search

petamoriken / float16

Stage 3 IEEE 754 half-precision floating-point ponyfill
https://github.com/tc39/proposal-float16array
MIT License
96 stars 7 forks source link
binary16 float16 float16array fp16 half-precision ieee754 javascript typescript

readme

float16

IEEE 754 half-precision floating-point ponyfill for JavaScript
See TC39 proposal or the archive of the ES Discuss Float16Array topic for details

npm downloads npm version jsr version
dependencies license codecov

Sauce Labs browser matrix

Install

Node.js

npm install @petamoriken/float16

Deno

[!NOTE] float16 features are supported since Deno v1.43.

deno add @petamoriken/float16

Bun

[!NOTE] float16 features are supported since Bun v1.1.23.

bun add @petamoriken/float16

Import

Node.js, Deno, Bun or Bundler

import {
  Float16Array, isFloat16Array, isTypedArray,
  getFloat16, setFloat16,
  f16round,
} from "@petamoriken/float16";

Browser

Deliver a browser/float16.mjs or browser/float16.js file in the npm package from your Web server with the JavaScript Content-Type HTTP header.

<!-- Module Scripts -->
<script type="module">
  import {
    Float16Array, isFloat16Array, isTypedArray,
    getFloat16, setFloat16,
    f16round,
  } from "DEST/TO/float16.mjs";
</script>
<!-- Classic Scripts -->
<script src="https://github.com/petamoriken/float16/raw/master/DEST/TO/float16.js"></script>
<script>
  const {
    Float16Array, isFloat16Array, isTypedArray,
    getFloat16, setFloat16,
    f16round,
  } = float16;
</script>
Or, you can use jsDelivr CDN. ```html ``` ```html ```

Support engines

This package only requires ES2015 features and does not use environment-dependent features (except for inspect/), so you can use it without any problems. It works fine with the current officially supported versions of Node.js.

Float16Array implemented by Proxy and Reflect, so IE11 is never supported even if you use polyfills.

Pre-transpiled JavaScript files (CommonJS, IIFE)

lib/ and browser/ directories in the npm package have JavaScript files already transpiled, and they have been tested automatically in the following environments:

API

Float16Array

Float16Array is similar to TypedArray such as Float32Array (MDN).

const array = new Float16Array([1.0, 1.1, 1.2, 1.3]);
for (const value of array) {
  // 1, 1.099609375, 1.2001953125, 1.2998046875
  console.log(value);
}

// Float16Array(4) [ 2, 2.19921875, 2.3984375, 2.599609375 ]
array.map((value) => value * 2);

isFloat16Array

[!WARNING] This API returns false for ECMAScript's native Float16Array

isFloat16Array is a utility function to check whether the value given as an argument is an instance of Float16Array or not.

const buffer = new ArrayBuffer(256);

// true
isFloat16Array(new Float16Array(buffer));

// false
isFloat16Array(new Float32Array(buffer));
isFloat16Array(new Uint16Array(buffer));
isFloat16Array(new DataView(buffer));

isTypedArray

isTypedArray is a utility function to check whether the value given as an argument is an instance of a type of TypedArray or not. Unlike util.types.isTypedArray in Node.js, this returns true for Float16Array.

const buffer = new ArrayBuffer(256);

// true
isTypedArray(new Float16Array(buffer));
isTypedArray(new Float32Array(buffer));
isTypedArray(new Uint16Array(buffer));

// false
isTypedArray(new DataView(buffer));

getFloat16, setFloat16

getFloat16 and setFloat16 are similar to DataView methods such as DataView#getFloat32 (MDN) and DataView#setFloat32 (MDN).

declare function getFloat16(view: DataView, byteOffset: number, littleEndian?: boolean): number;
declare function setFloat16(view: DataView, byteOffset: number, value: number, littleEndian?: boolean): void;
const buffer = new ArrayBuffer(256);
const view = new DataView(buffer);

view.setUint16(0, 0x1234);
getFloat16(view, 0); // 0.0007572174072265625

// You can append methods to DataView instance
view.getFloat16 = (...args) => getFloat16(view, ...args);
view.setFloat16 = (...args) => setFloat16(view, ...args);

view.getFloat16(0); // 0.0007572174072265625

view.setFloat16(0, Math.PI, true);
view.getFloat16(0, true); // 3.140625

f16round (alias: hfround)

f16round is similar to Math.fround (MDN). This function returns nearest half-precision float representation of a number.

declare function f16round(x: number): number;
Math.fround(1.337); // 1.3370000123977661
f16round(1.337); // 1.3369140625

Float16Array limitations (edge cases)

Float16Array has some limitations, because it is impossible to completely reproduce the behavior of TypedArray. Be careful when checking if it is a TypedArray or not by using ArrayBuffer.isView, and when using Web standards such as structuredClone and WebGL. ### Built-in functions Built-in `TypedArray` objects use "internal slots" for built-in methods. Some limitations exist because the `Proxy` object can't trap internal slots ([explanation](https://javascript.info/proxy#built-in-objects-internal-slots)). This package isn't polyfill, in other words, it doesn't change native global functions and static/prototype methods. E.g. `ArrayBuffer.isView` is the butlt-in method that checks if it has the `[[ViewedArrayBuffer]]` internal slot. It returns `false` for `Proxy` object such as `Float16Array` instance. ```js ArrayBuffer.isView(new Float32Array(10)); // true ArrayBuffer.isView(new Float16Array(10)); // false ``` ### The structured clone algorithm (Web Workers, IndexedDB, etc) The structured clone algorithm copies complex JavaScript objects. It is used internally when invoking `structuredClone()`, to transfer data between Web Workers via `postMessage()`, storing objects with IndexedDB, or copying objects for other APIs ([MDN](https://developer.mozilla.org/en-US/docs/Web/API/Web_Workers_API/Structured_clone_algorithm)). It can't clone `Proxy` object such as `Float16Array` instance, you need to convert it to `Uint16Array` or deal with `ArrayBuffer` directly. ```js const array = new Float16Array([1.0, 1.1, 1.2]); const cloned = structuredClone({ buffer: array.buffer }); ``` ### WebGL WebGL requires `Uint16Array` for buffer or texture data whose types are `gl.HALF_FLOAT` (WebGL 2) or `ext.HALF_FLOAT_OES` (WebGL 1 extension). Do not apply the `Float16Array` object directly to `gl.bufferData` or `gl.texImage2D` etc. ```js // WebGL 2 example const vertices = new Float16Array([ -0.5, -0.5, 0, 0.5, -0.5, 0, 0.5, 0.5, 0, ]); const buffer = gl.createBuffer(); gl.bindBuffer(gl.ARRAY_BUFFER, buffer); // wrap in Uint16Array gl.bufferData(gl.ARRAY_BUFFER, new Uint16Array(vertices.buffer), gl.STATIC_DRAW); gl.vertexAttribPointer(location, 3, gl.HALF_FLOAT, false, 0, 0); gl.bindBuffer(gl.ARRAY_BUFFER, null); gl.enableVertexAttribArray(location); ``` ### Others See JSDoc comments in `src/Float16Array.mjs` for details. If you don't write hacky code, you shouldn't have any problems.

Float16Array custom inspection

Provides custom inspection for Node.js and Deno, which makes the results of console.log more readable. ### Node.js ```js import { Float16Array } from "@petamoriken/float16"; import { customInspect } from "@petamoriken/float16/inspect"; Float16Array.prototype[Symbol.for("nodejs.util.inspect.custom")] = customInspect; ``` ### Deno ```ts import { Float16Array } from "https://deno.land/x/float16/mod.ts"; import { customInspect } from "https://deno.land/x/float16/inspect.ts"; // deno-lint-ignore no-explicit-any (Float16Array.prototype as any)[Symbol.for("Deno.customInspect")] = customInspect; ```

Development

Manual build and test ### Manual build This repository uses corepack for package manager manager. You may have to activate yarn in corepack. ```console corepack enable yarn ``` Download devDependencies. ```console yarn ``` Build `lib/`, `browser/` files. ```console yarn run build ``` Build `docs/` files (for browser test). ```console yarn run docs ``` ### Test This repository uses corepack for package manager manager. You may have to activate yarn in corepack. ```console corepack enable yarn ``` Download devDependencies. ```console yarn ``` #### Node.js test ```console NODE_ENV=test yarn build:lib yarn test ``` #### Browser test ```console NODE_ENV=test yarn build:browser yarn docs ``` Access `docs/test/index.html` with browsers. You can access current [test page](https://petamoriken.github.io/float16/test) ([power-assert version](https://petamoriken.github.io/float16/test/power)) in `master` branch.

License

MIT License

This software contains productions that are distributed under the Apache 2.0 License. Specifically, index.d.ts is modified from the original TypeScript lib files.