Function main

example/matrix_transpose.cpp:219–355

#endif

int main(int argc, char *argv[])
{
    // setup command line arguments
    po::options_description options("options");
    options.add_options()
        ("help", "show usage instructions")
        ("rows", po::value<uint_>()->default_value(4096), "number of matrix rows")
        ("cols", po::value<uint_>()->default_value(4096), "number of matrix columns")
    ;

    // parse command line
    po::variables_map vm;
    po::store(po::parse_command_line(argc, argv, options), vm);
    po::notify(vm);

    // check command line arguments
    if(vm.count("help")){
        std::cout << options << std::endl;
        return 0;
    }

    // get number rows and columns for the matrix
    const uint_ rows = vm["rows"].as<uint_>();
    const uint_ cols = vm["cols"].as<uint_>();

    // get the default device
    compute::device device = compute::system::default_device();

    // print out device name and matrix information
    std::cout << "Device: " << device.name() << std::endl;
    std::cout << "Matrix Size: " << rows << "x" << cols << std::endl;
    std::cout << "Grid Size: " << rows/TILE_DIM << "x" << cols/TILE_DIM << " blocks" << std::endl;
    std::cout << "Local Size: " << TILE_DIM << "x" << BLOCK_ROWS << " threads" << std::endl;
    std::cout << std::endl;

    // On OSX this example does not work on CPU devices
    #if defined(__APPLE__)
    if(device.type() & compute::device::cpu) {
        std::cout << "On OSX this example does not work on CPU devices" << std::endl;
        return 0;
    }
    #endif

    const size_t global_work_size[2] = {rows, cols*BLOCK_ROWS/TILE_DIM};
    const size_t local_work_size[2] = {TILE_DIM, BLOCK_ROWS};

    // setup input data on the host
    const uint_ size = rows * cols;
    std::vector<float> h_input(size);
    std::vector<float> h_output(size);
    std::vector<float> expectedResult(size);
    generate_matrix(h_input, expectedResult, rows, cols);

    // create a context for the device
    compute::context context(device);

    // device vectors
    compute::vector<float> d_input(size, context);