Given a code in which rand is called in a inner function, as in this simplified example:
module Test
import Random
export outer
RNG = Random.Xoshiro()
function inner(x)
x += rand(RNG)
end
function outer(x)
x += inner(x)
end
end
Which do you think is the most idiomatic way to allow RNG choice without causing type instabilities?
One possibility is to pass RNG as a parameter on every interface, of course. Do you see a better alternative? (without the const RNG the call to outer is obviously type-unstable).
The purpose of this is be able to use StableRNGs to test some functions with reproducible random states.
module Test
import Random
export outer
RNG = Random.Xoshiro()
function inner(x; rng=RNG)
x = rand(rng)
end
function outer(x; rng=RNG)
inner(x; rng)
end
end
which I guess is what you mean by “pass RNG as a parameter on every interface”?
I did this more as a matter of avoiding globaI state than a worry about type stability. I’m surprised though that having a non-const RNG would cause type instability. Doesn’t rand have a known output type? What if you wrote your own rand wrapper; something like
RNG = Random.Xoshiro()
function set_rng(rng)
global RNG
RNG = rng
end
function myrand()::Float64
global RNG
return rand(RNG)
end
but using that or annotating the output is not solving the problem (the annotation prevents the instability from propagating only):
julia> RNG = Random.Xoshiro()
function test(x::T) where {T}
x + rand(RNG, T)::T
end
test (generic function with 1 method)
julia> @btime test(0.0)
116.872 ns (1 allocation: 16 bytes)
0.2989532433595429
julia> function test2(x::T, RNG) where {T}
x = rand(RNG, T)
end
test2 (generic function with 1 method)
julia> @btime test2(0.0, $RNG)
3.010 ns (0 allocations: 0 bytes)
0.6171378061283105
I can live with passing RNG on every interface, but this is a case where I would be fine with a global state.