Function load_from_path_run_cuda

lite/example/cpp_example/mge/basic.cpp:450–501

#if LITE_WITH_CUDA
namespace {
bool load_from_path_run_cuda(const Args& args) {
    std::string network_path = args.model_path;
    std::string input_path = args.input_path;
    set_log_level(LiteLogLevel::DEBUG);
    //! config the network running in CUDA device
    lite::Config config{false, -1, LiteDeviceType::LITE_CUDA};
    //! set NetworkIO
    NetworkIO network_io;
    std::string input_name = "img0_comp_fullface";
    bool is_host = false;
    IO device_input{input_name, is_host};
    network_io.inputs.push_back(device_input);
    //! create and load the network
    std::shared_ptr<Network> network = std::make_shared<Network>(config, network_io);
    network->load_model(network_path);

    std::shared_ptr<Tensor> input_tensor = network->get_input_tensor(0);
    Layout input_layout = input_tensor->get_layout();

    //! read data from numpy data file
    auto src_tensor = parse_npy(input_path);

    //! malloc the device memory
    auto tensor_device = Tensor(LiteDeviceType::LITE_CUDA, input_layout);

    //! copy to the device memory
    tensor_device.copy_from(*src_tensor);

    //! Now the device memory if filled with user input data, set it to the
    //! input tensor
    input_tensor->reset(tensor_device.get_memory_ptr(), input_layout);

    //! forward
    {
        lite_example_helper::Timer ltimer("warmup");
        network->forward();
        network->wait();
        ltimer.print_used_time(0);
    }
    lite_example_helper::Timer ltimer("forward_iter");
    for (int i = 0; i < 10; i++) {
        ltimer.reset_start();
        network->forward();
        network->wait();
        ltimer.print_used_time(i);
    }
    //! get the output data or read tensor set in network_in
    size_t output_size = network->get_all_output_name().size();
    output_info(network, output_size);
    output_data_info(network, output_size);
    return true;
}
}  // namespace

REGIST_EXAMPLE("load_from_path_run_cuda", load_from_path_run_cuda);