Added model_recalculate_bounds_exact()

[paulmelis]

This computes a model bound based on (transformed) vertices, instead of using per-mesh AABBs. Main reason for this is to have a much tighter bound, especially for interaction volumes, at the cost of a longer time needed to compute bounds for models with many vertices.

Note that I haven't tested this from the C# side, but the C code does compile on Linux 😄

[paulmelis]

Btw, model_recalculate_bounds_exact() isn't called in any way, except when forced through Model.RecalculateBoundsExact(). Perhaps that call can even be offloaded to a thread/async, in order not to lock up the app.

[paulmelis]

Comparison on a model file where I noticed the (still) quite loose bounds. This is a GLB file that has a few different transforms (e.g. -90 rotation in X, -11 degree in Z) above several pieces of geometry.

Current: center [-0, 0.26, 0], dimensions [12.43, 0.52, 11.94]

Exact: center [-0, 0.26, -0], dimensions [9.28, 0.52, 8.5]

Edit: fixed wrong pasted bbox values

[maluoi]

This is a pretty neat feature! Do you know what level of "slower" this might be? I imagine it's probably not so bad that it'd really need to be async, but I'm curious!

I've also taken to using some of the DirectXMath functions directly for bounds stuff, I think its likely to have some good speed benefits in this loop! XMVectorMin/Max in particular, in case the compiler isn't auto-vectorizing that already. It'd be nice if you could add that here! Might also be worth hoisting the XMMATRIX conversion out to a higher scope like here, but I'm less certain about that.

I'm not genuinely concerned about micro-optimization here, just wanted to shout out that there may be some low hanging fruit :) I'll take it either way if you're not keen on that right now!

[paulmelis]

Good suggestion on the DirectXMath functions. Just updated the patch and exact bounds computation time went from 99ms to 33ms. This is with the C API and a large unoptimized (85 visuals, 5.4M vertices and 5.4M indices) GLB file.

[maluoi]

Wooh, that's a hefty model! Yeah, okay, I can see why that might be a good candidate for async still, I'll keep that in mind for later :) Though I'm glad to know the DXM functions helped that much!

[paulmelis]

I only now noticed that this model (one of a set I've been using for months) consists of unconnected triangles 😆 "Blame" it on modern hardware (GTX3090) and lean software like SK to make that hardly noticeable.