Hi Paul, On Sun, Aug 15, 2021 at 9:25 AM Paul Buxton <paulbuxton.m...@googlemail.com> wrote:
> Specifically any of > pow() > exp() > log() > pow(a, b) = exp(log(a) * b), so you only need exp/log. As for exp/log optimized versions, there's two approaches that may work for you. First, integer operation is critical (e.g. 16-bit signed, with all non-sign bits being fractional). Then, some people re-implement integer exp/log using a repeated linear approximation, same as what you'd do for sqrt. I've tried this in the past, it works and precision can be quite good, and SIMD'ed implementations will be faster than calling float exp/log. Precision vs. speed is a trade-off of number of repetitions. It's easily SIMD'ed. But... It tends to lose out vs. a much simpler (the second) approach - at least in pure software implementations: look-up tables.This can - optionally - be SIMD'ed using vpgatherdd in AVX2/512. You could in theory even use float output in LUTs, although the input (index) obviously still has to be integerized. Hope that makes sense, feel free to ask more specific questions if needed. Ronald _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
