Module native 

SIMD-native model format for zero-copy Trueno inference (spec §5) SIMD-Native Model Format (spec §5)

Provides types optimized for zero-copy SIMD inference with Trueno. Designed for maximum performance on CPU-based inference:

• 64-byte alignment: Compatible with AVX-512
• Contiguous storage: No pointer chasing
• Row-major ordering: Matches Trueno convention
• Cache-line optimization: Efficient prefetch

Performance Targets

• Linear (100 features, 1K samples): < 10 μs
• K-Means (10 clusters, 100d, 1K samples): < 50 μs
• Random Forest (100 trees, 1K samples): < 1 ms

Reference

[Intel Intrinsics Guide], [Fog 2023] “Optimizing Software in C++”

Structs

AlignedVec

64-byte aligned vector for SIMD operations (spec §5.2)

LayerData

Neural network layer data

ModelExtra

Additional model-specific data

TreeData

Decision tree structure data

TruenoNativeModel

Model format optimized for Trueno SIMD operations (spec §5.2)

Enums

LayerType

Neural network layer types

NativeModelError

Errors for native model operations