CN100405851C - Design Method of Motion Vector Prediction Multiplexing in Multimode Standard Decoder - Google Patents

Abstract

Translated fromChinese

一种多模标准解码器中运动矢量预测复用设计方法，(1)由固件发送给MVP模块一个avs_flag标志；(2)块运动矢量预测采用分块下嵌套子块的形式；(3)取相邻块信息模块中，利用对帧场自适应标志置数，使两种不同标准可以进行相同处理，实现复用；(4)取参考宏块信息的设计，并行宏块的所有运动信息总从参考图像宏块运动信息缓存中的第一个参考图像取，更新时也总是保存到第一个参考图像中；(5)对邻块数据流更新机制的设计，使复用所需的相邻块信息量降至最少，减少了读外存次数。其优点在于，实现了H.264与AVS两种视频编码标准MVP功能的兼容，不仅保证了正确性，且可以大幅度减少解码器硬件面积。经验证，可以减少75%的片上缓存，资源占用情况是单独同时实现两个标准的70%，能实现对高清图像的实时解码。A method for designing multiplexing of motion vector prediction in a multi-mode standard decoder, (1) sending an avs_flag sign to the MVP module by firmware; (2) block motion vector prediction adopts the form of nesting sub-blocks under the block; (3) In the adjacent block information module, the frame field adaptive flag is used to set the number, so that the two different standards can be processed in the same way to achieve multiplexing; (4) The design of taking reference macroblock information, all motion information of parallel macroblocks Always take from the first reference image in the reference image macroblock motion information buffer, and always save it in the first reference image when updating; (5) The design of the adjacent block data stream update mechanism makes the multiplexing required The amount of adjacent block information is minimized, reducing the number of times to read external memory. The advantage is that it realizes the compatibility of the MVP function of the two video coding standards of H.264 and AVS, which not only ensures the correctness, but also greatly reduces the hardware area of the decoder. It has been verified that the on-chip cache can be reduced by 75%, and the resource occupation is 70% of the two standards independently and simultaneously, enabling real-time decoding of high-definition images.

Description

Translated fromChinese

多模标准解码器中运动矢量预测复用设计方法Design Method of Motion Vector Prediction Multiplexing in Multimode Standard Decoder

技术领域technical field

本发明涉及一种运动矢量预测复用设计方法，特别涉及一种在H.264/AVS多模标准解码器中的运动矢量预测(MVP)复用设计方法。The invention relates to a motion vector prediction multiplexing design method, in particular to a motion vector prediction (MVP) multiplexing design method in an H.264/AVS multimode standard decoder.

背景技术Background technique

AVS是中国第一个具有自主知识产权的数字音视频编解码标准，全称是《信息技术先进音视频编码标准》。H.264是由ITU-T和ISO两个国际标准化组织的有关视频编码专家联合组成的视频联合工作组(JVT)开发制定的一个视频编码标准，这两种标准在编码效率上较之以往的视频压缩标准都有了显著的提高，但是与此同时硬件实现解码的复杂度也随之上升，给硬件设计带来一定的挑战。AVS is China's first digital audio and video codec standard with independent intellectual property rights, and its full name is "Advanced Audio and Video Coding Standard for Information Technology". H.264 is a video coding standard developed by the Joint Video Working Group (JVT) composed of relevant video coding experts from ITU-T and ISO two international standardization organizations. Video compression standards have been significantly improved, but at the same time, the complexity of hardware decoding has also increased, which brings certain challenges to hardware design.

在实时的高清视频解码运算中，需要大量的运算逻辑且有大量中间数据需要保存和读取。尤其是要实现同时支持AVS及H.264标准的高清解码器，如果简单地将两个解码器叠加，其面积将非常巨大。在这种情况下，如何充分利用标准的相关性来实现解码器的最大程度复用就变得越来越重要。In the real-time high-definition video decoding operation, a large amount of operation logic is required and a large amount of intermediate data needs to be saved and read. In particular, to implement a high-definition decoder that supports AVS and H.264 standards at the same time, if the two decoders are simply superimposed, the area will be very large. In this case, it becomes more and more important how to make full use of the correlation of the standard to realize the maximum multiplexing of the decoder.

发明内容Contents of the invention

本发明所要解决的技术问题是针对上述现有技术现状而提供一种能够同时支持AVS及H.264标准的高清解码器的运动矢量预测(MVP)复用设计方法。The technical problem to be solved by the present invention is to provide a motion vector prediction (MVP) multiplexing design method for a high-definition decoder that can simultaneously support AVS and H.264 standards in view of the above-mentioned current state of the art.

本发明解决上述技术问题所采用的技术方案为：一种多模标准解码器中运动矢量预测复用设计方法，其特征在于：(1)由固件发送给运动矢量预测模块一个avs_flag标志，作为区分AVS及H.264两种数字音视频编解码标准的依据，能使同一个模块处理所述两种标准；(2)块运动矢量预测的主循环采用分块循环下嵌套子块循环的形式，能兼容所述两种标准的算法；(3)取相邻块信息模块中，利用对帧场自适应标志置数，使两种不同标准可以进行相同处理，完全实现复用；(4)取参考宏块信息的设计，并行宏块的所有运动信息总从参考图像宏块运动信息缓存中的第一个参考图像取，更新时也总是保存到第一个参考图像中，简化了存取的逻辑，提高了存取的速度；(5)对邻块数据流更新机制的设计，仅更新紧密相邻的相邻块，使复用所需的相邻块信息量降至最少，更新时仅读取右上宏块对，存入左邻宏块的缓存中，减少了读外存次数。The technical solution adopted by the present invention to solve the above-mentioned technical problems is: a motion vector prediction multiplexing design method in a multimode standard decoder, characterized in that: (1) an avs_flag flag is sent to the motion vector prediction module by the firmware as a distinguishing The basis of AVS and H.264 two kinds of digital audio and video codec standards, can make same module process described two kinds of standards; , can be compatible with the algorithms of the two standards; (3) in the adjacent block information module, the frame field adaptive flag is used to set the number, so that two different standards can be processed in the same way, and the multiplexing is fully realized; (4) In the design of fetching reference macroblock information, all motion information of parallel macroblocks is always fetched from the first reference image in the macroblock motion information cache of the reference image, and is always saved to the first reference image when updating, which simplifies storage. The logic of fetching improves the speed of access; (5) The design of the update mechanism of the adjacent block data stream only updates the adjacent blocks that are closely adjacent to minimize the amount of adjacent block information required for multiplexing and update When only reading the upper right macroblock pair, it is stored in the cache of the left adjacent macroblock, which reduces the number of times of reading external memory.

与现有技术相比，本发明的优点在于，实现了H.264与AVS两种视频编码标准MVP功能的兼容，不仅保证了正确性，且可以大幅度减少解码器硬件面积。经验证，可以减少75％的片上缓存，资源占用情况是单独同时实现两个标准的70％，能实现对高清图像的实时解码。Compared with the prior art, the present invention has the advantage of realizing the compatibility of the MVP function of two video encoding standards of H.264 and AVS, not only ensuring the correctness, but also greatly reducing the hardware area of the decoder. It has been verified that the on-chip cache can be reduced by 75%, and the resource occupation is 70% of the two standards independently and simultaneously, and real-time decoding of high-definition images can be realized.

具体实施方式Detailed ways

以下结合附图实施例对本发明作进一步详细描述。The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

一种多模标准解码器中运动矢量预测复用设计方法，包括：A design method for motion vector prediction multiplexing in a multimode standard decoder, comprising:

1.MVP模块根据数据流的性质，可以划分成输入模块、运动矢量预测主循环模块、块运动预测模块、取相邻块信息模块、取参考宏块信息和输出模块，其复用设计方法：1. According to the nature of the data stream, the MVP module can be divided into an input module, a main loop module for motion vector prediction, a block motion prediction module, a module for obtaining adjacent block information, a module for obtaining reference macroblock information, and an output module. The multiplexing design method:

1)输入模块：包括″读命令模块(read cmd fifo)″、″读残差信息模块(read vldfifo)″，它们分别传送command和残差信息。在固件(firmware)发送序列级信息时首先发送一个avs_flag标志，在其后的所有模块都通过这个标志来区分两个标准中的相异部分，使同一模块处理两个标准。1) Input module: including "read command module (read cmd fifo)" and "read residual information module (read vldfifo)", which transmit command and residual information respectively. When the firmware (firmware) sends sequence-level information, it first sends an avs_flag flag, and all subsequent modules use this flag to distinguish the different parts of the two standards, so that the same module can handle the two standards.

2)运动矢量预测主循环模块：因为在整个凤芯二号系统划分时每个模块(例如MVP模块、IQ模块等)独立完成的最小任务是一个宏块的解码过程，而在宏块中通过partition和sub-partition会将模块划分成若干个小块。AVS中没有sub-partition的小块，也就是说没有8x8以下的划分，除此以外和H.264的宏块划分是完全一样的，因此可以将循环做成partition下嵌套sub-partition的形式，解决AVS没有子划分只要让AVS忽略子循环就可以了。这样设计能兼容两种标准的算法，比设计两套独立的循环机制减少33％的逻辑单元。2) Motion vector prediction main loop module: Because the minimum task independently completed by each module (such as MVP module, IQ module, etc.) is the decoding process of a macroblock when the entire Fengxin No. 2 system is divided, and in the macroblock through partition and sub-partition will divide the module into several small pieces. There is no sub-partition small block in AVS, that is to say, there is no division below 8x8. In addition, it is exactly the same as the H.264 macro block division, so the loop can be made into a sub-partition nested under the partition. , to solve the problem that AVS has no sub-division, just let AVS ignore the sub-cycle. This design can be compatible with two standard algorithms, and the logic unit is reduced by 33% compared with the design of two sets of independent loop mechanisms.

3)块运动预测模块(block mv prediction)：对两个标准的处理方法基本相同，只不过AVS在这里的计算量要大得多，无需特殊设计。3) Block motion prediction module (block mv prediction): The processing methods for the two standards are basically the same, except that the calculation amount of AVS here is much larger, and no special design is required.

4)取相邻块信息模块(GetNeighborBlkInfo:get neighbor block information)：只需要在AVS时将帧场自适应标志置为0，就可以做到和H.264相同的处理方法。从而，两种不同标准可以进行相同处理，完全实现复用，比单独设计减低了50％的逻辑单元。4) Get neighbor block information module (GetNeighborBlkInfo: get neighbor block information): You only need to set the frame field adaptive flag to 0 during AVS, and you can achieve the same processing method as H.264. Therefore, two different standards can be processed in the same way, fully realize the multiplexing, and reduce the logic unit by 50% compared with the separate design.

5)取参考宏块信息：包括三个模块，取图像级信息(GetColPicInfo)、取宏块层信息(getMBColocated)和取块级信息(get4x4Colocated)三个模块。因为H.264中并行参考图像可以多达5帧(就其main profile中level 4.0以上而言)，而AVS中仅仅使用一帧最近解码的后向参考帧，因此AVS中不存在取并行宏块时去寻找它属于哪幅参考图像的问题。也就是说可以不考虑宏块以上的信息保存的问题。所以复用时仅仅默认成并行宏块信息总从参考图像宏块信息缓存中的第一个参考图像取，而更新时也总是保存到第一个参考图像中。这样设计可以简化存取的逻辑，提高存取的速度。5) Obtaining reference macroblock information: including three modules, obtaining image level information (GetColPicInfo), obtaining macroblock layer information (getMBColocated) and obtaining block level information (get4x4Colocated). Because the parallel reference image in H.264 can be as many as 5 frames (in terms of level 4.0 or above in its main profile), and only one frame of the latest decoded backward reference frame is used in AVS, so there is no parallel macroblock in AVS to find out which reference image it belongs to. That is to say, the problem of saving information above the macroblock may not be considered. Therefore, when multiplexing, it is only defaulted that the parallel macroblock information is always taken from the first reference image in the reference image macroblock information cache, and is always saved to the first reference image when updating. This design can simplify the access logic and improve the access speed.

6)输出模块：包括输出到片外存储器(memory access)、参考图像取得模块(RDR)和环路滤波器模块(deblock loop filter)三部分，这三部分两种标准处理方法基本相同，无需进行特殊设计。6) Output module: including output to off-chip memory (memory access), reference image acquisition module (RDR) and loop filter module (deblock loop filter). The two standard processing methods of these three parts are basically the same, no need to special design.

2.邻块数据更新机制的设计，使复用时所需保存的相邻块的信息量降至最少，没有了冗余信息的保存；同时减小了去读取外存的次数。2. The design of the adjacent block data update mechanism minimizes the amount of information of adjacent blocks that need to be saved during multiplexing, without saving redundant information; at the same time, it reduces the number of times to read external memory.

1)复用的片内缓存大小统计：1) Multiplexed on-chip cache size statistics:

(1)相邻块运动矢量和参考索引的信息：(1) Information of adjacent block motion vector and reference index:

每个小块需要保存(也就是要去片外取)的信息：39bits×2＝78(bits)；The information that needs to be saved (that is, to be fetched outside the chip) for each small block: 39bits×2=78(bits);

相邻块需要保存：38×78＝2964(bits)Adjacent blocks need to be saved: 38×78=2964(bits)

当前宏块对需要保存：32×78＝2496(bits)The current macroblock pair needs to be saved: 32×78=2496 (bits)

总共需要：2964+2496＝5460(bits)Total need: 2964+2496＝5460(bits)

(2)参考列表的信息：(2) Information on the reference list:

需要保存当前条带的两个列表和参考并行宏块所在的条带的两个列表，总共需要：26×6×2×5×2＝3120(bits)Two lists of the current slice and two lists of the slice where the reference parallel macroblock is located need to be saved, and a total of: 26×6×2×5×2=3120 (bits) is required

(3)先进先出队列(fifo)的信息：(3) First-in-first-out queue (fifo) information:

需要至少保存：cmd_mvp_fifo：27×32＝864(bits)；Need to save at least: cmd_mvp_fifo: 27×32=864(bits);

vld_mvp_fifo：32×28＝896(bits)；vld_mvp_fifo: 32×28=896 (bits);

总共需要：864+894＝1760(bits)；Total needs: 864+894=1760(bits);

(4)总缓存大小为：5460+3120+1760＝10340(bits)(4) The total cache size is: 5460+3120+1760＝10340(bits)

如果对AVS与H.264单独设计片上缓存，那么所需的相邻块运动矢量和参考索引的信息与参考列表的信息要分别保存，总缓存的大小会是：5460*2+3120*1.2+1760＝16424(bits)，If the on-chip cache is designed separately for AVS and H.264, then the required adjacent block motion vector and reference index information and reference list information must be saved separately, and the total cache size will be: 5460*2+3120*1.2+ 1760=16424 (bits),

可见复用之后减少了37％的片上缓存。It can be seen that the on-chip cache is reduced by 37% after multiplexing.

2)更新的过程：2) The update process:

(1)当前宏块解码前的初始化过程：(1) The initialization process before the decoding of the current macroblock:

将上邻宏块对中的八个小块保存到左上邻宏块对的缓存(buffer)中；Preserving the eight small blocks in the upper-adjacent macroblock pair to the cache (buffer) of the left-upper-adjacent macroblock pair;

将右上邻宏块对中的八个小块读入上邻宏块对的缓存(buffer)中；Read the eight small blocks in the upper-right adjacent macroblock pair into the cache (buffer) of the upper-adjacent macroblock pair;

从外存中读取右上宏块对的八个小块的信息。Read the information of the eight small blocks of the upper right macroblock pair from the external memory.

(2)当前宏块解码完成后：(2) After the decoding of the current macroblock is completed:

将左邻宏块中可能用于参考的块保存到片外；Save the blocks that may be used for reference in the left adjacent macroblocks to off-chip;

将当前解码块中需要保存起来的块存放到左邻宏块的缓存(buffer)中。Store the blocks that need to be saved in the current decoded block into the buffer of the left adjacent macroblock.

在这种更新机制下，每个宏块读取外存仅仅为4个小块的相邻块信息(当为最左列宏块时需要读8个)，当为直接解码模式时还需要读取16个并行小块信息。在128位的带宽下每次读写2个小块的信息。Under this update mechanism, each macroblock reads the adjacent block information of only 4 small blocks in the external storage (8 pieces need to be read when it is the leftmost macroblock), and it also needs to be read when it is the direct decoding mode. Take 16 parallel small blocks of information. Read and write 2 small blocks of information at a time with a bandwidth of 128 bits.

因为是仅仅存取与当前宏块″紧密相邻″的小块信息，是实现正确解码所必需且最少的邻域信息，比存取一个宏块所有16个小块减少了75％的信息量，所以相应的访问外存的次数也减至25％。Because it only accesses the small block information "closely adjacent" to the current macroblock, it is the necessary and minimum neighborhood information for correct decoding, which reduces the amount of information by 75% compared to accessing all 16 small blocks of a macroblock , so the corresponding number of accesses to external memory is also reduced to 25%.

Claims (1)

Translated fromChinese

1.一种多模标准解码器中运动矢量预测复用设计方法，其特征在于：1. motion vector prediction multiplexing design method in a multimode standard decoder, is characterized in that:(1)由固件发送给运动矢量预测模块一个avs_flag标志，作为区分AVS及H.264两种数字音视频编解码标准的依据，能使同一个模块处理所述两种标准；(1) an avs_flag mark is sent to the motion vector prediction module by the firmware, as the basis for distinguishing AVS and H.264 two digital audio and video codec standards, the same module can be made to process the two standards;(2)块运动矢量预测的主循环采用分块循环下嵌套子块循环的形式，能兼容所述两种标准的算法；(2) The main loop of block motion vector prediction adopts the form of nested sub-block loop under the block loop, which can be compatible with the algorithms of the two standards;(3)取相邻块信息模块中，利用对帧场自适应标志置数，使所述两种不同标准可以进行相同处理，完全实现复用；(3) In the adjacent block information module, the frame field adaptive flag is used to set the number, so that the two different standards can be processed in the same way, and multiplexing is fully realized;(4)取参考宏块信息的设计，并行宏块的所有运动信息总从参考图像宏块运动信息缓存中的第一个参考图像取，更新时也总是保存到第一个参考图像中，简化了存取的逻辑，提高了存取的速度；(4) The design of taking reference macroblock information, all the motion information of parallel macroblocks is always taken from the first reference image in the macroblock motion information cache of the reference image, and is always stored in the first reference image when updating, Simplifies the logic of access and improves the speed of access;(5)对邻块数据流更新机制的设计，仅更新紧密相邻的相邻块，使复用所需的相邻块信息量降至最少，更新时仅读取右上宏块对，存入左邻宏块的缓存中，减少了读外存次数。(5) The design of the adjacent block data stream update mechanism only updates the closely adjacent adjacent blocks to minimize the amount of adjacent block information required for multiplexing. When updating, only the upper right macroblock pair is read and stored in In the cache of the left-adjacent macroblock, the number of times to read external memory is reduced.