Class WebGpuBackend

constructor

• new WebGpuBackend(
      fallback: (plan: PlanNode) => TableView | Promise<TableView>,
      device?: any,
      options?: { lossyF32?: boolean },
  ): WebGpuBackend

  Parameters

  • fallback: (plan: PlanNode) => TableView | Promise<TableView>
  • Optionaldevice: any
  • Optionaloptions: { lossyF32?: boolean }

  Returns WebGpuBackend

Readonlycapabilities

Readonlyname

Readonlyoptions

initialised

• get initialised(): boolean

  True once detection has run (successfully or not).

  Returns boolean

execute

• execute(plan: PlanNode, ctx?: ExecContext): Promise<TableView>

  ctx is optional for backward compatibility; backends that receive it should trace what they run.

  Parameters

  • plan: PlanNode
  • Optionalctx: ExecContext

  Returns Promise<TableView>

gpuFilterAnd

• gpuFilterAnd(
      predicates: GpuFilterPredicate[],
  ): Promise<Uint32Array<ArrayBufferLike>>

  Multi-predicate AND filter over typed predicates (see prepareGpuFilterPredicate); columns stay resident.

  Parameters

  • predicates: GpuFilterPredicate[]

  Returns Promise<Uint32Array<ArrayBufferLike>>

gpuFilterAndTimed

• gpuFilterAndTimed(
      predicates: GpuFilterPredicate[],
  ): Promise<{ computeMs: number; mask: Uint32Array; transferMs: number }>

  Same as gpuFilterAnd with host↔device transfer and submit→readback timings split out.

  Parameters

  • predicates: GpuFilterPredicate[]

  Returns Promise<{ computeMs: number; mask: Uint32Array; transferMs: number }>

gpuMap

• gpuMap(
      values: Float32Array,
      op: number,
      literal: number,
  ): Promise<Float32Array<ArrayBufferLike>>

  Parameters

  • values: Float32Array
  • op: number
  • literal: number

  Returns Promise<Float32Array<ArrayBufferLike>>

gpuMapTimed

• gpuMapTimed(
      values: Float32Array,
      op: number,
      literal: number,
  ): Promise<{ computeMs: number; transferMs: number; values: Float32Array }>

  Parameters

  • values: Float32Array
  • op: number
  • literal: number

  Returns Promise<{ computeMs: number; transferMs: number; values: Float32Array }>

gpuReduce

• gpuReduce(values: Float32Array, mode?: "sum" | "min" | "max"): Promise<number>

  Parameters

  • values: Float32Array
  • Optionalmode: "sum" | "min" | "max"

  Returns Promise<number>

gpuReduceSum

• gpuReduceSum(values: Float32Array): Promise<number>

  Parameters

  • values: Float32Array

  Returns Promise<number>

  Deprecated

  use gpuReduce(..., 'sum')

supports

• supports(_plan: PlanNode): boolean

  Parameters

  • _plan: PlanNode

  Returns boolean

waitReady

• waitReady(): Promise<boolean>

  Request the adapter / device (once). Nothing talks to navigator.gpu before this is called — importing the library or constructing the backend has no side effects. init() in columna calls it.

  Returns Promise<boolean>