No-jitter models loading

[sdumetz]

Reducing renderer jitter on model loading would enable loading smaller models seamlessly in a (yet to be developped) dynamic LOD system.

Using WebGLRenderer.compileAsync() from THREE r161 (#27098 seems to reduce the initial blocking time by half for a given model, without much effect on other metrics.

Additionally, WebGLRenderer.initTexture() does drastically reduce upload freeze for models with large textures and low (ish) vertex count, especially if it is able to load textures through multiple frames.

[sdumetz]

Detailed test results. Tests were made with dev build in standalone mode, mean of 3 runs. Looking for the sum of [Violation] 'requestAnimationFrame' handler took <x>ms reported during object loading. Results should be considered accurate within ~10% of their value.

Model / Test	Delay (mean)
10M Vert. 5*2k tex.
- initial	1380ms
- compileAsync	814ms
- initTexture	796ms
- compile+init	804ms
25k Vert. 1*8k tex.
- initial	675ms
- compileAsync	480ms
- initTexture	74ms
- compile+init	<50ms

Both model were purposefully chosen quite large to better show the performance bottlenecks. We see that it's possible to largely speed up texture loading but large geometries will stay a problem unless we find some additional optimizations.

The geometry limit is not that much of a problem because keeping the vertex count below the 1M (arbitrary) limit is easy enough in most cases.

[sdumetz]

Running the previous tweaks on large multi-model scene, we go from there :

To there

Which is better but shows how an unoptimized model is unlikely to load seamlessly.

Ensuring a serialization on renderer.initTexture() seems to help a bit : It makes intuitive sense as models loading in parallel might otherwise try to init their textures during the same animationFrame. load time improvements were however not clear enough to be sure about this.

Compressing the scene's larger models into 2k maps (from 30MB total size to 15MB) still yields some delayed frames which tends to indicate some other source of contention to uncover.

Some of this contention seems to come from the environment map being loaded more than once.

[sdumetz]

See https://github.com/Holusion/dpo-voyager/tree/seamless_load_proto for ongoing efforts

[sdumetz]

See Smithsonian/dpo-voyager#305, Smithsonian/dpo-voyager#332, Smithsonian/dpo-voyager#319, Smithsonian/dpo-voyager#288.