JVET-R0452 CCLM: common text for spec bugfixes [L. Li, X. Li, S. Liu (Tencent), A. Filippov, V. Rufitskiy, E. Alshina (Huawei)] [late]
This document provides a common CCLM text to fix several spec bugs in the current VVC specification. Several tickets have been reported in VVC bug tracker. It is asserted that the behavior of VTM-8.0 is correct, and spec should be aligned with SW.
Following aspects are proposed to change regarding VVC Draft 8:
Aspect 1: Wrong upper bounds of numTopRight and numLeftBelow (ticket #1002)
Aspect 2: Undefined luma neighbour samples are used in down-sample process (ticket #796)
Aspect 3: Remove unused one-dimensional filter coefficients array F1. (leftover from last meeting’s adoption JVET-Q0500)
It is proposed to change the spec and and aligning it with the software.
It is mentioned that the second mismatch pointed out in JVET-R0314 relates to aspect 2 here (tickets #1011 and #796 are somewhat overlapping), but JVET-R0314 covers yet another aspect.
This aspect was agreed in principle. Proponents of JVET-R0452, JVET-R0375 and JVET-R0314 were asked to sort out the editorially best solution and come with a combined text. The first mismatch of JVET-R0314 (relating to ticket #1012) is not included in JVET-R0452, and should also become part of the combined text.
v6 of the document was later provided with an integrated text, presented Monday 20 April 1650. It was generally agreed that this is a step in the correct direction to resolve the SW/text mismatch. It was however pointed out that still cases could occur with contradicting numSampL/numSampT from (353-356) versus availability checks, which could result in unassigned values (in particular whenever negative x in combination with positive y comes up). Also, the mismatch in terms of 4:2:2 (see below under JVET-R0471) was missing.
This was agreed by proponents, although further fixes were confirmed to be necessary.