Skip to content
WebGPU Learning Website

GPUComputePipeline

Create a GPUComputePipeline with device.createComputePipeline(descriptor). The async form is createComputePipelineAsync(). It has one programmable stage.

interface GPUComputePipelineDescriptor {
  label?: string;
  layout: GPUPipelineLayout | "auto";
  compute: {
    module: GPUShaderModule;
    entryPoint?: string;
    constants?: Record<string, number>;
  };
}

The entry point must be marked @compute and @workgroup_size. getBindGroupLayout(index) works the same way as on a render pipeline.

A compute stage has no vertex buffers and no color attachments. Inputs and outputs are bindings:

  • a "storage" buffer, or WGSL var<storage, read_write>
  • a "read-only-storage" buffer
  • a "uniform" buffer
  • a storage texture whose access is "write-only", "read-only", or "read-write"

"write-only" and "read-only" access for "rgba8unorm" are core. "read-write" requires the texture-formats-tier2 feature.

Dispatch uses GPUComputePassEncoder:

interface GPUComputePassEncoder {
  setPipeline(pipeline: GPUComputePipeline): undefined;
  setBindGroup(index: GPUIndex32, bindGroup: GPUBindGroup | null, dynamicOffsets?: number[]): undefined;
  dispatchWorkgroups(workgroupCountX: GPUSize32, workgroupCountY?: GPUSize32, workgroupCountZ?: GPUSize32): undefined;
  dispatchWorkgroupsIndirect(indirectBuffer: GPUBuffer, indirectOffset: GPUSize64): undefined;
  end(): undefined;
}

The arguments to dispatchWorkgroups() are workgroup counts. The thread count is that count multiplied by @workgroup_size. workgroupCountY and workgroupCountZ default to 1.

Section titled “Related pages”