Output distribution of rand(Float32) and rand(Float64), thread 2

[dlakelan]

@mbauman

This is why I call this the work of magicians and have been (perhaps overly) critical here. There’s so many ways you can get RNGs wrong.

You’re not wrong! Maybe the real solution is to provide an implementation that throws an error message telling the user to choose one of several implementations explicitly.

@mikmoore

probability that we round a Float32(rand(Float64)) == 1f0 is a whopping 2^{-25}2−252^{-25}. Definitely not 2^{-53}2−53

Yeah, I realized this after I wrote it. thanks for that clarification.

@mikmoore

This point aside, anything with a branch or loop is a performance killer, no matter how rarely it is used, because it is impossible to SIMD.

and this is a good point.