@plasius/gpu-worker

@plasius/gpu-worker

A WebGPU worker runtime that builds on @plasius/gpu-lock-free-queue to schedule WGSL workloads like ray tracing, physics, and acoustics.

Apache-2.0. ESM + CJS builds. WGSL assets are published in dist/.

Install

npm install @plasius/gpu-worker

Usage

import {
  assembleWorkerWgsl,
  createWorkerLoop,
  loadJobWgsl,
  loadWorkerWgsl,
} from "@plasius/gpu-worker";

const workerWgsl = await loadWorkerWgsl();
const jobType = await loadJobWgsl({
  wgsl: `
fn process_job(job_index: u32, job_type: u32, payload_words: u32) {
  // job logic here
}
`,
  label: "physics",
});

const shaderCode = await assembleWorkerWgsl(workerWgsl, { debug: true });
// Pass shaderCode to device.createShaderModule({ code: shaderCode })

loadJobWgsl registers job WGSL and returns the assigned job_type index. Call assembleWorkerWgsl again after registering new jobs to rebuild the combined WGSL. Job types are assigned in registration order, so keep the registration order stable across rebuilds if you need deterministic ids.

assembleWorkerWgsl also accepts an optional second argument to override the queue WGSL source: assembleWorkerWgsl(workerWgsl, { queueWgsl, queueUrl, fetcher }). By default it applies queue compatibility renames (for example JobMeta -> JobDesc); set queueCompat: false to disable that behavior. If you are concatenating WGSL manually, loadQueueWgsl provides the same compatibility renames by default: loadQueueWgsl({ url, fetcher, queueCompat: false }). Set queueMode: "dag" on loadQueueWgsl(...) or assembleWorkerWgsl(...) to assemble against the multi-root DAG-ready queue helpers from @plasius/gpu-lock-free-queue.

const shaderCode = await assembleWorkerWgsl(workerWgsl, {
  preludeWgsl,
  jobs,
  queueMode: "dag",
});

Worker WGSL now calls complete_job(job_index) after process_job(...). Queue assets from @plasius/gpu-lock-free-queue already provide that hook. If you pass a custom queue source without it, the assembler appends a no-op shim so existing flat queue integrations keep working.

To bypass the registry, pass jobs directly:

const shaderCode = await assembleWorkerWgsl(workerWgsl, {
  jobs: [{ wgsl: jobA }, { wgsl: jobB, label: "lighting" }],
  debug: true,
});

When assembling jobs, each job WGSL must define process_job(job_index, job_type, payload_words). The assembler rewrites each job's process_job to a unique name and generates a dispatcher based on job_type. Set debug: true to detect identifier clashes across appended WGSL.

To run the worker loop at the highest practical rate (or a target rate), use the helper:

const loop = createWorkerLoop({
  device,
  worker: { pipeline: workerPipeline, bindGroups: [queueBindGroup, simBindGroup] },
  jobs: [
    { pipeline: physicsPipeline, bindGroups: [queueBindGroup, simBindGroup], workgroups: physicsWorkgroups },
    { pipeline: renderIndirectPipeline, bindGroups: [queueBindGroup, simBindGroup], workgroups: 1 },
  ],
  workgroupSize: 64,
  maxJobsPerDispatch: queueCapacity,
  // rateHz: 120, // optional throttle; omit for animation-frame cadence
});
loop.start();

For opt-in local instrumentation, createWorkerLoop also accepts:

• frameId: a string or function returning the current frame correlation id.
• worker.owner, worker.queueClass, worker.jobType, worker.label, worker.workgroupSize: optional metadata for the dequeue pass.
• the same metadata fields on each job descriptor.
• telemetry.onDispatch(sample): called after submit for each worker/job dispatch.
• telemetry.onTick(summary): called after submit with the per-tick dispatch summary.

That allows direct integration with @plasius/gpu-debug without coupling the runtime to the package:

import { createGpuDebugSession } from "@plasius/gpu-debug";

const debug = createGpuDebugSession({ enabled: true });

const loop = createWorkerLoop({
  device,
  frameId: () => `frame-${frameNumber}`,
  worker: {
    pipeline: workerPipeline,
    bindGroups: [queueBindGroup],
    workgroups: [2, 1, 1],
    workgroupSize: 64,
    owner: "lighting",
    queueClass: "lighting",
    jobType: "worker.dequeue",
  },
  jobs: [
    {
      pipeline: directLightingPipeline,
      bindGroups: [queueBindGroup, lightingBindGroup],
      workgroupCount: [32, 18, 1],
      workgroupSize: [8, 8, 1],
      owner: "lighting",
      queueClass: "lighting",
      jobType: "lighting.direct",
    },
  ],
  telemetry: {
    onDispatch(sample) {
      debug.recordDispatch({
        owner: sample.owner,
        queueClass: sample.queueClass,
        jobType: sample.jobType,
        frameId: sample.frameId,
        workgroups: sample.workgroups,
        workgroupSize: sample.workgroupSize,
      });
    },
  },
});

@plasius/gpu-worker also now exposes a scene-preparation manifest helper for snapshot-driven chunk DAG planning:

import { createScenePreparationManifest } from "@plasius/gpu-worker";

const manifest = createScenePreparationManifest({
  snapshotId: "visual-snapshot-42",
  chunks: [
    {
      chunkId: "chunk-near-0",
      representationBand: "near",
      gameplayImportance: "critical",
      visible: true,
      playerRelevant: true,
    },
    {
      chunkId: "chunk-far-3",
      representationBand: "far",
      gameplayImportance: "medium",
      visible: false,
    },
  ],
});

console.log(manifest.graph.chunkRoots);
console.log(manifest.graph.priorityLanes[0]);

The helper publishes one chunk-local DAG per stable snapshot, keeps joins local to chunk stage boundaries, and rejects render-preparation manifests that try to mutate authoritative simulation state.

What this is

• A minimal GPU worker layer that combines a lock-free queue with user WGSL jobs.
• A helper to assemble WGSL modules with queue helpers included.
• A reference job format for fixed-size job dispatch (u32 indices).

DAG Queue Modes

@plasius/gpu-worker now supports two scheduler assembly modes:

• flat: the original lock-free FIFO/worklist execution flow.
• dag: a multi-root, priority-aware ready-queue flow where jobs can declare dependencies and join points through their package-owned manifests.

The worker runtime still stays lock-free and policy-light. It does not resolve budgets or decide priorities itself. Instead it exposes the queue mode and the completion hook needed by package manifests and @plasius/gpu-performance to coordinate DAG-shaped workloads.

Package manifests should be treated as explicit DAG node definitions, not just loose hints. In practice that means:

• multiple roots are allowed and expected,
• manifest labels act as dependency ids,
• priorities map to ready-queue lanes,
• downstream jobs are unlocked only when every upstream dependency completes.

Package Integration Model

@plasius/gpu-worker is the preferred execution plane for discrete GPU work across current and future @plasius/gpu-* compute packages.

Package authors should:

• register stable worker job types for each effect family,
• keep scheduling in terms of compact worklists and bounded dispatches,
• expose priority, dependencies, and schedulerMode metadata in manifests when jobs form ordered DAG stages instead of a flat queue,
• let @plasius/gpu-performance adjust worker budgets instead of building separate package-local governors,
• expose optional local instrumentation through createWorkerLoop(..., { telemetry }) and route that into @plasius/gpu-debug when clients enable it.

This pattern is intended to scale across post-processing, cloth, fluids, lighting refresh, voxel generation, and additional GPU job families without splitting scheduling policy across packages.

Demo

The demo enqueues physics and render jobs on the GPU, builds per-type worklists, runs the physics kernel, and uses an indirect draw for the particle pass. Install dependencies first so the lock-free queue package is available for the browser import map.

npm install
npm run demo

Then open http://localhost:8000/demo/.

HTTPS demo

WebGPU requires a secure context. For non-localhost access, run the HTTPS demo server.

mkdir -p demo/certs
mkcert -key-file demo/certs/localhost-key.pem -cert-file demo/certs/localhost.pem localhost 127.0.0.1 ::1
# or
openssl req -x509 -newkey rsa:2048 -nodes -keyout demo/certs/localhost-key.pem -out demo/certs/localhost.pem -days 365 -subj "/CN=localhost" -addext "subjectAltName=DNS:localhost,IP:127.0.0.1"
npm run demo:https

Then open https://localhost:8443/demo/. If you use a different hostname/IP, generate a certificate for that name and set DEMO_HOST, DEMO_PORT, DEMO_TLS_CERT, and DEMO_TLS_KEY as needed.

Build Outputs

npm run build emits dist/index.js, dist/index.cjs, and dist/worker.wgsl.

Development Checks

npm run lint
npm run typecheck
npm run test:coverage
npm run build
npm run pack:check

Files

• demo/index.html: Loads the WebGPU demo.
• demo/main.js: WebGPU setup, queue jobs, physics worklists, and indirect draw.
• demo/jobs/common.wgsl: Shared WGSL definitions for demo jobs.
• demo/jobs/physics.job.wgsl: Physics job kernel (worklist + integration).
• demo/jobs/render.job.wgsl: Render job kernel (worklist + indirect args).
• src/worker.wgsl: Minimal worker entry point template (dequeue + process_job hook).
• src/index.js: Helper functions to load/assemble WGSL.
• docs/adrs/*: architectural decisions for worker runtime and scheduling.
• docs/tdrs/*: technical design records for multi-package worker integration.
• docs/design/*: design notes for package integration, DAG queue modes, and future expansion.

Job shape

Jobs are variable-length payloads stored in a caller-managed buffer. Each job supplies job_type, payload_offset, and payload_words metadata plus a payload stored in the input payload buffer. For simple cases, use a single-word payload containing an index into your workload array.

Set output_stride in queue params to the maximum payload size you want copied out for a job; job_type can be used by schedulers to route work to different kernels. The queue mirrors input metadata into output_jobs and optionally copies payloads into output_payloads.