帧内预测在多核系统上的并行实现

由于H.264的帧内预测造成了I帧的宏块具有数据依赖性,导致多核系统无法并行处理。针对这种情况,分别对intra4×4和intra16×16类型的宏块进行处理。如果当前宏块是inra4×4类型,则解码完毕任一处于最右边一列的子宏块后就通过邮箱把数据发送给下一个宏块以开始其帧内预测。如果当前宏块是intra16×16类型,则首先对最右边的一列数据进行帧内预测,然后和残差数据相加,得到解码后的数据并通过邮箱传送给下一个宏块以开始其帧内预测。分别对相邻宏块的4种情况做了详细的分析,给出了其并行方法。仿真结果表明,使用该方法后,单个宏块解码时间减少,DSP访问存储器的次数减少。多核系统的并行度得到提高。

Parallel Intra Prediction on Multi-core Platform

Because of the data dependence caused by intra prediction of H.264, multi-core platform could not deal with the macro-block data in parallel. Aiming at this problem, the paper processes the intra4x4 and intra16x16 macro-block separately. If current macro-block is intra4x4, the decoded data of right four sub-blocks is sent to next macro-block through mailbox to start intra prediction. If current macro-block is intra16x16, the right line does the intra prediction first and add the data with residual. Then data is sent to next macro-block through mailbox to start intra predicting. The paper analyses the 4 combination of two macro-blocks which are close to each other and points out the parallelism method separately. The simulation result shows that after using the proposed method, less time is used to decode one macro-block and DSPs visits the memory less. The parallel degree of multi-core platform is increased.