Fix multiple issues with CSG module.

[madmiraal]

Replaces BuildPoly with Build2DFaces, which creates faces as each pair of face intersections are processed, instead of trying to create them after all the intersections are processed. Ensures that faces are merged when possible, and removes degenerate triangles.

Treats the child as inside the parent when faces are coplanar.

Includes a general clean up of csg.h and csg.cpp.

Fixes #21125
Fixes #33117
Fixes #33121
Fixes #33665

[reduz]

that is impressive

[Calinou]

Thanks a lot for working on this 🙂

Correctness is important to have, but is performance as good or better than the current implementation?

[madmiraal]

@Calinou Do you have a benchmark we can test it against?

[Calinou]

@madmiraal I made one: test_csg.zip

It may not be representative of a real-world use case, but it's a start. Try moving the CSGSphere node called "MoveMe" and see how fast the editor reacts (compare the current master branch and your PR).

[madmiraal]

@Calinou I've tested it against your scene (and many of my own scenes during debugging) and the subjective experience is the same.

[madmiraal]

In the absence of an existing benchmark, I've created the attached scene to push the CSG calculations beyond what they're designed for:
CSG Benchmark.zip

It consists of a CSGTorus, which I've increased to 16 sides and 16 ring sides and a script that creates 5 (but this is configurable) default CSGSpheres (12 segments and 6 rings) in subtraction mode, which are positioned evenly around the torus and configured to revolve along the torus every frame creating around 350,000 triangle intersections per frame.

On my laptop I get the following frame times:
Current CSG: 127ms ±3ms
Fixed CSG: 141ms ±3ms

So the fixed version is marginally slower, but I wouldn't say significantly so.

[Calinou]

@madmiraal Sounds good to me 🙂

In the future, we could add a debouncing delay to CSG mesh rebuilds to make the editor more responsive.

[akien-mga]

Thanks!