Wan teacher "H.265 / HEVC" Chapter 10: Code Parallel Processing

This blog is conducted in parallel processing of H.265 / HEVC. First understand the thread: thread is part of the process. The process can be understood here as a video coding, that is, the process encodes a video, then threads can be considered as encoding a GOP, encoding a slice, encoding a CTU, etc. It can be considered that the process contains many threads. 1 Parallel processing technology H.265's high complexity source: more intra mode, more flexible division, TU block introduction. Parallel method: Multiple processors, multi-core parallelism can be doubled to increase the decoding speed. Function parallel: prediction, transformation, filtering, entropy encoding function module together. Data parallel: The data is completely independent, TILE is carried out together; in addition, the CTU level is parallel. The function in H.265 is parallel: Entropy decoding and the remaining modules are parallel to decode serial. The data in H.265 is parallel: Tile is parallel and the pair of pairs of pairs. Parallel treatment defect: the yield soaring problem: At the same time, many contents can result in a lot of bits, and the code rate is not stable. Load imbalance: There are many tasks to do, but there are very few threads to do, this imbalance is in parallel, and the SLICE level will appear. 2 specific application 1, GOP level parallel: A thread is used to coordinate control, multiple thread parallel processing GOP, this method is more suitable for encoding ends. 2, image level parallel: Due to inter prediction, it can be applied under the same level B frame. 3, SLICE level parallel: Each slice is independent of each other, so it is entirely possible. 4, TILE level parallel: Tile is a rectangular CTU group, introducing mainly for parallel processing, which changes the scanning order of the CTU, reducing inter-frame cache, but ignores correlation can result in performance loss. 5, the CTU level is parallel: according to the line of CTU, multi-line CTU is handled simultaneously, but the latter line is slow 2 CTUs than the previous line to ensure the original performance. The so-called parallel parallel (WPP) refers to the parallel processing of the CTU is not independent and needs to use each other. If the intra prediction requires the first line of the left and one column reference pixel, the inter prediction requires the MV information of the upper side block and the left block, and the loop filtering process will cross the CTU and various boundaries. 6, dependent sheet parallel: divide Slice into different regions, packaged into separate NAL, respectively. Decoding the current dependency sheet only needs to wait for a part of the reference to complete the decoding.

Our other product:

Professional lightning protection system for Tower	Professional FM Radio Station Equipment Package	FMUSER FU-1000C 1000W FM Transmitter With USB Audio Input Interface
FMUSER 200W FM Radio Broadcasting Package	Hotel IPTV Solution	FMUSER FSN-1000T V5.0 1KW FM Radio Broadcast Transmitter