Francis Murtagh | b87f544 | 2021-09-23 13:20:53 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2021 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include <armnn/ArmNN.hpp> |
| 7 | |
| 8 | #include <iostream> |
| 9 | |
| 10 | // A simple example application to show the usage of Memory Management Pre Importing of Inputs and Outputs. In this |
| 11 | // sample, the users single input number is added to itself using an add layer and outputted to console as a number |
| 12 | // that is double the input. The code does not use EnqueueWorkload but instead uses runtime->Execute |
| 13 | |
| 14 | int main() |
| 15 | { |
| 16 | using namespace armnn; |
| 17 | |
| 18 | float number; |
| 19 | std::cout << "Please enter a number: " << std::endl; |
| 20 | std::cin >> number; |
| 21 | |
| 22 | // Turn on logging to standard output |
| 23 | // This is useful in this sample so that users can learn more about what is going on |
| 24 | armnn::ConfigureLogging(true, false, LogSeverity::Info); |
| 25 | |
| 26 | armnn::IRuntime::CreationOptions options; |
| 27 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 28 | |
| 29 | armnn::NetworkId networkIdentifier1 = 0; |
| 30 | |
| 31 | armnn::INetworkPtr testNetwork(armnn::INetwork::Create()); |
| 32 | auto inputLayer1 = testNetwork->AddInputLayer(0, "input 1 layer"); |
| 33 | auto inputLayer2 = testNetwork->AddInputLayer(1, "input 2 layer"); |
| 34 | auto addLayer = testNetwork->AddAdditionLayer("add layer"); |
| 35 | auto outputLayer = testNetwork->AddOutputLayer(2, "output layer"); |
| 36 | |
| 37 | // Set the tensors in the network. |
| 38 | TensorInfo tensorInfo{{4}, armnn::DataType::Float32}; |
| 39 | |
| 40 | inputLayer1->GetOutputSlot(0).Connect(addLayer->GetInputSlot(0)); |
| 41 | inputLayer1->GetOutputSlot(0).SetTensorInfo(tensorInfo); |
| 42 | inputLayer2->GetOutputSlot(0).Connect(addLayer->GetInputSlot(1)); |
| 43 | inputLayer2->GetOutputSlot(0).SetTensorInfo(tensorInfo); |
| 44 | |
| 45 | addLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0)); |
| 46 | addLayer->GetOutputSlot(0).SetTensorInfo(tensorInfo); |
| 47 | |
| 48 | // Set preferred backend to CpuRef |
| 49 | std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef }; |
| 50 | |
| 51 | // To hold an eventual error message if loading the network fails |
| 52 | std::string er; |
| 53 | |
| 54 | // Initialize network properties with asyncEnabled and MemorySources != MemorySource::Undefined |
| 55 | armnn::INetworkProperties networkProperties(true, MemorySource::Malloc, MemorySource::Malloc); |
| 56 | |
| 57 | // Optimize and Load the network into runtime |
| 58 | runtime->LoadNetwork(networkIdentifier1, |
| 59 | Optimize(*testNetwork, backends, runtime->GetDeviceSpec()), |
| 60 | er, |
| 61 | networkProperties); |
| 62 | |
| 63 | // Create structures for input & output |
| 64 | std::vector<float> inputData1(4, number); |
| 65 | std::vector<float> inputData2(4, number); |
| 66 | ConstTensor inputTensor1(tensorInfo, inputData1.data()); |
| 67 | ConstTensor inputTensor2(tensorInfo, inputData2.data()); |
| 68 | |
| 69 | std::vector<float> outputData1(4); |
| 70 | Tensor outputTensor1{tensorInfo, outputData1.data()}; |
| 71 | |
| 72 | // ImportInputs separates the importing and mapping of InputTensors from network execution. |
| 73 | // Allowing for a set of InputTensors to be imported and mapped once, but used in execution many times. |
| 74 | // ImportInputs is not thread safe and must not be used while other threads are calling Execute(). |
| 75 | // Only compatible with AsyncEnabled networks |
| 76 | |
| 77 | // PreImport inputTensors giving pre-imported ids of 1 and 2 |
| 78 | std::vector<ImportedInputId> importedInputVec = runtime->ImportInputs(networkIdentifier1, |
| 79 | {{0, inputTensor1}, {1, inputTensor2}}); |
| 80 | |
| 81 | // Create a new unique WorkingMemHandle object. Create multiple handles if you wish to have |
| 82 | // overlapped Execution by calling this function from different threads. |
| 83 | auto memHandle = runtime->CreateWorkingMemHandle(networkIdentifier1); |
| 84 | |
| 85 | // Execute evaluates a network using input in inputTensors and outputs filled into outputTensors. |
| 86 | // This function performs a thread safe execution of the network. Returns once execution is complete. |
| 87 | // Will block until this and any other thread using the same workingMem object completes. |
| 88 | // Execute with PreImported inputTensor1 as well as Non-PreImported inputTensor2 |
| 89 | runtime->Execute(*memHandle.get(), {}, {{2, outputTensor1}}, importedInputVec /* pre-imported ids */); |
| 90 | |
| 91 | // ImportOutputs separates the importing and mapping of OutputTensors from network execution. |
| 92 | // Allowing for a set of OutputTensors to be imported and mapped once, but used in execution many times. |
| 93 | // This function is not thread safe and must not be used while other threads are calling Execute(). |
| 94 | // Only compatible with AsyncEnabled networks |
| 95 | // Provide layerBinding Id to outputTensor1 |
| 96 | std::pair<LayerBindingId, class Tensor> output1{2, outputTensor1}; |
| 97 | // PreImport outputTensor1 |
| 98 | std::vector<ImportedOutputId> importedOutputVec = runtime->ImportOutputs(networkIdentifier1, {output1}); |
| 99 | |
| 100 | // Execute with Non-PreImported inputTensor1 as well as PreImported inputTensor2 |
| 101 | runtime->Execute(*memHandle.get(), {{0, inputTensor1}}, {{2, outputTensor1}}, {1 /* pre-imported id */}); |
| 102 | |
| 103 | // Clear the previously PreImportedInput with the network Id and inputIds returned from ImportInputs() |
| 104 | // Note: This will happen automatically during destructor of armnn::LoadedNetwork |
| 105 | runtime->ClearImportedInputs(networkIdentifier1, importedInputVec); |
| 106 | |
| 107 | // Clear the previously PreImportedOutputs with the network Id and outputIds returned from ImportOutputs() |
| 108 | // Note: This will happen automatically during destructor of armnn::LoadedNetwork |
| 109 | runtime->ClearImportedOutputs(networkIdentifier1, importedOutputVec); |
| 110 | |
| 111 | // Execute with Non-PreImported inputTensor1, inputTensor2 and the PreImported outputTensor1 |
| 112 | runtime->Execute(*memHandle.get(), {{0, inputTensor1}, {1, inputTensor2}}, {{2, outputTensor1}}); |
| 113 | |
| 114 | std::cout << "Your number was " << outputData1.data()[0] << std::endl; |
| 115 | |
| 116 | return 0; |
| 117 | } |