remove extra renderpass when using depth sensing

[cabanier]

The current implementation of depth sensing introduces an extra renderpass.
I was tracing some WebXR experiences with depth sensing and noticed that if there's also a shadowmap, the WebXR texture ends up being flushed. In addition to the extra overhead, this also removes foveation (because it is no longer applied after a flush)

This change makes the painting of the depth values part of the regular draw pass.

Output of the render pass profiler with old code:

Surface 0    | 3360x1760 | color 32bit, depth 24bit, stencil 64bit, MSAA 4, Mode: 1 (HwBinning) | 72  288x320 bins ( 56  rendered) |  4.93 ms | 174 stages : Binning : 0.268ms Render : 0.807ms StoreColor : 0.442ms Blit : 0.005ms Preempt : 2.14ms StoreDepthStencil : 0.672ms
Surface 1    | 4096x4096 | color 32bit, depth 24bit, stencil 0 bit, MSAA 1, Mode: 0 (Direct)    | 1   4096x4096 bins ( 1   rendered) |  1.39 ms | 4   stages : Render : 0.798ms Blit : 0.006ms Preempt : 0.581ms
Surface 2    | 3360x1760 | color 32bit, depth 24bit, stencil 64bit, MSAA 4, Mode: 1 (HwBinning) | 63  384x256 bins ( 63  rendered) |  6.11 ms | 189 stages : Binning : 0.074ms LoadColor : 0.102ms Render : 1.723ms StoreColor : 0.099ms Preempt : 3.385ms LoadDepthStencil : 0.112ms

Surface 0 = clear + drawing of the depth pixels
Surface 1 = shadowmap (4kx4k?)
Surface 2 = drawing of the regular scene

Notice saving and restoring of the depth stencil buffer which adds overhead

Output of the render pass profiler with new code:

Surface 0    | 3360x1760 | color 32bit, depth 24bit, stencil 64bit, MSAA 4, Mode: 0 (Direct)    | 1   3360x1760 bins ( 1   rendered) |  0.00 ms | 1   stages : Render : 0.003ms
Surface 1    | 4096x4096 | color 32bit, depth 24bit, stencil 0 bit, MSAA 1, Mode: 0 (Direct)    | 1   4096x4096 bins ( 1   rendered) |  0.85 ms | 2   stages : Render : 0.844ms Blit : 0.005ms
Surface 2    | 3360x1760 | color 32bit, depth 24bit, stencil 64bit, MSAA 4, Mode: 1 (HwBinning) | 72  288x320 bins ( 46  rendered) |  8.49 ms | 95  stages : Binning : 0.101ms Render : 2.033ms StoreColor : 0.403ms Blit : 0.005ms Preempt : 5.594ms

Surface 0 = clear
Surface 1 = shadowmap
Surface 2 = drawing of the depth pixels + drawing of the regular scene

This contribution is funded by Meta

[github-actions]

📦 Bundle size

Full ESM build, minified and gzipped.

Filesize dev	Filesize PR	Diff
676.3 kB (168.2 kB)	676.4 kB (168.2 kB)	+113 B

🌳 Bundle size after tree-shaking

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

Filesize dev	Filesize PR	Diff
457.1 kB (111 kB)	457.3 kB (111 kB)	+113 B

[cabanier]

@mrdoob thoughts? This will reduce overhead on mobile gpu's.

I'm still trying to find out where that invalidOperation error is coming from. :-\

[mrdoob]

I'm thinking of refactoring the current implementation.
I'll work on it tomorrow.

[cabanier]

I'm thinking of refactoring the current implementation. I'll work on it tomorrow.

Sounds good! That gl warning is throwing me for a loop so maybe it will go away with the refactor.

[cabanier]

If you want to debug renderpasses, you could use ovrgpuprofiler
If you see depth being stored/loaded, there is a problem.

[mrdoob]

Hmm... Does the Quest Pro support depth sensing?

[cabanier]

Hmm... Does the Quest Pro support depth sensing?

It does not unfortunately.

[mrdoob]

Ah! Won't be able to work on this until Monday then...

[cabanier]

I'm thinking of refactoring the current implementation. I'll work on it tomorrow.

Did you end up refactoring it?

[Mugen87]

Fixed via #28530 🎉 .