Scalable-Grain Parallelization Method For Media Applications On MPSoC

Recently, multi-core processors keep increasing development, and have a wide range of applications in personal computer, embedded system, high performance computing, etc. MPSoC(Multiprocessor System-on-Chip) has multiple cores executing currently to improve efficiency of some applications, and satisfies the requirement of multimedia, communication, network security, etc. MPSoC is considered as the new solution in SoC.Contrary to the hardware development in MPSoC, the software programming problem, especially the parallelization of sequential C codes for specific hardware has become increasingly prominent. There is no automatic parallel programming model right now. Instead, the common way is to develop new program or modify the existing legacy sequential C code to generate parallel routine, which extends the development and cycle heavily, and has a high requirement of programmer.Targeting the above issues, this thesis proposes a source-to-source parallelization method, which transforms C code from sequential to parallel. This method defines numerous and powerful analysis, transformation modules, and make optimization for media applications so as to reduce the involvement of programmers and explore more parallelism of media applications.The proposed method, firstly perform static/dynamic analysis on the original sequential C code to generate SGDFG (Scalable Grain Data Flow Graph). Then, scalable grain tasks are hierarchically scheduled to generate various parallelization plans of different grains by successive refinement. And the final parallelization result is chosen to satisfy the constraints and requirements of target hardware architecture and media applications. At last, T264 is chosen to be parallelized for verifying the proposed method. The experiment shows that the speedup 1.93 is gained on the multi-core platform with three processors, which proves that the parallelization method proposed can achieve a good performance.