Module simd 

SIMD kernels behind the EdgeSimd trait.

The optimiser core (glm::levenberg) calls f64::dot_simd(..) and f64::exp_in_place_simd(..) directly, since likelihoods and optimisers work in f64 regardless of the crate’s T: EdgeFloat data layer. The vector width lives here and nowhere else; no intrinsics or wide types appear outside this file.

Each operation exists at four widths: scalar, 128-bit, 256-bit and 512-bit. detect_simd_level caches what the machine supports and the trait impls dispatch on it. A tier vectorises what it can at its own width, then hands the remainder down to the tier below. That ladder matters here because the hot caller is linear_predictor, which dots design rows of two to ten elements: a single-width kernel would drop those straight to scalar.

The reductions run one accumulator, not several. Multiple accumulators break the fadd dependency chain and pay from a few hundred elements up, but the block loop that drives them needs lanes * accumulators elements before it can fire at all, which at two to ten coefficients it never does. Dropping the four-accumulator version moved glm_fit by under three per cent either way at two and four coefficients and made it five per cent faster at eight, so it was buying code and nothing else.

Enums

SimdLevel

Widest vector width available on the running machine.

Traits

EdgeSimd

SIMD-accelerated primitives over slices of a float type.

Functions

detect_simd_level

Detects the widest vector width this machine supports.