SIMD dispatch

Documentation index

The optional simd feature enables four-label AVX2 dominance scans. simd-avx512 includes it and enables eight-label AVX-512F scans on Rust 1.89+. Both features are off by default. Runtime detection keeps binaries portable: AVX-512F (when compiled in), AVX2, then the portable implementation. Scans shorter than 16 remaining labels use the portable implementation.

With simd, rank-one local-search bitset accumulation also uses AVX2 on supported x86-64 CPUs for differences spanning at least 16 u64 words. The kernel widens integer AND/OR operations through compiler vectorization; slice bounds checks remain in place, and floating-point gain calculations retain their order. Short bitsets and other CPUs use the portable loop. simd-avx512 includes this AVX2 bitset path; only dominance uses AVX-512F.

The AVX-512F kernel uses ordered comparisons, retains the exact EPSILON addition/subtraction order, and returns the first matching label or prefix key stop in storage order. It widens eight packed u16 NG masks into eight u64 lanes without requiring AVX-512BW, VL, DQ or VPOPCNTDQ. All loaded columns are checked before pointer loads; tails use the scalar kernel. The comparison tests exercise direct AVX2/AVX-512 paths and runtime dispatch against an independent scalar reference, including NaNs, infinities, subnormals, signed zero, epsilon boundaries, offsets and incomplete blocks.

Build configuration

See the build guide for Cargo features and saved Make settings. SIMD is optional and disabled by default; kernel timings alone do not establish an improvement in complete solver performance.