Efficient Floating-Point Division for Digital Signal Processing Application [Tips & Tricks].

Abstract

Floating-point division is an expensive operation for processors in digital signal processing (DSP). The basis of the division operation is finding the reciprocal of the divisor. We present a reciprocal algorithm with four multiplications for single accuracy and six for double accuracy. The algorithm specifically includes bithack (known as magic constant) operations and floating-point addition, multiplication, and the fused multiply-add (fma) operation. The proposed approach improves two characteristics of the division process: accuracy and number of steps. This second characteristic has a direct impact on the division operation performance time and on the amount of equipment required for hardware implementation.