Razaekel
commented
7 years ago This would be excellent to have! However, as far as I understand it (and in my testing) the current type signature of NoiseModule<T, U> would prevent being able to do this because of the fn get(&self, point: T) -> U signature, and the way some of the NoiseModules use that signature as is, and others use fn get(&self, point: PointN<T>) -> U'. If the trait definition wasfn get(&self, &[T]) -> U, I'm pretty sure it would do bad things to the composability of the modules. Not sure what would happen if theimpls used&[T]` instead.
A number of the modules use fn(&self, point: T) -> U directly, since they're agnostic about the dimensionality of the point, ex. Abs or Const.
If we had const generics, the signature could be simplified to NoiseModule<T> with a fn get(&self, [T; N]) -> T and const N.
I know generators/checkerboard.rs uses a &[T] internally, but it still uses the PointN<T> interface for compatibility. This may be an acceptable intermediate stage.
amaranth
tayloraswift
cheako
Cazadorro
As mentioned in #134 one way to get the 5D and 6D noise functions would be to, as much as possible, make the library generic over dimensions. I've implemented this as a proof of concept using generic-array for Perlin noise (TODO: publish this and link to it) and it's at worst 200% the run time of the specialized implementations but I believe I can improve on that. If not we can always keep the existing implementations too, we'll just have four implementations of NoiseModule for each generator then (Point, Point2, Point3, Point4). On nightly Rust we could even use impl specialization to get the specialized implementations when using Point.
We can easily do N-dimensional versions of Perlin, Simplex, and Value noise as well as all of the modifiers, fractals, and checkboard and such. The complicated/impossible ones are OpenSimplex, SuperSimplex, and Worley. Those might be possible with more research or with much slower generic implementations. Worst case we can have Point\<U2>, Point\<U3>, and Point\<U4> versions that call to the specialized implementations and the compiler will give a type error if someone tries to get 5D SuperSimplex noise.
As an aside, while working on the N-dimensional Perlin version I discovered using an N-dimensional PermTable via a
get(&[T]) -> usizemethod results in pretty dramatic speedups for the 64x64 benchmarks. Even if we don't do full N-dimensional I'm going to try to make as much of the helper machinery (PermTable, math, per-noise helpers) N-dimensional and see if that has a similar effect on performance.