simd_math

A Rust library providing SIMD-accelerated mathematical functions for games, graphics, robotics, and other spatial computing applications.

// See tests/integration_test.rs test_readme_example() for full runnable code.
// See tests/integration_test.rs for more examples.

use simd_math::prelude::*;
use std::f32::consts::PI;

// Create a 3D vector
let vector = SimdVec3::from([1.0, 0.0, 0.0]);

// Create a rotation quaternion (90 degrees around Z-axis)
let rotation = SimdUnitQuat::from_axis_angle(SimdVec3::from([0.0, 0.0, 1.0]), PI / 2.0);

// Apply rotation to vector
let rotated_vector = rotation * vector;

// Components are accessed by index; the prelude provides X/Y/Z/W constants
assert!((rotated_vector[Y] - 1.0).abs() < 1e-6);

We ensure that our algebraic types are convertible to and from Rust arrays. Use arrays in your APIs for simple, math-library-agnostic interfaces. Cf Mujoco.

In the same spirit, we do not expose .x, .y, .z fields on vectors or quaternions. Use indexing: v[0], v[1], v[2], or equivalently v[X], v[Y], v[Z] with the named index constants from simd_math::prelude.

This decision further eschews conventions about what is front, up, and right in 3D space, which vary between applications. It also makes it easier to use vectors for non-spatial data, e.g. color.

Vectors are column vectors. Matrix multiplication is Matrix * Vector.

Matrices are constructed in column-major order, i.e. each argument to SimdMat3x4::new is a column. Matrix dimensions are named rows×columns, as in mathematics, HLSL, and Vulkan — so SimdMat3x4 has 3 rows and 4 columns: three transformed basis-vector columns plus a translation column, with an implicit [0, 0, 0, 1] bottom row, mapping homogeneous 4-vectors to 3-vectors. (Beware: GLSL and GLM name matrix types columns-first and would call this same shape mat4x3.)

Quaternions are represented as (s, x, y, z) where s is the scalar part. Think s + iv.

We assume a right-handed coordinate system.

Spherical coordinates are (azimuth, elevation, radius) with Y up: azimuth is measured in the XZ plane from +X towards +Z in [-π, π], and elevation is measured towards +Y in [-π/2, π/2].

3-component vectors are stored in 4-lane registers. The extra lane is an internal detail: it always holds zero, every operation preserves that invariant, and equality, Debug, indexing, and reductions all ignore it. Indexing a SimdVec3 with v[3] panics.

cargo bench runs a small suite comparing the SIMD types against scalar baselines. It uses the nightly libtest bench harness, so there are no extra dependencies.