Does supersampling always invoke a linear computational cost? (Eg. 4x computation for 2x2 supersampling)?