GMP-WASM

Arbitrary-precision Integer, Rational and Float types based on the GMP and MPFR libraries.

Features

• Supports all modern browsers, web workers, Node.js and Deno
• Includes an easy-to-use, high-level wrapper, but low-level functions are also exposed
• Has a lot more features, and in some cases, it's faster than the built-in BigInt type
• The WASM binary is bundled as a compressed base64 string (no problems with linking)
• Works even without Webpack or other bundlers
• Includes TypeScript type definitions, check API here.
• Zero dependencies
• Full minified and gzipped bundle has a size of
• It also packages a mini bundle without Float/MPFR operations
• 100% open source & transparent build process

Installation

npm i gmp-wasm

It can also be used directly from HTML (via jsDelivr):

<!-- loads the full, minified library into the global `gmp` variable -->
<script src="https://cdn.jsdelivr.net/npm/gmp-wasm"></script>

<!-- or loads the non-minified library -->
<script src="https://cdn.jsdelivr.net/npm/gmp-wasm/dist/index.umd.js"></script>

<!-- or loads the minified library without Float/MPFR functions -->
<script src="https://cdn.jsdelivr.net/npm/gmp-wasm/dist/mini.umd.min.js"></script>

Usage

gmp-wasm also provides a high-level wrapper over the GMP functions. There are three major components:

• g.Integer() - Wraps integers (MPZ)
• g.Rational() - Wraps rational numbers (MPQ)
• g.Float() - Wraps floating-point numbers (MPFR)

const gmp = require('gmp-wasm');

gmp.init().then(({ calculate }) => {
  // calculate() automatically deallocates all objects created within the callback function
  const result = calculate((g) => {
    const six = g.Float(1).add(5);
    return g.Pi().div(six).sin(); // sin(Pi/6) = 0.5
  });
  console.log(result);
});

It is also possible to delay deallocation through the getContext() API:

const gmp = require('gmp-wasm');

gmp.init().then(({ getContext }) => {
  const ctx = getContext();
  let x = ctx.Integer(1);
  for (let i = 2; i < 16; i++) {
    x = x.add(i);
  }
  console.log(x.toString());
  setTimeout(() => ctx.destroy(), 50);
});

The precision and the rounding modes can be set by passing a parameter to the context or to the Float constructor.

const roundingMode = gmp.FloatRoundingMode.ROUND_DOWN;
const options = { precisionBits: 10, roundingMode };

const result = calculate(g => g.Float(1).div(3), options);
// or
const result2 = calculate(g => g.Float(1, options).div(3));
// or
const ctx = getContext(options);
const result3 = ctx.Float(1).div(3).toString();

Predefined constants

• Pi
• EulerConstant
• EulerNumber
• Log2
• Catalan

Advanced usage

High-level wrapper can be combined with low-level functions:

const sum = calculate((g) => {
  const a = g.Float(1);
  const b = g.Float(2);
  const c = g.Float(0);
  // c = a + b
  binding.mpfr_add(c.mpfr_t, a.mpfr_t, b.mpfr_t, 0);
  return c;
});

If you want more control and performance you can use the original GMP / MPFR functions even without high-level wrappers.

const gmp = require('gmp-wasm');

gmp.init().then(({ binding }) => {
  // Create first number and initialize it to 30
  const num1Ptr = binding.mpz_t();
  binding.mpz_init_set_si(num1Ptr, 30);
  // Create second number from string. The string needs to be copied into WASM memory
  const num2Ptr = binding.mpz_t();
  const strPtr = binding.malloc_cstr('40');
  binding.mpz_init_set_str(num2Ptr, strPtr, 10);
  // Calculate num1Ptr + num2Ptr, store the result in num1Ptr
  binding.mpz_add(num1Ptr, num1Ptr, num2Ptr);
  // Get result as integer
  console.log(binding.mpz_get_si(num1Ptr));
  // Deallocate memory
  binding.free(strPtr);
  binding.mpz_clears(num1Ptr, num2Ptr);
  binding.mpz_t_frees(num1Ptr, num2Ptr);
});

Sometimes, it's easier and faster to deallocate everything by reinitializing the WASM bindings:

// Deallocate all memory objects created by gmp-wasm
await binding.reset();

Performance

In some cases, this library can provide better performance than the built-in BigInt type.

For example, calculating 8000 digits of Pi using the following formula provides better results:

PI = 3
  + 3 * (1/2) * (1/3) * (1/4)
  + 3 * ((1 * 3)/(2 * 4)) * (1/5) * (1 / (4^2))
  + 3 * ((1 * 3 * 5) / (2 * 4 * 6)) * (1/7) * (1 / (4^3))
  + ...

| Test | Avg. time | Speedup | | ----------------------------------------------------------------------------------- | --------- | -------- | | With JS built-in BigInt type | 129 ms | 1x | | gmp-wasm Integer() high-level wrapper | 88 ms | 1.47x | | Same as previous with delayed memory deallocation | 78 ms | 1.65x | | gmp-wasm MPZ low-level functions | 53 ms | 2.43x | | decimal.js 10.3.1 with integer division | 443 ms | 0.29x | | big-integer 1.6.51 | 129 ms | 1x | | ---------------------------- | -------- | -------- | | gmp-wasm Float() high-level wrapper | 175 ms | 0.74x | | Same as previous with delayed memory deallocation | 169 ms | 0.76x | | gmp-wasm MPFR low-level functions | 118 ms | 1.09x | | decimal.js 10.3.1 with float division | 785 ms | 0.16x | | ---------------------------- | -------- | -------- | | gmp-wasm Float(1).atan().mul(4) | 0.6 ms | 215x | | gmp-wasm Float('0.5').asin().mul(6) | 17 ms | 7.59x |

* These measurements were made with Node.js v16.14 on an Intel Kaby Lake desktop CPU. Source code is here.