A QHD-capable parallel H.264 decoder

Abstract

Video coding follows the trend of demanding higher performance every new generation, and therefore could utilize many-cores. A complete parallelization of H.264, which is the most advanced video coding standard, was found to be difficult due to the complexity of the standard. In this paper a parallel implementation of a complete H.264 decoder is presented. Our parallelization strategy exploits functionlevel as well as data-level parallelism. Function-level parallelism is used to pipeline the H.264 decoding stages. Datalevel parallelism is exploited within the two most time consuming stages, the entropy decoding stage and the macroblock decoding stage. The parallelization strategy has been implemented and optimized on three platforms with very different memory architectures, namely an 8-core SMP, a 64-core cc-NUMA, and an 18-core Cell platform. Evaluations have been performed using 4k×2k QHD sequences. On the SMP platform a maximum speedup of 4.5× is achieved. The SMP-implementation is reasonably performance portable as it achieves a speedup of 26.6× on the cc-NUMA system. However, to obtain the highest performance (speedup of 33.4× and throughput of 200 QHD frames per second), several cc-NUMA specific optimizations are necessary such as optimizing the page placement and statically assigning threads to cores. Finally, on the Cell platform a near ideal speedup of 16.5× is achieved by completely hiding the communication latency.

Amdahl's law for predicting the future of multicores considered harmful

“…Gustafson assumes that the amount of work that can be parallelized grows linearly with the number of cores p. This, on the other hand, can be considered overly optimistic. For example, although video coding is a domain that can take benefit from multi-and many-cores [3], the resolution and computational requirements will not grow indefinitely.…”

“…More cores are used to solve larger and more complex problems. For example, more cores are used to perform weather forecasting on a larger area (e.g., [12]) or for video decoding with higher resolutions [3]. Furthermore, for many applications, when the problem size scales the parallel part scales faster than the serial part.…”

Improving the parallelization efficiency of HEVC decoding

“…In previous video codecs, like H.264/AVC, picture partitions were only possible with slices with a high cost in terms of coding efficiency. For scalable parallel H.264/AVC decoding it is necessary to combine macroblock-level parallelism for picture reconstruction and frame-level parallelism for entropy decoding [3]. his approach, however, provides limited reduction in picture latencies and results in high memory usage.…”

Programming parallel embedded and consumer applications in OpenMP superscalar

“…Grouping the tasks, however, reduces the parallelism, which in turn limits the performance at higher core counts. In the Pthreads version of h264dec the synchronization is highly optimized using a line decoding strategy [1] and, therefore, grouping of tasks is not necessary.…”