Correctly rounded floating-point division for DSP-enabled FPGAs

Floating-point division is a very costly operation in FPGA designs. High-frequency implementations of the classic digit-recurrence algorithms for division have long latencies (of the order of the number fraction bits) and consume large amounts of logic. Additionally, these implementations require important routing resources, making timing closure difficult in complete designs. In this paper we present two multiplier-based architectures for division which make efficient use of the DSP resources in recent Altera FPGAs. By balancing resource usage between logic, memory and DSP blocks, the presented architectures maintain high frequencies is full designs. Additionally, compared to classical algorithms, the proposed architectures have significantly lower latencies. The architectures target faithfully rounded results, similar to most elementary functions implementations for FPGAs but can also be transformed into correctly rounded architectures with a small overhead. The presented architectures are built using the Altera DSP Builder Advanced framework and will be part of the default blockset.