harness.cpp

#include <armnnDeserializer/IDeserializer.hpp>
#include <armnn/IRuntime.hpp>

#include <cstdint>
#include <fstream>
#include <iostream>
#include <iterator>
#include <string>
#include <vector>

void PrintOutputBinding(
    const armnnDeserializer::IDeserializer& deserializer,
    const std::string& name)
{
    const armnnDeserializer::BindingPointInfo binding =
        deserializer.GetNetworkOutputBindingInfo(0, name);
    const armnn::TensorShape& shape = binding.m_TensorInfo.GetShape();

    std::cout << name << " binding=" << binding.m_BindingId << " shape=[";
    for (unsigned int i = 0; i < shape.GetNumDimensions(); ++i)
    {
        if (i != 0)
        {
            std::cout << ",";
        }
        std::cout << shape[i];
    }
    std::cout << "] elements=" << binding.m_TensorInfo.GetNumElements() << "\n";
}

void PrintRuntimeOutput(
    const armnn::IRuntime& runtime,
    armnn::NetworkId networkId,
    armnn::LayerBindingId bindingId)
{
    const armnn::TensorInfo& info =
        runtime.GetOutputTensorInfo(networkId, bindingId);
    const armnn::TensorShape& shape = info.GetShape();

    std::cout << "runtime output " << bindingId << " shape=[";
    for (unsigned int i = 0; i < shape.GetNumDimensions(); ++i)
    {
        if (i != 0)
        {
            std::cout << ",";
        }
        std::cout << shape[i];
    }
    std::cout << "] elements=" << info.GetNumElements() << "\n";
}

int main(int argc, char** argv)
{
    if (argc != 2)
    {
        std::cerr << "usage: " << argv[0] << " MODEL.armnn\n";
        return 2;
    }

    std::ifstream input(argv[1], std::ios::binary);
    if (!input)
    {
        std::cerr << "could not open " << argv[1] << "\n";
        return 2;
    }

    std::vector<uint8_t> bytes(
        (std::istreambuf_iterator<char>(input)),
        std::istreambuf_iterator<char>());

    try
    {
        auto deserializer = armnnDeserializer::IDeserializer::Create();
        auto network = deserializer->CreateNetworkFromBinary(bytes);
        std::cout << "loaded " << bytes.size() << " bytes\n";
        PrintOutputBinding(*deserializer, "output-0");
        PrintOutputBinding(*deserializer, "output-1");

        auto runtime = armnn::IRuntime::Create(armnn::IRuntime::CreationOptions());
        auto optimized = armnn::Optimize(
            *network, {armnn::Compute::CpuRef}, runtime->GetDeviceSpec());
        if (!optimized)
        {
            std::cerr << "optimization failed\n";
            return 1;
        }

        armnn::NetworkId networkId = -1;
        std::string errorMessage;
        if (runtime->LoadNetwork(networkId, std::move(optimized), errorMessage)
            != armnn::Status::Success)
        {
            std::cerr << "runtime load failed: " << errorMessage << "\n";
            return 1;
        }

        PrintRuntimeOutput(*runtime, networkId, 0);
        PrintRuntimeOutput(*runtime, networkId, 1);

        auto inputBinding =
            deserializer->GetNetworkInputBindingInfo(0, "input");
        auto output0Binding =
            deserializer->GetNetworkOutputBindingInfo(0, "output-0");
        auto output1Binding =
            deserializer->GetNetworkOutputBindingInfo(0, "output-1");

        inputBinding.m_TensorInfo.SetConstant(true);
        std::vector<float> inputData{1.0F, 2.0F, 3.0F, 4.0F};
        std::vector<float> output0Data(
            output0Binding.m_TensorInfo.GetNumElements());
        std::vector<float> output1Data(
            output1Binding.m_TensorInfo.GetNumElements());

        armnn::InputTensors inputTensors{{
            inputBinding.m_BindingId,
            armnn::ConstTensor(inputBinding.m_TensorInfo, inputData.data())
        }};
        armnn::OutputTensors outputTensors{
            {
                output0Binding.m_BindingId,
                armnn::Tensor(output0Binding.m_TensorInfo, output0Data.data())
            },
            {
                output1Binding.m_BindingId,
                armnn::Tensor(output1Binding.m_TensorInfo, output1Data.data())
            }
        };

        const armnn::Status enqueueStatus =
            runtime->EnqueueWorkload(networkId, inputTensors, outputTensors);
        std::cout << "enqueue status="
                  << (enqueueStatus == armnn::Status::Success ? "success" : "failure")
                  << "\n";
        std::cout << "allocated output elements="
                  << output0Data.size() << "," << output1Data.size() << "\n";
        return network ? 0 : 1;
    }
    catch (const std::exception& error)
    {
        std::cerr << "load failed: " << error.what() << "\n";
        return 1;
    }
}