Environment

Ubuntu 18.04.1 LTS
cuda == 10.0
libsdl2-dev = 2.0.8
libsdl2-ttf-dev = 2.0.14

library installation method in Ubuntu 18.04.1:

$ sudo apt install -y libsdl2-dev libsdl2-ttf-dev

How to use

$ mkdir build && cd build
$ cmake ..
$ make -j8 && make install
$ cd bin
$ ./NBodySimulation [random | stable | symmetric | breathe]

(Arguments in [] are optional).

Release Notes

cpu version - Oct. 28

Simple implementation with boundary detection. 60fps.

gpu version 1.0 - Oct. 29

This is a naive version with limited scalability, which can process at most 1024 bodies. That's because we simply map i/j in serial version to blockIdx/thredIdx, which has a maximum of 1024.

Features:

60 fps when doing 1024 bodies simulation with trail effect
support trails using frame decay, this part is writen with cuda kernel
applied cuda kernel to acclerate rasterization
implemented boundary detection
some optimization details:
- Avoid redundat cudaMalloc and cudaFree operation when doing rasterization
- slightly optimized reduction in time step, when calculate acceleration

TODOs:

write more flexible kernel to compute time step, which has better scalability (support more bodies)
do more optimizations to achieve better effiency, e.g. AOS(array of structure) -> SOA
add more features, e.g. carefully designed initial conditions, force fields, user interaction of effect control

gpu version 2.0 - Nov. 1st

This version is scalable and has high peroformance.

Features

60fps when doing 8096 bidies simulation on 980Ti (12-14 fps on K620).
support trail effect using frame decay (frame = beta * frame + newState) over time. This effect is implemented with cuda kernel.
applied cuda kernel to accelerate rasterization
boundary detection and bounce back effct
elaborate initial state. Support command line arguments to specify pattern, i.e. choose different initial states
- usage: ./NBodySimulation [random | stable | symmetric | breathe]

Optimization details

allocate/Free memory collectively, avoid redudantly frequent allocation and free
keep a shared memory copy of input data to reduce global memory accessing
surprisingly, we found rsqrt() is far more fast than pow() (at least about 6x - 8x)
store intermediate results to save float-point computations.

Acknowlegement

Thanks Hu Wenbo for providing a cmake project file for my referrence and pointing out the problem of using pow() and rsqrt().

rayleizhu / CSCI5390-Nbody