Curve trait for general interoperation

[mweatherley]

This is a PR containing an implementation of the curve API described in this RFC.

See #13213 for Curves Working Group tasks.

Objective

Motivation is described in the RFC (linked above) but briefly summarized here.

The goal is to create a trait (Curve<T>) to provide a general functional API surface for working with curves of arbitrary origin. This allows different areas in Bevy to avoid reinventing the wheel, and paves the way for things like the following:

• Making animation generic over curves that describe motion.
• Using general curves to describe gradients of colors.
• Performing manipulations in geometry of general curves valued in Vec3 or similar.

Put another way, the goal is to allow many consumers to be made effectively generic over curves rather than needing to rely on specific concrete constructions (e.g. CubicCurve, VariableCurve, and so on). With a trait, this can be done without sacrificing runtime performance.

Solution

The interface itself is described in detail in the RFC itself. Here is a brief summary.

The Curve trait is generic in a single parameter T. In general, arbitrary T are allowed, but T: StableInterpolate guarantees access to some niceties. This bound will generally not hold for types from user-space.

The fundamentals of a Curve are as follows:

• It has a domain, which is an Interval (this type represents nonempty closed intervals that are possibly unbounded in either direction).
• It can be sampled (via sample) at f32 parameter values to spit out values of type T.

The sampling has to return data for any f32 input, but its behavior outside of the domain interval is implementation-specific. There are additional sampling functions which do actually use the domain explicitly:

• sample_checked: This is like sample, but it returns an Option<T> taking None values when sampling outside of the domain.
• sample_clamped: This is like sample, but it clamps the input parameter to the domain interval before sampling.

Furthermore, there are a number of operations afforded to Curve as part of its API:

• map: Given a Curve<S> and a function f: S -> T, this produces a Curve<T> whose samples are those of the former curve mapped through the latter.
• reparametrize: Given a Curve<T> and a function f: f32 -> f32, this produces a Curve<T> whose samples at parameter t are those of the given curve at parameter f(t).
• graph: Given a Curve<T>, this produces a Curve<(f32, T)> whose samples are those of the given curve coupled with the sampling parameter.

In addition to these functional operations, the API also supports a couple of resampling operations. In their general form, these depend on an explicitly given interpolation I: Fn(&T, &T, f32) -> T.
The resampling operations are as follows:

• resample: Given a Curve<T>, a number of samples to use, and an explicit interpolation function, this produces a SampleCurve<T, I>, which is a Curve<T> whose sampled values are determined by interpolating between evenly-spaced samples.
• resample_uneven: Given a Curve<T>, an iterator over f32 parameter values, and an explicit interpolation function, this produces an UnevenSampleCurve<T, I>, which is a Curve<T> whose sampled values are determined by interpolating between samples taken from the original curve at the given parameter values. (That is, this works like keyframes.)

These SampleCurve types are supported by a cores submodule that is intended for use in user-space (and especially by people like library authors) to define curves that are determined by sample interpolation. It provides a number of structs and access patterns that encapsulate, for example, the notions of evenly-spaced sampling and keyframe sampling; these can just be dropped into structs in user-space together with interpolation metadata.

The EvenCore documentation gives an example of how this works:

enum InterpolationMode {
    Linear,
    Step,
}

trait LinearInterpolate {
    fn lerp(&self, other: &Self, t: f32) -> Self;
}

fn step<T: Clone>(first: &T, second: &T, t: f32) -> T {
    if t >= 1.0 {
        second.clone()
    } else {
        first.clone()
    }
}

struct MyCurve<T> {
    core: EvenCore<T>,
    interpolation_mode: InterpolationMode,
}

impl<T> Curve<T> for MyCurve<T>
where
    T: LinearInterpolate + Clone,
{
    fn domain(&self) -> Interval {
        self.core.domain()
    }
     
    fn sample(&self, t: f32) -> T {
        match self.interpolation_mode {
            InterpolationMode::Linear => self.core.sample_with(t, <T as LinearInterpolate>::lerp),
            InterpolationMode::Step => self.core.sample_with(t, step),
        }
    }
}

The API has a few more bells and whistles than this, but this hits the major points.

[IQuick143]

What if the interval is infinite?

[mweatherley]

Good catch!

[IQuick143]

As I've discussed on discord, I don't think this bound should be placed on Curve<T> itself, only at the part implementing resample (about which we could also debate some more).

[mweatherley]

I think we all pretty much agree that that would be better, so I've limited it to just the resample methods and the curve constructs that require it now (SampleCurve and UnevenSampleCurve).

[IQuick143]

What does/should sample do when it's not "checked" like this?

[mweatherley]

Yeah, that's a good question. Generally speaking, the idea is that sampling behavior outside of the domain interval is implementation-specific. We might want to consider renaming some of the methods to discourage using unchecked sampling in user-space, but it's pretty important to have access to it in the API, since doing otherwise introduces required overhead.

[IQuick143]

Are we sure this cannot go out of bounds due to numerical error in f32 calculations?

ie:

t = domain.end;
// step could be slightly shorter due to numerical error
step = (domain.end - domain.start) / subdivs as f32;
// t_shifted could be slightly too much due to numerical error
t_shifted = t - domain.start = domain.end - domain.start

// t_shifted / step could be therefore slightly bigger than subdivs, using .ceil() on it could therefore give the next integer, causing and out of bounds access.
upper_index = (t_shifted / step).ceil()

[mweatherley]

I rewrote this to make it less prone to this kind of thing and easier to reason about.

[mweatherley]

Closing in favor of breaking this up to be more easily reviewed — #14630