Improving the Performance of the Divide-Add Fused Operation Using Variable Latency Quotient Generation

Abstract

Dedicated floating point units for divide-add fused operation (division followed by addition/subtraction) can be used to increase the performance of the interval Newton's method. The key issue regarding these units is represented by the number of quotient bits generated. A high number leads to better accuracy, but also to low performance. The required number of quotient bits is determined by the exponents' difference and the number of leading zeros. In this paper, we propose a divide-add fused unit which generates a variable number of quotient bits. This way, the latency for the cases when few quotient bits are needed is reduced, without loss in precision. Thus, the average performance of the divide-add fused operation is improved, while the area overhead is around 1%.