Module kernel_fusion 

Kernel Fusion Engine

Automatically detects and fuses element-wise / pointwise kernel sequences to eliminate intermediate memory allocations and round-trips. This mirrors the kernel fusion passes in TensorRT, XLA, and TVM.

Fusion rules:

1. Element-wise ops (add, mul, relu, etc.) can always fuse.
2. Reduction followed by broadcast can fuse (vertical fusion).
3. Producer-consumer pairs with matching shapes can fuse (horizontal).

Structs

FusedKernel

A fused kernel — a sequence of operations executed as one kernel.

FusionAnalyzer

Fusion analysis engine that detects fusable patterns.

FusionNode

A node in the fusion graph.

FusionResult

Result of fusion analysis.

Enums

BinaryOp

Binary element-wise operations.

FusableOp

An operation that can be part of a fused kernel.

MemOp

Memory operations.

PrecisionType

Precision types for cast operations.

ReduceOp

Reduction operations.

UnaryOp

Unary element-wise operations.