ACCELERATED IMAGE PROCESSING THROUGH IMPLY-BASED NOCARRYAPPROXIMATED ADDERS

Abstract

Conventional approaches to meeting the demands of computing power are finding it difficult to keep up with the sharp rise in demand. Alternative computer paradigms have therefore proliferated in an effort to address this discrepancy. An emerging technique for increasing speed, space efficiency, and energy consumption in error-resilient applications like computer vision and machine learning is approximate computing, or AxC. Accuracy is sacrificed in exchange for these improvements. Because of their low power consumption and intrinsic non-volatility, which make them appropriate for In-Memory Computation (IMC), memristors have attracted a lot of attention from a technological standpoint. In order to address the discrepancy between performance progress and demand increase, another computer paradigm has emerged. We use Material Implication (IMPLY), a memristive stateful in-memory logic, in this study. In the framework of AxC, we study sophisticated adder topologies with the goal of fusing the advantages of both cutting-edge computing paradigms. For every adder topology based on IMPLY, we provide two estimated methods. Compared to the comparable exact full adders, they lower the number of steps by 6% to 54% and the energy consumption by 7% to 54% when integrated into a Ripple Carry Adder (RCA).We compare our work with State-of-the-Art (SoA) circuit-level approximations that improve speed and energy efficiency by up to 72% and 34%, respectively, and lower the Normalized Median Error Distance (NMED) by up to 81%. We assess our adders in four widely used image processing applications and give two more test datasets. In most cases, our proposed adders may reduce the number of image processing steps and energy usage by up to 60% and 57%, respectively, while improving quality metrics over the SoA.