Analysis of Pel Decimation and Technology Choices to Reduce Energy on SAD Calculation

As the number of pixels per frame tends to increase in new high definition video coding standards such as HEVC and VP9, pel decimation appears as a viable means of increasing the energy efficiency of Sum of Absolute Differences (SAD) calculation. First, we analyze the quality costs of pel decimation using a video coding software. Then we present and evaluate two VLSI architectures to compute the SAD of 4x4 pixel blocks: one that can be configured with 1:1, 2:1 or 4:1 sampling ratios and a non-configurable one, to serve as baseline in comparisons. The architectures were synthesized for 90nm, 65nm and 45nm standard cell libraries assuming both nominal and Low-Vdd/High-Vt (LH) cases for maximum and for a given target throughput. The impacts of both subsampling and LH on delay, power and energy efficiency are analyzed. In a total of 24 syntheses, the 45nm/LH configurable SAD architecture synthesis achieved the highest energy efficiency for target throughput when operating in pel decimation 4:1, spending only 2.05pJ for each 4×4 block. This corresponds to about 13.65 times less energy than the 90nm/nominal configurable architecture operating in full sampling mode and maximum throughput and about 14.77 times less than the 90nm/nominal non-configurable synthesis for target throughput. Aside the improvements achieved by using LH, pel decimation solely was responsible for energy reductions of 40% and 60% when choosing 2:1 and 4:1 subsampling ratios, respectively, in the configurable architecture. Finally, it is shown that the configurable architecture is more energy-efficient than the non-configurable one.