Adaptive multithreaded H.264/AVC decoding

Abstract

The current trend towards multi-core processors imposes the necessity of finding viable strategies to exploit the additional computational resources in media processing. Among the challenges for video decoding are the appropriate partitioning of decoder steps, efficient tracking of dependencies and resource allocation/synchronization for multiple threads with respect to the resulting overhead. In this paper, we propose two variants of multithreading with distributed synchronization. The first method is optimized for minimum latency decoding, necessary for conversational applications. The second method aims to maximize the total throughput at the cost of a higher latency. In addition, we propose a method of dynamic core usage in order to reduce the total allocated processing resources due to inter-process communication overhead. This method is based on a coarse grained complexity estimation. To implicitly adapt to different combinations of processor architectures, associated memory interfaces and power-saving states, the scheme is feedback assisted. By correlating the initial estimate with the actual required processing time, a sufficiently accurate prediction of the required number of cores for the image processing part can be obtained. Experimental results demonstrate the scaling abilities of up to factor 3.5 on a quad-core machine, as well as the limits of the proposed approach regarding the complexity of sequential bitstream processing. We demonstrate that real-time 4k resolution decoding is feasible on current mid-range PC hardware. For less demanding streams, the adaptive mode reduces the total required CPU resources by up to 10% compared to the greedy approach

An end-to-end testbed for scalable video streaming to mobile devices over HTTP

“…Several parallelism strategies have been proposed, e.g., at macroblock (MB), frame, and GoP levels [10,11]. There are data dependencies among groups of video packets.…”