JVET-Q0387 AHG14: Throughput and coding efficiency report of JVET-P0300 on VTM-7.0 [H. Kirchhoffer, B. Bross, T. Nguyen, D. Marpe, H. Schwarz, T. Wiegand (HHI)]
This contribution evaluates the throughput and coding efficiency of the high throughput CABAC mode proposed in JVET-P0300 based on VTM-7.0. The method was implemented into the throughput evaluation testbed of JVET-L1025 (CE5 subtest 2) in order to evaluate the coding engine outside of VTM-7.0. In this testbed, the coding engine of JVET-P0300 increases the throughput by approximately 70 to 100% compared to the VTM-7.0 engine. For three generalized bypass probability values (1/4, 1/8, and 1/16) and QPs 2-17, the BD-rates of the VTM-7.0-based implementation are 4.55% for AI, 5.22% for RA, 4.53% for LB, and 4.66% for LP. The corresponding throughput of the decoder (in mbit per second) is increased by 20% for AI, 18% for RA, 16% for LB, and 18% for LP.
Throughput of CABAC engine in current VVC is worse than HEVC, due to the modified CABAC engine.
It has also to be considered that the rate increases by the high-throughput method, which makes the throughput lower again.
It is further mentioned that throughput is not only dependent on the core engine, but also the context derivation has to be taken into account.
Also, it is mentioned that the gain of VVC over HEVC is significantly lower in the low QP range – so a loss of 5% may already be very significant there.
Questions raised:
- would a second entropy coding be desirable?
- if yes, what should be the advantage in terms of throughput?
- could some of these problems be resolved by parallelism, e.g. tiles?
- what are the use cases, and would VVC be competitive with HEVC at all for these cases?
Throughput is also highly dependent on hardware architecture.
Further study was requested – requirements need to be clarified. This would probably have no impact for VVC V1.