diff options
Diffstat (limited to 'ArmnnPreparedModel.hpp')
| -rw-r--r-- | ArmnnPreparedModel.hpp | 63 |
1 files changed, 55 insertions, 8 deletions
diff --git a/ArmnnPreparedModel.hpp b/ArmnnPreparedModel.hpp index 89f6226f..b9d89d4b 100644 --- a/ArmnnPreparedModel.hpp +++ b/ArmnnPreparedModel.hpp @@ -11,6 +11,7 @@ #include <NeuralNetworks.h> #include <armnn/ArmNN.hpp> +#include <armnn/Threadpool.hpp> #include <string> #include <vector> @@ -38,7 +39,11 @@ public: armnn::IRuntime* runtime, const HalModel& model, const std::string& requestInputsAndOutputsDumpDir, - const bool gpuProfilingEnabled); + const bool gpuProfilingEnabled, + const bool asyncModelExecutionEnabled = false, + const unsigned int numberOfThreads = 1, + const bool importEnabled = false, + const bool exportEnabled = false); virtual ~ArmnnPreparedModel(); @@ -56,18 +61,60 @@ public: bool ExecuteWithDummyInputs(); private: + + template<typename CallbackContext> + class ArmnnThreadPoolCallback : public armnn::IAsyncExecutionCallback + { + public: + ArmnnThreadPoolCallback(ArmnnPreparedModel<HalVersion>* model, + std::shared_ptr<std::vector<::android::nn::RunTimePoolInfo>>& pMemPools, + std::shared_ptr<armnn::InputTensors>& inputTensors, + std::shared_ptr<armnn::OutputTensors>& outputTensors, + CallbackContext callbackContext) : + m_Model(model), + m_MemPools(pMemPools), + m_InputTensors(inputTensors), + m_OutputTensors(outputTensors), + m_CallbackContext(callbackContext) + {} + + void Notify(armnn::Status status, armnn::InferenceTimingPair timeTaken) override; + + ArmnnPreparedModel<HalVersion>* m_Model; + std::shared_ptr<std::vector<::android::nn::RunTimePoolInfo>> m_MemPools; + std::shared_ptr<armnn::InputTensors> m_InputTensors; + std::shared_ptr<armnn::OutputTensors> m_OutputTensors; + CallbackContext m_CallbackContext; + }; + template <typename TensorBindingCollection> void DumpTensorsIfRequired(char const* tensorNamePrefix, const TensorBindingCollection& tensorBindings); - armnn::NetworkId m_NetworkId; - armnn::IRuntime* m_Runtime; - HalModel m_Model; + /// schedule the graph prepared from the request for execution + template<typename CallbackContext> + void ScheduleGraphForExecution( + std::shared_ptr<std::vector<::android::nn::RunTimePoolInfo>>& pMemPools, + std::shared_ptr<armnn::InputTensors>& inputTensors, + std::shared_ptr<armnn::OutputTensors>& outputTensors, + CallbackContext m_CallbackContext); + + armnn::NetworkId m_NetworkId; + armnn::IRuntime* m_Runtime; + HalModel m_Model; // There must be a single RequestThread for all ArmnnPreparedModel objects to ensure serial execution of workloads // It is specific to this class, so it is declared as static here - static RequestThread<ArmnnPreparedModel, HalVersion, CallbackContext_1_0> m_RequestThread; - uint32_t m_RequestCount; - const std::string& m_RequestInputsAndOutputsDumpDir; - const bool m_GpuProfilingEnabled; + static RequestThread<ArmnnPreparedModel, + HalVersion, + CallbackContext_1_0> m_RequestThread; + uint32_t m_RequestCount; + const std::string& m_RequestInputsAndOutputsDumpDir; + const bool m_GpuProfilingEnabled; + // Static to allow sharing of threadpool between ArmnnPreparedModel instances + static std::unique_ptr<armnn::Threadpool> m_Threadpool; + std::shared_ptr<armnn::IWorkingMemHandle> m_WorkingMemHandle; + const bool m_AsyncModelExecutionEnabled; + const bool m_EnableImport; + const bool m_EnableExport; }; } |