Module kernel 

Kernel dispatch layer — safe public API + runtime CPU feature detection.

For application code (safe, no unsafe)

Use only:

• axpy — y[i] ^= c * x[i] (equal lengths; non-overlapping buffers)
• scale — y[i] = c * x[i] (equal lengths; non-overlapping; use scale_inplace for in-place)
• scale_inplace — y[i] = c * y[i]
• axpy_multi — adaptive blocked/fused multi-source AXPY for encoding
• dot — GF(2⁸) dot product (GFNI accelerated when available)
• active_kernel_name — which SIMD tier is active

Length and overlap are asserted in release on the safe wrappers. Tier modules (x86, arm, wasm) are pub(crate) — not part of the external API.

Dispatch

All SIMD variants are compiled in. On the first call to axpy / scale / scale_inplace (with std), the best tier is selected via is_x86_feature_detected! and cached in OnceLock<KernelSet>.

Dispatch priority (x86_64)

Tier	Runtime check	Width	Instruction
1	gfni + avx512f + avx512bw	512-bit	GF2P8MULB zmm
2	gfni + avx2	256-bit	GF2P8MULB ymm
3	gfni + sse4.2	128-bit	GF2P8MULB xmm
4	avx512f + avx512bw + ssse3	512-bit	vpshufb zmm
5	avx2 + ssse3	256-bit	vpshufb ymm
6	ssse3	128-bit	pshufb xmm
7	neon (AArch64)	128-bit	vqtbl1q
8	wasm simd128 (compile-time)	128-bit	i8x16_swizzle
9	scalar (universal fallback)	1 byte	table lookup

no_std targets

OnceLock requires std. Without std, dispatch uses compile-time #[cfg(target_feature)] selection (bare-metal / embedded).

Functions

active_kernel_name

Returns the name of the currently active kernel tier.

axpy

y[i] ^= c * x[i] in GF(2⁸) — the fundamental RLNC primitive.

axpy_multi

Multi-source fused AXPY with improved cache behaviour: for each chunk of the destination, apply y ^= c_i * source_i for all i.

dot

sum(a[i] * b[i]) in GF(2⁸).

scale

y[i] = c * x[i] in GF(2⁸).

scale_inplace

In-place scale: y[i] = c * y[i] in GF(2⁸).