Collision: more accurate computation with acceleration and long dtime

[kno10]

Refreshed version of #15029 (reverted for 5.10.0 by #15400, includes #15378)

This fixes inaccuracies in the old collision code which in essence assumed a linear movement, and in case of a collision still used the final speed.

This version:

• uses a parabolic computation with acceleration (postponed for now, did not get it to work perfectly)
• more accurately computes the time of collision
• sorts node boxes to (hopefully) find collisions faster (dropped, did not improve enough)
• adds several additional unit tests (not all are perfectly passed by the previous nor the current code yet)

It is still an approximation to not need to iterate too much - consider a ball bouncing on a surface, and we have a long time step: the ball might do multiple bounces; this code will still only allow one per timestep: to avoid issues when standing on a floor, any collision will zero the axis velocity (unless bouncy) and zero the acceleration (always) for now.
The ultimate solution would be to make everything so fast that we can do much smaller dtime steps and all of this does not matter anymore.

The following cases should be tested:

• bouncy nodes are bouncy
• slippery nodes are slippery Slippery nodes are no longer slippery #15375
• falling nodes turn back solid Falling nodes do not turn back to solid blocks when connected to a server #15381
• walking up stairs works
• not glitching into the floor after stairs
• not stuck when placing nodes right below oneself

[sfan5]

You can rebase now.

[kno10]

@appgurueu: working on taking acceleration into account directly.

[kno10]

@sfan5 @appgurueu this version now is probably good for reviewing and testing.

[JosiahWI]

I'm not doing a thorough review, but I looked into the MinGW CI failure. Please include the <algorithm> header since you used std::sort; it's transitively depending on an include all the way down in an irrlicht/ header, and that also makes the compiler errors much more confusing.

[SmallJoker]

Did a few rudimentary performance comparisons. In a new devtest world.

Walking on floor, in circles. Full range view enabled.

• master: 10.5 us
• PR: 13.0 us

Holding the forward button in this scene:

• master: 18 us
• PR: 24 us

This is actually pretty good considering the introduced sqrt calculations.

[SmallJoker]

First review iteration. Works well from what I can see so far.

Tested:

• singleplayer: walking
• local server with dedicated_server_step = 0.6: falling nodes

[kno10]

Giving up on the parabolic movement for now. I could not get rid of all the glitches.
If you walk diagonally against a wall, you would frequently get a collision with the next node, too.
Hence I reset this pull to the basic version for now. For this one I could not notice obvious glitches; it is closer to the current version, but improves the accuracy of the collision time computation.
We can then try to tackle the collision detection with acceleration (parabolic, not linear movement) in a subsequent step/pull potentially in a future version.

[kno10]

rebased, no changes. Ready for review.

[SmallJoker]

Slippery test

Tested with this script: #15375 (comment) (N° 173 for my reference)

• PR sand: 11.7 tiles
• PR grass: 15 tiles
• master sand: 11.6 tiles
• master grass: 14.9 tiles

☑️ This is good enough.

Bouncy test

Physical entitiy code:

diff --git a/games/devtest/mods/testentities/visuals.lua b/games/devtest/mods/testentities/visuals.lua
index 6bb1b8282..7517f7a3e 100644
--- a/games/devtest/mods/testentities/visuals.lua
+++ b/games/devtest/mods/testentities/visuals.lua
@@ -52,7 +52,11 @@ core.register_entity("testentities:mesh", {
                textures = {
                        "testnodes_mesh_stripes2.png"
                },
+               physical = true
        },
+       on_activate = function(self)
+               self.object:set_acceleration({ x = 0, y = -9.81, z = 0 })
+       end,
 })
 
 core.register_entity("testentities:mesh_unshaded", {

Setup: sand pillar + testnodes:bouncy100

1. Local server
2. Spawned with the Entity Spawner tool -> testentities:mesh.
3. Executed the command while in downwards movement: //lua local t0 = os.time() while os.time() < t0+2 do end (from Bounciness issues with large dtime steps #15027 (comment))
4. The entity shoots off into the sky, hence does not fix Bounciness issues with large dtime steps #15027.

[kno10]

That last test goes beyond the scope of this pull request, as it requires computing the parabolic movement.
As this did not yet work sufficiently well, and there was little interest, I have currently given up on this.
The case it tries to fix is catapulting entities even with next to no fall velocity into the sky nevertheless.
Think of the villagers in Mineclonia etc. standing on beds and suddenly bouncing through (!) the roof with some dtime lag (e.g., because the village emerge code is still working, or because all entities want to pathfind now).
Try the same without the sand pillar on master and on this branch.