Cascaded Shadow Maps

[x2048]

Upgrades the current shadow mapping implementation to cascaded shadow maps.

Benefits:

• Affine transformation for the shadow = simpler math, faster shader, less maintenance
• Fixes shadows for large objects/meshes such as e.g. in Little Lady or shadows for clouds in the future
• Smoother updates of shadows (every cascade updates at its own pace, cascade 0 updates every frame)
• Much less peter-panning
• Less VRAM consumption for similar shadow quality/scale
• Visible area covered with shadows in most cases even with fast camera movement

Other fixes:

• Removed flicker when mapblock is updated (e.g. flowing water)
• Removed flicker when camera offset changes (every 200 nodes)
• Added bilinear filter for SHADOW_FILTER == 1 for playable shadow quality with more performance

Notes:

• 3 cascades
• Cascade 0 is fixed small size in world and updates every frame, giving reasonable shadow quality close to the player
• Cascade 2 is updated every shadow_update_frames frame (up to 32) and cascade 1 somewhere in-between
• Entity shadows on cascades 0 and 1
• Colored shadows work
• Soft shadows work (may need some tuning)
• Shadow performance with client_mesh_chunk_size = 1 may be lower because of increased number of drawcalls covering larger area. Increasing shadow_update_frames should help spread out the updates.
• shadow_map_texture_size now controls resolution of each of 3 cascades (2 for entities) and should be divided by 2 when migrating

To do

This PR is Ready for Review.

• Testing needed
• Tuning/defaults feedback needed

How to test

1. Turn on the shadows
2. Enter a game that supports shadows
3. Check the shadows both at close and long distance, when flying etc.

[lhofhansl]

Nice!

Some observations... There's a visible change between Cascade 0 and 1 (i.e. as approach a shadow of a larger object there's a visible discontinuity). I also saw some "phantom shadows", as you approach an area without shadow there appear circle shaped darker areas. (I'll see if I can do a short video of that).
It seems plant shadow are only in Cascade 0...?

[lhofhansl]

I think the cascade default distances might be too small. And I cannot easily figure out where the distances are actually set for each cascade.

[lhofhansl]

Just so I remember later... The cascade max distance is set here:

void DirectionalLight::update_frustum(const Camera *cam, Client *client, bool force, u8 max_cascades)
{
	client->getEnv().getClientMap().allocateDrawListShadowCascades(getCascadesCount());

	f32 scale = 20. * BS / getFarValue();
	f32 scale_factor = 1./pow(scale, 1./(cascades.size() - 1.));

	for (auto &cascade: cascades) {
		cascade.farPlane = farPlane;
		cascade.scale = scale;
		scale *= scale_factor;
	}

So if viewing_range would 100, the first cascade would end at 20, the next one at around 45, and the last at 100.
For 1000, it be again at 20, then 140, then 1000. Etc.

If we wanted the smaller cascades to reach farther we'd change it to
f32 scale = 40. * BS / getFarValue(); .. or something.

Or maybe we'd start with scale just set to 0.1 (i.e. 1/10 of the viewing range), or 0.2.

[lhofhansl]

With a changed distance calculation I'm more happy with this.

Still have to look into the "phantom shadow"/moire effects with cascade 0 - as time permits.

[lhofhansl]

I need to slog through a lot of code to understand which bias I need to tweak to eliminate any fake shadows.

After this PR:

Notice the round-ish extra dark artifact on the cliff and that the overall scene is darker.

Before this PR:

[lhofhansl]

Setting the z perspective bias to 0.25 instead of 0.5 gets rid of the "fake" shadows.

[lhofhansl]

I'll file a separate PR, starting with x2048 and my changes on top as separate commits.

[lhofhansl]

Continued in #13956
We can close this one. The initial commit there is of course credited to @x2048 !