Massive performance drop caused by long thin triangles

[stephomi]

I subtract a 1M mesh with a <10k box (screenshot).

The ugly triangulation on the box (front/back faces) is causing massive performance slow-down.
With long thin triangles: 2472ms
With uniform box: 482ms

It also uses lot of memory according to the xcode memory chart (not sure how precise that is)
With long thin triangles: 1GB -> 5GB spike
With uniform box: 1GB -> 1.71GB spike

note: the ugly triangulation is only on faces that never intersect with the other mesh

[pca006132]

This is probably due to how our collider currently works. Long thin triangles will make the bounding box larger, and more collision pairs are recorded. We plan to optimize the collider (not exactly this issue, but that should help with memory usage and give much better performance in general), but it will take some time to get to that.

[elalish]

Yeah, we use AABB (axis-aligned bounding boxes) for our BVH, so long triangles, especially if they aren't axis-aligned, will definitely hurt performance a lot. In general I was figuring this wouldn't be a huge problem because most times when you have large triangles you also only have a few. You could try RefineToLength() - curious if that would help or hurt?

[stephomi]

Since I control this geometry, for now I can simply make a regular box.
I was just surprised by how big the difference was.

[elalish]

Yeah - and it's a bit hard to tell, but those long triangles look roughly axis-aligned. Is that true? Because that's even more surprising that it makes that much difference. Can you give us the models so we can repro?

[stephomi]

Here's the model
high.obj.zip