Module backends 

Backend implementations for different SIMD instruction sets

This module contains the actual SIMD implementations for each backend. All backends implement the same trait-based interface to ensure API consistency.

Safety

All unsafe code is isolated within backend implementations. The public API remains 100% safe.

Backends

• scalar: Portable baseline implementation (no SIMD)
• sse2: x86_64 baseline SIMD (128-bit)
• avx2: x86_64 advanced SIMD (256-bit with FMA)
• avx512: x86_64 maximum SIMD (512-bit)
• neon: ARM SIMD (128-bit)
• wasm: WebAssembly SIMD128

Modules

avx2

AVX2 backend implementation (x86_64 advanced SIMD)

avx512

AVX-512 backend implementation (x86_64 advanced SIMD)

gpu

GPU backend using wgpu (Vulkan/Metal/DX12/WebGPU)

q4k

Fused Q4_K Matrix-Vector Multiply (F-GPU-130)

q6k

Fused Q6_K Matrix-Vector Multiply

scalar

Scalar (non-SIMD) backend implementation

sse2

SSE2 backend implementation (x86_64 baseline SIMD)

Traits

VectorBackend

Backend trait defining common operations