GPU FLIP water in a box: the grid sits on the entity (floor at its height, centred on it horizontally), the water starts as the Fill box (fractions of the domain), and box and sphere colliders inside the domain are walls
A boids swarm on the GPU: the sample GPU ECS world. Fields Position and Velocity are GPU-owned; a counting sort into a uniform grid (clear, count, scan, scatter) feeds a 27-cell neighbour search, then the agents move and
FLIP water on the GPU: particles carry it, a MAC grid solves pressure with red-black Gauss-Seidel sweeps, the walls of the domain and the scene's colliders (a signed distance field per cell) bound it.
The water's surface from the per-cell particle counts: a smoothed density field, contoured with surface nets (the dual of marching cubes: one vertex per cell the iso-surface crosses, at the mean of its edge crossings, an
A swarm simulated entirely on the GPU (boids: separation, alignment, cohesion) and drawn from the same buffers through the GPU-culled instanced path. The agents live in a cube of side 2 x Half Size around the entity and
Who writes a buffer, and so when the CPU may touch it. Gpu — component state the kernels own (positions, velocities). The CPU writes it once, before the first step, and afterwards sees it only through SyncToCpu (block
A system: an HLSL compute kernel and the buffers it binds, in binding order (binding i = bindings[i], set 0). Push constants are the world's parameter block, the same for every kernel of a step.
Component arrays in GPU buffers (structure of arrays: one buffer per field) and the kernels that step them, in registration order with a compute barrier between each. A step is recorded into the frame being built (Record