JVET-P0400 AhG16/Non-CE8: Removal of shared merge list [Y. Wang, L. Zhang, H. Liu, K. Zhang, J. Xu, Y. Wang (Bytedance)]
This contribution proposes to increase the MV throughput of small blocks with IBC enabled. In the proposed method, the shared merge list is removed, and the derivation of spatial merge candidates for 4×4 blocks is skipped. Therefore, the pruning operations among BV candidates are totally removed. Additionally, updating of HMVP tables for certain block sizes (no greater than 16) is skipped as well. It is asserted that the proposed method could further reduce the complexity and increase the MV throughput. Simulation results reportedly show that BD-rate changes are {0.01%, -0.01%, -0.03%} and {0.06%, 0.02%, -0.11%} in {AI, RA, LDB} configurations for class F and class TGM, respectively.
HMVP table is never updated for case of 4x4 blocks. In the hypothetical case of only 4x4 blocks, the HMVP table would never be filled. This is no problem at the decoder; an encoder typically would not choose such a case.
The proposal simplifies the specification by removing the processes related to shared merge list. It is however introducing another condition of checking for larger block size in various steps of the merge list construction. It is argued that this would enforce duplicating the entire merge list construction logic, where the argument behind that is not fully clear.
It is agreed that this is a clear simplification which does not have significant impact on performance.
Confirmed by cros-checkers.
Decision: Adopt JVET-P0400.