Knowing the axis at compile time
A problem with using recursion on an kd tree is that every time you recurse, you have to access a different axis, so you might have branches in your code. A branch predictor might have problems seeing the pattern that the axis alternate with each call. One way to avoid this would be to handle the nodes in bfs order. Then you only have to alternate the axis once every level. But this way you lose the nice divide and conquer aspect of splitting the problem into two and handling those two problems concurrently. So to avoid, this, the axis of a particular recursive call is known at compile time. Each recursive call, will call the next with the next axis type. This way all branching based off of the axis is known at compile time. A downside to this is that the starting axis of the tree must be chosen at compile time. It is certainly possible to create a wrapper around two specialized versions of the tree, one for each axis, but this would leads to alot of generated code, for benefit. Not much time is spent handling the root node anyway, so even if the suboptimal starting axis is picked it is not that big of a deal.