Method configure

src/gpu/cl/kernels/ClCastKernel.cpp:95–159

}

void ClCastKernel::configure(const CLCompileContext &compile_context,
                             const ITensorInfo      *src,
                             ITensorInfo            *dst,
                             ConvertPolicy           policy)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);

    const DataType src_dtype = src->data_type();
    const DataType dst_dtype = dst->data_type();

    // Auto initialize dst shape if not initialized (We can only auto-configure the shape, datatype must be given)
    set_shape_if_empty(*dst, src->tensor_shape());

    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, policy));

    auto padding_info = get_padding_info({src, dst});

    // Get data sizes
    const size_t src_size = data_size_from_type(src_dtype);
    const size_t dst_size = data_size_from_type(dst_dtype);

    // Get number of elements to process per iterations
    const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / src->element_size(), src->dimension(0));

    // Set build options

    CLBuildOptions build_opts;
    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
    build_opts.add_option("-DVEC_SIZE_LEFTOVER=" +
                          support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration));
    build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(src_dtype));
    build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(dst_dtype));
    // Conversions from float always SATURATE as out-of-bounds conversion from float->integer is implementation defined
    build_opts.add_option_if(is_data_type_float(src_dtype) || policy == ConvertPolicy::SATURATE, "-DSATURATE");
    build_opts.add_option_if(dst_dtype == DataType::QASYMM8 && is_data_type_quantized_per_channel(src_dtype),
                             "-DQSYMM8_PER_CHANNEL_TO_QASYMM8");

    // Create kernel
    const std::string kernel_name = (src_size >= dst_size) ? "cast_down" : "cast_up";
    _kernel                       = create_kernel(compile_context, kernel_name, build_opts.options());

    // Configure kernel
    Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration));
    ICLKernel::configure_internal(win);

    // Collapse window
    const Window &full_window      = window();
    Window        collapsed_window = full_window.collapse_if_possible(full_window, Window::DimZ);
    ICLKernel::configure_internal(collapsed_window);

    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));

    // Set config_id for enabling LWS tuning
    _config_id = kernel_name;
    _config_id += "_";
    _config_id += lower_string(string_from_data_type(src_dtype));
    _config_id += "_";
    _config_id += support::cpp11::to_string(src->dimension(0));