A fast and flexible accuracy-guaranteed fractional bit-widths optimization approach

This paper presents a novel approach to optimize fractional bit-widths during floating-point to fixed-point transformation of digital signal processing and communication algorithms. The approach guarantees output accuracy while minimizing total area cost. Quantization-operation-error (QOE) model is proposed to formulate signals' quantization error bound propagations. Based on QOE, a fast and flexible greedy algorithm is developed to find near-optimal solution. Case studies demonstrate that the approach runs 99 to 315 times faster than the existing one: MiniBit.