Method configure

src/gpu/cl/kernels/ClScaleKernel.cpp:106–231

}

void ClScaleKernel::configure(const CLCompileContext &compile_context,
                              ITensorInfo            *src,
                              ITensorInfo            *dst,
                              const ScaleKernelInfo  &info)
{
    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, info));
    auto padding_info = get_padding_info({src, dst});

    // Info required for the static tuning
    _data_layout = info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : info.data_layout;

    const bool is_nhwc = _data_layout == DataLayout::NHWC;

    float scale_x              = 0.f;
    float scale_y              = 0.f;
    std::tie(scale_x, scale_y) = calculate_scale_factors(src, dst, _data_layout, info.align_corners);

    // Area interpolation behaves as Nearest Neighbour in case of up-sampling
    auto interpolation_policy_to_use = info.interpolation_policy;
    if (info.interpolation_policy == InterpolationPolicy::AREA && scale_x <= 1.f && scale_y <= 1.f)
    {
        interpolation_policy_to_use = InterpolationPolicy::NEAREST_NEIGHBOR;
    }

    // Create kernel
    const int          idx_width         = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
    const int          idx_height        = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
    const int          idx_channel       = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
    const unsigned int src_width         = src->dimension(idx_width);
    const unsigned int src_height        = src->dimension(idx_height);
    const unsigned int dst_width         = dst->dimension(idx_width);
    const unsigned int dst_channels      = dst->dimension(idx_channel);
    unsigned int       vec_size          = 0;
    unsigned int       vec_size_leftover = 0;

    CLBuildOptions build_opts;
    if (_data_layout == DataLayout::NHWC)
    {
        vec_size          = adjust_vec_size(src->data_type() == DataType::F32 ? 4 : 8, dst_channels);
        vec_size_leftover = dst_channels % vec_size;
        build_opts.add_option("-DSRC_TENSOR_TYPE=BUFFER");
        build_opts.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(src->data_type()));
        build_opts.add_option("-DDST_TENSOR_TYPE=BUFFER");
        build_opts.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(dst->data_type()));
        build_opts.add_option("-DCONSTANT_VALUE=" +
                              string_from_pixel_value(info.constant_border_value, src->data_type()));
        build_opts.add_option("-DN0=" + support::cpp11::to_string(vec_size));
        build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(vec_size_leftover));
        build_opts.add_option("-DSCALE_" + string_from_interpolation_policy(interpolation_policy_to_use));
        build_opts.add_option_if(src->num_dimensions() > 3, "-DBATCHED_EXECUTION");
        build_opts.add_option_if(info.border_mode == BorderMode::REPLICATE, "-DBORDER_MODE_REPLICATE");
        build_opts.add_option_if(info.border_mode == BorderMode::CONSTANT, "-DBORDER_MODE_CONSTANT");
        build_opts.add_option_if(info.align_corners, "-DALIGN_CORNERS");
        build_opts.add_option_if(is_data_type_float(src->data_type()), "-DIS_FLOATING_POINT");
        build_opts.add_option_if_else(info.sampling_policy == SamplingPolicy::CENTER, "-DSAMPLING_POLICY_CENTER",
                                      "-DSAMPLING_POLICY_TOP_LEFT");
    }
    else if (_data_layout == DataLayout::NCHW)