TSL: Support layout functions with uniforms shared across materials

[shotamatsuda]

Related issue: #31325

Description

This PR adds support for using layout functions with uniforms shared across different materials, by caching them per material when they contain uniform nodes.

Layout functions are built and cached per backend for the first material, and associated uniforms become invalid for other materials. This causes the following situation to fail.

const x = uniform(1)

const f = Fn(() => x).setLayout({ type: 'float', name: 'f', inputs: [] })

const material1 = new NodeMaterial()
material1.fragmentNode = f()

const material2 = new NodeMaterial()
material2.fragmentNode = f()

With this PR, the example above works. Nested layout functions that includes uniforms are also supported.

const x = uniform(1)

const f = Fn(() => x).setLayout({ type: 'float', name: 'f', inputs: [] })
const g = Fn(() => f()).setLayout({ type: 'float', name: 'g', inputs: [] })
const h = Fn(() => f()).setLayout({ type: 'float', name: 'h', inputs: [] })

const material1 = new NodeMaterial()
material1.fragmentNode = h().add(g())

const material2 = new NodeMaterial()
material2.fragmentNode = f().add(g())

[github-actions]

📦 Bundle size

Full ESM build, minified and gzipped.

	Before	After	Diff
WebGL	365.43 86.79	365.43 86.79	+0 B +0 B
WebGPU	661.82 182.63	662.29 182.74	+472 B +105 B
WebGPU Nodes	659.86 182.33	660.34 182.44	+472 B +103 B

🌳 Bundle size after tree-shaking

Minimal build including a renderer, camera, empty scene, and dependencies.

	Before	After	Diff
WebGL	498.49 121.58	498.49 121.58	+0 B +0 B
WebGPU	733.06 197.45	733.53 197.56	+472 B +104 B
WebGPU Nodes	682.28 184.86	682.75 184.96	+472 B +99 B

[Mugen87]

TBH, I find it not ideal that an additional flag would have to be used to avoid this issue. Can't the layout functions be cached per node builder by default?

[shotamatsuda]

I find it not ideal that an additional flag would have to be used to avoid this issue.

I agree, but it will have broader performance ramifications, I suppose, when layout functions are cached per material by default, since all the complex functions like those in mx_noise.js must be rebuilt for each material.

[shotamatsuda]

I roughly measured the performance difference in the first frame involving node building in webgpu_materialx_noise:

Cache per backend: approx. 100-105ms
Cache per material: approx. 120-125ms

[shotamatsuda]

Although it needs further testing, by detecting whether a layout function has uniforms, it appears that the user-set "shared" flag will be unnecessary.