[RFC] Generic functions

[iwoplaza]

Generic parameters with slots & derived

const paramSlot = tgpu.slot<d.Vec2f | d.Vec3f | d.Vec4f>();

const myAdd = tgpu.derived(() => {
  const T = paramSlot.$;
  
  return tgpu.fn([T, T], T)((a, b) => {
    return std.add(a, b);
  });
});

// Using `myAdd`
myAdd.with(paramSlot, d.vec2f).$(d.vec2f(1, 2), d.vec2f(3, 4));

Function overloads (new design space)

const myAdd = tgpu.overload([
  tgpu.fn([d.vec2f, d.vec2f], d.vec2f),
  tgpu.fn([d.vec3f, d.vec3f], d.vec3f),
  tgpu.fn([d.vec4f, d.vec4f], d.vec4f),
])((a, b) => {
  return std.add(a, b);
});

// Using `myAdd` will automatically detect which overload to use based on the parameter types
// NOTE: This only works in JS
myAdd(d.vec2f(1, 2), d.vec2f(3, 4));

// For use in WGSL, we need to extract one of the overloads
const myAdd2 = myAdd.collapse(tgpu.fn([d.vec2f, d.vec2f], d.vec2f));
//    ^? TgpuFn<(a: v2f, b: v2f) => v2f>

[reczkok]

Is there no way around the .collapse? Don't we have enough information to pick the correct overload at resolve time?

[iwoplaza]

Is there no way around the .collapse? Don't we have enough information to pick the correct overload at resolve time?

In the case of TGSL there is, but it gets tricky for WGSL-implemented functions without semantic analysis

const foo = tgpu.fn([])`() {
  let a = vec3f();
  let b = vec3f(4);
  // Which overload do we inject here?
  myAdd(a, b);
}`.$uses({ myAdd });

[iwoplaza]

This has been solved in a nicer way with shell-less functions 🚀