Column simulation

Source: src/gpu/pipelines/compute-step.ts

The simulation is a compute pass: one GPU thread per column, advancing the Column[] storage buffer. It’s the only pass that writes simulation state.

One thread per column

The compute shader runs with a workgroup size of 64; each invocation handles one column, guarded against overrun:

const columnCount = floor(resolution.x / cellSize);
if (gid.x >= columnCount) return;
const column = columns.$[gid.x];

Advancing the head

Each step moves the head down by the column’s speed (in cells):

$\text{headY}' = \text{headY} + \text{speed}$

stepRate (Hz) controls how often this runs, decoupled from the render frame rate — the renderer interpolates between steps via stepProgress.

Respawn

A column respawns when its head has fallen off the bottom and a per-step random roll passes:

const offscreen = nextHeadY * cellSize > resolution.y;
if (offscreen && randf.sample() > density) {
  nextHeadY = 0;
  nextSeed = randf.sample() * U32_MAX;
}

Two things worth noting:

• The density semantic is inverted from intuition. density is the probability a column does not respawn this step. Higher density → respawn rarely → the screen empties out (sparser). At density = 1, randf.sample() > 1 is never true, so columns never respawn and the field drains. This matches the original 2D reference’s behavior.
• Respawn rerolls the seed. A new seed means a new set of glyphs and a new brightness-jitter pattern for that column’s next descent (see Glyph rendering).

Per-thread randomness

The RNG is seeded per thread, per frame from the column’s stored seed and the current time:

randf.seed2(vec2f(column.seed / U32_MAX, uniforms.$.time));

Seeding on time is deliberate — it’s why the respawn roll differs frame to frame. (It’s also why the paused settle loop advances time per iteration: a frozen time makes every roll identical and the field drains to black.) randf comes from @typegpu/noise; its functions are GPU-marked, so they compose into our 'use gpu' compute shader.

What it does not touch

speed, depth, and tailLength are set once at column creation (on the CPU) and never change during simulation — only headY and seed mutate here. The per-column constants are what give each column its place in the parallax depth field.