Skip to main content

ComputePipeline

info

ComputePipeline is only available on WebGPU.

A ComputePipeline holds a compiled compute shader plus the ComputeShaderLayout that describes its bindings.

Bindings and bind groups​

ComputePipeline.setBindings() accepts either:

  • flat bindings
  • grouped bindGroups

Grouped bindings are keyed by bind-group index, and sparse groups are valid.

Usage​

const source = /* wgsl */ `
@group(0) @binding(0) var<storage, read_write> data: array<i32>;

struct SceneUniforms {
  addend: i32
};

@group(2) @binding(0) var<uniform> sceneUniforms: SceneUniforms;

@compute @workgroup_size(1)
fn main(@builtin(global_invocation_id) id: vec3<u32>) {
  let i = id.x;
  data[i] = data[i] + sceneUniforms.addend;
}
`;

const shader = webgpuDevice.createShader({source});
const computePipeline = webgpuDevice.createComputePipeline({
  shader,
  shaderLayout: {
    bindings: [
      {name: 'data', type: 'storage', group: 0, location: 0},
      {name: 'sceneUniforms', type: 'uniform', group: 2, location: 0}
    ]
  }
});

computePipeline.setBindings({
  0: {data: workBuffer},
  2: {sceneUniforms}
});

Flat bindings work as well:

computePipeline.setBindings({
  data: workBuffer,
  sceneUniforms
});

When flat bindings are used, luma.gl partitions them into groups using the shader layout.

ComputePipelineProps​

Important properties:

  • shader: Shader
  • entryPoint?: string
  • constants?: Record<string, number>
  • shaderLayout?: ComputeShaderLayout | null

Unlike RenderPipeline, grouped bindings are not configured through ComputePipelineProps. They are supplied through setBindings().

Methods​

setBindings(bindingsOrBindGroups)​

computePipeline.setBindings(bindingsOrBindGroups);

Accepts either:

  • Bindings
  • BindingsByGroup

destroy()​

Releases the pipeline resources immediately.

Performance​

Creating compute pipelines can be expensive. Applications that create compatible pipelines repeatedly should prefer PipelineFactory over device.createComputePipeline() directly.