Hi Kurt
I'm not quite sure about what 'pulling from another source' means here. So I can only reply you based on my own understanding.

This BVH is something I wrote myself when I was learning the PBR-book. Actually PBR book provides an excellent implementation, but I thought if I just copied the code then there wouldn't be much I could learn. This code is based on the understanding of mine for the contents in the PBR-book, therefore the idea behind the BVH implementation should closely resemble to that of PBR-book's.

And, maybe, if I understand correctly, you are asking about whether we are able to use BVH algorithms implemented by others or load precomputed BVH into this renderer. The answer is yes but it comes at a cost: we need to implement the correct interface, which converts the BVH outputs from whatever code it is, to Taichi lang ti.dataclass. This is the only modification my renderer requires, that is, you just need to correctly feed the constructed BVH and everything is good to go.

However, note that this BVH implementation is suboptimal. Here is the situation: Taichi lang has limited ability to dynamically allocate memory inside of its kernel function, but BVH tree traversal (the optimal one) requires building a stack (otherwise, use recursive calls). I didn't find a easy way to build a stack in the kernel function (actually, dynamic node in Taichi lang might help but I'm not sure whether it can be a fast implementation), so the tree traversal is done linearly. There is going-backwards: if the lchild is accessed first and pruned, I can directly jump to the rchild and this would accelerate the rendering since we do skip some impossible ray-intersections, but I can't do this the other way around, i.e I can't check rchild first then jump back to lchild. The implementation could get messy so I stuck to the easier and more intuitive way.