Module portable_simd 

Portable SIMD Scan Kernels (Task 6)

Provides a family of SIMD kernels that avoid gather pathologies and work across diverse hardware:

1. AVX-512: Gather or permute-based
2. AVX2: Byte LUT via shuffle + partial sums
3. NEON: Table lookup primitives
4. Scalar: Universal fallback

Design Principles

• Prefer layouts that allow structured loads (SoA)
• Use int16/int32 accumulation to reduce bandwidth
• Minimize unpredictable memory refs
• Maximize instruction-level parallelism (ILP)

Math/Algorithm

Inner loop is Θ(N_scanned). Performance is dominated by:

• Memory access patterns
• Instruction throughput

Kernel design minimizes cache misses and maximizes ILP.

Structs

Avx2Kernel

CpuFeatures

Detected CPU features

KernelDispatcher

Global kernel dispatcher that selects best implementation

ScalarKernel

Scalar fallback implementation (works everywhere)

ScanOps

High-level scan operations using best available SIMD

Enums

SimdLevel

SIMD level for kernel dispatch

Traits

DistanceKernel

Trait for portable distance kernels