The default version (interpolate8x8_ia64.s) of the interpolate functions suffer from some rounding errors, but are fast. I our oppinion there is no need to be such exact here. Just in the case we provide another set of function with exact rounding: interpolate8x8_ia64_exact.s