JVET-S0232 Removal of redundant clipping operations on inter prediction samples in forward luma mapping [X. Xiu, Y.-W. Chen, T.-C. Ma, H.-J. Jhu, X. Wang (Kwai), Tzu-Der Chuang, Ching-Yeh Chen, Chih-Wei Hsu, Yu-Wen Huang (MediaTek)]
When LMCS is applied, for inter and CIIP modes, forward luma mapping is applied to convert the inter prediction samples of luma component before they are added with the corresponding residual samples. In this contribution, one mismatch is identified between VVC draft 9 and VTM-9.0 software on the clipping operations that are applied to luma inter prediction samples when the LMCS is applied. Specifically, in VVC draft 9, the luma inter prediction samples of the CIIP mode are clipped to the full range of internal bit-depth after the forward luma mapping is applied. Meanwhile, such clipping operation is not applied to luma prediction samples of inter modes. On the other hand, in VTM-9.0, the inter prediction samples of luma component for both inter and CIIP modes are always clipped after the forward luma mapping is applied. It is asserted that all the above clipping operations are redundant. For a more consistent and clean design, it is proposed to remove all those unnecessary clipping operations from the current specification and the reference software. Simulation results prove the redundancy of the removed clipping operations by providing bit-exact BD-rate performance.
This relates to ticket #1027.
It was agreed by other experts that the clipping is not needed in the forward mapping. There are bitstream constraints in the spec that prescribe that the values are never out of range, so clipping would only be required in case of a non-conformant bitstream.
This was further discussed at 2045 on 29 June (chaired by GJS & JRO).
After study, it was determined that the clipping is needed in some corner case for 16 bit video. This does not affect the defined profiles.
Editor action item (editorial cleanup spec – no effect on v1 profiles): It is suggested to align the clipping in the forward mapping of LM to align the spec with the software (perhaps conditioning it on the bit depth to make it clear that this does not affect v1 profiles).