In TComInterpolationFilter::filter(), when isLast is true, the interpolated samples are clipped with the following code:

if ( isLast )
{

val = ( val < 0 ) ? 0 : val;
val = ( val > maxVal ) ? maxVal : val;

}

This clipping operation is not in "8.5.3.3.3 Fractional sample interpolation process". In rare cases, an interpolated sample can be negative prior to the above clipping, due to negative filter coefficients.

In the case of uni-directional prediction, the above clipping is innocuous, because there is equivalent clipping specified during uni-directional weighting.

In the case of bi-directional prediction, 2 interpolated samples are added and the result is clipped; see "8.5.3.3.4 Weighted sample prediction process", expressions 8-239 and 8-252. If either interpolated sample is negative, there can be a different result depending on if the interpolated sample was clipped in filter() or not. If an encoder "pre-clips" and a decoder does not, there can be drift.