@Phil at the moment we need to test more on ATI hardware (our lab is all NVIDIA’s at the moment). I do have at least one friend with an ATI card but would appreciate any help on that front.

Again very impressed with what you are doing with OpenCL. I am looking forward to your new updates.

About my previous kd-tree statement and spatial hashing, there are a few relevant papers anyone interested could check out:

Popov et al. Stackless kd-tree traversal for high performance GPU ray tracing. Computer Graphics Forum (2007) vol. 26 (3) pp. 415-424

Rozen et al. GPU bucket sort algorithm with applications to nearest-neighbour search. Journal of the 16th Int. Conf. in Central Europe on Computer Graphics, Visualization and Computer Vision (February 2008)

This is great work! I have just “discovered” OpenCL and i am totally excited about the idea of using it in combination with the simulation nodes :D
Atm the these node trees are using “classic” multithreading on the CPU and an ad-hoc batch processing model: data from the scene (currently only vertex positions and some basic particle attributes) is copied into fixed-size batches, which are then processed per node by available worker threads. It is already very efficient as far as one can tell from the simple arithmetic nodes i made so far, but i’d gladly replace this with OpenCL for increased efficiency and better parallel design.

Currently i’m further improving the node interface capabilities to combine them with the RNA, so that basically all the properties in Blenders API can be used inside the node trees with a generic GetData/SetData node. When this works i’ll start digging deeper into OpenCL (your tutorials are awesome!) Any help would of course be appreciated, maybe we can actually join forces some day?!

Cheers
Lukas

PS: i love the turtle ;)

I still have a lot to learn! I do hope we can see more OCL adoption in Blender as time goes on.

As for the collisions, yes that is right, only the bounding boxes are checked against until one is hit, then it checks against the triangles inside. I have a lot more reading to do on that subject, apparently you can do some cool stuff with spatial hashing (but things like kd-trees don’t port well to the GPU)

I mean we have Nicholas doing bullet physic integration, and next version of it. 3.X will have OCL acceleration. so for the viewport speed ups we might get it for free. :)

anyways, you mentioned Gordon the advisor, helping you speed the collisions when using custom shapes (ie. not bounding boxes) using triangular meshes. But doing the speed up by somehow using the bounding boxes, is that like … using the bounding boxes as calculations , until the particle that’s gonna hit it, enters the objects bounding box then it switches over and goes through the triangles to find the collision point?.

anyways 5x speedups sounds awesome!