notes:ieee_754-1985

**This is an old revision of the document!**

The IEEE 754 standard defines a representation of floating point numbers and rules for manipulating them. The standard was published in 1985 and revised in 2008 - this page currently describes only the original 1985 standard.

Briefly, floating point numbers are a way to represent a wide numeric range in a limited set of bits by storing a limited number of significant digits. As the absolute size of the number increases, the absolute precision decreases.

The principle is similar to exponential notation for numbers (e.g. $3.523 \times 10^6$) except that IEEE floating point values use powers of 2 instead of 10.

The standard defines two precisions - **single** is a **32-bit** representation and **double** is a **64-bit** representation. These typically correspond to the C types `float`

and `double`

on platforms where the underlying hardware uses IEEE 754 (which is to say most current hardware architectures).

Each value is split into three chunks of bits, with byte boundaries ignored:

Section | Single | Double | Meaning |
---|---|---|---|

Sign | 1 bit | 1 bit | Set if the number is negative |

Exponent | 8 bits | 11 bits | The power of 2 to multiply |

Significand | 23 bits | 52 bits | The sigificant digits of the number |

Note that the **significand** is also known as the **mantissa** in some texts, although this is discouraged by the IEEE 754 standards committee and others because of confusion with other uses of the term.

This single bit is **0 for positive** numbers and **1 for negative** numbers. Note that the sign bit is valid in most cases even for special values - for example, the standard differentiates between positive and negative zero.

The exponent indicates the power of 2 by which the significand is multiplied. It is stored in **biased form**, which is an easy way to store a signed value in an unsigned field by simply adding a fixed value. The range of the significant, and the bias value which is added to it to obtain the actual unsigned value stored, is:

Precision | Bias | Range of valid exponents |
---|---|---|

Single | 127 | -126 – 127 |

Double | 1023 | -1022 – 1023 |

Note that the range is missing the values at each end — this is because a zero exponent and an exponent with all bits set both have special meanings.

This portion of the value stores the significant binary digits. To save space, there is assumed to be a leading **1** digit. For example, the significand **1.0100111…** is stored as **0100111…**. This form is said to be **normalised**.

Note that this is a simplification as there is also a **denormalised** form for values near zero, which is described below.

The most common form of IEEE floating point numbers is the normalised form — this is where the exponent has a value in the valid range (once the bias has been subtracted from the unsigned value stored). As explained above, the significand stores only the digits after the leading **1**, which is implicit.

If the standard C library is available, the `frexp()`

function normalises a floating point value such that the fractional part will be in the range $0.5 \le \times < 1.0$. Multiplying this value by **2** and reducing the exponent by **1** yields a value in the desired range $1.0 \le \times < 2.0$. At this point the leading digit can then be discarded as the implicit leading **1** (see the Significand section for details).

If `frexp()`

is available then it should be used, as it will likely use the underlying hardware representation to avoid expensive loops. However, a naive implementation can quite simply mimic its functionality — the following version demonstrates the principle, but a production version would also need to check for special values (zero, NaN, infinities) as well as catching under- and overflows:

double sample_frexp(double value, int *exponent) { *exponent = 0; if (value < 1.0) { while (value <= 0.5) { value *= 2.0; --(*exponent); } } else { while (value >= 1.0) { value /= 2.0; ++(*exponent); } } return value; }

At the lower end of the scale, very small numbers can be stored in **denormalised** form, where the implicit leading digit is a **0** instead of 1. In IEEE 754 this is represented by an exponent field of all zeroes and a non-zero significand. The actual exponent that this value represents is one higher than would be expected from a zero exponent field:

Precision | Denormalised Exponent |
---|---|

Single | -126 |

Double | -1022 |

At first sight it appears that this introduces overlap with the normalised numbers, as these are the lowest value exponents for a normalised value. However, the leading zero in the significand means that in fact there's no overlap.

Aside from normalised and denormalised numbers, there are a variety of values represented by specific bit patterns in the representation.

Sign bit | Any |
---|---|

Exponent | Zero |

Significand | Zero |

A value of exactly zero is represented by a exponent and significand of zero. The sign bit may be set or unset and IEEE 754 has the concept of both a positive and negative zero. For standard comparisons, however, these will both compare equal with zero, so the comparison $-0.0 < 0.0$ yields **false**.

To determine the sign of a floating point value including zero, the `copysign()`

function can be used with a non-zero value, or the `signbit()`

macro can be used more directly on some platforms (not available on WinCE, for example).

Sign bit | Any |
---|---|

Exponent | All bits set |

Significand | Zero |

If all bits are set in the exponent and the signficand is zero then the value represented is either positive or negative infinity, depending on the sign bit.

Sign bit | Any |
---|---|

Exponent | All bits set |

Significand | Non-zero |

If all bits are set in the exponent and the significand is non-zero then the value represented is ** not a number** often abbreviated to

Since any non-zero significand value is permitted, this allows a range of values to be specified. The sign bit may also be set or unset, which could be used to differentiate between different types of NaN. The main distinction is between a **quiet NaN** which has the most-significant bit of the significand set and a **signalling NaN** which has the MSB of the significand clear^{1)} (although the overall value must still be non-zero).

The intention of the **signalling NaN** is that this will raise some sort of exception, and then go on to yield a **quiet NaN** if a result is required. This means that each error is only raised once. However, the support for this on various platforms seems to differ.

There are three types of operations which can produce a NaN value:

- Operations which are provided an existing NaN value as an argument.
- Operations whose results are mathematically indeterminate - some examples are listed below:
**0.0 / 0.0**and**±∞ / ±∞****0.0 x ±∞****∞ – ∞**and equivalents

- Operations which yield complex results - some examples are listed below:
**√–n****log(–n)****sin⁻¹(x)**or**cos⁻¹(x)**where**x < –1**or**x > 1**

Values in the table below apply regardless of sign, since the sign bit is an independent quantity. Base 10 values are all rounded to one decimal place.

Limit | Single precision | Double precision | ||
---|---|---|---|---|

Base 2 | Base 10 | Base 2 | Base 10 | |

Smallest denormal | 2⁻²³ x 2⁻¹²⁶ | 1.4 x 10⁻⁴⁵ | 2⁻⁵² x 2⁻¹⁰²² | 4.9 x 10⁻³²⁴ |

Middle denormal | 2⁻¹ x 2⁻¹²⁶ | 5.9 x 10⁻³⁹ | 2⁻1 x 2⁻¹⁰²² | 1.1 x 10⁻³⁰⁸ |

Largest denormal | (1–2⁻²³) x 2⁻¹²⁶ | 1.2 x 10⁻³⁸ | (1–2⁻⁵²) x 2⁻¹⁰²² | 2.2 x 10⁻³⁰⁸ |

Smallest normal | 1 x 2⁻¹²⁶ | 1.2 x 10⁻³⁸ | 1 x 2⁻¹⁰²² | 2.2 x 10⁻³⁰⁸ |

Middle normal | 1 x 2⁶³ | 9.2 x 10¹⁸ | 1 x 2⁵¹¹ | 6.7 x 10¹⁵³ |

Largest normal | (2–2⁻²³) x 2¹²⁷ | 3.4 x 10³⁸ | (2–2⁻⁵²) x 2¹⁰²3 | 1.8 x 10³⁰⁸ |

notes/ieee_754-1985.1361664738.txt.gz · Last modified: 2013/02/24 00:12 by andy