JVET-R0176 On chroma LFNST [C. Rosewarne, J. Gan (Canon)]
With the adoption of JVET-Q0784 LFNST is only able to be applied in the luma channel for slices using a shared coding tree. This results in LFNST application in chroma only applicable to chroma branches of separate coding trees, which are only possible in Intra slices (occurring once every IntraPeriod in random access configuration). To reduce the computational burden of LFNST, this contribution proposes to remove LFNST from chroma altogether. As a consequence, the lfnst_idx can be moved to just after the luma residual last significant position, for latency benefit. BD-rate effect for AI is 0.27%, 1.24%, 1.42%, and for RA is 0.16%, 0.37%, 0.48% for Y, Cb, and Cr channels, respectively. With the number of luma intra modes for full RDO increased by 2 for each block size, the results are: AI is 0.13%, 1.20%, 1.40%, RA is 0.08%, 0.29%, 0.47% for Y, Cb, Cr channels, respectively.
There is not latency or complexity problem with current design. The current design probably has some ugliness, but implementation does not seem to be a problem. The proposal introduces loss which is not a good tradeoff versus the benefit of removing some of the ugliness.
No action was taken on this.