Vectorization and SIMD

Scalar code touches one number at a time; hot loops touch four. Vec4f and Vec4i are 128-bit SIMD value types: four lanes in one register, no heap, no ARC traffic, plain operators.

Constructing vectors

Four lanes from four values, or one value broadcast to all four with splat. Vectors are values: they pass as function arguments and return from functions on every backend:

import { Vec4f } from "@std/simd";

fn dbl(v: Vec4f): Vec4f {
    return v * Vec4f.splat(2.0);
}

print(dbl(new Vec4f(1.0, 2.0, 3.0, 4.0)).get(2));

Vec4i mirrors the shape for 32-bit integer lanes (new Vec4i(1, 2, 3, 4), Vec4i.splat(3)), with wrapping integer arithmetic.

Lane arithmetic

+ - * / combine lane-wise, and / on integer lanes traps on zero divisors like scalar division. Read lanes back with x()/y()/z()/w() or the indexed get(i) (out-of-range lanes trap):

import { Vec4f } from "@std/simd";

let a = new Vec4f(1.0, 2.0, 3.0, 4.0);
let b = Vec4f.splat(10.0);
let c = a + b;
print(c.get(0), c.get(3));

Reductions collapse lanes to scalars: dot sums pairwise products, min/max take per-lane extrema, and sqrt roots each lane (negative lanes yield NaN per IEEE-754):

import { Vec4f } from "@std/simd";

let v = new Vec4f(4.0, 9.0, 16.0, 25.0);
print(v.sqrt().get(3), v.dot(Vec4f.splat(1.0)));

Float discipline still applies

Lane values are strict Floats. Mixing a FastFloat lane source without .asFast() is the same E305 error as scalar code, and Int lane readouts convert with Int(...). The process entry point still returns Int — vectors travel through helpers, and lanes cross the final boundary.

Summary

• Vec4f/Vec4i: construct, splat, lane operators, x/y/z/w/get.
• Reductions: dot, min, max, sqrt with IEEE-754 edge semantics.
• First-class function values on all backends; Int at the entry point.