N0196 – Six tap interpolation filter for affine prediction

[bheng]

N0196 changes the interpolation filter to 6-tap for affine blocks.

However, the following non-normative optimization skips the affine prediction process and treats a PU as regular inter prediction if all the control point vectors are identical. Unfortunately, this optimization now changes from the new 6-tap affine filter back to the standard 8-tap filter, and therefore the behavior is no longer equivalent.

The following optimization could just be removed. Or if it is desired to keep this optimization, the N0196 code should be modified to handle this non-normative optimization case as well.

if ( (pu.cu->affineType == AFFINEMODEL_6PARAM && _mv[0] == _mv[1] && _mv[0] == _mv[2])

(pu.cu->affineType == AFFINEMODEL_4PARAM && _mv[0] == _mv[1])

)

{

Mv mvTemp = _mv[0];
clipMv( mvTemp, pu.cu->lumaPos(),

pu.cu->lumaSize(),
*pu.cs->sps );

xPredInterBlk( compID, pu, refPic, mvTemp, dstPic, bi, clpRng

, false
, false
);

return;

}

[XiangLi]

Fixed in ​https://vcgit.hhi.fraunhofer.de/jvet/VVCSoftware_VTM/merge_requests/511