shadow: introduce a tighter caster culling method

[show50726]

Inspired by this post:
https://x.com/SebAaltonen/status/1717455941551919261?s=20

The basic idea is to look at the view frustum from the light, and find those planes that construct a "silhouette". All the renderables outside this silhouette will never cast the shadow to the view frustum so we can safely cull them. Note that in our approach, we construct a table to easily look up for each back planes combination, what would be the silhouette vertices, so we can avoid some calculation.

The basic idea is to look at the view frustum from the light's perspective and find the planes that construct a "silhouette". Any renderables outside this silhouette will never cast a shadow into the view frustum, so we can safely cull them. Note that in our approach, we construct a lookup table for each combination of back-facing planes to determine the silhouette vertices, allowing us to avoid some calculations.

Implementation reference:
https://github.com/godotengine/godot/pull/82584/changes#diff-8babe54c0b184d1c71d515dc1a35aedcb772169ef06ade4fe5d7f162086aaa2a

Comparison

I tested it with gltf_viewer using the Bistro scene.

Without tighter caster culling:

With tighter caster culling:

We can observe that the frame time is slightly lower in the image below.

[pixelflinger]

Is there a reason to not always do this?
I mean, should this be a config like you did, or just a feature_flag, that we eventually remove?

What are the downsides of always doing this?

[show50726]

I think people might want to leave this feature disabled when they are at CPU bound instead of GPU bound.

Additionally, as discussed offline, this shouldn't be used alongside techniques like VSM or mipmapping. Rather than disabling culling automatically, I added a warning in the configuration so the user can decide.

[show50726]

I've made it a per-view option instead