WebGPURenderer's Matrix4 side-effects break tree-shaking and WebGLRenderer

[CodyJasonBennett]

WebGPURenderer has side-effects that directly modify THREE.Matrix4 that were introduced in #20283. This is a bit problematic since those side-effects can't be tree-shaken, and they will bundle alongside threejs in client/server bundles (these can be code-split to be imported from multiple pages/bundles). This can break apps that utilize both WebGLRenderer and WebGPURenderer (like a blog or somewhere you'd have a lot of demos).

I discovered this by importing a flatbundle of three/examples/jsm (from #23368) but you can reproduce by importing the renderer individually:

import { WebGPURenderer } from 'three/examples/jsm/renderers/webgpu/WebGPURenderer'

[Mugen87]

I've outlined an alternative approach that would also solve the side-effects issue: #20276 (comment)

[LeviPesin]

Looking at https://github.com/mrdoob/three.js/blob/dev/examples/jsm/renderers/webgpu/WebGPURenderer.js#L251-L252, I think the way to solve this (including #24383) would to do the following:

1. Add Camera.ndcProjectionMatrix property,
2. Add a parameter ndc = true/false to Frustum.setFromProjectionMatrix method.

Looking on these lines again, maybe we even can move NDC-related changes entirely into 2 without doing 1...

[CodyJasonBennett]

What happens if WebGPU camera matrices are used in WebGL? I'm not sure I understand the relationship between the back-end's clipping space and the resulting projection matrices.

[WestLangley]

What happens if WebGPU camera matrices are used in WebGL?

Near-plane clipping will occur at the wrong distance.

[Mugen87]

Regarding #25823 this issue is not only problematic for tree-shaking, there are also side effects if an app wants to switch from WebGLRenderer to WebGPURenderer and vice versa.

A global like like suggested in #20276 (comment) would support this use case but not when both renderers are used at the same time (e.g. when rendering to different canvases).

[LeviPesin]

I think the approach described in #23372 (comment) is the best. Maybe we can try it?