minifloat

Minifloat

Minifloat is a simple library for dealing with small floating point numbers, extrapolating from the IEEE 754 standard as much as possible.

See Minifloat on Wikipedia for more information.

It is capable of working with numbers up to single precision (i.e. 32 bits).

Usage

To use this library, you must instantiate a Minifloat object (see the API Reference below for details on arguments).

import {Minifloat} from 'minifloat';
const float8 = new Minifloat(1, 4, 3);
console.log(float8.fromBits(48)); // logs 0.5
console.log(float8.toBits(0.5));  // logs 48

Note that this library is implemented in WASM. For now, the compiled module is inlined directly as a base64 string for maximum ease of use, but this does exclude it from use in environments where compiling WASM is not allowed, and it is slightly less efficient than loading the file directly (though at 1.5k of WASM, it's not likely to matter).

Note also that this library is distributed as an ESM module that makes use of top-level await (to load the WASM). As a result, it cannot be used synchronously in CommonJS modules (i.e. you cannot require('minifloat') in Node). Instead, it must be imported (though note that CommonJS can use import('minifloat').then(...) if necessary).

API Reference

Minifloat constructor

new Minifloat(signBits, exponentBits, significandBits, exponentBias?, defaultRoundingRule?)

Args:

• signBits: must be 0 or 1, determines whether floats are signed or not
• exponentBits: must be between 1 and 8, determines the range
• significandBits: must be between 1 and 23, determines the precision
• exponentBias: a signed integer used to shift the exponent; must be such that the minimum exponent is at least -1023 and the maximum exponent is at most 1024; by default, it's half-way between 0 and (1 << exponentBits) - 1, rounded down
• defaultRoundingRule: an element of the RoundingRule enum indicating how toBits will round by default (defaults to NearestEven)

Returns:

• a new Minifloat instance

Note that instances are returned frozen.

Minifloat#toBits

instance.toBits(float, rounding?)

Args:

• float: a floating point number
• rounding: an optional rounding rule, with default set by the instance

Returns:

• an unsigned integer with the bits

Minifloat#fromBits

instance.fromBits(bits)

Args:

• bits: an unsigned integer with bits to unpack

Returns:

• a double-precision floating point number

Minifloat#LABEL

String property containing the a human-readable label for this system. The format is just the first 3 (or 4) constructor arguments separated by decimal points.

Minifloat#SIGNED

Boolean property indicating whether this minifloat system is signed.

Minifloat#BITS

Number property indicating the total number of bits in the underlying binary representation.

Minifloat#CARDINALITY

Number property indicating the total number of representable numbers in this minifloat system (counting each individual NaN, despite the fact that we don't actually handle them individually). Note that this is always just 2 ** BITS.

Minifloat#MAX_VALUE

Number property indicating the maximum representable finite number in this minifloat system. This is just fromBits(toBits(Infinity) - 1).

Minifloat#MIN_VALUE

Number property indicating the minimum representable positive number in this minifloat system. This is just fromBits(1).

Minifloat#EPISLON

NOTE: This is probably not what you're looking for! Please see this stack overflow answer for an explanation of why EPSILON is not actually all that useful.

Number property indicating the difference between 1 and the next-larger representable number. If 1 is not representable, then either MIN_VALUE or MAX_VALUE is used instead, depending on which is closest to 1.

This is included as an analogue to Number.EPSILON, but generally should not be used for anything.

RoundingRule

An enum representing the five possible rounding rules, which fall into two categories.

The first two rules round to the nearest representable value. They differ only in how ties are broken, when an input is exactly halfway between two values:

• NearestEven: the default rule, breaks ties toward evens (i.e. the value whose least significant binary digit is zero)
• NearestAway: breaks ties away from zero

The other three rules are directed roundings. They will always round in the same direction, and will never return a result in the other direction. This can in some cases return results quite far from the input:

• Trucate: rounds toward zero
• Ceiling: rounds up, toward +Infinity
• Floor: rounds down, toward -Infinity

Unsigned floats and negative values

When rounding negative values to an unsigned float, the result will always be either NaN or +0. If the signed result would have been -0, we treat it as +0, since IEEE 754 considers these values equal.

We take a strict approach with the directed rounding rules, since they must never give a result in the "wrong" direction. For Ceiling and Truncate, we return +0 for all negative inputs. For Floor, we return NaN for all negative inputs except -0, which rounds to +0 per the previous paragraph.

For the "nearest" roundings, we treat the rounding as if a sign bit were available. If the signed result would round up to -0 then we return +0. Otherwise, we return NaN.

For use cases where it's desirable to round all negative values to zero (rather than NaN), it's recommended to call Math.max(0, arg) before calling round or toBits, which transforms all negative args to zero ahead of time. Note that round(Math.max(0, NaN)) will still return NaN.

License

Minifloat is Copyright 2023 Stephen Hicks, and is released under the MIT license.