Back to Search Document details
23rd Meeting: by teleconference, July 2021 2021-07-05 09:42
AHG8: Removal of a prevention mechanism of extended precision for low bit-depth
Abstract
In the latest WD for VVC ver.2, there is a prevention mechanism to deactivate extended precision even when sps-level enabled flag (extended_precision_processing_flag) is set to 1. This prevention mechanism loses an encoder flexibility to use extended precision for low bit-depth. In addition, extended precision provides minor gain (roughly -0.1% for AI, -0.1% for RA and -0.1% for LB) for YUV420 10-bit sequences at low QP. This contribution proposes to remove its prevention mechanism, which gives encoders flexibility to use extended precision for low bit-depth (proposal#1).
JVET-W0109 AHG8: Removal of a prevention mechanism of extended precision for low bit-depth [T. Tsukuba, M. Ikeda, T. Suzuki (Sony)]

In the latest WD for VVC ver.2, there is a prevention mechanism to deactivate extended precision even when sps-level enabled flag (extended_precision_processing_flag) is set to 1. This prevention mechanism loses an encoder flexibility to use extended precision for low bit-depth. In addition, extended precision provides minor gain (roughly -0.1% for AI, -0.1% for RA and -0.1% for LB) for YUV420 10-bit sequences at low QP. This contribution proposes to remove its prevention mechanism, which gives encoders flexibility to use extended precision for low bit-depth (proposal#1).

If necessary to disable extended precision for low bit-depth coding, rather than having a prevention mechanism, this contribution suggests to:

  • Impose a bitstream constraint which indicates such a tool shall be disabled for bit-depth less than or equal to 10 in the profile like HEVC or semantics (proposal#2), or
  • Omit signalling of extended_precision_processing_flag depending on BitDepth to avoid unnecessary signalling (proposal#3).

The contribution shows that there is a small benefit of extended precision usage and Rice parameter modification even for 10 bit video in low QP range.

Certain application domain standards prescribe using specific profiles. This could be incurring cases where content with lower bit depth is coded in a higher bit depth profile.

The current design prevents usage of the extended precision mechanism for bit depth <= 10 (even if set on by the encoder) by a block-level check. A smart decoder could just overwrite such a flag. Contributions JVET-W0070 and JVET-W0109 try to resolve this contradiction by either a bitstream restriction, or a high-level mechanism.

After some discussion, it is concluded that imposing a semantics constraint (extended precision flag shall be equal to zero when bit depth is <= 10) is the most appropriate solution. This would be approximately “proposal 2 case 1” from JVET-W0109. At the same time, removal of the block-level condition becomes purely editorial.

Decision: Adopt JVET-W0109 “proposal 2 case 1”, up to discretion of editors to align with block-level.

A similar approach should be taken for the Rice parameter extension mechanism (or potentially other mechanisms of v2 that currently have low-level check for <= 10 bits).

Proponents of JVET-W0070 and JVET-W0109 were asked to discuss this offline and propose an appropriate text. See JVET-W0178.

Decisions
adopted
Adopt JVET-W0109 “proposal 2 case 1”, up to discretion of editors to align with block-level
Citation