Tunable Inexact Subtracters for Division in Image Processing Applications

Error resilient applications have paved the way for energy-efficient approximate computing circuits. Among the various computation blocks, the divider is known for large latency and high energy consumption. We present novel energy- quality tunable inexact dividers built using Tunable Inexact Subtractors (TISs). TISs allow designers to dynamically trade-off between energy and quality depending upon the requirements. We present two Tunable Inexact Subtracters, TISI and TISI, that have a maximum bounded error. TISI and TIS2 show an overall improvement of 17.9% and 18.5% in EDP when used in approximate mode, but have an overhead of 11% and 4% in EDP when used in exact mode as compared to the exact subtractor respectively. We have used TISI and TIS2 to implement Tunable Inexact Restoring Array Dividers (TIRADs). To the best of our knowledge, these proposed designs are the first- ever configurable subtractors that equip inexact restoring array divider circuits with the feature of configurability. The 16- by-8 TIRADs employing TISI and TIS2 have been used in three image processing applications - change detection, background removal and JPEG compression. We see an overall reduction of 8%, 7%, and 11% in energy consumption for the three applications using TIRAD1 as compared to exact divider. For TIRAD2, we see 25%, 26%, and 21% energy reduction for the same as compared to exact divider.