Module reduction 

Reduction Primitives

Provides GPU reduction operations and multi-phase kernel synchronization for iterative algorithms like PageRank, K-Means, and graph analytics.

Reduction Modes

• Single-pass: Simple reductions completed in one kernel launch
• Multi-phase: Complex reductions requiring intermediate storage
• Cooperative: GPU-wide synchronization using cooperative groups

Sync Modes

• Cooperative: Use CUDA cooperative groups for grid-wide sync
• SoftwareBarrier: Software-based barrier with atomic operations
• MultiLaunch: Separate kernel launches per phase

Example

use rustkernel_core::memory::reduction::{InterPhaseReduction, SyncMode};

let reduction = InterPhaseReduction::<f64>::new(1000, SyncMode::Cooperative);

// Phase 1: Local reduction
reduction.phase_start(0);
// ... kernel execution ...
reduction.phase_complete(0);

// Phase 2: Global reduction
reduction.phase_start(1);
// ... kernel execution ...
let result = reduction.finalize();

Structs

CooperativeBarrier

Helper to create a cooperative sync barrier

GlobalReduction

Global reduction tracker for K2K coordination

InterPhaseReduction

Inter-phase reduction state

ReductionBuilder

Builder for reduction operations

ReductionConfig

Configuration for inter-phase reduction

Enums

PhaseState

Phase state for multi-phase reductions

ReductionError

Reduction errors

ReductionOp

Reduction operation type

SyncMode

Synchronization mode for multi-phase reductions