Estimate the rounding error so that we can have a useful margin of
error when checking the asm against the C code even when the float
operations get reordered due to -ffast-math.