JVET-Q0167 Non-CE5: Multiplication simplification for ALF and CC-ALF [J. Ström, Z. Zhang, K. Andersson (Ericsson)]
This document proposes a restriction of ALF coefficients. Instead of allowing all coefficient values from 127 to 127, only values that are pure powers of two +/-{0, 1, 2, 4, 8, …} and values that are 3 times pure powers of two +/-{3, 6, 12, 24, …} are proposed to be allowed. It is claimed that this gives a substantial complexity reduction in terms of surface area of hardware implementations of ALF since every per-sample multiplication can be then replaced by an addition and shifts. Reported BDR figures are:
Test 1: Y-BDR: 0.02% (AI), 0.02% (RA), 0.00% (LDB)
YUV-BDR: 0.01% (AI), 0.01% (RA) -0.07% (LDB)
A variant (test 2) which, in addition to allowing the values of test 1, also allows coefficient values that are 5 times a pure power of two +/-{5, 10, 20, 40, …} is also presented, and it is claimed that this can also be implemented in hardware with the same number of additions as test 1. Reported BDR figures are:
Test 2: Y-BDR: 0.00% (AI), 0.00% (RA), -0.04% (LDB)
YUV-BDR: -0.01% (AI) -0.01% (RA) -0.08% (LDB).
Combinations with CC-ALF enabled are also presented for information.
Question: Is the quantiztion considered in the encoder? Yes, similar way as currently in VTM.
Question: What would happen without change of encoding? Less than 0.1%
Would only be beneficial for hardware – software would use direct multiplication.
Benefit for hardware would probably depend on architecture.
Too many changes at this late stage. The implementation benefit is not large enough to justify this.