This program simulate the flow of a liquid around a solid object using the Lattice-Bolzmann method. Particular attention was dedicated to the technology aspect (design and implementation) rather than to the algorithm itself.
This project started as a fork of 2d-lbm-cuda and the initial development efforts were done here. Now it has become an independent identity with very ambitious goals.
- heavily refactor the original implementation to decouple the computation from the rendering
- design a coherent API
- allow multiple lbm implementations (cpu based, compute shaders, opencl)
- port the application to opengl 4.3
- add imgui integration
It is particularly difficult to port the lbm functions to compute shaders. Here is a list of problems:
- cumbersome declaration of buffer objects
- parameters of kernels have to be manually passed as uniforms
- kernel compilation errors are obscure
- Port the lbm computation to compute shaders
- Experiment with different work group size to maximize performance
- Interactivity
- start/stop the iterations
- reload the experiment
- count the FPS
- ImGui integration
- opengl 4.3
- glfw3 (libglfw3-dev for ubuntu and debian systems)
- make
Out of the box: make run
At the time of 40db3b the serial and gpu implementation share a substantial part of code, I want to remove this duplication by using this mechanism:
- the lbm file implements higher level API functions, and it calls internal functions that are implementation dependent
- each implementation implements these functions and also some higher level ones
- at link time you would have to link both the lbm object and the desired implementation object
Update: I'm at peace with having duplication in lbm.c and gpu.c, will get back to this problem once I get a good idea