Function main

samples/CustomMemoryAllocatorSample.cpp:64–177

A simple example application to show the usage of a custom memory allocator. In this sample, the users single input number is multiplied by 1.0f using a fully connected layer with a single neuron to produce an output number that is the same as the input. All memory required to execute this mini network is allocated with the provided custom allocator. Using a Custom Allocator is required for use w

//       are being imported instead of copied. (Import must be enabled when using a Custom Allocator)
//       You might find this useful for comparison.
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
    ConfigureLogging(true, false, LogSeverity::Info);

    // Construct ArmNN network
    NetworkId networkIdentifier;
    INetworkPtr network = INetwork::Create();
    FullyConnectedDescriptor fullyConnectedDesc;
    float weightsData[] = {1.0f}; // Identity
    TensorInfo weightsInfo(TensorShape({1, 1}), DataType::Float32, 0.0f, 0, true);
    weightsInfo.SetConstant(true);
    ConstTensor weights(weightsInfo, weightsData);

    IConnectableLayer* inputLayer   = network->AddInputLayer(0);
    IConnectableLayer* weightsLayer = network->AddConstantLayer(weights, "Weights");
    IConnectableLayer* fullyConnectedLayer =
            network->AddFullyConnectedLayer(fullyConnectedDesc, "fully connected");
    IConnectableLayer* outputLayer  = network->AddOutputLayer(0);

    inputLayer->GetOutputSlot(0).Connect(fullyConnectedLayer->GetInputSlot(0));
    weightsLayer->GetOutputSlot(0).Connect(fullyConnectedLayer->GetInputSlot(1));
    fullyConnectedLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
    weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsInfo);

    // Create ArmNN runtime:
    //
    // This is the interesting bit when executing a model with a custom allocator.
    // You can have different allocators for different backends. To support this
    // the runtime creation option has a map that takes a BackendId and the corresponding
    // allocator that should be used for that backend.
    // Only GpuAcc supports a Custom Allocator for now
    //
    // Note: This is not covered in this example but if you want to run a model on
    //       protected memory a custom allocator needs to be provided that supports
    //       protected memory allocations and the MemorySource of that allocator is
    //       set to MemorySource::DmaBufProtected
    IRuntime::CreationOptions options;
    auto customAllocator = std::make_shared<SampleClBackendCustomAllocator>();
    options.m_CustomAllocatorMap = {{"GpuAcc", std::move(customAllocator)}};
    IRuntimePtr runtime = IRuntime::Create(options);

    //Set the tensors in the network.
    TensorInfo inputTensorInfo(TensorShape({1, 1}), DataType::Float32);
    inputLayer->GetOutputSlot(0).SetTensorInfo(inputTensorInfo);

    unsigned int numElements = inputTensorInfo.GetNumElements();
    size_t totalBytes = numElements * sizeof(float);

    TensorInfo outputTensorInfo(TensorShape({1, 1}), DataType::Float32);
    fullyConnectedLayer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);