Module ndarray_fusion 

NDArray expression fusion pass.

Identifies chains of element-wise ndarray operations that produce single-use temporaries, and fuses them into a single loop. For example:

// Before fusion:
let t1 = a.add(&b);    // allocates + loops
let t2 = t1.mul(&c);   // allocates + loops
let result = t2.sub(&d); // allocates + loops

// After fusion (conceptually):
let result = fused_loop(a, b, c, d, |a, b, c, d| (a + b) * c - d);

This eliminates intermediate allocations and reduces memory traffic from 3 full passes over data to 1 pass. The energy savings are substantial: each eliminated pass saves ~5 pJ per cacheline of data.

The fusion pass works at the MIR level by:

1. Identifying Terminator::Call sequences targeting ndarray element-wise functions
2. Building a FusedExpr tree from the chain
3. Replacing the chain with a single fused call

The C codegen then emits the fused call as one loop with the composed expression.

Structs

FusionChain

A fusion opportunity: a chain of element-wise ops that can be fused.

Enums

FusedBinOp

Binary operations that can be fused

FusedExpr

A fused expression tree representing composed element-wise operations.

FusedUnaryOp

Unary operations that can be fused

Functions

find_fusion_chains

Analyze a function’s MIR to find fusion opportunities.

fused_expr_to_c

Generate the C expression string for a fused expression tree. The sources are accessed as src0->data[i], src1->data[i], etc.