Skip to content

Fixed-precision decimal numbers based on IEEE 754 decimal64

License

Notifications You must be signed in to change notification settings

anz-bank/decimal

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

83 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

decimal

This library implements fixed-precision decimal numbers based on the IEEE 754-2019 standard; https://ieeexplore.ieee.org/document/8766229. More info can be found at: http://speleotrove.com/decimal/

Features

  • Decimal64, partial implementation of the ieee-754R standard
  • Half up and half even rounding
  • Up to 3 times faster than arbitrary precision decimal libraries in Go

Goals

Installation and use

Run go get github.com/anz-bank/decimal

package main

import (
	"fmt"

	"github.com/anz-bank/decimal"
)

func main() {
	var a decimal.Decimal64
	b := decimal.MustParse64("0.1")
	c := decimal.MustParse64("0.3")
	d := decimal.New64FromInt64(123456)

	fmt.Println(a, b, c, d)
}

Usage notes

Formatting

Decimal64 provides numerous ways to present numbers for human and machine consumption.

Decimal64 implements the following conventional interfaces:

  • fmt: Formatter, Scanner and Stringer
  • json: Marshaller and Unmarshaller
  • encoding: BinaryMarshaler, BinaryUnmarshaler, TextMarshaler and TextUnmarshaler
  • encoging/gob: GobEncoder and GobDecoder

The following methods provide more direct access to the internal methods used to implement fmt.Formatter. (For maximum control, use fmt.Printf &co or invoke the fmt.Formatter interface directly.)

  • Decimal64.Append formats straight into a []byte buffer.
  • Decimal64.Text formats in the same way, but returns a string.

Debugging

tl;dr: Use the decimal_debug compiler tag during debugging to greatly ease runtime inspection of Decimal64 values.

Debugging with the decimal package can be challeging because a Decimal64 number is encoded in a uint64 and the values it holds are inscrutable even to the trained eye. For example, the number one decimal.One64 is represented internally as the number 3450757314565799936 (2fe38d7ea4c68000 in hexadecimal).

To ease debugging, Decimal64 holds an optional debugInfo structure that contains a string representation and unpacked components of the uint64 representation for every Decimal64 value.

This feature is enabled through the decimal_debug compiler tag. This is done at compile time instead of through runtime flags because having the structure there even if not used would double the size of each number and greatly increase the cost of using it. The size and runtime cost of this feature is zero when the compiler tag is not present.

Docs

https://godoc.org/github.com/anz-bank/decimal

Why decimal?

Binary floating point numbers are fundamentally flawed when it comes to representing exact numbers in a decimal world. Just like 1/3 can't be represented in base 10 (it evaluates to 0.3333333333 repeating), 1/10 can't be represented in binary. The solution is to use a decimal floating point number. Binary floating point numbers (often just called floating point numbers) are usually in the form Sign * Significand * 2 ^ exp and decimal floating point numbers change this to Sign * Significand * 10 ^ exp. The use of base 10 eliminates the decimal fraction problem.

Why fixed precision?

Most implementations of a decimal floating point datatype implement an arbitrary precision type, which often uses an underlying big int. This gives flexibility in that as the number grows, the number of bits assigned to the number grows ( hence the term "arbitrary precision"). This library is different. It uses a 64-bit decimal datatype as specified in the IEEE-754R standard. This sacrifices the ability to represent arbitrarily large numbers, but is much faster than arbitrary precision libraries. There are two main reasons for this:

  1. The fixed-size data type is a uint64 under the hood and never requires heap allocation.
  2. All the algorithms can hard-code assumptions about the number of bits to work with. In fact, many of the operations work on the entire number as a single unit using 64-bit integer arithmetic and, on the occasions it needs to use more, 128 bits always suffices.