Module reduction 

Global Reduction Primitives

This module provides traits and types for GPU-accelerated reduction operations. Reductions aggregate values across all GPU threads using operations like sum, min, max, etc.

Use Cases

• PageRank: Sum dangling node contributions across all nodes
• Graph algorithms: Compute convergence metrics, global norms
• Scientific computing: Vector norms, dot products, energy calculations

Architecture

Reductions use a hierarchical approach for efficiency:

1. Warp-level: Use shuffle instructions for fast intra-warp reduction
2. Block-level: Tree reduction in shared memory with __syncthreads()
3. Grid-level: Atomic accumulation from block leaders, then broadcast

Example

use ringkernel_core::reduction::{ReductionOp, GlobalReduction};

// In kernel code (DSL):
let my_contrib = if out_degree[idx] == 0 { rank } else { 0.0 };
let dangling_sum = reduce_and_broadcast(my_contrib, &accumulator);
let new_rank = base + damping * (incoming + dangling_sum / n);

Structs

ReductionConfig

Configuration for reduction operations.

Enums

ReductionOp

Reduction operation types.

Traits

GlobalReduction

Trait for GPU runtimes that support global reduction operations.

ReductionHandle

Handle to a reduction buffer for streaming operations.

ReductionScalar

Trait for scalar types that support reduction operations.