From 0a2dfabd76a45c58d0a14567f0503369c4e6fbf3 Mon Sep 17 00:00:00 2001 From: Sadik Armagan Date: Wed, 6 Oct 2021 16:41:44 +0100 Subject: IVGCVSW-5636 'Implement NNAPI caching functions' * Cached serialized ArmNN model. !armnn:6384 Signed-off-by: Sadik Armagan Signed-off-by: Kevin May Change-Id: I78120a7f8ea892a28c0ff25f1b54e67a4f912574 --- 1.3/ArmnnDriver.hpp | 86 +++++++---- 1.3/ArmnnDriverImpl.cpp | 391 ++++++++++++++++++++++++++++++++++++++++++++++-- 1.3/ArmnnDriverImpl.hpp | 29 +++- 3 files changed, 453 insertions(+), 53 deletions(-) (limited to '1.3') diff --git a/1.3/ArmnnDriver.hpp b/1.3/ArmnnDriver.hpp index fd4aa74c..6d2e0b7a 100644 --- a/1.3/ArmnnDriver.hpp +++ b/1.3/ArmnnDriver.hpp @@ -21,6 +21,8 @@ #include "../1.0/ArmnnDriverImpl.hpp" #include "../1.0/HalPolicy.hpp" +#include + #include namespace armnn_driver @@ -31,6 +33,7 @@ namespace hal_1_3 class ArmnnDriver : public ArmnnDevice, public V1_3::IDevice { public: + using HidlToken = android::hardware::hidl_array; ArmnnDriver(DriverOptions options) : ArmnnDevice(std::move(options)) @@ -39,9 +42,7 @@ public: } ~ArmnnDriver() {} - using HidlToken = android::hardware::hidl_array; -public: Return getCapabilities(V1_0::IDevice::getCapabilities_cb cb) override { ALOGV("hal_1_3::ArmnnDriver::getCapabilities()"); @@ -131,10 +132,13 @@ public: cb); } - Return prepareModel_1_2(const V1_2::Model& model, V1_1::ExecutionPreference preference, - const android::hardware::hidl_vec&, - const android::hardware::hidl_vec&, const HidlToken&, - const android::sp& cb) + Return prepareModel_1_2( + const V1_2::Model& model, + V1_1::ExecutionPreference preference, + const android::hardware::hidl_vec& modelCacheHandle, + const android::hardware::hidl_vec& dataCacheHandle, + const HidlToken& token, + const android::sp& cb) { ALOGV("hal_1_3::ArmnnDriver::prepareModel_1_2()"); @@ -151,6 +155,9 @@ public: m_ClTunedParameters, m_Options, model, + modelCacheHandle, + dataCacheHandle, + token, cb, model.relaxComputationFloat32toFloat16 && m_Options.GetFp16Enabled()); @@ -174,14 +181,15 @@ public: cb); } - Return prepareModel_1_3(const V1_3::Model& model, - V1_1::ExecutionPreference preference, - V1_3::Priority priority, - const V1_3::OptionalTimePoint&, - const android::hardware::hidl_vec&, - const android::hardware::hidl_vec&, - const HidlToken&, - const android::sp& cb) + Return prepareModel_1_3( + const V1_3::Model& model, + V1_1::ExecutionPreference preference, + V1_3::Priority priority, + const V1_3::OptionalTimePoint&, + const android::hardware::hidl_vec& modelCache, + const android::hardware::hidl_vec& dataCache, + const HidlToken& token, + const android::sp& cb) { ALOGV("hal_1_3::ArmnnDriver::prepareModel_1_3()"); @@ -199,11 +207,13 @@ public: return V1_3::ErrorStatus::INVALID_ARGUMENT; } - return ArmnnDriverImpl::prepareArmnnModel_1_3(m_Runtime, m_ClTunedParameters, m_Options, model, + modelCache, + dataCache, + token, cb, model.relaxComputationFloat32toFloat16 && m_Options.GetFp16Enabled(), @@ -219,10 +229,13 @@ public: Return getNumberOfCacheFilesNeeded(getNumberOfCacheFilesNeeded_cb cb) { - ALOGV("hal_1_3::ArmnnDriver::getSupportedExtensions()"); - - // Set both numbers to be 0 for cache not supported. - cb(V1_0::ErrorStatus::NONE, 0, 0); + ALOGV("hal_1_3::ArmnnDriver::getNumberOfCacheFilesNeeded()"); + unsigned int numberOfCachedModelFiles = 0; + for (auto& backend : m_Options.GetBackends()) + { + numberOfCachedModelFiles += GetNumberOfCacheFiles(backend); + } + cb(V1_0::ErrorStatus::NONE, numberOfCachedModelFiles, 1ul); return Void(); } @@ -250,26 +263,35 @@ public: } Return prepareModelFromCache( - const android::hardware::hidl_vec&, - const android::hardware::hidl_vec&, - const HidlToken&, - const android::sp& callback) + const android::hardware::hidl_vec& modelCacheHandle, + const android::hardware::hidl_vec& dataCacheHandle, + const HidlToken& token, + const android::sp& cb) { ALOGV("hal_1_3::ArmnnDriver::prepareModelFromCache()"); - callback->notify_1_2(V1_0::ErrorStatus::GENERAL_FAILURE, nullptr); - return V1_0::ErrorStatus::GENERAL_FAILURE; + return hal_1_2::ArmnnDriverImpl::prepareModelFromCache(m_Runtime, + m_Options, + modelCacheHandle, + dataCacheHandle, + token, + cb); } Return prepareModelFromCache_1_3( const V1_3::OptionalTimePoint&, - const android::hardware::hidl_vec&, - const android::hardware::hidl_vec&, - const HidlToken&, - const android::sp& callback) + const android::hardware::hidl_vec& modelCacheHandle, + const android::hardware::hidl_vec& dataCacheHandle, + const HidlToken& token, + const android::sp& cb) { - ALOGV("hal_1_3::ArmnnDriver::prepareModelFromCache()"); - callback->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); - return V1_3::ErrorStatus::GENERAL_FAILURE; + ALOGV("hal_1_3::ArmnnDriver::prepareModelFromCache_1_3()"); + + return ArmnnDriverImpl::prepareModelFromCache_1_3(m_Runtime, + m_Options, + modelCacheHandle, + dataCacheHandle, + token, + cb); } Return allocate(const V1_3::BufferDesc& /*desc*/, diff --git a/1.3/ArmnnDriverImpl.cpp b/1.3/ArmnnDriverImpl.cpp index 3ecd2f82..e1d65f92 100644 --- a/1.3/ArmnnDriverImpl.cpp +++ b/1.3/ArmnnDriverImpl.cpp @@ -8,8 +8,12 @@ #include "../ModelToINetworkConverter.hpp" #include "../SystemPropertiesUtils.hpp" +#include + #include +#include + namespace { const char *g_RelaxedFloat32toFloat16PerformanceExecTime = "ArmNN.relaxedFloat32toFloat16Performance.execTime"; @@ -100,6 +104,9 @@ Return ArmnnDriverImpl::prepareArmnnModel_1_3( const armnn::IGpuAccTunedParametersPtr& clTunedParameters, const DriverOptions& options, const V1_3::Model& model, + const android::hardware::hidl_vec& modelCacheHandle, + const android::hardware::hidl_vec& dataCacheHandle, + const HidlToken& token, const android::sp& cb, bool float32ToFloat16, V1_3::Priority priority) @@ -138,8 +145,13 @@ Return ArmnnDriverImpl::prepareArmnnModel_1_3( // Serialize the network graph to a .armnn file if an output directory // has been specified in the drivers' arguments. + std::vector dataCacheData; + bool serializeToFile = dataCacheHandle.size() < 1 ? false : true; auto serializedNetworkFileName = - SerializeNetwork(*modelConverter.GetINetwork(), options.GetRequestInputsAndOutputsDumpDir()); + SerializeNetwork(*modelConverter.GetINetwork(), + options.GetRequestInputsAndOutputsDumpDir(), + dataCacheData, + serializeToFile); // Optimize the network armnn::IOptimizedNetworkPtr optNet(nullptr, nullptr); @@ -147,12 +159,42 @@ Return ArmnnDriverImpl::prepareArmnnModel_1_3( OptOptions.m_ReduceFp32ToFp16 = float32ToFloat16; OptOptions.m_ProfilingEnabled = options.IsGpuProfilingEnabled(); + int cachedFd = -1; + bool saveCachedNetwork = options.SaveCachedNetwork(); + + unsigned int numberOfCachedModelFiles = 0; + if (modelCacheHandle.size() > 0) + { + unsigned int index = 0; + for (auto& backend : options.GetBackends()) + { + // modelCacheHandle size should be equal to numberOfCachedModelFiles + // modelCacheHandle vector should be in same order as backends + auto numberOfCacheFiles = GetNumberOfCacheFiles(backend); + if (numberOfCacheFiles > 0) + { + numberOfCachedModelFiles += numberOfCacheFiles; + if (modelCacheHandle[index]->numFds == 1) + { + // For GpuAcc numberOfCachedFiles is 1 + if (backend == armnn::Compute::GpuAcc) + { + cachedFd = modelCacheHandle[index]->data[0]; + saveCachedNetwork = true; + } + } + index += numberOfCachedModelFiles; + } + } + } + armnn::BackendOptions gpuAcc("GpuAcc", { { "FastMathEnabled", options.IsFastMathEnabled() }, - { "SaveCachedNetwork", options.SaveCachedNetwork() }, + { "SaveCachedNetwork", saveCachedNetwork }, { "CachedNetworkFilePath", options.GetCachedNetworkFilePath() }, - { "MLGOTuningFilePath", options.GetClMLGOTunedParametersFile() } + { "MLGOTuningFilePath", options.GetClMLGOTunedParametersFile() }, + { "CachedFileDescriptor", cachedFd } }); armnn::BackendOptions cpuAcc("CpuAcc", @@ -203,7 +245,11 @@ Return ArmnnDriverImpl::prepareArmnnModel_1_3( std::string msg; armnn::INetworkProperties networkProperties(options.isAsyncModelExecutionEnabled(), MemorySource::Undefined, - MemorySource::Undefined); + MemorySource::Undefined, + options.IsGpuProfilingEnabled()); + + auto numInputs = getMainModel(model).inputIndexes.size(); + auto numOutputs = getMainModel(model).outputIndexes.size(); try { if (runtime->LoadNetwork(netId, move(optNet), msg, networkProperties) != armnn::Status::Success) @@ -239,28 +285,345 @@ Return ArmnnDriverImpl::prepareArmnnModel_1_3( // Run a single 'dummy' inference of the model. This means that CL kernels will get compiled (and tuned if // this is enabled) before the first 'real' inference which removes the overhead of the first inference. - if (!preparedModel->ExecuteWithDummyInputs()) + // Only run this if the GpuAcc backend has been added to options + if (std::find(options.GetBackends().begin(), + options.GetBackends().end(), + armnn::Compute::GpuAcc) != options.GetBackends().end()) { - return FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, "Network could not be executed", cb); + if (!preparedModel->ExecuteWithDummyInputs(numInputs, numOutputs)) + { + return FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, "Network could not be executed", cb); + } + + if (clTunedParameters && + options.GetClTunedParametersMode() == armnn::IGpuAccTunedParameters::Mode::UpdateTunedParameters) + { + // Now that we've done one inference the CL kernel parameters will have been tuned, + // so save the updated file. + try + { + clTunedParameters->Save(options.GetClTunedParametersFile().c_str()); + } + catch (std::exception& error) + { + ALOGE("ArmnnDriverImpl::prepareModel: Failed to save CL tuned parameters file '%s': %s", + options.GetClTunedParametersFile().c_str(), error.what()); + } + } + } + size_t hashValue = 0; + // Cache the model + if (dataCacheHandle.size() > 0) + { + // Cache the Arm NN model + if (dataCacheHandle.size() != 1) + { + NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); + return V1_3::ErrorStatus::NONE; + } + + if (dataCacheHandle[0]->numFds != 1) + { + ALOGW("ArmnnDriverImpl::prepareArmnnModel_1_3: Cannot cache the data, numFds != 1."); + NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); + return V1_3::ErrorStatus::NONE; + } + int dataCacheFileAccessMode = fcntl(dataCacheHandle[0]->data[0], F_GETFL) & O_ACCMODE; + if (dataCacheFileAccessMode != O_RDWR) + { + ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3(): Invalid Access Mode."); + NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); + return V1_3::ErrorStatus::NONE; + } + + write(dataCacheHandle[0]->data[0], dataCacheData.data(), dataCacheData.size()); + hashValue = CacheDataHandlerInstance().Hash(dataCacheData); } - if (clTunedParameters && - options.GetClTunedParametersMode() == armnn::IGpuAccTunedParameters::Mode::UpdateTunedParameters) + // Cache the model data + if (modelCacheHandle.size() > 0) { - // Now that we've done one inference the CL kernel parameters will have been tuned, so save the updated file. - try + if (modelCacheHandle.size() != numberOfCachedModelFiles) { - clTunedParameters->Save(options.GetClTunedParametersFile().c_str()); + NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); + return V1_3::ErrorStatus::NONE; } - catch (std::exception& error) + + for (uint32_t i = 0; i < modelCacheHandle.size(); ++i) { - ALOGE("ArmnnDriverImpl::prepareModel: Failed to save CL tuned parameters file '%s': %s", - options.GetClTunedParametersFile().c_str(), error.what()); + if (modelCacheHandle[i]->numFds == 1) + { + int modelCacheFileAccessMode = fcntl(modelCacheHandle[i]->data[0], F_GETFL) & O_ACCMODE; + if (modelCacheFileAccessMode != O_RDONLY) + { + struct stat statBuffer; + if (fstat(modelCacheHandle[i]->data[0], &statBuffer) == 0) + { + long modelDataSize = statBuffer.st_size; + if (modelDataSize > 0) + { + std::vector modelData(modelDataSize); + pread(modelCacheHandle[i]->data[0], modelData.data(), modelData.size(), 0); + hashValue ^= CacheDataHandlerInstance().Hash(modelData); + } + } + } + } } } + if (hashValue != 0) + { + CacheDataHandlerInstance().Register(token, hashValue, dataCacheData.size()); + } NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); + return V1_3::ErrorStatus::NONE; +} + +Return ArmnnDriverImpl::prepareModelFromCache_1_3( + const armnn::IRuntimePtr& runtime, + const DriverOptions& options, + const android::hardware::hidl_vec& modelCacheHandle, + const android::hardware::hidl_vec& dataCacheHandle, + const HidlToken& token, + const android::sp& cb) +{ + ALOGV("ArmnnDriverImpl::prepareModelFromCache_1_3()"); + if (token.size() != ANEURALNETWORKS_BYTE_SIZE_OF_CACHE_TOKEN) + { + cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); + return V1_3::ErrorStatus::GENERAL_FAILURE; + } + + if (cb.get() == nullptr) + { + ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3: Invalid callback passed to prepareModelFromCache_1_3"); + return V1_3::ErrorStatus::INVALID_ARGUMENT; + } + + if (!runtime) + { + ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3: Device unavailable"); + return V1_3::ErrorStatus::DEVICE_UNAVAILABLE; + } + + // DataCacheHandle size should always be 1 + // Arm NN model + if (dataCacheHandle.size() != 1) + { + cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); + return V1_3::ErrorStatus::GENERAL_FAILURE; + } + + // Check if model files cached they match the expected value + unsigned int numberOfCachedModelFiles = 0; + for (auto& backend : options.GetBackends()) + { + numberOfCachedModelFiles += GetNumberOfCacheFiles(backend); + } + if (modelCacheHandle.size() != numberOfCachedModelFiles) + { + cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); + return V1_3::ErrorStatus::GENERAL_FAILURE; + } + + if (dataCacheHandle[0]->numFds != 1) + { + ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3(): Cannot read from the cache data, numFds != 1."); + cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); + return V1_3::ErrorStatus::GENERAL_FAILURE; + } + + int dataCacheFileAccessMode = fcntl(dataCacheHandle[0]->data[0], F_GETFL) & O_ACCMODE; + if (dataCacheFileAccessMode != O_RDWR) + { + cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); + return V1_3::ErrorStatus::GENERAL_FAILURE; + } + + auto dataSize = CacheDataHandlerInstance().GetCacheSize(token); + if (dataSize == 0) + { + ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3: Invalid data to deserialize!"); + cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); + return V1_3::ErrorStatus::GENERAL_FAILURE; + } + + int offset = 0; + { + struct stat statBuffer; + if (fstat(dataCacheHandle[0]->data[0], &statBuffer) == 0) + { + unsigned long bufferSize = statBuffer.st_size; + if (bufferSize <= 0) + { + ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3: Invalid data to deserialize!"); + cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); + return V1_3::ErrorStatus::GENERAL_FAILURE; + } + if (bufferSize > dataSize) + { + offset = bufferSize - dataSize; + } + } + } + std::vector dataCacheData(dataSize); + pread(dataCacheHandle[0]->data[0], dataCacheData.data(), dataCacheData.size(), offset); + auto hashValue = CacheDataHandlerInstance().Hash(dataCacheData); + + int gpuAccCachedFd = -1; + bool saveCachedNetwork = false; + if (modelCacheHandle.size() > 0) + { + unsigned int index = 0; + for (auto& backend : options.GetBackends()) + { + // modelCacheHandle size should be equal to numberOfCachedModelFiles + // modelCacheHandle vector should be in same order as backends + auto numberOfCacheFiles = GetNumberOfCacheFiles(backend); + if (numberOfCacheFiles > 0) + { + if (modelCacheHandle[index]->numFds != 1) + { + ALOGW( + "ArmnnDriverImpl::prepareModelFromCache_1_3(): Cannot read from the model cache, numFds != 1."); + cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); + return V1_3::ErrorStatus::GENERAL_FAILURE; + } + auto cachedFd = modelCacheHandle[index]->data[0]; + + int modelCacheFileAccessMode = fcntl(cachedFd, F_GETFL) & O_ACCMODE; + if (modelCacheFileAccessMode != O_RDWR) + { + cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); + return V1_3::ErrorStatus::GENERAL_FAILURE; + } + + struct stat statBuffer; + if (cachedFd != -1 && fstat(cachedFd, &statBuffer) == 0) + { + long modelDataSize = statBuffer.st_size; + if (modelDataSize > 0) + { + std::vector modelData(modelDataSize); + pread(cachedFd, modelData.data(), modelData.size(), 0); + hashValue ^= CacheDataHandlerInstance().Hash(modelData); + + // For GpuAcc numberOfCachedFiles is 1 + if (backend == armnn::Compute::GpuAcc) + { + gpuAccCachedFd = cachedFd; + } + } + } + index += numberOfCacheFiles; + } + } + } + + if (!CacheDataHandlerInstance().Validate(token, hashValue)) + { + ALOGW("ArmnnDriverImpl::prepareModelFromCache_1_3: ValidateHash() failed!"); + cb->notify_1_3(V1_3::ErrorStatus::GENERAL_FAILURE, nullptr); + return V1_3::ErrorStatus::GENERAL_FAILURE; + } + + // Deserialize the network.. + auto network = armnnDeserializer::IDeserializer::Create()->CreateNetworkFromBinary(dataCacheData); + + // Optimize the network + armnn::IOptimizedNetworkPtr optNet(nullptr, nullptr); + armnn::OptimizerOptions OptOptions; + OptOptions.m_ReduceFp32ToFp16 = options.GetFp16Enabled(); + OptOptions.m_ProfilingEnabled = options.IsGpuProfilingEnabled(); + + armnn::BackendOptions gpuAcc("GpuAcc", + { + {"FastMathEnabled", options.IsFastMathEnabled()}, + {"SaveCachedNetwork", saveCachedNetwork}, + {"CachedNetworkFilePath", options.GetCachedNetworkFilePath()}, + {"MLGOTuningFilePath", options.GetClMLGOTunedParametersFile()}, + {"CachedFileDescriptor", gpuAccCachedFd} + }); + + armnn::BackendOptions cpuAcc("CpuAcc", + { + {"FastMathEnabled", options.IsFastMathEnabled()}, + {"NumberOfThreads", options.GetNumberOfThreads()} + }); + OptOptions.m_ModelOptions.push_back(gpuAcc); + OptOptions.m_ModelOptions.push_back(cpuAcc); + + std::vector errMessages; + try + { + optNet = armnn::Optimize(*network.get(), + options.GetBackends(), + runtime->GetDeviceSpec(), + OptOptions, + errMessages); + } + catch (std::exception& e) + { + std::stringstream message; + message << "Exception (" << e.what() << ") caught from optimize."; + FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, message.str(), cb); + return V1_3::ErrorStatus::NONE; + } + + // Check that the optimized network is valid. + if (!optNet) + { + std::stringstream message; + message << "Invalid optimized network"; + for (const std::string& msg : errMessages) + { + message << "\n" << msg; + } + FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, message.str(), cb); + return V1_3::ErrorStatus::NONE; + } + + // Export the optimized network graph to a dot file if an output dump directory + // has been specified in the drivers' arguments. + std::string dotGraphFileName = ExportNetworkGraphToDotFile(*optNet, + options.GetRequestInputsAndOutputsDumpDir()); + + // Load it into the runtime. + armnn::NetworkId netId = 0; + std::string msg; + armnn::INetworkProperties networkProperties(options.isAsyncModelExecutionEnabled(), + MemorySource::Undefined, + MemorySource::Undefined, + options.IsGpuProfilingEnabled()); + + try + { + if (runtime->LoadNetwork(netId, move(optNet), msg, networkProperties) != armnn::Status::Success) + { + return FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, msg, cb); + } + } + catch (std::exception& e) + { + std::stringstream message; + message << "Exception (" << e.what() << ") caught from LoadNetwork."; + FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, message.str(), cb); + return V1_3::ErrorStatus::NONE; + } + + std::unique_ptr> preparedModel( + new ArmnnPreparedModel_1_3(netId, + runtime.get(), + options.GetRequestInputsAndOutputsDumpDir(), + options.IsGpuProfilingEnabled(), + V1_3::Priority::MEDIUM, + options.isAsyncModelExecutionEnabled(), + options.getNoOfArmnnThreads(), + true)); + + NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release()); return V1_3::ErrorStatus::NONE; } diff --git a/1.3/ArmnnDriverImpl.hpp b/1.3/ArmnnDriverImpl.hpp index 3c094fe5..a482edac 100644 --- a/1.3/ArmnnDriverImpl.hpp +++ b/1.3/ArmnnDriverImpl.hpp @@ -7,6 +7,7 @@ #include +#include "../CacheDataHandler.hpp" #include "../DriverOptions.hpp" #include @@ -31,13 +32,27 @@ namespace hal_1_3 class ArmnnDriverImpl { public: - static Return prepareArmnnModel_1_3(const armnn::IRuntimePtr& runtime, - const armnn::IGpuAccTunedParametersPtr& clTunedParameters, - const DriverOptions& options, - const V1_3::Model& model, - const android::sp& cb, - bool float32ToFloat16 = false, - V1_3::Priority priority = V1_3::Priority::MEDIUM); + using HidlToken = android::hardware::hidl_array; + + static Return prepareArmnnModel_1_3( + const armnn::IRuntimePtr& runtime, + const armnn::IGpuAccTunedParametersPtr& clTunedParameters, + const DriverOptions& options, + const V1_3::Model& model, + const android::hardware::hidl_vec& modelCacheHandle, + const android::hardware::hidl_vec& dataCacheHandle, + const HidlToken& token, + const android::sp& cb, + bool float32ToFloat16 = false, + V1_3::Priority priority = V1_3::Priority::MEDIUM); + + static Return prepareModelFromCache_1_3( + const armnn::IRuntimePtr& runtime, + const DriverOptions& options, + const android::hardware::hidl_vec& modelCacheHandle, + const android::hardware::hidl_vec& dataCacheHandle, + const HidlToken& token, + const android::sp& cb); static Return getCapabilities_1_3(const armnn::IRuntimePtr& runtime, V1_3::IDevice::getCapabilities_1_3_cb cb); -- cgit v1.2.1