Added workaround for Nvidia: take into account that 3-channel vector type takes 4*elem_size in local memory.