reduceWGSL¶

Overview¶

A template that performs a reduce operation on a single workgroup. The value will be valid at local_id.x === 0.

@author Jonathan Olson <jonathan.olson@colorado.edu>

Type reduceWGSLOptions¶

import type { reduceWGSLOptions } from 'scenerystack/alpenglow';

• value: WGSLVariableName
  the "input" and "output" variable name
• scratch: WGSLVariableName
  name for var<workgroup> array<T, workgroupSize> TODO: consider abstracting, so we could run multiple reduces TODO: concurrently
• binaryOp: BinaryOp<T>
• mapScratchIndex?: ( index: WGSLExpressionU32 ) => WGSLExpressionU32
  allows overriding the index used for the scratch array, so that we can run multiple smaller loads in the same workgroup
• convergent?: boolean
  Whether we should reduce in a convergent order. This will reduce control divergence when running, and will potentially allow warps to exit early. This should result in a speed-up, but the data either needs to have a commutative combine operation, OR the order should be in a "convergent" order. That would mean that for each data chunk read by each workgroup, the indices are bit-reversed (e.g. if we have a workgroup size of 256, then we are reversing the last 8 bits of the index, thus the first element should be stored at index 0, the second element at index 128, the third element at index 64, etc.). See get_convergent_index for more information. For instance, the order of reduction of the first 16 hex digits (in a convergent order) would be 084c2a6e195d3b7f.
• scratchPreloaded?: boolean
  If true, we won't need to load the value INTO the scratch array
• valuePreloaded?: boolean
  If true, we won't need to load the value FROM the scratch array
• & WorkgroupSizable & LocalIndexable

Source Code¶

See the source for reduceWGSL.ts in the alpenglow repository.