CUDA test 6: 3D SPH simulations

3D XSPH examples (Dam break, dam break with obstacles, falling drop and Rayleigh-Taylor inestability) using nVIDIA CUDA on a GTS250. This isn't a real time video, but the application run at 15 fps using 20k-30k particles.

I used a cubic spline for the kernel, Tait's equation of state and a very simple speed correction. In some simulations the color indicate the speed module (scaled to the max) using a modified "jet color map" (extracted from matlab), in multifluid the color indicate each component.

The particle representation was made using allegro libraries and a very simple algorithm to 3D/2D projection. The final video was created joining 7 video test using avidemux.

Category:

Science & Technology

Tags:

• SPH
• fluid simulation
• nVIDIA CUDA

[jahkr]

OK, in my CPU code I'm using a tree search algorithm, and it's similar to linked-list. In the GPU code I'm using shared memory because it's very fast for access to all data at same time, the problem is the shared memory it's only 16k (GTS 250).

About the reference material, I began with the OpenCL and CUDA online reference and after tried to understand and modify the example codes in CUDA SDK.

[trimtrim1980]

Thanks for your reply. I never tried GPU, but there are some algorithms maybe you can use (such as link-list) to improve the searching speedup. By the way, in your code, which type of memory you are using for the searching neighboring partilces: global memory or shared memory. Is there any reference material recommend for particle programming in GPU?

[jahkr]

Thanks for comment.

I have used float4 data types to improve the performance. The application run at 240 Gflops with 25k particles, but the main problem is that number of operations increase with n^2 (with n=number of particles). In the CPU version I'm using a neighbour search optimization and the number of operations increase with n*log10(n), this means that using large particle numbers CPU is faster than GPU. I'm trying to adapt the neighbour serach to GPU, but isn't easy.

[trimtrim1980]

Excellent simulation. How about the cuda speed up and how about data type used? single float to double precious?