GPU Tree Code examples and applications

I haven't been very active using CUDA recently. I need a lot of energy to actually work on CUDA programs because my cuda coding environment is based in windows xp and everything is running really slowly with MSVC 2010. Also I have problems tracking down errors via the compiler, so sometimes a project will have an error that won't show up in the list of build incidents.

Anyway I thought I'd share some information about GPU tree codes.

A good resource for tree based N-Body is here
http://castle.strw.leidenuniv.nl/software/bonsai-gpu-tree-code.html

This is more like a multi-grid (I think). The multi-grid method is useful in computational fluid dynamics (CFD) and computational cosmology (CC). It is often convenient to only process N-body interactions between particles or vortons inhabiting the same grid sector, or nearby grid sectors, because the center of gravity or center of mass will be a good enough approximation for distant grid cells. For this reason these algorithms have become very popular on CUDA, as they allow users to compute real-time systems with many more particles than a pure brute force N-body calculation.

Another thing to search for is the Barnes-Hut algorithm on GPU. This is a famous multi-grid tree algorithm. In fact I think the code in the link above is based on Barnes-Hut.

Another interesting CUDA application for simulation is named OpenCurrent (see google). Basically this is an industrial strength CFD solver built in CUDA, and so its a seriously useful building block for any CFD simulations, including smoke, fire and others. Incidently, CudaFire is built upon OpenCurrent - I spent a long time trying to compile CudaFire and eventually started a project based on a copy of the CUDA project "FluidsGL", because the title was mentioned in the cuda fire build, and added all the files into the project. I had to change the directory structure, had to download TinyXML and then I had to copy the entire OpenCurrent source codes into the project directory and add those. Then I had to build without debug information - because my computer ran out of memory for building the .exe. The program had a couple of errors too, a thrust functor was constructed with 3 floats as the template argument list instead of 3 float4 objects. Also there was a big linking problem that I solved with some trickery, including forcing it to build with multiply defined symbols. Eventually the program worked, but broke instantly.  I saw a bunch of particles in a quad, then the screen went blank, I could cycle through modes to see the slice renderer, the cube etc, but there was nothing going on - i.e. no simulation just a blank space.

Gotta go for a while, anyway watch this space for more information.