NPM Package

Fast 3d math library for webgpu

Why another 3d math library?


import {
} from 'https://wgpu-matrix.org/dist/3.x/wgpu-matrix.module.js';

const fov = 60 * Math.PI / 180
const aspect = width / height;
const near = 0.1;
const far = 1000;
const perspective = mat4.perspective(fov, aspect, near, far);

const eye = [3, 5, 10];
const target = [0, 4, 0];
const up = [0, 1, 0];
const view = mat4.lookAt(eye, target, up);

Note: for translation, rotation, and scaling there are 2 versions of each function. One generates a translation, rotation, or scaling matrix. The other translates, rotates, or scales a matrix.

const t = mat4.translation([1, 2, 3]);    // a translation matrix
const r = mat4.rotationX(Math.PI * 0.5);  // a rotation matrix
const s = mat4.scaling([1, 2, 3]);        // a scaling matrix
const m = mat4.identity();
const t = mat4.translate(m, [1, 2, 3]);    // m * translation([1, 2, 3])
const r = mat4.rotateX(m, Math.PI * 0.5);  // m * rotationX(Math.PI * 0.5)
const s = mat4.scale(m, [1, 2, 3]);        // m * scaling([1, 2, 3])

Functions take an optional destination to hold the result.

const m = mat4.create();            // m = new mat4
mat4.identity(m);                   // m = identity
mat4.translate(m, [1, 2, 3], m);    // m *= translation([1, 2, 3])
mat4.rotateX(m, Math.PI * 0.5, m);  // m *= rotationX(Math.PI * 0.5)
mat4.scale(m, [1, 2, 3], m);        // m *= scaling([1, 2, 3])

There is also the minified version

import {
} from 'https://wgpu-matrix.org/dist/3.x/wgpu-matrix.module.min.js';

// ... etc ...

and a UMD version

<script src="https://wgpu-matrix.org/dist/3.x/wgpu-matrix.js"></script>
const { mat4, vec3 } = wgpuMatrix;
const m = mat4.identity();

or UDM min version

<script src="https://wgpu-matrix.org/dist/3.x/wgpu-matrix.min.js"></script>

or via npm

npm install --save wgpu-matrix

then using a build process

import {vec3, mat3} from 'wgpu-matrix';

// ... etc ...




mat4.perspective, mat4.ortho, and mat4.frustum all return matrices with Z clip space from 0 to 1 (unlike most WebGL matrix libraries which return -1 to 1)

mat4.create makes an all zero matrix if passed no parameters. If you want an identity matrix call mat4.identity


mat3 uses the space of 12 elements

// a mat3
  xx, xy, xz, ?
  yx, yy, yz, ?
  zx, zy, zz, ?

This is because WebGPU requires mat3s to be in this format and since this library is for WebGPU it makes sense to match so you can manipulate mat3s in TypeArrays directly.

vec3 in this library uses 3 floats per but be aware that an array of vec3 in a Uniform Block or other structure in WGSL, each vec3 is padded to 4 floats! In other words, if you declare

struct Foo {
  bar: vec3<f32>[3];

then bar[0] is at byte offset 0, bar[1] at byte offset 16, bar[2] at byte offset 32.

See the WGSL spec on alignment and size.

Columns vs Rows

WebGPU follows the same conventions as OpenGL, Vulkan, Metal for matrices. Some people call this “column major”. The issue is the columns of a traditional “math” matrix are stored as rows when declaring a matrix in code.

  x1, x2, x3, x4,  // <- column 0
  y1, y2, y3, y4,  // <- column 1
  z1, z2, z3, z4,  // <- column 2
  w1, w2, w3, w4,  // <- column 3

To put it another way, the translation vector is in elements 12, 13, 14

  xx, xy, xz, 0,  // <- x-axis
  yx, yy, yz, 0,  // <- y-axis
  zx, zy, zz, 0,  // <- z-axis
  tx, ty, tz, 1,  // <- translation

This issue has confused programmers since at least the early 90s 😌

Performance vs Convenience

Most functions take an optional destination as the last argument. If you don’t supply it, a new one (vector, matrix) will be created for you.

// convenient usage

const persp = mat4.perspective(fov, aspect, near, far);
const camera = mat4.lookAt(eye, target, up);
const view = mat4.inverse(camera);
// performant usage

// at init time
const persp = mat4.create();
const camera = mat4.create();
const view = mat4.create();

// at usage time
mat4.perspective(fov, aspect, near, far, persp);
mat4.lookAt(eye, target, up, camera);
mat4.inverse(camera, view);

For me, most of the stuff I do in WebGPU, the supposed performance I might lose from using the convenient style is so small as to be unmeasurable. I’d prefer to stay convenient and then, if and only if I find a performance issue, then I might bother to switch to the performant style.

As the saying goes premature optimization is the root of all evil. 😉


2.x -> 3.x

In JavaScript there should be no difference in the API except for the removable of setDefaultType.

In TypeScript, 3.x should mostly be type compatible with 2.x. 3.x is an attempt to fix the casting that was necessary in 2.x.

// 2.x
device.queue.writeData(buffer, 0, mat4.identity() as Float32Array);  // sadness! 😭

// 3.x
device.queue.writeData(buffer, 0, mat4.identity());  // Yay! 🎉

In TypeScript the differences are as follows

Functions have a default type but return what is passed to them as the dst.

In 3.x each function has a default type but if you pass it a destination it returns the type of the destination

mat4.identity()                       // returns Float32Array
mat4.identity(new Float32Array(16));  // returns Float32Array
mat4.identity(new Float64Array(16));  // returns Float32Array
mat4.identity(new Array(16));         // returns number[]

Types are specific

const a: Mat4 = ...;    // a = Float32Array
const b: Mat4d = ...;   // b = Float64Array
const c: Mat4n = ...;   // c = number[]

This is means code like this

const position: Mat4 = [10, 20, 30];

No longer works because Mat4 is a Float32Array.

BUT, functions take any of the normal types as an argument just like they used to

const position = [10, 20, 30];          // number[]
const target = vec3.create(1, 2, 3);    // Float32Array
const up = new Float64Array([0, 1, 0]); // Float64Array

// Works fine, even those types are different, just like 2.x did
const view = mat4.lookAt(position, target, up);  // Float32Array

If you really want types for each concrete type there’s

There are 3 sets of functions, each one returning a different default

mat4.identity()   // returns Float32Array
mat4d.identity()  // returns Float64Array
mat4n.identity()  // returns number[]

Similarly there’s mat3d, mat3n, quatd, quatn, vec2d, vec2n, vec3d, vec3n, vec4d, vec4n.

Note: that in general you’re unlikely to need any of these. Just use the same ones you were using in 2.x

1.x -> 2.x


git clone https://github.com/greggman/wgpu-matrix.git
cd wgpu-matrix
npm i
npm run build
npm test

You can run tests in the browser by starting a local server

npx servez

Now go to wherever your server serves pages. In the case of servez that’s probably http://localhost:8080/test/.

By default the tests test the minified version. To test the source use src=true as in http://localhost:8080/test/?src=true.

To limit which tests are run use grep=<regex>. For example [http://localhost:8080/test/?src=true&grep=mat3.?translate](http://localhost:8080/test/src=true&grep=mat3.?translate) runs only tests with mat3 followed by translate in the name of test.