Multicore Runtime for Dynamic Dataflow Video Decoders
Résumé
The emergence of massively parallel architectures, along with the increasing complexity of signal processing applications, has revived the interest in the dataflow programming paradigm. Indeed, dataflow programming is well-suited to describe complex and modular applications while expressing parallelism explicitly. Paradoxically, most of the studies stay focused on static dataflow programming, even if a pragmatic development process requires the expressiveness and the practicality offered by dynamic dataflow programming. However, dynamic dataflow programs come up with the need of runtime scheduling decisions because of their unpredictable behavior. This paper introduces a run-time mapping/scheduling system dedicated to the execution of dynamic dataflow programs on multi-core platforms.
The program behavior is profiled and analyzed on-the-fly, thus allowing the system to adapt to changing environments (e.g. different video streams) on real time thanks to dynamic reconfiguration. Our system has been implemented into an open-source tool and demonstrated on well-known video decoders including one based on the new High Efficiency Video Coding (HEVC) standard. The acceleration factor reaches 7,75 on a 8-core x86 processor over single-core execution showing the scalability of our approach.