opencl-nbody

This repository contains all the OpenCL code and the report for the lab course "Parallel Programmeren" by Joost Vennekens and Wiebe Van Ranst (campus De Nayer, KU Leuven).

Who are we?

• Simon Vandevelde
• Dylan Van Assche

N-body problem

The N-body problem is a well known physics problem where we want to predict the individual motions of a group of celestial objects interacting with each other gravitationally. You can find more information about this on Wikipedia.org

To calculate all these motions a lot of computer power is needed if you have a lot of bodies. When using OpenCL we can use the GPU to calculate all the motions of the bodies in parallel.

Performance

We tested several methods and OpenCL kernels, we hit the limit with our current implementation at 50 000 bodies.

• Kernel 0: CPU version, fails to start
• Kernel 1 & 4: OpenCL kernel where only a small for lus is running on the GPU, fails to start (calculates the new position)
• Kernel 3: OpenCL kernel wehere a big for lus is running on the GPU (calculates the new speed)
• Kernel 5: Combines the OpenCL kernels 1 & 3 without optimising the data transfer (due time limit)

As you can see, the 3rd and 5th kernels are still running smoothly with only a frame time of 0.9 seconds! The CPU calculated 5 000 bodies in 0.6 seconds. You can find more benchmarks in report.pdf (Dutch only).

💡 If we had more time, we could optimize several things for the 5th kernel like: data transfers, memory types, ... to increase the performance.

⚠️ Not all files are included in this repository due the fact that they are not written by us and we don't have the permission to publish them.