This is the same as what was done for the fguv function, to reduce
the amount of space used for it (and also simplifying the calling
code).

This gives no significant slowdown for the case currently benchmarked
by checkasm, while shrinking the code produced by film_grain.S by
320 bytes.