Combining Pel Decimation with Partial Distortion Elimination to increase SAD energy efficiency

Abstract

The most energy-hungry step of Video Coding (VC) is the Block Matching Algorithm (BMA), even when a simple similarity metric such as the Sum of Absolute Differences (SAD) is employed. Moreover, with the increasing resolutions supported by state-of-the-art VC standards (H.264/AVC, HEVC and VP9), the SAD must be as energy-efficient as possible to increase the battery lifetime in portable mobile devices. Two well-known techniques to decrease the number of operations in SAD calculation are Pel Decimation and Partial Distortion Elimination (PDE). The energy savings provided by the former are dictated by the chosen decimation ratio and comes with a cost in coding efficiency. For the latter, energy savings have no cost in coding efficiency but are dictated by the video content and search parameters. In this work we present two configurable SAD4×4 architectures: one designed to dynamically operate using one among four Pel Decimation ratios (1∶1, 4∶3, 2∶1 or 4∶1) and the other one able to use PDE in addition to Pel Decimation. We simulated Pel Decimation and PDE behavior during motion estimation using 22 video samples from the Common Test Conditions (CTC) encoded using 4 different quantization parameters (QPs). Thus, this simulation was performed over 5.82×1012 PDE SADs. The Pel Decimation impacts are shown in terms of Bjontegaard Delta (BD)-Rate, ranging from 3.16% (1∶1 ratio) up to 21.94% (4∶1). In addition, we found that by using PDE solely (i.e., without Pel Decimation) one can reduce from 10 to 6.38 (in average) the number of required cycles to calculate one SAD. To show the improvements in terms of energy, we synthesized both presented architectures using a 45nm standard cell library. Finally, the use of PDE can improve energy efficiency more than Pel Decimation alone, without coding efficiency degradation.