpuAcc); |
| 486 | OptOptions.m_ModelOptions.push_back(cpuAcc); |
| 487 | |
| 488 | std::vector<std::string> errMessages; |
| 489 | try |
| 490 | { |
| 491 | optNet = armnn::Optimize(*network.get(), |
| 492 | options.GetBackends(), |
| 493 | runtime->GetDeviceSpec(), |
| 494 | OptOptions, |
| 495 | errMessages); |
| 496 | } |
| 497 | catch (std::exception& e) |
| 498 | { |
| 499 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) << e.what(); |
| 500 | } |
| 501 | |
| 502 | // Check that the optimized network is valid. |
| 503 | if (!optNet) |
| 504 | { |
| 505 | std::stringstream message; |
| 506 | message << "Invalid optimized network"; |
| 507 | for (const std::string& msg : errMessages) |
| 508 | { |
| 509 | message << "\n" << msg; |
| 510 | } |
| 511 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) << message.str(); |
| 512 | } |
| 513 | |
| 514 | // Export the optimized network graph to a dot file if an output dump directory |
| 515 | // has been specified in the drivers' arguments. |
| 516 | std::string dotGraphFileName = ExportNetworkGraphToDotFile(*optNet, |
| 517 | options.GetRequestInputsAndOutputsDumpDir()); |
| 518 | |
| 519 | // Load it into the runtime. |
| 520 | armnn::NetworkId netId = 0; |
| 521 | std::string msg; |
| 522 | armnn::INetworkProperties networkProperties(options.isAsyncModelExecutionEnabled(), |
| 523 | MemorySource::Undefined, |
| 524 | MemorySource::Undefined, |
| 525 | options.IsGpuProfilingEnabled()); |
| 526 | try |
| 527 | { |
| 528 | if (runtime->LoadNetwork(netId, move(optNet), msg, networkProperties) != armnn::Status::Success) |
| 529 | { |
| 530 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) << "Network could not be loaded"; |
| 531 | } |
| 532 | } |
| 533 | catch (std::exception& e) |
| 534 | { |
| 535 | std::stringstream message; |
| 536 | message << "Exception (" << e.what()<< ") caught from LoadNetwork."; |
| 537 | return NN_ERROR(ErrorStatus::GENERAL_FAILURE) << message.str(); |
| 538 | } |
| 539 | |
| 540 | m_NetworkIDs.push_back(netId); |
| 541 | return std::make_shared<const ArmnnPreparedModel>(netId, |
| 542 | runtime.get(), |
| 543 | options.GetRequestInputsAndOutputsDumpDir(), |
| 544 | options.IsGpuProfilingEnabled(), |
| 545 | Priority::MEDIUM, |
| 546 | true); |
| 547 | } |
| 548 | |
| 549 | const Capabilities& ArmnnDriverImpl::GetCapabilities(const armnn::IRuntimePtr& runtime) |
| 550 | { |
| 551 | VLOG(DRIVER) << "ArmnnDriverImpl::GetCapabilities()"; |
| 552 | static const Capabilities theCapabilities = GenerateCapabilities(); |
| 553 | return theCapabilities; |
| 554 | } |
| 555 | |
| 556 | void ArmnnDriverImpl::ClearNetworks() |
| 557 | { |
| 558 | m_NetworkIDs.clear(); |
| 559 | } |
| 560 | |
| 561 | } // namespace armnn_driver |