Low communication overhead dynamic mapping of multiple HEVC video stream decoding on NoCs

Abstract

The High Efficiency Video Coding (HEVC) standard offers several parallelisation tools such as wave-front parallel processing (WPP) and Tiles (independent frame regions) to better manage the computationally expensive workloads on modern multicore/many-core platforms. However, poor allocation of tile-level HEVC decoding tasks to processing elements may result in increased latency and energy consumption due to data-communication overhead between dependent tiles. In this work, we discuss the difficulties in decoding multiple HEVC bitstreams with highly varying resolutions and data-dependency characteristics as seen in HEVC coded video streams with random-access, adaptive group of pictures (GoP) structures. Secondly, in order to address the above challenges, we introduce a runtime tile allocation scheme that help to reduce the energy usage during HEVC decoding. Evaluations against a bin-packing algorithm, show that the proposed workload mapping technique is able to maintain reasonably acceptable latency results, whilst reducing communication overhead (8-10%) and increasing the mean processor idle periods (~30%) to support dynamic power management.