[Libre-soc-dev] big-int goldschmidt div

[lkcl]

On April 24, 2022 11:23:01 AM UTC, lkcl <luke.leighton at gmail.com> wrote:

>compared to Knuth Algorithm D the "downtime" of the SIMD Mul backend is
>*really* mimimal. yes there are 2 big-int multiplies but guess what,
>there will be *less iterations* and certainly less qhat/rhat increment
>adjustment (i.e. none).

https://www.researchgate.net/publication/44384708_Improved_algorithms_and_hardware_designs_for_division_by_convergence

apparently, larger multipliers are required for goldschmidt, but a BigInt Mul that just means twice the Vector Length, not twice the hardware size, which in turn would mean twice as long to complete each BigInt Mul but i *still* think it would be less iterations overall.

l.