Module loop_fusion 

Loop fusion optimization pass.

This module provides optimization passes that fuse multiple loops/reductions over the same axes to improve cache locality and reduce memory traffic.

Overview

Loop fusion combines multiple consecutive operations that iterate over the same axis into a single fused operation. This optimization:

• Reduces memory traffic (fewer intermediate tensors)
• Improves cache locality (better temporal locality)
• Reduces loop overhead (fewer loop iterations)

Fusion Criteria

Two loops can be fused if:

1. They iterate over the same axis/axes
2. They have compatible domains
3. There are no dependencies that prevent fusion
4. The fused operation doesn’t exceed memory constraints

Examples

use tensorlogic_compiler::passes::fuse_loops;
use tensorlogic_ir::EinsumGraph;

let graph = EinsumGraph::new();
let (fused_graph, stats) = fuse_loops(&graph);

Structs

LoopFusionConfig

Configuration for loop fusion optimization.

LoopFusionStats

Statistics from loop fusion optimization.

Functions

estimate_fusion_benefit

Estimate the benefit of fusing a group of nodes.

fuse_loops

Fuse loops in an einsum graph.

fuse_loops_with_config

Fuse loops with custom configuration.