# simd

Float geometry plus integer-widening lane operations. Four modules:

- `simd/vec3` — `Vec3` (with a lane-3-zero invariant): `dot`, `cross`, `length`, `normalize`, `reflect`, `refract`, `lerp`, `clamp`, and more.
- `simd/vec4` — a full 4-lane vector with `raw()` / `from_raw()` for matrix code.
- `simd/mat4x4` — a column-major `[Vec4; 4]` with `mul_vec` (four `fma <4 x float>` ops) and `mul`.
- `simd/integer` — integer-widening lane helpers built from the compiler's Tier-1 SIMD primitives (`widen` / `low` / `high` / `swizzle`): widening multiply, multiply-accumulate, pairwise add, and `dot_i32`, a 16-lane signed-byte dot product accumulated in `i32`. This is the lane surface quantized kernels build on.

```cplus
import "simd/vec3" as vec3;

let a = vec3::Vec3::new(1.0f32, 2.0f32, 3.0f32);
let b = vec3::Vec3::new(4.0f32, 5.0f32, 6.0f32);
let d: f32 = a.dot(b);     // 32.0
let c = a.cross(b);        // (-3, 6, -3)
```

Use these when the SIMD shape should be visible at the source level. On Apple Silicon the scalar FMA codegen can beat explicit `Vec3` SIMD when the 4th lane is wasted, so measure before betting on SIMD types.