Module simd 

Explicit-SIMD batch kernels for transform math — the AGDF compute-heavy hot loop lever.

A component column is stored as an array of whole structs (AoS-within-the- component), which is fine for random access but leaves SIMD lanes empty on a batched math sweep: a Vec3/Quaternion only fills part of a vector register. The lever is to feed the heavy loops field-split data — every field its own contiguous f32 stream (TrsBatchSoa) — and process eight entities at a time with [wide::f32x8]. f32x8 maps 1:1 to an 8-wide AoSoA tile, so the index split is trivial and the compiler emits one aligned vector op per field instead of eight scalar ones.

Why explicit SIMD rather than trusting auto-vectorisation: the heavy parts of this math are a quaternion-to-matrix expansion and a 1/√(x²+y²+z²+w²) normalisation. The normalisation’s horizontal sum is a floating-point reduction, and FP addition is not associative — the compiler is not allowed to reorder it into independent lanes, so it serialises and the loop stays scalar. Writing the lanes by hand recovers the throughput the auto-vectoriser leaves on the table.

The kernels are layout-agnostic in the LLAMA sense: callers fill a TrsBatchSoa from whatever their storage is, and the kernel owns the physical f32x8 tiling underneath. Each kernel has a scalar twin (used for the ragged tail and as the equivalence oracle in tests) so the SIMD path is guaranteed to match the scalar result to the last representable bit of the same operations.

Structs

TrsBatchSoa

A field-split (struct-of-arrays) batch of translate/rotate/scale inputs.

Constants

LANES

SIMD lane width. Eight f32 lanes = one 256-bit register (AVX) and the AoSoA tile width the TrsBatchSoa kernels stride by.

Functions

compose_trs_to_mat4

Composes out[i] = T · R · S for every entity in batch, the same column-major affine matrix Mat4::from_translation(t) * Mat4::from_quat(q) * Mat4::from_scale(s) produces — eight entities per [f32x8] tile, scalar for the ragged tail.

compose_trs_to_mat4_scalar

Scalar twin of compose_trs_to_mat4 — identical result, one entity at a time, no SIMD. This is the fallback a caller should prefer for small batches (where the field-split transpose costs more than the vector op saves) and the equivalence oracle the SIMD path is tested against.

normalize_quat_batch

Normalises a batch of quaternions in place to unit length, eight at a time.