Why wouldn’t it? Isn’t HLSL turing-complete?

A lot of implementations just use a precomputed image because the noise doesn’t have to be truly random, it’s okay if it’s just perceived as such.