Block Structures and Parallelism Features in HEVC - High Efficiency Video Coding (HEVC) - page 87

Fig. 3.13 Ultra-low delay operation mode combined with high-level parallelism tools of tiles ( left )
and WPP ( right ) with end of DU indicated by crosses

buffer (DPB) removal for the whole access unit, a timestamp for each DU. The DU
timing gives a separate, possibly earlier timing for decoding of the DU compared
to the whole picture. This has the merits that parts of the picture may be encoded
by the sender, transmitted to the receiver and decoded at the receiver earlier than if
the whole access unit is processed at once. This means that this indication of earlier
decoding times of sub-pictures allows to reduce the whole end-to-end delay of a
capturing, coding, transmission, decoding and presentation chain [ 8 ].
In ultra-low delay applications, such as remote-video or broadcast contribution
where delays below picture durations are required, the encoder needs to output a
picture partition in the form of a DU to the transmission chain as soon as encoding
is finished.
HEVC further provides the ability of using high-level parallelization techniques
in order to reduce processing demands in multiprocessing unit environments.
Therefore, HEVC allows for subdividing the picture into tiles or WPP substreams,
i.e., rows of CTUs as described above in Sect. 3.3.2 . Either of the methods may be
used with ultra-low latency operations, where the WPP case can only be achieved
using dependent slice segments, as discussed in Sect. 3.3.1 .
The ultra-low latency mode for high-level parallelism is shown in Fig. 3.13 .The
left part of the Fig. 3.13 shows the coding process of tiles, where the first three tiles
(marked by cross) are bound to the same decoding unit. In the right part of Fig. 3.13 ,
six CTU rows are provided, where each CTU row belongs to a single decoding unit,
each consisting of a single slice segment (and marked by a cross).

3.3.4 Summary of High-Level Parallelization Tools

It is clear from the previous sections that tiles and WPP have different pros and cons.
WPP is generally well-suited for the parallelization of the encoder and decoder
because it allows a high number of picture partitions with relatively low coding-
efficiency losses. Additionally, WPP does not need an additional pass of in-loop
filtering in comparison to slices and tiles [ 41 ]. WPP can also be used for low-delay
applications, especially those requiring sub-picture delay (also called ultra-low
delay), as described in Sect. 3.3.3 .