Why doesn't Random.seed! allocate for StableRNGs.LehmerRNG?

General Usage
1

I ran the following code on my machine:

import StableRNGs
using BenchmarkTools, Random

const T = Float64

rng = StableRNGs.LehmerRNG(0)
rng2 = Random.Xoshiro(0)
rng3 = Random.MersenneTwister(0)

@btime randn($rng, T)
@btime randn($rng2, T)
@btime randn($rng3, T)

@btime Random.seed!($rng, 123)
@btime Random.seed!($rng2, 123)
@btime Random.seed!($rng3, 123)

My REPL shows:

2.420 ns (0 allocations: 0 bytes)
2.440 ns (0 allocations: 0 bytes)
3.300 ns (0 allocations: 0 bytes)
2.060 ns (0 allocations: 0 bytes)
291.890 ns (7 allocations: 368 bytes)
6.358 μs (9 allocations: 424 bytes)

The documentation for StableRNGs.jl is that it focuses on stable streams. Is it an unintended benefit for StableRNGs.LehmerRNG to not allocate when used with Random.seed!?

2

seed! doesn’t allocate with StableRNG because the algorithm is simpler (and arguably too simple); it essentially just uses the seed as is to initialize the RNG state.

With Xoshiro and MersenneTwister, the seed is hashed (currently with SHA256) before initializing the RNG state, and this takes time and allocations. This gives confidence that close seeds give uncorrelated RNGs (like Xoshiro(1) and Xoshiro(2)).