mainScanWGSL¶

Overview¶

A raked scan implementation capable of non-commutable cases, where:

1. All threads load state into workgroup memory in a coalesced fashion
2. All threads perform an inclusive sequential scan on their data (of grainSize elements)
3. All threads perform an inclusive scan of the "reuced" values for each thread (Hillis-Steele)
4. The remaining values are filled in with the previous scanned value.workgroup
5. The workgroup memory is written to the main output in a coalesced fashion

Based on the described coarsened scan in "Programming Massively Parallel Processors" by Hwu, Kirk and Hajj, chap11.

Additionally, reductions can be stored in a separate array (for use in multi-level scans), AND/OR scanned reductions can be added into final results (also for use in multi-level scans).

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

Type mainScanWGSLOptions¶

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

• binaryOp: BinaryOp<T>
• inPlace?: boolean
• storeReduction?: boolean
• addScannedReduction?: boolean
• addScannedDoubleReduction?: boolean
• areScannedReductionsExclusive?: boolean
• input?: BufferSlot<T[]> | null
  Iff inPlace:false
• output?: BufferSlot<T[]> | null
• data?: BufferSlot<T[]> | null
  Iff inPlace:true
• reduction?: BufferSlot<T[]> | null
  iff storeReduction:true
• scannedReduction?: BufferSlot<T[]> | null
  iff addScannedReduction:true
• scannedDoubleReduction?: BufferSlot<T[]> | null
  iff addScannedDoubleReduction:true
• getAddedValue?: scanComprehensiveWGSLOptions<T>[ 'getAddedValue' ]
  only if addScannedReduction:false
• & RakedSizable & Pick<scanComprehensiveWGSLOptions<T>, "exclusive" | "lengthExpression" | "inputOrder" | "inputAccessOrder" | "factorOutSubexpressions" | "stripeReducedOutput">

Source Code¶

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