Skip to main content

RenderKernel

Import RenderKernel alongside Kernel from @luma.gl/graph. Neither class is exported by engine or core.

RenderKernel owns compiled shaders and a cached render pipeline. It accepts already assembled GLSL on WebGL or WGSL on WebGPU. Shader assembly, uniforms, geometry ownership, and graph scheduling stay with the caller.

import {RenderKernel} from '@luma.gl/graph';

const kernel = new RenderKernel(device, {
  vertexSource,
  fragmentSource,
  shaderLayout,
  bufferLayout,
  colorAttachmentFormats: ['rgba32float']
});
const vertexArray = device.createVertexArray({
  shaderLayout: kernel.pipeline.shaderLayout,
  bufferLayout
});
kernel.draw(renderPass, {bindings, vertexArray, vertexCount: 3});

Backend behavior​

  • Direct and indexed draws work on WebGL and WebGPU. Supply a complete binding set and vertex array for each draw.
  • WebGL requires a fragment shader, including for transform-feedback pipelines; a trivial fragment shader can be used with a discard render pass.
  • WebGL supports transform feedback, compatibility uniforms, and per-draw topology/parameters through the existing render-pass draw options. Supplying instanceCount selects instanced drawing unless isInstanced is explicitly provided.
  • Indexed and non-indexed indirect draws remain WebGPU-only and reject WebGL before changing pass state.
  • WebGPU supports vertex-only pipelines. Set attachment formats to match the target render pass; unlike Model, RenderKernel does not infer them from a framebuffer.
  • createAsync() uses the existing shader/pipeline factory compilation lifecycle. draw() returns false if the backend is not ready to draw.

The caller owns vertex arrays, buffers, textures, passes, and command submission. destroy() releases the kernel's shader and pipeline references and is idempotent.

Graph use​

Flow-particle simulation uses RenderKernel directly on both backends. It retains its existing shaders, two state textures, uniform buffer, and ping-pong stepping logic, and supplies the current field/state bindings per draw. It no longer depends on GPUVectorModel.