Dynamic LOD switching

[sdumetz]

So there is this idea_ I worked on for the past few months that makes it possible to open a ~200 models scene with each model having a 4k diffuse (in High quality) on a mobile device.

It's still a very rough prototype but I wanted to share it and begin discussions on what requirements we'd have to meet for this to get merged into voyager.

This branch temporarily includes #266 to allow more camera mobility and a lot of performance patches that should be (or have been) extracted to their own PRs. It also needs to be rebased on Voyager's latest release. I'll do all this down the line as I clean everything up.

A working example is visible there:

https://lodtest.ecorpus.holusion.net/ui/scenes/Lascaux_retake2/view

Same model, without LOD (everything is squashed together and downscaled to a 8k diffuse map) is publicly available on the french ministry of culture website and takes longer to load initially, has a lot of blur and doesn't even work on most mobile devices.

I also have a simpler demo example that should clearly show what is happening:

https://lodtest.ecorpus.holusion.net/ui/scenes/simple_room/view

Reasoning

I started with 2 facts :

• as stated in Derivative quality "High" is always used #254 even lower-end devices are fully capable of showing High quality (4k textures) derivatives and there is no real point in trying to serve larger models even for high-end devices
• Quality is selected at the scene level. Voyager has the ability to use heterogeneous quality internally but no way to exploit this.

My objective was to be able to handle spatially large models. Things like a large painting, multi-figure sculpture or a whole room. In such an environment, I want what's right under my nose to have a very high texture resolution (abt. 4k for something filling the whole screen is a good upper bound) but on the other hand I don't care about the resolution of pretty much everything else.

Having a scene with 10s of objects, each with 4k diffuses tends to overload even desktop computers with low end discrete GPUs. I want to be able to have 100s of them.

How it works

I repurposed the existing Derivative system to dynamically switch models quality depending on the "perceived importance".

Pretty much everything happens in the new CVDerivativeController class. The scene is initially loaded with Thumbquality (as it was before). Then the controller regularly sorts every model in the scene by importance (on-screen size and angle to camera are taken into account) and increments the quality of the most important model(s). If/When too much models are upgraded, it downgrades models as needed starting with the least important ones.

Heuristics

Tradeofs that had to be made. Can and should be adjusted.

Performance

Estimating a performance budget is hard (see #254). Even if we could reliably determine a device's graphic power (we can't), we can't know if the user wants us to use 100% of his computer's power.

I settled somewhat arbitrarily on an initial budget of 2 High resolution models that is then further decreased if we detect a low-end device (CPU count, RAM size, is a mobile device).

Model weight

Models are weighted based on their size in camera space, their distance to the camera and their angle relative to the Camera.forward.

The Distance modifier is somewhat redundant with size but we kinda want far away models to be less important even if they are very big. In fact most LOD algorithms are purely distance-based.