Investigation of reconfigurable-accuracy approximate adder designs for image processing applications

Abstract

In the last decades, integrated circuits with CMOS technology show progressive scaling challenges of both increased power density and power dissipation. Meanwhile, high-performance requirements of current and future application operations show rapid demands of computing resources like power. This design conflict has pushed much effort to search for high performance and energy efficient design approach, such as approximate computing. Approximate computing exploits the error resilience of compute- intensive applications such as image processing applications to implement approximation design techniques with different levels of abstractions and scalability. The basic principle is to relax the strict accuracy requirements in favour of a lower design complexity, thereby achieving more computational performance (i.e., speed) and energy saving. The adder arithmetic unit is considered one of the essential computational blocks in most of the applications. As such, much effort has explored new designs of an efficient approximate adder design. This thesis presents an investigation into design enhancement, novel approximate adder designs and implementation approaches. The first approach introduces a modification to the error detection technique of a popular configurable-accuracy approximate adder design. The proposed lightweight error detection technique reduces the required gates of the error detection circuit, thus, mitigating the design area overhead. Furthermore, at the error correction process of the adder, we have proposed an extensive error detection while activating more than one correction stage concurrently. As a result, this ensures achieving an optimum accuracy of outputs for the worst case of quality requirements. In general, approximate (speculative) adder designs use the seg- mentation technique to divide the adder into multiple short length sub-adders which operate in parallel. Hence, this would limit the long chains of carry propagation and result in a better performance operations. However, the use of overlapped parts of sub-adders regarding a better carry speculation and then more accuracy be- comes a significant challenge of a large design area overhead. The second approach continues mitigating this challenge by present- ing a novel and simpler adder dividing technique to a number of sub-adders. The new method uses what is known as the carry-kill signal for both limiting the carry propagation and applying adder segmentation. Further, between every two adjacent sub-adders, one AND gate and one XOR gate are used for carry speculation and error (i.e., carry propagation) detection respectively. Thus, a significant reduction of the design overhead has been achieved, yet, with acceptable levels of output results accuracy. In the third final approach, simple logic OR gates are used to build the approximate adder while compensating the conventional full adders operation. The resulted approximate adder design presents very low complex- ity, high speed, and low power consumption. Furthermore, instead of augmenting error recovery circuit, short bit-length exact adders are used as correction stages to control the general level of output quality (i.e., without error detection overhead). At the final correc- tion stage, the proposed design would operate the same as an exact adder. To validate the efficiency of these approaches, a number of adders with different bit-widths are designed and synthesized showing considerable reductions in the critical delay, silicon area and more savings in energy consumption, compared to other existing ap- proaches. In addition to acceptable levels or output errors, which are extensively analysed for each proposed design. In this study, the proposed configurable adder designs exhibit energy/quality trade-offs at a different number of correction stages. These trade-offs can be effectively exploited to implement adders in applications, where energy can be gracefully minimised within the envelope of quality requirements. As such, designs implemen- tation in an image processing application known as Gaussian blur filter was introduced, demonstrating the loss in the image quality at each error correction stage. The output images showed promis- ing results to use the proposed designs for more energy-efficient applications, where output quality requirements can be relaxed.