Re:Fast Approximate Logarithms, Part III: The Formulas
by brandmeyer on Saturday June 27, 2015 @08:56PM
(#9792027)
Actually, there is quite a bit of available ILP in the forms that use division - as high as four insns. The numerator and denominator can be evaluated in parallel, and the coefficient loads can be interleaved with the FMAs to hide some of their latency.
That said, with such a low variation in the timings I wonder if some other overhead was dominating the costs. For example, floating-point division is typically much slower than either addition or multiplication. According to Agner Fog, div is 2-4x slower in latency than add and mul while completely hogging the pipeline.
That said, with such a low variation in the timings I wonder if some other overhead was dominating the costs. For example, floating-point division is typically much slower than either addition or multiplication. According to Agner Fog, div is 2-4x slower in latency than add and mul while completely hogging the pipeline.