Variables

Storage and uniform variables

When using TypeGPU for resolution, there is no need to ever define either var<uniform> or var<storage[, address_space]>, as such definitions are added automatically.

For example, in the code snippet below, the external counter is automatically included as a @group(0) @binding(0) var<storage, read_write> counter: vec3f. In this case, the group and index match the automatically generated “catchall” bind group, used for fixed, “bindless” resources.

const counter = root.createMutable(d.vec3f, d.vec3f(0, 1, 0));

const increment = tgpu.computeFn({
  in: { num: d.builtin.numWorkgroups },
  workgroupSize: [1],
})((input) => {
  const tmp = counter.$.x;
  counter.$.x = counter.$.y;
  counter.$.y += tmp;
  counter.$.z += d.f32(input.num.x);
});

const pipeline = root.createComputePipeline({ compute: increment });

console.log(tgpu.resolve([pipeline]));

// resolved WGSL
@group(0) @binding(0) var<storage, read_write> counter: vec3f;

struct increment_Input {
  @builtin(num_workgroups) num: vec3u,
}

@compute @workgroup_size(1) fn increment(input: increment_Input) {
  var tmp = counter.x;
  counter.x = counter.y;
  counter.y += tmp;
  counter.z += f32(input.num.x);
}

Private and workgroup variables

For variables of private and workgroup address spaces, TypeGPU provides tgpu.privateVar() and tgpu.workgroupVar() constructor functions.

const threadCounter = tgpu.privateVar(d.u32, 0);

const getNext = () => {
  'use gpu';
  threadCounter.$ += 1;
  return [42, 418, 23][threadCounter.$];
};

const myComputeFn = tgpu.computeFn({
  workgroupSize: [64],
})(() => {
  const a = getNext();
  const b = getNext();
  const c = getNext();
  // ...
});

Const variables

The meaning of const differs between JS and WGSL. In JS, a const is a reference that cannot be reassigned, while in WGSL, it means a value known at shader-creation time. Therefore, to use the WGSL const, TypeGPU provides tgpu.const(), an interface analogous to tgpu.privateVar() and tgpu.workgroupVar(), with the only difference being that the previously optional init value is now required.

const Boid = d.struct({
  pos: d.vec3f,
  vel: d.vec3u,
});

const boid = tgpu.const(Boid, {
  pos: d.vec3f(1, 2, 3),
  vel: d.vec3u(4, 5, 6),
});

const func = () => {
  'use gpu';
  const pos = boid.$;
  const vel = boid.$.vel;
  const velX = boid.$.vel.x;
};

Caution

TypeGPU does not support defining tgpu.const() variables inside a TypeGPU function.