Smart search space reduction for approximate computing: A low energy HEVC encoder case study

Abstract

International audienceThe approximate computing paradigm provides methods to optimize algorithms while considering both application quality of service and computational complexity. Approximate computing can be applied at different levels of abstraction, from algorithm level to application level. Approximate computing at algorithm level reduces the computational complexity by approximating or skipping computational blocks. A number of applications in the signal and image processing domain integrate algorithms based on discrete optimization techniques. These techniques minimize a cost function by exploring an application parameter search space. In this paper, a new methodology is proposed that exploits the computation-skipping approximate computing concept. The methodology, named Smart Search Space Reduction(SSSR), explores at design time the Pareto relationship between computational complexity and application quality. At run time, an approximation manager can then early select a good candidate configuration. SSSR reduces the run time search space and, in turn, reduces computational complexity. An efficient SSSR technique adjusts at design time the configuration selectivity while selecting at run time the most suitable functions to skip. The real time High Efficiency Video CodingHEVC encoder in All Intra(AI) profile is used as a case study to illustrate the benefits of SSSR. In this application, two discrete optimizations are performed. They explore different coding parameters and select the values leading to the minimal cost in terms of a tradeoff between bitrate, quality and computational energy by acting on both the HEVC coding-tree partitioning and the intra-modes. Combining two SSSRs iterations on this use case, the energy consumption is reduced by up to 77%. Moreover, the combination of the two SSSRs iterations in comparison to using only one reduces the BD-BR bitrate/quality metric by 4% for the same energy consumption. © 201

Low-Power and High-Throughput Approximated Architecture for AV1 FME Interpolation

“…Multiple related works propose hardware architectures employing approximate computing for video coding, like [9] [13] [14], but those works are focused in the HEVC standard [15]. On the other hand, some works in the literature target hardware designs for the AV1 interpolation filters.…”

Adaptive simulation-based framework for error characterization of inexact circuits

“…or carefully tune the precision in optimizing the bit-widths of the data [2], [3]. At algorithmic level, the processing complexity can be reduced, for instance, by skipping or approximating a fraction of the computations [4]- [6]. Finally, at hardware level, approximations have been exploited in different manners such as overclocking [7] or by modifying the logic structure of original exact operators into an inexact version [8]- [12] with a lower logic complexity or shorter critical paths.…”

Cassis: Characterization with Adaptive Sample- Size Inferential Statistics Applied to Inexact Circuits

“…Approximations in circuits have been introduced in different manners such as overclocking [1] at circuit level or computation skipping [2] at algorithmic level. The logic structure of original exact operators can also be modified into an inexact version [3]- [7].…”