International organisation for standardisation organisation internationale de normalisation

5.13.5Tiles

A BoG (coordinated by M. Horowitz) was requested to review contributions in the tiles and wavefronts categories).

5.13.5.1.1.1.1.1.1JCTVC-H0727 BoG report: tiles and wavefront parallel processing [M. Horowitz]

This BoG report contained notes including recommendations to JCT-VC on the topic of tiles, wavefront parallel processing and entropy slices. Some of the BoG recommendations are recorded in notes for particular documents as found below.

In later discussion, the following additional actions were taken.

• 
  Decision: Wavefront syntax for PPS should be placed in SPS with ability to override in PPS (not just placed in the PPS).
  
• 
  Decision: Regarding WPP and tile mixtures, send two flags in SPS with three allowed combinations
  
  • 
    0,0 = multiple tiles not allowed and no WPP allowed
    
  • 
    0,1 = multiple tiles allowed, no WPP allowed
    
  • 
    1,0 = multiple tiles not allowed, WPP is allowed
    

5.13.5.1.1.1.1.1.2JCTVC-H0737 Text for tiles, WPP and entropy slices [Y. -K. Wang (Qualcomm), A. Segall (Sharp), M. Horowitz (eBrisk), Hendry (LG Electronics), W. Wade (Broadcom), F. Henry (Orange), T. Lee (Samsung)]

The relation to the H0556 related aspects was discussed.

5.13.5.1.1.1.1.1.3JCTVC-H0179 Storage reduction for dependent tiles [E. François (Canon), S. Pautet]

The BoG recommended adoption, and the JCT-VC agreed, subject to the assumption that dependent tiles remained in the design. However, it was subsequently decided that dependent tiles would not remain in the design (as noted below), so the adoption decision became moot.

5.13.5.1.1.1.1.1.4JCTVC-H0658 AHG4: Crosscheck for Canon storage reduction for dependent tiles in JCTVC-H0179 [T.-D. Chuang, Y.-W. Huang (MediaTek)] [late]
5.13.5.1.1.1.1.1.5JCTVC-H0203 AHG4: Further latency reduction of CABAC initialization for dependent Tiles [Hendry, S. Jeong, S. Park, B. Jeon (LG)]

The BoG did not reach a consensus on adoption of this. It was designated for further study.

5.13.5.1.1.1.1.1.6JCTVC-H0414 Cross-check of "AHG4: Further latency reduction of CABAC initialization for dependent Tiles" (JCTVC-H203) [M. Horowitz, S. Xu (eBrisk)]
5.13.5.1.1.1.1.1.7JCTVC-H0329 AHG4: Crosscheck for LG further latency reduction of CABAC initialization for dependent tiles in JCTVC-H0203 [C.-W. Hsu, Y.-W. Huang (MediaTek)] [late]
5.13.5.1.1.1.1.1.8JCTVC-H0348 Simplification on tiles and slices [T. Lee, B. Choi, J. Park (Samsung)]

The BoG did not reach a consensus on adoption of this. No action was taken on it.

5.13.5.1.1.1.1.1.9JCTVC-H0593 Cross-verification of JCTVC-H0348 on simplification on tiles and slices [V. Seregin, M. Coban (Qualcomm)] [late]
5.13.5.1.1.1.1.1.10JCTVC-H0270 AHG4: Avoiding lower left LCUs at dependent tile column boundaries [C.-W. Hsu, C.-Y. Tsai, Y.-W. Huang, S. Lei (MediaTek)]

The BoG recommended adoption. The software actually currently does what is proposed, but the WD says differently. The proposal is to do what the software does. The JCT-VC agreed with the BoG, pending a decision on dependent tile support and test results comparing coding efficiency with and without the proposed constraint. However, it was subsequently decided that dependent tiles would not remain in the design (as noted below), so the adoption decision became moot.

5.13.5.1.1.1.1.1.11JCTVC-H0345 Constrained tile for parallel decoding [H. Sasai, K. Uchibayashi, T. Nishi (Panasonic)]

No action was taken on this.

5.13.5.1.1.1.1.1.12JCTVC-H0463 Comments on specifying tile support in HEVC Profiles and Levels [W. Wan, T. Hellman (Broadcom)]

The BoG recommended the adoption of the constraint that a picture shall not be composed of N tiles and M wavefronts where N > 1 and M > 1. Decision: Agreed.

Most BoG participants were not supportive of allowing "dependent" tiles at all, while some thought they should be considered for some "non-main" profile(s). The gain from this feature seemed insufficient to justify letting it become a sticking point. Decision: Remove this feature from the draft.

The decision was initially made pending review of evidence of coding efficiency effect, and was later confirmed. Coding efficiency test results for dependent tiles were later reported:

• 
  In most tests, the gain from dependent tiles was reported as not too significant.
  
• 
  For Classes A and B with four tiles structured as columns, the gain was reported as averaging about 0.4-0.5%.
  
• 
  For Class E Low-Delay B (720p60 as for videoconferencing) with 12 tiles per picture, the reported gain was reported as averaging 4-5%. In discussion, was remarked that due to the high frame rate and low scene activity for these sequences, the bit rate at which they are encoded is extremely low.
  
• 
  The use cases for tiles were discussed. It was remarked that typical use would be to use encoder parallelism in a way that would not use dependent tiles, and that typical use would not use very many tiles per picture. The use case for dependent tiles would not be parallel encoding, but rather for single-chip memory bandwidth reduction for motion search with higher coding efficiency. Since the purpose of dependent tiles is primarily a matter of coding efficiency.
  
• 
  Overall, this did not seem to be a sufficient justification for retaining the feature.
  

The BoG recommended the study of whether in-loop deblocking is sufficient to eliminate visual artefacts when SAO and ALF are active, but do not filter across vertical and/or horizontal tile boundaries. The JCT-VC agreed.

The BoG recommended that a maximum number of tile columns constraint be considered when defining profile(s) and levels. The JCT-VC agreed.

With respect to the availability of the slice header in the context of a single-core decoder wishing to reconstruct LCUs in picture scan order, two prospective solutions were identified by the BoG:

1. 
   To mandate that all LCUs in a slice shall lie at or "beyond" the first LCU of the slice in the picture raster scan (i.e. this constraint would not prevent multiple tiles to be contained in one slice),

   Note: It was remarked that the presence of an "entry point" identifying the start of each independent tile in the slice should be mandated.

   
   
2. 
   Or to mandate that slices shall not cross tile boundaries (i.e., a constraint that would prevent multiple tiles from being contained in one slice). Note that this constraint implies the constraint above.
   

The BoG recommended constraint #1. The BoG was unable to reach consensus regarding a position on constraint #2, and the BoG recommended further discussion on this topic.

In later Track A discussion, it was suggested to mandate that either all slices within a tile shall be complete or all tiles within a slice shall be complete. Decision: The JCT-VC adopted this suggestion rather than the above input from the BoG. (Note that this does not prohibit a bitstream from having both characteristics mixed within a picture.)

It was remarked that our encoding software needs to be able to operate with this constraint in effect – e.g. there could be a parameter that controls the encoder to encode slices in units of tiles. M. Horowitz volunteered to at least scope the effort for such a modification.

5.13.5.1.1.1.1.1.13JCTVC-H0520 Tile groups [Y. -K. Wang, A. K. Ramasubramonian, Y. Chen, M. Coban, M. Karczewicz (Qualcomm)]

No action was taken on this.

5.13.5.1.1.1.1.1.14JCTVC-H0521 Flexible tile dependency and loop filter control [Y. -K. Wang, J. Chon, Y. Chen, I. S. Chong, M. Coban, M. Karczewicz (Qualcomm)]

No action was taken on this.

5.13.5.1.1.1.1.1.15JCTVC-H0556 AHG4: Harmonized method for signalling entry points of tiles and WPP substreams [Hendry, S. O. Jeong, S. W. Park, B. M. Jeon (LG), K. Misra, A. Segall (Sharp)]

The BoG recommended adoption. Some aspects were further discussed (byte alignment, need for type flag), and the outcome was documented by H0737. Decision: Adopted (details clarified/refined as documented in H0737).

5.13.5.1.1.1.1.1.16JCTVC-H0370 Crosscheck of LGE/Sharp proposal for byte-aligned substreams and harmonized entry points (JTCVC-H0556) [F. Henry, G. Clare (Orange)]

5.13.5.1.1.1.1.1.17JCTVC-H0293 Syntax on parallel processing information signalling [S. S. Jeong, Y. Park, C. Kim, J. H. Kim, J. H. Park (Samsung)]

The BoG recommended investigating the placing of WPP parameters into the SPS in a manner consistent with the way tiles parameters are signalled. Such further study was encouraged.