I’m working on 3D scan matching which needs a fast nearest neighbour search approach. I tried GLSL with a grid based brute force algorithm. Now I need to apply some more advanced structure to make improvement. But it’s difficult to implement complicated algorithms on GLSL.
Now I wanna build a kd-tree by CUDA. Look my kd-tree is based on 3D points, meaning, each intermediate node contains a point. So it’s quite similar to what photon mapping is using.
Does anyone have experience of building a GPU based kd-tree? Or adding GPU based NNS? I really can’t figure out how to make the procedure like SIMD style.