InstanceNode: Honor updateRanges.

[Mugen87]

Fixed #31814.

Description

The PR makes sure the internal attributes in InstanceNode honor the update ranges setup of the app level attributes.

@AndreyPrakhov With this PR, the bufferSubData() performance in your fiddle improves significantly. Do you mind testing if everything works on the functional side as expected?

[github-actions]

📦 Bundle size

Full ESM build, minified and gzipped.

	Before	After	Diff
WebGL	338.91 79.11	338.91 79.11	+0 B +0 B
WebGPU	581.32 160.08	581.53 160.13	+211 B +47 B
WebGPU Nodes	579.93 159.84	580.14 159.88	+211 B +47 B

🌳 Bundle size after tree-shaking

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

	Before	After	Diff
WebGL	470.78 113.9	470.78 113.9	+0 B +0 B
WebGPU	650.92 175.81	651.13 175.86	+211 B +51 B
WebGPU Nodes	605.02 164.95	605.23 165	+211 B +48 B

[Mugen87]

Side note: When the instanced mesh only holds a small number of instances, InstanceNode uses buffer() which results in a UBO where update ranges are not supported yet. However, the update ranges mechanism is especially important when large attribute data should receive (small) partial updates. This use case should be supported now.

[Mugen87]

@sunag What was the reason for using UBOs for instanceMatrixNode in the first place? Why not just always using attributes?

[AndreyPrakhov]

Everything is great. It works as expected.

[sunag]

@sunag What was the reason for using UBOs for instanceMatrixNode in the first place? Why not just always using
attributes?

The idea was to eliminate some extra computations; the official WebGPU instancing example also uses a UBO, but the performance issue can be much broader, as the PR itself shows. I don’t think anything prevents us from using only attributes if that yields better results.

[Mugen87]

In this case, I would say let's keep the status quo since the UBO usage only works with a small number of instances. There is probably no measurable performance difference in updating the entire small buffer or only a portion but loosing the optimization from #28726. If someone shows us the difference, we can of course always use attributes.