From b87f5443f903a875cafd6affcbee5b01e21116f6 Mon Sep 17 00:00:00 2001 From: Francis Murtagh Date: Thu, 23 Sep 2021 13:20:53 +0100 Subject: IVGCVSW-6314 Pre-Importing - Create example app Signed-off-by: Francis Murtagh Change-Id: Ibfbc9088aa09d363d67010a695866e68cef823fc --- include/armnn/IRuntime.hpp | 2 +- samples/CMakeLists.txt | 5 ++ samples/PreImportMemorySample.cpp | 117 ++++++++++++++++++++++++++++++++++++++ 3 files changed, 123 insertions(+), 1 deletion(-) create mode 100644 samples/PreImportMemorySample.cpp diff --git a/include/armnn/IRuntime.hpp b/include/armnn/IRuntime.hpp index 47bfef588a..014481357d 100644 --- a/include/armnn/IRuntime.hpp +++ b/include/armnn/IRuntime.hpp @@ -246,7 +246,7 @@ public: /// Allowing for a set of OutputTensors to be imported and mapped once, but used in execution many times. /// This function is not thread safe and must not be used while other threads are calling Execute(). /// Only compatible with AsyncEnabled networks - std::vector ImportOutputs(NetworkId networkId, const OutputTensors& outputTensors); + std::vector ImportOutputs(NetworkId networkId, const OutputTensors& outputTensors); /// Un-import and delete the imported InputTensor/s /// This function is not thread safe and must not be used while other threads are calling Execute(). diff --git a/samples/CMakeLists.txt b/samples/CMakeLists.txt index 7af8b7265a..210dc9c864 100644 --- a/samples/CMakeLists.txt +++ b/samples/CMakeLists.txt @@ -16,3 +16,8 @@ if(BUILD_SAMPLE_APP AND ARMCOMPUTECL) target_link_libraries(CustomMemoryAllocatorSample armnn ${CMAKE_THREAD_LIBS_INIT}) endif() +if(BUILD_SAMPLE_APP AND ARMNNREF) + add_executable(PreImportMemorySample PreImportMemorySample.cpp) + target_link_libraries(PreImportMemorySample armnn ${CMAKE_THREAD_LIBS_INIT}) +endif() + diff --git a/samples/PreImportMemorySample.cpp b/samples/PreImportMemorySample.cpp new file mode 100644 index 0000000000..cf196fd20e --- /dev/null +++ b/samples/PreImportMemorySample.cpp @@ -0,0 +1,117 @@ +// +// Copyright © 2021 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include + +#include + +// A simple example application to show the usage of Memory Management Pre Importing of Inputs and Outputs. In this +// sample, the users single input number is added to itself using an add layer and outputted to console as a number +// that is double the input. The code does not use EnqueueWorkload but instead uses runtime->Execute + +int main() +{ + using namespace armnn; + + float number; + std::cout << "Please enter a number: " << std::endl; + std::cin >> number; + + // Turn on logging to standard output + // This is useful in this sample so that users can learn more about what is going on + armnn::ConfigureLogging(true, false, LogSeverity::Info); + + armnn::IRuntime::CreationOptions options; + armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); + + armnn::NetworkId networkIdentifier1 = 0; + + armnn::INetworkPtr testNetwork(armnn::INetwork::Create()); + auto inputLayer1 = testNetwork->AddInputLayer(0, "input 1 layer"); + auto inputLayer2 = testNetwork->AddInputLayer(1, "input 2 layer"); + auto addLayer = testNetwork->AddAdditionLayer("add layer"); + auto outputLayer = testNetwork->AddOutputLayer(2, "output layer"); + + // Set the tensors in the network. + TensorInfo tensorInfo{{4}, armnn::DataType::Float32}; + + inputLayer1->GetOutputSlot(0).Connect(addLayer->GetInputSlot(0)); + inputLayer1->GetOutputSlot(0).SetTensorInfo(tensorInfo); + inputLayer2->GetOutputSlot(0).Connect(addLayer->GetInputSlot(1)); + inputLayer2->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + addLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0)); + addLayer->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + // Set preferred backend to CpuRef + std::vector backends = { armnn::Compute::CpuRef }; + + // To hold an eventual error message if loading the network fails + std::string er; + + // Initialize network properties with asyncEnabled and MemorySources != MemorySource::Undefined + armnn::INetworkProperties networkProperties(true, MemorySource::Malloc, MemorySource::Malloc); + + // Optimize and Load the network into runtime + runtime->LoadNetwork(networkIdentifier1, + Optimize(*testNetwork, backends, runtime->GetDeviceSpec()), + er, + networkProperties); + + // Create structures for input & output + std::vector inputData1(4, number); + std::vector inputData2(4, number); + ConstTensor inputTensor1(tensorInfo, inputData1.data()); + ConstTensor inputTensor2(tensorInfo, inputData2.data()); + + std::vector outputData1(4); + Tensor outputTensor1{tensorInfo, outputData1.data()}; + + // ImportInputs separates the importing and mapping of InputTensors from network execution. + // Allowing for a set of InputTensors to be imported and mapped once, but used in execution many times. + // ImportInputs is not thread safe and must not be used while other threads are calling Execute(). + // Only compatible with AsyncEnabled networks + + // PreImport inputTensors giving pre-imported ids of 1 and 2 + std::vector importedInputVec = runtime->ImportInputs(networkIdentifier1, + {{0, inputTensor1}, {1, inputTensor2}}); + + // Create a new unique WorkingMemHandle object. Create multiple handles if you wish to have + // overlapped Execution by calling this function from different threads. + auto memHandle = runtime->CreateWorkingMemHandle(networkIdentifier1); + + // Execute evaluates a network using input in inputTensors and outputs filled into outputTensors. + // This function performs a thread safe execution of the network. Returns once execution is complete. + // Will block until this and any other thread using the same workingMem object completes. + // Execute with PreImported inputTensor1 as well as Non-PreImported inputTensor2 + runtime->Execute(*memHandle.get(), {}, {{2, outputTensor1}}, importedInputVec /* pre-imported ids */); + + // ImportOutputs separates the importing and mapping of OutputTensors from network execution. + // Allowing for a set of OutputTensors to be imported and mapped once, but used in execution many times. + // This function is not thread safe and must not be used while other threads are calling Execute(). + // Only compatible with AsyncEnabled networks + // Provide layerBinding Id to outputTensor1 + std::pair output1{2, outputTensor1}; + // PreImport outputTensor1 + std::vector importedOutputVec = runtime->ImportOutputs(networkIdentifier1, {output1}); + + // Execute with Non-PreImported inputTensor1 as well as PreImported inputTensor2 + runtime->Execute(*memHandle.get(), {{0, inputTensor1}}, {{2, outputTensor1}}, {1 /* pre-imported id */}); + + // Clear the previously PreImportedInput with the network Id and inputIds returned from ImportInputs() + // Note: This will happen automatically during destructor of armnn::LoadedNetwork + runtime->ClearImportedInputs(networkIdentifier1, importedInputVec); + + // Clear the previously PreImportedOutputs with the network Id and outputIds returned from ImportOutputs() + // Note: This will happen automatically during destructor of armnn::LoadedNetwork + runtime->ClearImportedOutputs(networkIdentifier1, importedOutputVec); + + // Execute with Non-PreImported inputTensor1, inputTensor2 and the PreImported outputTensor1 + runtime->Execute(*memHandle.get(), {{0, inputTensor1}, {1, inputTensor2}}, {{2, outputTensor1}}); + + std::cout << "Your number was " << outputData1.data()[0] << std::endl; + + return 0; +} -- cgit v1.2.1