Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2020 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "ArmnnDriverImpl.hpp" |
| 7 | #include "../ArmnnPreparedModel_1_3.hpp" |
| 8 | #include "../ModelToINetworkConverter.hpp" |
| 9 | #include "../SystemPropertiesUtils.hpp" |
| 10 | |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 11 | #include <armnnDeserializer/IDeserializer.hpp> |
| 12 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 13 | #include <log/log.h> |
| 14 | |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 15 | #include <sys/stat.h> |
| 16 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 17 | namespace |
| 18 | { |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 19 | const char *g_RelaxedFloat32toFloat16PerformanceExecTime = "ArmNN.relaxedFloat32toFloat16Performance.execTime"; |
| 20 | const char *g_RelaxedFloat32toFloat16PerformancePowerUsage = "ArmNN.relaxedFloat32toFloat16Performance.powerUsage"; |
| 21 | |
Kevin May | 2eaa119 | 2020-04-15 16:50:57 +0100 | [diff] [blame] | 22 | const char *g_ifPerformanceExecTime = "ArmNN.ifPerformance.execTime"; |
| 23 | const char *g_ifPerformancePowerUsage = "ArmNN.ifPerformance.powerUsage"; |
| 24 | |
| 25 | const char *g_whilePerformanceExecTime = "ArmNN.whilePerformance.execTime"; |
| 26 | const char *g_whilePerformancePowerUsage = "ArmNN.whilePerformance.powerUsage"; |
| 27 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 28 | const char *g_OperandTypeTensorFloat32PerformanceExecTime = "Armnn.operandTypeTensorFloat32Performance.execTime"; |
| 29 | const char *g_OperandTypeTensorFloat32PerformancePowerUsage = "Armnn.operandTypeTensorFloat32Performance.powerUsage"; |
| 30 | |
| 31 | const char *g_OperandTypeFloat32PerformanceExecTime = "Armnn.operandTypeFloat32Performance.execTime"; |
| 32 | const char *g_OperandTypeFloat32PerformancePowerUsage = "Armnn.operandTypeFloat32Performance.powerUsage"; |
| 33 | |
| 34 | const char *g_OperandTypeTensorFloat16PerformanceExecTime = "Armnn.operandTypeTensorFloat16Performance.execTime"; |
| 35 | const char *g_OperandTypeTensorFloat16PerformancePowerUsage = "Armnn.operandTypeTensorFloat16Performance.powerUsage"; |
| 36 | |
| 37 | const char *g_OperandTypeFloat16PerformanceExecTime = "Armnn.operandTypeFloat16Performance.execTime"; |
| 38 | const char *g_OperandTypeFloat16PerformancePowerUsage = "Armnn.operandTypeFloat16Performance.powerUsage"; |
| 39 | |
| 40 | const char *g_OperandTypeTensorQuant8AsymmPerformanceExecTime = |
| 41 | "Armnn.operandTypeTensorQuant8AsymmPerformance.execTime"; |
| 42 | const char *g_OperandTypeTensorQuant8AsymmPerformancePowerUsage = |
| 43 | "Armnn.operandTypeTensorQuant8AsymmPerformance.powerUsage"; |
| 44 | |
| 45 | const char *g_OperandTypeTensorQuant8AsymmSignedPerformanceExecTime = |
| 46 | "Armnn.operandTypeTensorQuant8AsymmSignedPerformance.execTime"; |
| 47 | const char *g_OperandTypeTensorQuant8AsymmSignedPerformancePowerUsage = |
| 48 | "Armnn.operandTypeTensorQuant8AsymmSignedPerformance.powerUsage"; |
| 49 | |
| 50 | const char *g_OperandTypeTensorQuant16SymmPerformanceExecTime = |
| 51 | "Armnn.operandTypeTensorQuant16SymmPerformance.execTime"; |
| 52 | const char *g_OperandTypeTensorQuant16SymmPerformancePowerUsage = |
| 53 | "Armnn.operandTypeTensorQuant16SymmPerformance.powerUsage"; |
| 54 | |
| 55 | const char *g_OperandTypeTensorQuant8SymmPerformanceExecTime = |
| 56 | "Armnn.operandTypeTensorQuant8SymmPerformance.execTime"; |
| 57 | const char *g_OperandTypeTensorQuant8SymmPerformancePowerUsage = |
| 58 | "Armnn.operandTypeTensorQuant8SymmPerformance.powerUsage"; |
| 59 | |
| 60 | const char *g_OperandTypeTensorQuant8SymmPerChannelPerformanceExecTime = |
| 61 | "Armnn.operandTypeTensorQuant8SymmPerChannelPerformance.execTime"; |
| 62 | const char *g_OperandTypeTensorQuant8SymmPerChannelPerformancePowerUsage = |
| 63 | "Armnn.operandTypeTensorQuant8SymmPerChannelPerformance.powerUsage"; |
| 64 | |
| 65 | |
| 66 | const char *g_OperandTypeTensorInt32PerformanceExecTime = "Armnn.operandTypeTensorInt32Performance.execTime"; |
| 67 | const char *g_OperandTypeTensorInt32PerformancePowerUsage = "Armnn.operandTypeTensorInt32Performance.powerUsage"; |
| 68 | |
| 69 | const char *g_OperandTypeInt32PerformanceExecTime = "Armnn.operandTypeInt32Performance.execTime"; |
| 70 | const char *g_OperandTypeInt32PerformancePowerUsage = "Armnn.operandTypeInt32Performance.powerUsage"; |
| 71 | |
| 72 | |
Sadik Armagan | 188675f | 2021-02-12 17:16:42 +0000 | [diff] [blame] | 73 | void NotifyCallbackAndCheck(const android::sp<V1_3::IPreparedModelCallback>& callback, |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 74 | V1_3::ErrorStatus errorStatus, |
Sadik Armagan | 188675f | 2021-02-12 17:16:42 +0000 | [diff] [blame] | 75 | const android::sp<V1_3::IPreparedModel>& preparedModelPtr) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 76 | { |
| 77 | Return<void> returned = callback->notify_1_3(errorStatus, preparedModelPtr); |
| 78 | // This check is required, if the callback fails and it isn't checked it will bring down the service |
| 79 | if (!returned.isOk()) |
| 80 | { |
| 81 | ALOGE("ArmnnDriverImpl::prepareModel: hidl callback failed to return properly: %s ", |
| 82 | returned.description().c_str()); |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | Return<V1_3::ErrorStatus> FailPrepareModel(V1_3::ErrorStatus error, |
| 87 | const std::string& message, |
Sadik Armagan | 188675f | 2021-02-12 17:16:42 +0000 | [diff] [blame] | 88 | const android::sp<V1_3::IPreparedModelCallback>& callback) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 89 | { |
| 90 | ALOGW("ArmnnDriverImpl::prepareModel: %s", message.c_str()); |
| 91 | NotifyCallbackAndCheck(callback, error, nullptr); |
| 92 | return error; |
| 93 | } |
| 94 | |
| 95 | } // anonymous namespace |
| 96 | |
| 97 | namespace armnn_driver |
| 98 | { |
| 99 | namespace hal_1_3 |
| 100 | { |
| 101 | |
| 102 | Return<V1_3::ErrorStatus> ArmnnDriverImpl::prepareArmnnModel_1_3( |
| 103 | const armnn::IRuntimePtr& runtime, |
| 104 | const armnn::IGpuAccTunedParametersPtr& clTunedParameters, |
| 105 | const DriverOptions& options, |
| 106 | const V1_3::Model& model, |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 107 | const android::hardware::hidl_vec<android::hardware::hidl_handle>& modelCacheHandle, |
| 108 | const android::hardware::hidl_vec<android::hardware::hidl_handle>& dataCacheHandle, |
| 109 | const HidlToken& token, |
Sadik Armagan | 188675f | 2021-02-12 17:16:42 +0000 | [diff] [blame] | 110 | const android::sp<V1_3::IPreparedModelCallback>& cb, |
Narumol Prangnawarat | cad4e91 | 2020-06-02 12:07:43 +0100 | [diff] [blame] | 111 | bool float32ToFloat16, |
| 112 | V1_3::Priority priority) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 113 | { |
| 114 | ALOGV("ArmnnDriverImpl::prepareArmnnModel_1_3()"); |
| 115 | |
| 116 | if (cb.get() == nullptr) |
| 117 | { |
| 118 | ALOGW("ArmnnDriverImpl::prepareModel: Invalid callback passed to prepareModel"); |
| 119 | return V1_3::ErrorStatus::INVALID_ARGUMENT; |
| 120 | } |
| 121 | |
| 122 | if (!runtime) |
| 123 | { |
| 124 | return FailPrepareModel(V1_3::ErrorStatus::DEVICE_UNAVAILABLE, "Device unavailable", cb); |
| 125 | } |
| 126 | |
| 127 | if (!android::nn::validateModel(model)) |
| 128 | { |
| 129 | return FailPrepareModel(V1_3::ErrorStatus::INVALID_ARGUMENT, "Invalid model passed as input", cb); |
| 130 | } |
| 131 | |
| 132 | // Deliberately ignore any unsupported operations requested by the options - |
| 133 | // at this point we're being asked to prepare a model that we've already declared support for |
| 134 | // and the operation indices may be different to those in getSupportedOperations anyway. |
| 135 | std::set<unsigned int> unsupportedOperations; |
| 136 | ModelToINetworkConverter<HalPolicy> modelConverter(options.GetBackends(), |
| 137 | model, |
| 138 | unsupportedOperations); |
| 139 | |
| 140 | if (modelConverter.GetConversionResult() != ConversionResult::Success) |
| 141 | { |
| 142 | FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, "ModelToINetworkConverter failed", cb); |
| 143 | return V1_3::ErrorStatus::NONE; |
| 144 | } |
| 145 | |
Sadik Armagan | b302143 | 2021-01-13 15:56:51 +0000 | [diff] [blame] | 146 | // Serialize the network graph to a .armnn file if an output directory |
| 147 | // has been specified in the drivers' arguments. |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 148 | std::vector<uint8_t> dataCacheData; |
| 149 | bool serializeToFile = dataCacheHandle.size() < 1 ? false : true; |
Sadik Armagan | b302143 | 2021-01-13 15:56:51 +0000 | [diff] [blame] | 150 | auto serializedNetworkFileName = |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 151 | SerializeNetwork(*modelConverter.GetINetwork(), |
| 152 | options.GetRequestInputsAndOutputsDumpDir(), |
| 153 | dataCacheData, |
| 154 | serializeToFile); |
Sadik Armagan | b302143 | 2021-01-13 15:56:51 +0000 | [diff] [blame] | 155 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 156 | // Optimize the network |
| 157 | armnn::IOptimizedNetworkPtr optNet(nullptr, nullptr); |
| 158 | armnn::OptimizerOptions OptOptions; |
| 159 | OptOptions.m_ReduceFp32ToFp16 = float32ToFloat16; |
Kevin May | daf7dd0 | 2021-10-22 11:57:30 +0100 | [diff] [blame] | 160 | OptOptions.m_ProfilingEnabled = options.IsGpuProfilingEnabled(); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 161 | |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 162 | int cachedFd = -1; |
| 163 | bool saveCachedNetwork = options.SaveCachedNetwork(); |
| 164 | |
| 165 | unsigned int numberOfCachedModelFiles = 0; |
| 166 | if (modelCacheHandle.size() > 0) |
| 167 | { |
| 168 | unsigned int index = 0; |
| 169 | for (auto& backend : options.GetBackends()) |
| 170 | { |
| 171 | // modelCacheHandle size should be equal to numberOfCachedModelFiles |
| 172 | // modelCacheHandle vector should be in same order as backends |
| 173 | auto numberOfCacheFiles = GetNumberOfCacheFiles(backend); |
| 174 | if (numberOfCacheFiles > 0) |
| 175 | { |
| 176 | numberOfCachedModelFiles += numberOfCacheFiles; |
| 177 | if (modelCacheHandle[index]->numFds == 1) |
| 178 | { |
| 179 | // For GpuAcc numberOfCachedFiles is 1 |
| 180 | if (backend == armnn::Compute::GpuAcc) |
| 181 | { |
| 182 | cachedFd = modelCacheHandle[index]->data[0]; |
| 183 | saveCachedNetwork = true; |
| 184 | } |
| 185 | } |
| 186 | index += numberOfCachedModelFiles; |
| 187 | } |
| 188 | } |
| 189 | } |
| 190 | |
Mike Kelly | 7ed56dd | 2020-09-30 20:22:56 +0100 | [diff] [blame] | 191 | armnn::BackendOptions gpuAcc("GpuAcc", |
| 192 | { |
Sadik Armagan | f36e10b | 2021-01-11 16:34:01 +0000 | [diff] [blame] | 193 | { "FastMathEnabled", options.IsFastMathEnabled() }, |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 194 | { "SaveCachedNetwork", saveCachedNetwork }, |
Finn Williams | f5ca16c | 2021-02-12 14:26:23 +0000 | [diff] [blame] | 195 | { "CachedNetworkFilePath", options.GetCachedNetworkFilePath() }, |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 196 | { "MLGOTuningFilePath", options.GetClMLGOTunedParametersFile() }, |
| 197 | { "CachedFileDescriptor", cachedFd } |
Mike Kelly | 7ed56dd | 2020-09-30 20:22:56 +0100 | [diff] [blame] | 198 | }); |
Finn Williams | f5ca16c | 2021-02-12 14:26:23 +0000 | [diff] [blame] | 199 | |
Mike Kelly | 7ed56dd | 2020-09-30 20:22:56 +0100 | [diff] [blame] | 200 | armnn::BackendOptions cpuAcc("CpuAcc", |
| 201 | { |
Matthew Sloyan | cd639c9 | 2021-02-11 16:57:38 +0000 | [diff] [blame] | 202 | { "FastMathEnabled", options.IsFastMathEnabled() }, |
| 203 | { "NumberOfThreads", options.GetNumberOfThreads() } |
Mike Kelly | 7ed56dd | 2020-09-30 20:22:56 +0100 | [diff] [blame] | 204 | }); |
| 205 | OptOptions.m_ModelOptions.push_back(gpuAcc); |
| 206 | OptOptions.m_ModelOptions.push_back(cpuAcc); |
| 207 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 208 | std::vector<std::string> errMessages; |
| 209 | try |
| 210 | { |
| 211 | optNet = armnn::Optimize(*modelConverter.GetINetwork(), |
| 212 | options.GetBackends(), |
| 213 | runtime->GetDeviceSpec(), |
| 214 | OptOptions, |
| 215 | errMessages); |
| 216 | } |
| 217 | catch (std::exception& e) |
| 218 | { |
| 219 | std::stringstream message; |
| 220 | message << "Exception (" << e.what() << ") caught from optimize."; |
| 221 | FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, message.str(), cb); |
| 222 | return V1_3::ErrorStatus::NONE; |
| 223 | } |
| 224 | |
| 225 | // Check that the optimized network is valid. |
| 226 | if (!optNet) |
| 227 | { |
| 228 | std::stringstream message; |
| 229 | message << "Invalid optimized network"; |
| 230 | for (const std::string& msg : errMessages) |
| 231 | { |
| 232 | message << "\n" << msg; |
| 233 | } |
| 234 | FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, message.str(), cb); |
| 235 | return V1_3::ErrorStatus::NONE; |
| 236 | } |
| 237 | |
| 238 | // Export the optimized network graph to a dot file if an output dump directory |
| 239 | // has been specified in the drivers' arguments. |
| 240 | std::string dotGraphFileName = ExportNetworkGraphToDotFile(*optNet, |
| 241 | options.GetRequestInputsAndOutputsDumpDir()); |
| 242 | |
| 243 | // Load it into the runtime. |
| 244 | armnn::NetworkId netId = 0; |
Finn Williams | d8fb540 | 2021-05-19 20:52:00 +0100 | [diff] [blame] | 245 | std::string msg; |
| 246 | armnn::INetworkProperties networkProperties(options.isAsyncModelExecutionEnabled(), |
| 247 | MemorySource::Undefined, |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 248 | MemorySource::Undefined, |
| 249 | options.IsGpuProfilingEnabled()); |
| 250 | |
| 251 | auto numInputs = getMainModel(model).inputIndexes.size(); |
| 252 | auto numOutputs = getMainModel(model).outputIndexes.size(); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 253 | try |
| 254 | { |
Finn Williams | d8fb540 | 2021-05-19 20:52:00 +0100 | [diff] [blame] | 255 | if (runtime->LoadNetwork(netId, move(optNet), msg, networkProperties) != armnn::Status::Success) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 256 | { |
| 257 | return FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, "Network could not be loaded", cb); |
| 258 | } |
| 259 | } |
| 260 | catch (std::exception& e) |
| 261 | { |
| 262 | std::stringstream message; |
| 263 | message << "Exception (" << e.what()<< ") caught from LoadNetwork."; |
| 264 | FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, message.str(), cb); |
| 265 | return V1_3::ErrorStatus::NONE; |
| 266 | } |
| 267 | |
Sadik Armagan | b302143 | 2021-01-13 15:56:51 +0000 | [diff] [blame] | 268 | // Now that we have a networkId for the graph rename the exported files to use it |
| 269 | // so that we can associate the graph file and the input/output tensor exported files |
| 270 | RenameExportedFiles(serializedNetworkFileName, |
| 271 | dotGraphFileName, |
| 272 | options.GetRequestInputsAndOutputsDumpDir(), |
| 273 | netId); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 274 | |
| 275 | std::unique_ptr<ArmnnPreparedModel_1_3<hal_1_3::HalPolicy>> preparedModel( |
| 276 | new ArmnnPreparedModel_1_3<hal_1_3::HalPolicy>( |
| 277 | netId, |
| 278 | runtime.get(), |
| 279 | model, |
| 280 | options.GetRequestInputsAndOutputsDumpDir(), |
Narumol Prangnawarat | cad4e91 | 2020-06-02 12:07:43 +0100 | [diff] [blame] | 281 | options.IsGpuProfilingEnabled(), |
Finn Williams | d8fb540 | 2021-05-19 20:52:00 +0100 | [diff] [blame] | 282 | priority, |
Finn Williams | ca3a3e0 | 2021-06-11 15:04:02 +0100 | [diff] [blame] | 283 | options.isAsyncModelExecutionEnabled(), |
| 284 | options.getNoOfArmnnThreads())); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 285 | |
| 286 | // Run a single 'dummy' inference of the model. This means that CL kernels will get compiled (and tuned if |
| 287 | // this is enabled) before the first 'real' inference which removes the overhead of the first inference. |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 288 | // Only run this if the GpuAcc backend has been added to options |
| 289 | if (std::find(options.GetBackends().begin(), |
| 290 | options.GetBackends().end(), |
| 291 | armnn::Compute::GpuAcc) != options.GetBackends().end()) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 292 | { |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 293 | if (!preparedModel->ExecuteWithDummyInputs(numInputs, numOutputs)) |
| 294 | { |
| 295 | return FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, "Network could not be executed", cb); |
| 296 | } |
| 297 | |
| 298 | if (clTunedParameters && |
| 299 | options.GetClTunedParametersMode() == armnn::IGpuAccTunedParameters::Mode::UpdateTunedParameters) |
| 300 | { |
| 301 | // Now that we've done one inference the CL kernel parameters will have been tuned, |
| 302 | // so save the updated file. |
| 303 | try |
| 304 | { |
| 305 | clTunedParameters->Save(options.GetClTunedParametersFile().c_str()); |
| 306 | } |
| 307 | catch (std::exception& error) |
| 308 | { |
| 309 | ALOGE("ArmnnDriverImpl::prepareModel: Failed to save CL tuned parameters file '%s': %s", |
| 310 | options.GetClTunedParametersFile().c_str(), error.what()); |
| 311 | } |
| 312 | } |
| 313 | } |
| 314 | size_t hashValue = 0; |
| 315 | // Cache the model |
| 316 | if (dataCacheHandle.size() > 0) |
| 317 | { |
| 318 | // Cache the Arm NN model |
| 319 | if (dataCacheHandle.size() != 1) |
| 320 | { |
| 321 | NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); |
| 322 | return V1_3::ErrorStatus::NONE; |
| 323 | } |
| 324 | |
| 325 | if (dataCacheHandle[0]->numFds != 1) |
| 326 | { |
| 327 | ALOGW("ArmnnDriverImpl::prepareArmnnModel_1_3: Cannot cache the data, numFds != 1."); |
| 328 | NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); |
| 329 | return V1_3::ErrorStatus::NONE; |
| 330 | } |
Sadik Armagan | ee6818b | 2021-11-05 14:41:52 +0000 | [diff] [blame] | 331 | |
| 332 | if (dataCacheHandle[0]->data[0] < 0) |
| 333 | { |
| 334 | ALOGW("ArmnnDriverImpl::prepareArmnnModel_1_3: Cannot cache the data, fd < 0"); |
| 335 | NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); |
| 336 | return V1_3::ErrorStatus::NONE; |
| 337 | } |
| 338 | |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 339 | int dataCacheFileAccessMode = fcntl(dataCacheHandle[0]->data[0], F_GETFL) & O_ACCMODE; |
| 340 | if (dataCacheFileAccessMode != O_RDWR) |
| 341 | { |
| 342 | ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3(): Invalid Access Mode."); |
| 343 | NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); |
| 344 | return V1_3::ErrorStatus::NONE; |
| 345 | } |
| 346 | |
| 347 | write(dataCacheHandle[0]->data[0], dataCacheData.data(), dataCacheData.size()); |
| 348 | hashValue = CacheDataHandlerInstance().Hash(dataCacheData); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 349 | } |
| 350 | |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 351 | // Cache the model data |
| 352 | if (modelCacheHandle.size() > 0) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 353 | { |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 354 | if (modelCacheHandle.size() != numberOfCachedModelFiles) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 355 | { |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 356 | NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); |
| 357 | return V1_3::ErrorStatus::NONE; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 358 | } |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 359 | |
| 360 | for (uint32_t i = 0; i < modelCacheHandle.size(); ++i) |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 361 | { |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 362 | if (modelCacheHandle[i]->numFds == 1) |
| 363 | { |
| 364 | int modelCacheFileAccessMode = fcntl(modelCacheHandle[i]->data[0], F_GETFL) & O_ACCMODE; |
| 365 | if (modelCacheFileAccessMode != O_RDONLY) |
| 366 | { |
| 367 | struct stat statBuffer; |
| 368 | if (fstat(modelCacheHandle[i]->data[0], &statBuffer) == 0) |
| 369 | { |
| 370 | long modelDataSize = statBuffer.st_size; |
| 371 | if (modelDataSize > 0) |
| 372 | { |
| 373 | std::vector<uint8_t> modelData(modelDataSize); |
| 374 | pread(modelCacheHandle[i]->data[0], modelData.data(), modelData.size(), 0); |
| 375 | hashValue ^= CacheDataHandlerInstance().Hash(modelData); |
| 376 | } |
| 377 | } |
| 378 | } |
| 379 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 380 | } |
| 381 | } |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 382 | if (hashValue != 0) |
| 383 | { |
| 384 | CacheDataHandlerInstance().Register(token, hashValue, dataCacheData.size()); |
| 385 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 386 | |
| 387 | NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 388 | return V1_3::ErrorStatus::NONE; |
| 389 | } |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 390 | |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 391 | Return<V1_3::ErrorStatus> ArmnnDriverImpl::prepareModelFromCache_1_3( |
| 392 | const armnn::IRuntimePtr& runtime, |
| 393 | const DriverOptions& options, |
| 394 | const android::hardware::hidl_vec<android::hardware::hidl_handle>& modelCacheHandle, |
| 395 | const android::hardware::hidl_vec<android::hardware::hidl_handle>& dataCacheHandle, |
| 396 | const HidlToken& token, |
| 397 | const android::sp<V1_3::IPreparedModelCallback>& cb) |
| 398 | { |
| 399 | ALOGV("ArmnnDriverImpl::prepareModelFromCache_1_3()"); |
| 400 | |
| 401 | if (token.size() != ANEURALNETWORKS_BYTE_SIZE_OF_CACHE_TOKEN) |
| 402 | { |
| 403 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 404 | return V1_3::ErrorStatus::GENERAL_FAILURE; |
| 405 | } |
| 406 | |
| 407 | if (cb.get() == nullptr) |
| 408 | { |
| 409 | ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3: Invalid callback passed to prepareModelFromCache_1_3"); |
| 410 | return V1_3::ErrorStatus::INVALID_ARGUMENT; |
| 411 | } |
| 412 | |
| 413 | if (!runtime) |
| 414 | { |
| 415 | ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3: Device unavailable"); |
| 416 | return V1_3::ErrorStatus::DEVICE_UNAVAILABLE; |
| 417 | } |
| 418 | |
| 419 | // DataCacheHandle size should always be 1 |
| 420 | // Arm NN model |
| 421 | if (dataCacheHandle.size() != 1) |
| 422 | { |
| 423 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 424 | return V1_3::ErrorStatus::GENERAL_FAILURE; |
| 425 | } |
| 426 | |
| 427 | // Check if model files cached they match the expected value |
| 428 | unsigned int numberOfCachedModelFiles = 0; |
| 429 | for (auto& backend : options.GetBackends()) |
| 430 | { |
| 431 | numberOfCachedModelFiles += GetNumberOfCacheFiles(backend); |
| 432 | } |
| 433 | if (modelCacheHandle.size() != numberOfCachedModelFiles) |
| 434 | { |
| 435 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 436 | return V1_3::ErrorStatus::GENERAL_FAILURE; |
| 437 | } |
| 438 | |
| 439 | if (dataCacheHandle[0]->numFds != 1) |
| 440 | { |
| 441 | ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3(): Cannot read from the cache data, numFds != 1."); |
| 442 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 443 | return V1_3::ErrorStatus::GENERAL_FAILURE; |
| 444 | } |
| 445 | |
Sadik Armagan | ee6818b | 2021-11-05 14:41:52 +0000 | [diff] [blame] | 446 | if (dataCacheHandle[0]->data[0] < 0) |
| 447 | { |
| 448 | ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3(): Cannot read from the cache data, fd < 0"); |
| 449 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 450 | return V1_3::ErrorStatus::GENERAL_FAILURE; |
| 451 | } |
| 452 | |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 453 | int dataCacheFileAccessMode = fcntl(dataCacheHandle[0]->data[0], F_GETFL) & O_ACCMODE; |
| 454 | if (dataCacheFileAccessMode != O_RDWR) |
| 455 | { |
| 456 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 457 | return V1_3::ErrorStatus::GENERAL_FAILURE; |
| 458 | } |
| 459 | |
| 460 | auto dataSize = CacheDataHandlerInstance().GetCacheSize(token); |
| 461 | if (dataSize == 0) |
| 462 | { |
| 463 | ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3: Invalid data to deserialize!"); |
| 464 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 465 | return V1_3::ErrorStatus::GENERAL_FAILURE; |
| 466 | } |
| 467 | |
| 468 | int offset = 0; |
| 469 | { |
| 470 | struct stat statBuffer; |
| 471 | if (fstat(dataCacheHandle[0]->data[0], &statBuffer) == 0) |
| 472 | { |
| 473 | unsigned long bufferSize = statBuffer.st_size; |
Sadik Armagan | ee6818b | 2021-11-05 14:41:52 +0000 | [diff] [blame] | 474 | if (bufferSize != dataSize) |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 475 | { |
| 476 | ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3: Invalid data to deserialize!"); |
| 477 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 478 | return V1_3::ErrorStatus::GENERAL_FAILURE; |
| 479 | } |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 480 | } |
| 481 | } |
| 482 | std::vector<uint8_t> dataCacheData(dataSize); |
| 483 | pread(dataCacheHandle[0]->data[0], dataCacheData.data(), dataCacheData.size(), offset); |
| 484 | auto hashValue = CacheDataHandlerInstance().Hash(dataCacheData); |
| 485 | |
| 486 | int gpuAccCachedFd = -1; |
| 487 | bool saveCachedNetwork = false; |
| 488 | if (modelCacheHandle.size() > 0) |
| 489 | { |
| 490 | unsigned int index = 0; |
| 491 | for (auto& backend : options.GetBackends()) |
| 492 | { |
| 493 | // modelCacheHandle size should be equal to numberOfCachedModelFiles |
| 494 | // modelCacheHandle vector should be in same order as backends |
| 495 | auto numberOfCacheFiles = GetNumberOfCacheFiles(backend); |
| 496 | if (numberOfCacheFiles > 0) |
| 497 | { |
| 498 | if (modelCacheHandle[index]->numFds != 1) |
| 499 | { |
| 500 | ALOGW( |
| 501 | "ArmnnDriverImpl::prepareModelFromCache_1_3(): Cannot read from the model cache, numFds != 1."); |
| 502 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 503 | return V1_3::ErrorStatus::GENERAL_FAILURE; |
| 504 | } |
| 505 | auto cachedFd = modelCacheHandle[index]->data[0]; |
| 506 | |
| 507 | int modelCacheFileAccessMode = fcntl(cachedFd, F_GETFL) & O_ACCMODE; |
| 508 | if (modelCacheFileAccessMode != O_RDWR) |
| 509 | { |
| 510 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 511 | return V1_3::ErrorStatus::GENERAL_FAILURE; |
| 512 | } |
| 513 | |
| 514 | struct stat statBuffer; |
| 515 | if (cachedFd != -1 && fstat(cachedFd, &statBuffer) == 0) |
| 516 | { |
| 517 | long modelDataSize = statBuffer.st_size; |
Sadik Armagan | ee6818b | 2021-11-05 14:41:52 +0000 | [diff] [blame] | 518 | if (modelDataSize <= 0) |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 519 | { |
Sadik Armagan | ee6818b | 2021-11-05 14:41:52 +0000 | [diff] [blame] | 520 | ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3(): Wrong cached model size!"); |
| 521 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 522 | return V1_3::ErrorStatus::NONE; |
| 523 | } |
| 524 | std::vector<uint8_t> modelData(modelDataSize); |
| 525 | pread(cachedFd, modelData.data(), modelData.size(), 0); |
| 526 | hashValue ^= CacheDataHandlerInstance().Hash(modelData); |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 527 | |
Sadik Armagan | ee6818b | 2021-11-05 14:41:52 +0000 | [diff] [blame] | 528 | // For GpuAcc numberOfCachedFiles is 1 |
| 529 | if (backend == armnn::Compute::GpuAcc) |
| 530 | { |
| 531 | gpuAccCachedFd = cachedFd; |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 532 | } |
| 533 | } |
| 534 | index += numberOfCacheFiles; |
| 535 | } |
| 536 | } |
| 537 | } |
| 538 | |
Sadik Armagan | ee6818b | 2021-11-05 14:41:52 +0000 | [diff] [blame] | 539 | if (!CacheDataHandlerInstance().Validate(token, hashValue, dataCacheData.size())) |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 540 | { |
| 541 | ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3: ValidateHash() failed!"); |
| 542 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 543 | return V1_3::ErrorStatus::GENERAL_FAILURE; |
| 544 | } |
| 545 | |
| 546 | // Deserialize the network.. |
Sadik Armagan | ee6818b | 2021-11-05 14:41:52 +0000 | [diff] [blame] | 547 | armnn::INetworkPtr network = armnn::INetworkPtr(nullptr, [](armnn::INetwork*){}); |
| 548 | try |
| 549 | { |
| 550 | network = armnnDeserializer::IDeserializer::Create()->CreateNetworkFromBinary(dataCacheData); |
| 551 | } |
| 552 | catch (std::exception&) |
| 553 | { |
| 554 | ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3: Exception caught from Deserializer!"); |
| 555 | cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); |
| 556 | return V1_3::ErrorStatus::GENERAL_FAILURE; |
| 557 | } |
Sadik Armagan | 0a2dfab | 2021-10-06 16:41:44 +0100 | [diff] [blame] | 558 | |
| 559 | // Optimize the network |
| 560 | armnn::IOptimizedNetworkPtr optNet(nullptr, nullptr); |
| 561 | armnn::OptimizerOptions OptOptions; |
| 562 | OptOptions.m_ReduceFp32ToFp16 = options.GetFp16Enabled(); |
| 563 | OptOptions.m_ProfilingEnabled = options.IsGpuProfilingEnabled(); |
| 564 | |
| 565 | armnn::BackendOptions gpuAcc("GpuAcc", |
| 566 | { |
| 567 | {"FastMathEnabled", options.IsFastMathEnabled()}, |
| 568 | {"SaveCachedNetwork", saveCachedNetwork}, |
| 569 | {"CachedNetworkFilePath", options.GetCachedNetworkFilePath()}, |
| 570 | {"MLGOTuningFilePath", options.GetClMLGOTunedParametersFile()}, |
| 571 | {"CachedFileDescriptor", gpuAccCachedFd} |
| 572 | }); |
| 573 | |
| 574 | armnn::BackendOptions cpuAcc("CpuAcc", |
| 575 | { |
| 576 | {"FastMathEnabled", options.IsFastMathEnabled()}, |
| 577 | {"NumberOfThreads", options.GetNumberOfThreads()} |
| 578 | }); |
| 579 | OptOptions.m_ModelOptions.push_back(gpuAcc); |
| 580 | OptOptions.m_ModelOptions.push_back(cpuAcc); |
| 581 | |
| 582 | std::vector<std::string> errMessages; |
| 583 | try |
| 584 | { |
| 585 | optNet = armnn::Optimize(*network.get(), |
| 586 | options.GetBackends(), |
| 587 | runtime->GetDeviceSpec(), |
| 588 | OptOptions, |
| 589 | errMessages); |
| 590 | } |
| 591 | catch (std::exception& e) |
| 592 | { |
| 593 | std::stringstream message; |
| 594 | message << "Exception (" << e.what() << ") caught from optimize."; |
| 595 | FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, message.str(), cb); |
| 596 | return V1_3::ErrorStatus::NONE; |
| 597 | } |
| 598 | |
| 599 | // Check that the optimized network is valid. |
| 600 | if (!optNet) |
| 601 | { |
| 602 | std::stringstream message; |
| 603 | message << "Invalid optimized network"; |
| 604 | for (const std::string& msg : errMessages) |
| 605 | { |
| 606 | message << "\n" << msg; |
| 607 | } |
| 608 | FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, message.str(), cb); |
| 609 | return V1_3::ErrorStatus::NONE; |
| 610 | } |
| 611 | |
| 612 | // Export the optimized network graph to a dot file if an output dump directory |
| 613 | // has been specified in the drivers' arguments. |
| 614 | std::string dotGraphFileName = ExportNetworkGraphToDotFile(*optNet, |
| 615 | options.GetRequestInputsAndOutputsDumpDir()); |
| 616 | |
| 617 | // Load it into the runtime. |
| 618 | armnn::NetworkId netId = 0; |
| 619 | std::string msg; |
| 620 | armnn::INetworkProperties networkProperties(options.isAsyncModelExecutionEnabled(), |
| 621 | MemorySource::Undefined, |
| 622 | MemorySource::Undefined, |
| 623 | options.IsGpuProfilingEnabled()); |
| 624 | |
| 625 | try |
| 626 | { |
| 627 | if (runtime->LoadNetwork(netId, move(optNet), msg, networkProperties) != armnn::Status::Success) |
| 628 | { |
| 629 | return FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, msg, cb); |
| 630 | } |
| 631 | } |
| 632 | catch (std::exception& e) |
| 633 | { |
| 634 | std::stringstream message; |
| 635 | message << "Exception (" << e.what() << ") caught from LoadNetwork."; |
| 636 | FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, message.str(), cb); |
| 637 | return V1_3::ErrorStatus::NONE; |
| 638 | } |
| 639 | |
| 640 | std::unique_ptr<ArmnnPreparedModel_1_3<hal_1_3::HalPolicy>> preparedModel( |
| 641 | new ArmnnPreparedModel_1_3<hal_1_3::HalPolicy>(netId, |
| 642 | runtime.get(), |
| 643 | options.GetRequestInputsAndOutputsDumpDir(), |
| 644 | options.IsGpuProfilingEnabled(), |
| 645 | V1_3::Priority::MEDIUM, |
| 646 | options.isAsyncModelExecutionEnabled(), |
| 647 | options.getNoOfArmnnThreads(), |
| 648 | true)); |
| 649 | |
| 650 | NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 651 | return V1_3::ErrorStatus::NONE; |
| 652 | } |
| 653 | |
| 654 | Return<void> ArmnnDriverImpl::getCapabilities_1_3(const armnn::IRuntimePtr& runtime, |
| 655 | V1_3::IDevice::getCapabilities_1_3_cb cb) |
| 656 | { |
| 657 | ALOGV("hal_1_3::ArmnnDriverImpl::getCapabilities()"); |
| 658 | |
| 659 | V1_3::Capabilities capabilities; |
| 660 | |
| 661 | float defaultValue = .1f; |
| 662 | |
| 663 | if (runtime) |
| 664 | { |
| 665 | capabilities.relaxedFloat32toFloat16PerformanceScalar.execTime = |
| 666 | ParseSystemProperty(g_RelaxedFloat32toFloat16PerformanceExecTime, defaultValue); |
| 667 | |
Kevin May | 2eaa119 | 2020-04-15 16:50:57 +0100 | [diff] [blame] | 668 | capabilities.relaxedFloat32toFloat16PerformanceScalar.powerUsage = |
| 669 | ParseSystemProperty(g_RelaxedFloat32toFloat16PerformancePowerUsage, defaultValue); |
| 670 | |
| 671 | capabilities.relaxedFloat32toFloat16PerformanceTensor.execTime = |
| 672 | ParseSystemProperty(g_RelaxedFloat32toFloat16PerformanceExecTime, defaultValue); |
| 673 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 674 | capabilities.relaxedFloat32toFloat16PerformanceTensor.powerUsage = |
| 675 | ParseSystemProperty(g_RelaxedFloat32toFloat16PerformancePowerUsage, defaultValue); |
| 676 | |
Kevin May | 2eaa119 | 2020-04-15 16:50:57 +0100 | [diff] [blame] | 677 | capabilities.ifPerformance.execTime = |
| 678 | ParseSystemProperty(g_ifPerformanceExecTime, defaultValue); |
| 679 | |
| 680 | capabilities.ifPerformance.powerUsage = |
| 681 | ParseSystemProperty(g_ifPerformancePowerUsage, defaultValue); |
| 682 | |
| 683 | capabilities.whilePerformance.execTime = |
| 684 | ParseSystemProperty(g_whilePerformanceExecTime, defaultValue); |
| 685 | |
| 686 | capabilities.whilePerformance.powerUsage = |
| 687 | ParseSystemProperty(g_whilePerformancePowerUsage, defaultValue); |
| 688 | |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 689 | // Set the base value for all operand types |
| 690 | capabilities.operandPerformance = nonExtensionOperandPerformance<HalVersion::V1_3>({FLT_MAX, FLT_MAX}); |
| 691 | |
| 692 | // Load supported operand types |
| 693 | update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_FLOAT32, |
| 694 | { |
| 695 | .execTime = ParseSystemProperty(g_OperandTypeTensorFloat32PerformanceExecTime, defaultValue), |
| 696 | .powerUsage = ParseSystemProperty(g_OperandTypeTensorFloat32PerformancePowerUsage, defaultValue) |
| 697 | }); |
| 698 | |
| 699 | update(&capabilities.operandPerformance, V1_3::OperandType::FLOAT32, |
| 700 | { |
| 701 | .execTime = ParseSystemProperty(g_OperandTypeFloat32PerformanceExecTime, defaultValue), |
| 702 | .powerUsage = ParseSystemProperty(g_OperandTypeFloat32PerformancePowerUsage, defaultValue) |
| 703 | }); |
| 704 | |
| 705 | update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_FLOAT16, |
| 706 | { |
| 707 | .execTime = ParseSystemProperty(g_OperandTypeTensorFloat16PerformanceExecTime, defaultValue), |
| 708 | .powerUsage = ParseSystemProperty(g_OperandTypeTensorFloat16PerformancePowerUsage, defaultValue) |
| 709 | }); |
| 710 | |
| 711 | update(&capabilities.operandPerformance, V1_3::OperandType::FLOAT16, |
| 712 | { |
| 713 | .execTime = ParseSystemProperty(g_OperandTypeFloat16PerformanceExecTime, defaultValue), |
| 714 | .powerUsage = ParseSystemProperty(g_OperandTypeFloat16PerformancePowerUsage, defaultValue) |
| 715 | }); |
| 716 | |
| 717 | update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_QUANT8_ASYMM, |
| 718 | { |
| 719 | .execTime = ParseSystemProperty(g_OperandTypeTensorQuant8AsymmPerformanceExecTime, defaultValue), |
| 720 | .powerUsage = ParseSystemProperty(g_OperandTypeTensorQuant8AsymmPerformancePowerUsage, defaultValue) |
| 721 | }); |
| 722 | |
| 723 | update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_QUANT8_SYMM, |
| 724 | { |
| 725 | .execTime = ParseSystemProperty(g_OperandTypeTensorQuant8SymmPerformanceExecTime, defaultValue), |
| 726 | .powerUsage = ParseSystemProperty(g_OperandTypeTensorQuant8SymmPerformancePowerUsage, defaultValue) |
| 727 | }); |
| 728 | update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_QUANT8_ASYMM_SIGNED, |
| 729 | { |
| 730 | .execTime = ParseSystemProperty(g_OperandTypeTensorQuant8AsymmSignedPerformanceExecTime, |
| 731 | defaultValue), |
| 732 | .powerUsage = ParseSystemProperty(g_OperandTypeTensorQuant8AsymmSignedPerformancePowerUsage, |
| 733 | defaultValue) |
| 734 | }); |
| 735 | |
| 736 | update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_QUANT16_SYMM, |
| 737 | { |
| 738 | .execTime = ParseSystemProperty(g_OperandTypeTensorQuant16SymmPerformanceExecTime, defaultValue), |
| 739 | .powerUsage = ParseSystemProperty(g_OperandTypeTensorQuant16SymmPerformancePowerUsage, defaultValue) |
| 740 | }); |
| 741 | |
| 742 | update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL, |
| 743 | { |
| 744 | .execTime = |
| 745 | ParseSystemProperty(g_OperandTypeTensorQuant8SymmPerChannelPerformanceExecTime, defaultValue), |
| 746 | .powerUsage = |
| 747 | ParseSystemProperty(g_OperandTypeTensorQuant8SymmPerChannelPerformancePowerUsage, defaultValue) |
| 748 | }); |
| 749 | |
| 750 | update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_INT32, |
| 751 | { |
| 752 | .execTime = ParseSystemProperty(g_OperandTypeTensorInt32PerformanceExecTime, defaultValue), |
| 753 | .powerUsage = ParseSystemProperty(g_OperandTypeTensorInt32PerformancePowerUsage, defaultValue) |
| 754 | }); |
| 755 | |
| 756 | update(&capabilities.operandPerformance, V1_3::OperandType::INT32, |
| 757 | { |
| 758 | .execTime = ParseSystemProperty(g_OperandTypeInt32PerformanceExecTime, defaultValue), |
| 759 | .powerUsage = ParseSystemProperty(g_OperandTypeInt32PerformancePowerUsage, defaultValue) |
| 760 | }); |
| 761 | |
| 762 | cb(V1_3::ErrorStatus::NONE, capabilities); |
| 763 | } |
| 764 | else |
| 765 | { |
Kevin May | 2eaa119 | 2020-04-15 16:50:57 +0100 | [diff] [blame] | 766 | capabilities.relaxedFloat32toFloat16PerformanceScalar.execTime = 0; |
| 767 | capabilities.relaxedFloat32toFloat16PerformanceScalar.powerUsage = 0; |
| 768 | capabilities.relaxedFloat32toFloat16PerformanceTensor.execTime = 0; |
| 769 | capabilities.relaxedFloat32toFloat16PerformanceTensor.powerUsage = 0; |
| 770 | capabilities.ifPerformance.execTime = 0; |
| 771 | capabilities.ifPerformance.powerUsage = 0; |
| 772 | capabilities.whilePerformance.execTime = 0; |
| 773 | capabilities.whilePerformance.powerUsage = 0; |
Kevin May | 42477c1 | 2020-03-26 13:34:14 +0000 | [diff] [blame] | 774 | |
| 775 | // Set the base value for all operand types |
| 776 | capabilities.operandPerformance = nonExtensionOperandPerformance<HalVersion::V1_3>({0.f, 0.0f}); |
| 777 | |
| 778 | cb(V1_3::ErrorStatus::DEVICE_UNAVAILABLE, capabilities); |
| 779 | } |
| 780 | |
| 781 | return Void(); |
| 782 | } |
| 783 | |
| 784 | } // namespace hal_1_3 |
| 785 | } // namespace armnn_driver |