Optimized leading zero anticipators for faster fused multiply-adds
David Lutz · 2017
Leading zero anticipators (LZAs) predict the number of leading zeros in a difference, and have long been used to speed up floating-point add and fused-multiply add (FMA) operations. Typically the LZA prediction must be corrected for small exponents and possible carries, and only the corrected value is useful for normalizing the difference and computing rounding. We present new techniques to speed up LZA correction and rounding, allowing differences to be computed in only 60% of their previous best latency.