Why is this faster than DoubleFloats.jl

[oscardssmith]

I would expect Float64x2 to be basically the same as DoubleF64. Do you know where the performance difference in the readme comes from?

[dzhang314]

Hey @oscardssmith! Two things: (1) DoubleFloats.jl does Inf/NaN checks which means their arithmetic algorithms aren't branch-free, and (2) MultiFloats.jl v3.0 (used in the README benchmarks) employs novel, faster algorithms (in particular the Float64x2 addition algorithm) not previously known in the floating-point literature.

I'll be making the MultiFloats.jl v3.0 public release very soon, once I'm done writing some paper revisions, so you'll be able to evaluate the performance difference on your own hardware.

[oscardssmith]

thanks! Also once you release, you should probably re-run the Quadmath benchmarks. I've been busy JuliaMath/Quadmath.jl#92 :)