profiling

[joevenzon]

Ah cool, very interesting. Their raycast is actually way faster than the one I was doing, so getting it faster still would be really cool.

We could probably use some frame coherence to speed up the bezier raycasts. Right now they are accelerated using an r-tree sort of data structure (aabb_space_partitioning.h) and we could probably start from the last frame leaf and traverse up the tree and then back down to the new leaf. Also, this data structure needs to be cache-optimized. I never dynamically add objects to it, and each level has only 2 branches, so it could easily be stored as a vector (like is commonly done with binary heaps). This would also speed up the object culling, which uses the same data structure.