A novel method for the approximation of multiplierless constant matrix vector multiplication

Since human beings have limited perceptual abilities, in many digital signal processing (DSP) applications, e.g., image and video processing, the outputs do not need to be computed accurately. Instead, their computation can be approximated so that the area, delay, and/or power dissipation of the design can be reduced. This paper presents an approximation algorithm, called aura, for the multiplierless design of the constant matrix vector multiplication (CMVM) which is a ubiquitous operation in DSP systems. aura aims to tune the constants such that the resulting matrix leads to a CMVM design which requires the fewest adders/subtractors, satisfying the given error constraints. This paper also introduces its modified version, called aura-dc, which can reduce the delay of the CMVM operation with a small increase in the number of adders/subtractors. Experimental results show that the proposed algorithms yield significant reductions in the number of adders/subtractors with respect to the original realizations without violating the error constraints, and consequently lead to CMVM designs with less area, delay, and power dissipation. Moreover, they can generate alternative CMVM designs under different error constraints, enabling a designer to choose the one that fits best in an application.