JVET-Q0500 AHG16: On derivation of CCLM predictors [Y.-W. Chen, X. Xiu, T.-C. Ma, H.-J. Jhu, X. Wang (Kwai Inc.)]
In the current CCLM design, different filters are used to generate the down-sampled luma samples depending on whether the luma samples are available or not, which is not friendly for certain parallel processing such as single instruction multiple data (SIMD) operation. In this proposal, it is proposed to perform repetitive padding for those unavailable luma samples, and apply the same 6-taps filter to generate all the down-sampled luma samples. The results show 0.00% (AI), 0.01%(RA) and 0.00%(LDB) BD-rate impact with negligible run time changes under CTC. It is also asserted that with the proposed method, the VVC spec can be reduced by around 100 lines in specifying the decoding process of CCLM mode.
Loss in chroma is slightly higher (not severe), which is obvious as CCLM is a chroma tool).
Cross-checkers express that the software still needs to check for picture border and boundary, the benefit in implementation is not really obvious, and may depend on architecture. SIMD advantage, but would only be effective at CTU boundary.
Main advantage would be the simplification of spec text. Cross-checker confirms that SW and text are matching.
Some support expressed.
Decision (cleanup): Adopt JVET-Q0500.