#923 closed defect (worksforme)

Editorial modification: GEO size constraint on 8x64/64x8

Reported by: chunchic Owned by:
Priority: minor Milestone:
Component: spec Version: VVC D8 vB
Keywords: Cc: ksuehring, XiangLi, fbossen, jvet@…, bbross

Description

According to the meeting notes which records "Decision: Stay with the 64-combination method from the CE, except disable GEO for 8x64 and 64x8", the decision made for GEO CU size constraint is to disable 8x64 and 64x8 CUs.
However, the both the spec (verison D) and VTM-8.0 implemeted this adoption differentely as:

[JVET-2001-vD]

  • Section 7.3.10.7:
    if( sps_ciip_enabled_flag  &&  sps_gpm_enabled_flag  &&
    	slice_type  = =  B  &&
    	cu_skip_flag[ x0 ][ y0 ]  = =  0  &&  cbWidth  >=  8  &&  cbHeight  >=  8  &&
    	'''cbWidth  <  ( 8 * cbHeight )  &&  cbHeight  <  ( 8 * cbWidth )  &&'''
    	cbWidth < 128  &&  cbHeight < 128 )
    

[VTM-8.0]

  • CABACReader.cpp at Line 2319 and CABACWriter.cpp at Line 1993:
    const bool geoAvailable = pu.cu->cs->slice->getSPS()->getUseGeo() && pu.cu->cs->slice->isInterB() && pu.cu->cs->picHeader->getMaxNumGeoCand() > 1
                                                                      && pu.cu->lwidth() >= GEO_MIN_CU_SIZE && pu.cu->lheight() >= GEO_MIN_CU_SIZE
                                                                      && pu.cu->lwidth() <= GEO_MAX_CU_SIZE && pu.cu->lheight() <= GEO_MAX_CU_SIZE
                                                                     ''' && pu.cu->lwidth() < 8 * pu.cu->lheight() && pu.cu->lheight() < 8 * pu.cu->lwidth();'''
    

Although CU's width-to-height ratio of either 1:8 or 8:1 covers the 8x64 and 64x8 CUs, these ratios may not be identical to "8x64/64x8" that are recorded in the meeting notes.
It seems the meeting notes may refer to JVET-Q0123 for adoption, although the it does not record the document number explicitly.
To reflect to the adoption of GEO size constraint recorded in the meeting notes, both spec and VTM-8.0 are supposed to be changed based on JVET-Q0123, which is "( cbWidth + cbHeight ) != 72".
As the spec and VTM software did similarly for disallowing 8x4/4x8 bi-prediction by explicitly specifying "The value of ( cbWidth + cbHeight ) is equal to 12", here it may be more adequate to follow this same editorial fashion to disable 8x64/64x8 GEO (that is cbWidth + cbHeight != 72 as in JVET-Q0123), as follows.

[JVET-2001-vD]

  • Section 7.3.10.7:
    if( sps_ciip_enabled_flag  &&  sps_gpm_enabled_flag  &&
    	slice_type  = =  B  &&
    	cu_skip_flag[ x0 ][ y0 ]  = =  0  &&  cbWidth  >=  8  &&  cbHeight  >=  8  &&
    	'''( cbWidth + cbHeight ) != 72 &&'''
    	cbWidth < 128  &&  cbHeight < 128 )
    

[VTM-8.0]

  • CABACReader.cpp at Line 2319 and CABACWriter.cpp at Line 1993:
    const bool geoAvailable = pu.cu->cs->slice->getSPS()->getUseGeo() && pu.cu->cs->slice->isInterB() && pu.cu->cs->picHeader->getMaxNumGeoCand() > 1
                                                                      && pu.cu->lwidth() >= GEO_MIN_CU_SIZE && pu.cu->lheight() >= GEO_MIN_CU_SIZE
                                                                      && pu.cu->lwidth() <= GEO_MAX_CU_SIZE && pu.cu->lheight() <= GEO_MAX_CU_SIZE
                                                                      '''&& (pu.cu->lwidth() + pu.cu->lheight()) != 72;'''
    

Change history (4)

comment:1 follow-up: Changed 19 months ago by LiZhang

The current text and SW are aligned, and the way the syntax condition is expressed is better for understanding that the purpose is to disallow the 8x64 and 64x8 sizes. And the meeting minutes were clear of adopting the joint design/text in JVET-Q0806, which contains the same as in the current draft text in this regard.

comment:2 in reply to: ↑ 1 Changed 19 months ago by chunchic

I still got confused about what the JVET-O0806 proposes and what the meeting notes records. Just to make it clear.
In JVET-O0806, the proposed technology is "JVET-Q0123/JVET-Q309: Remove 64x8 and 8x64 for GEO".
In meeting notes (version D), the two decisions below are made for GEO.
(1) Decision: adopt JVET-Q0806. Relative to v3, remove notes in 8.5.7.2. This includes all decisions made on JVET-Q0059, JVET-Q0077, JVET-Q0123, JVET-Q0188, JVET-Q0242, JVET-Q0309, JVET-Q0365 and JVET-Q0370.
(2) Decision: Stay with the 64-combination method from the CE, except disable GEO for 8x64 and 64x8.

Based on these, my understanding is that the adoption is to disable 8x64/64x8 GEO, but not the 1:8/8:1 CUs.
Editorial-wise, I think spec/sw are supposed to be implemented in the same way as proposed in a proposal, if not otherwise stated or serious bugs found.
I failed to find any information regarding the mentioned "the syntax condition is expressed is better" neither from the meeting notes nor from JVET-Q0806.
Would you please help indicate out where it is documented?

Replying to LiZhang:

The current text and SW are aligned, and the way the syntax condition is expressed is better for understanding that the purpose is to disallow the 8x64 and 64x8 sizes. And the meeting minutes were clear of adopting the joint design/text in JVET-Q0806, which contains the same as in the current draft text in this regard.

comment:3 Changed 19 months ago by yk

Understanding based on the comments above, this seems to be purely editorial issue, and thus can be decided by editors at their discretion. While I agree with Li's comment, I work for the same company as Li. Therefore I suggest that either Ben or Shan (Jianle works for the other company involved in the discussion here :-)) to take a look and decide.

comment:4 Changed 15 months ago by bbross

  • Resolution set to worksforme
  • Status changed from new to closed
Note: See TracTickets for help on using tickets.