void downscale(ImageBuffer& im_src)
{
    ImageBuffer im_dst;
    im_dst.wid = im_src.wid / 2;
    im_dst.hei = im_src.hei / 2;
    im_dst.buffer.resize(uint64_t(im_dst.wid) * im_dst.hei * 4);

    const unsigned W = im_src.wid * 4;

    #pragma omp parallel for
    for(unsigned y=0; y<im_dst.hei; y++) {
        for(unsigned x=0; x<im_dst.wid; x++) {
            unsigned src_x = x * 2;
            unsigned src_y = y * 2;
            unsigned src_idx = (src_x + src_y * im_src.wid) * 4;
            unsigned dst_idx = (x + y*im_dst.wid) * 4;

            for(unsigned n=0; n<4; ++n)
            {
                im_dst.buffer[dst_idx+n] = (0u +
                        uint8_t(im_src.buffer[src_idx+n]) +
                        uint8_t(im_src.buffer[src_idx+n+4]) +
                        uint8_t(im_src.buffer[src_idx+n+W]) +
                        uint8_t(im_src.buffer[src_idx+n+W+4]))/4;
            }
        }
    }

    std::swap(im_src, im_dst);
}