Method configure

src/gpu/cl/kernels/ClDequantizeKernel.cpp:81–133

}

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

    // Output tensor auto initialization if not yet initialized
    auto_init_if_empty(*dst, src->tensor_shape(), 1, DataType::F32);

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

    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst));

    const int  vec_size_x     = 16 / dst->element_size();
    const int  output_width_x = dst->tensor_shape().x();
    const bool multi_access_x = (output_width_x / vec_size_x > 0);

    const bool  is_quantized_per_channel = is_data_type_quantized_per_channel(src->data_type());
    std::string kernel_name              = "dequantization_layer";

    // Create kernel
    CLBuildOptions build_opts;
    if (!is_quantized_per_channel)
    {
        const UniformQuantizationInfo qinfo   = src->quantization_info().uniform();
        const int                     qoffset = is_data_type_quantized_asymmetric(src->data_type()) ? qinfo.offset : 0;
        build_opts.add_option("-DSCALE=" + float_to_string_with_full_precision(qinfo.scale));
        build_opts.add_option("-DOFFSET=" + support::cpp11::to_string(qoffset));
    }
    else
    {
        kernel_name += "_per_channel";
        kernel_name += src->data_layout() == DataLayout::NCHW ? "_nchw" : "_nhwc";
    }

    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
    build_opts.add_option("-DDATA_TYPE_SRC=" + get_cl_type_from_data_type(src->data_type()));
    build_opts.add_option("-DDATA_TYPE_DST=" + get_cl_type_from_data_type(dst->data_type()));
    build_opts.add_option_if(multi_access_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(
                                                                        std::max<int>(output_width_x - vec_size_x, 0)));

    // Create kernel name
    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());

    // Configure kernel window
    Window win = calculate_max_window(*dst);
    if (multi_access_x)
    {
        win.set(Window::DimX,
                Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x));
    }
    ICLKernel::configure_internal(win);

    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
}

Status ClDequantizeKernel::validate(const ITensorInfo *src, const ITensorInfo *dst)
{