Real‑time 2D fluid simulation on a grid, written in C++. The solver is built on Jos Stam's 2003 paper -- stable, well‑understood, and a solid starting point. The main challenge was making it scale. With some targeted optimizations, it runs 256×256 grids at a consistent 60 FPS on a single machine.
Physics
The engine uses an Eulerian grid with Stam's semi‑Lagrangian advection -- unconditionally stable, so no timestep restrictions. Projection is handled by Gauss‑Seidel relaxation: simple, fast, and converges well. I also added vorticity confinement to counteract the artificial damping that naturally occurs, keeping the fluid lively and swirling instead of turning into sludge.
Performance (the real engineering)
To push a 256×256 grid at 60 Hz, every unnecessary cost had to go:
- OpenMP – Grid operations are embarrassingly parallel, so I split the 2D array work across all CPU cores.
- GPU batching – Instead of rendering pixel by pixel on the CPU, I batch the grid into an
sf::VertexArrayand send it directly to the GPU. - Precomputed LUT – The Viridis colormap gives a clean, perceptual look, but computing RGB per cell per frame is wasteful. So I pre‑baked the map into a 1D texture. The CPU only tracks density; the GPU does the color mapping at zero extra cost.
The CMake configuration already sets -O3, -ffast-math, and enables OpenMP -- so a standard build will compile with performance in mind.
mkdir build && cd build
cmake ..
make -j$(nproc)
./fluid- Click and drag – injects dye and velocity. The injection rate scales with mouse speed, which feels natural.
- Real‑Time Fluid Dynamics for Games, Jos Stam (GDC 2003)