Fixed-point numbers
The fixed crate provides fixed-point numbers.
FixedI8
andFixedU8
are eight-bit fixed-point numbers.FixedI16
andFixedU16
are 16-bit fixed-point numbers.FixedI32
andFixedU32
are 32-bit fixed-point numbers.FixedI64
andFixedU64
are 64-bit fixed-point numbers.FixedI128
andFixedU128
are 128-bit fixed-point numbers.
An n-bit fixed-point number has f = Frac
fractional
bits where 0 ≤ f ≤ n, and
n − f integer bits. For example,
FixedI32<U24> is a 32-bit signed fixed-point number with
n = 32 total bits, f = 24 fractional bits, and
n − f = 8 integer bits.
FixedI32<U0> behaves like i32
, and
FixedU32<U0> behaves like u32
.
The difference between any two successive representable numbers is constant throughout the possible range for a fixed-point number: Δ = 1/2f. When f = 0, like in FixedI32<U0>, Δ = 1 because representable numbers are integers, and the difference between two successive integers is 1. When f = n, Δ = 1/2n and the value lies in the range −0.5 ≤ x < 0.5 for signed numbers like FixedI32<U32>, and in the range 0 ≤ x < 1 for unsigned numbers like FixedU32<U32>.
In version 1 the typenum crate is used for the fractional bit count Frac
;
the plan is to to have a major version 2 with const generics when the Rust
compiler’s generic_const_exprs
feature is ready and stabilized. An alpha
version is already available.
The main features are
- Representation of binary fixed-point numbers up to 128 bits wide.
- Conversions between fixed-point numbers and numeric primitives.
- Comparisons between fixed-point numbers and numeric primitives.
- Parsing from strings in decimal, binary, octal and hexadecimal.
- Display as decimal, binary, octal and hexadecimal.
- Arithmetic and logic operations.
This crate does not provide decimal fixed-point numbers. For example 0.001 cannot be represented exactly, as it is 1/103. It is binary fractions like 1/24 (0.0625) that can be represented exactly, provided there are enough fractional bits.
This crate does not provide general analytic functions.
- No algebraic functions are provided, for example no
pow
. - No trigonometric functions are provided, for example no
sin
orcos
. - No other transcendental functions are provided, for example no
log
orexp
.
These functions are not provided because different implementations can have different trade-offs, for example trading some correctness for speed. Implementations can be provided in other crates.
- The cordic crate provides various functions implemented using the CORDIC algorithm.
The conversions supported cover the following cases.
- Infallible lossless conversions between fixed-point numbers and numeric
primitives are provided using
From
andInto
. These never fail (infallible) and do not lose any bits (lossless). - Infallible lossy conversions between fixed-point numbers and numeric
primitives are provided using the
LossyFrom
andLossyInto
traits. The source can have more fractional bits than the destination. - Checked lossless conversions between fixed-point numbers and numeric
primitives are provided using the
LosslessTryFrom
andLosslessTryInto
traits. The source cannot have more fractional bits than the destination. - Checked conversions between fixed-point numbers and numeric primitives are
provided using the
FromFixed
andToFixed
traits, or using thefrom_num
andto_num
methods and their checked versions. - Additionally,
az
casts are implemented for conversion between fixed-point numbers and numeric primitives. - Fixed-point numbers can be parsed from decimal strings using
FromStr
, and from binary, octal and hexadecimal strings using thefrom_str_binary
,from_str_octal
andfrom_str_hex
methods. The result is rounded to the nearest, with ties rounded to even. - Fixed-point numbers can be converted to strings using
Display
,Binary
,Octal
,LowerHex
,UpperHex
,LowerExp
andUpperExp
. The output is rounded to the nearest, with ties rounded to even. - All fixed-point numbers are plain old data, so
bytemuck
bit casting conversions can be used.
What’s new
Version 1.28.0 news (2024-07-25)
- The crate now requires rustc version 1.79.0 or later.
- The following methods were added to all fixed-point numbers and to the
Fixed
trait: - For all fixed-point numbers and the
Fixed
trait, the following methods were renamed. The old method names are deprecated.round_ties_to_even
renamed toround_ties_even
checked_round_ties_to_even
renamed tochecked_round_ties_even
saturating_round_ties_to_even
renamed tosaturating_round_ties_even
wrapping_round_ties_to_even
renamed towrapping_round_ties_even
unwrapped_round_ties_to_even
renamed tounwrapped_round_ties_even
overflowing_round_ties_to_even
renamed tooverflowing_round_ties_even
- The following constants were added to the
consts
module and as associated constants to fixed-point types: - The experimental feature
nightly-float
was added. - For the experimental feature
num-traits
, the following traits were implemented for all fixed-point numbers:
Other releases
Details on other releases can be found in RELEASES.md.
Quick examples
use I20F12;
// 19/3 = 6 1/3
let six_and_third = I20F12 from_num / 3;
// four decimal digits for 12 binary digits
assert_eq!;
// find the ceil and convert to i32
assert_eq!;
// we can also compare directly to integers
assert_eq!;
The type I20F12
is a 32-bit fixed-point signed number with 20 integer bits
and 12 fractional bits. It is an alias to FixedI32<U12>. The
unsigned counterpart would be U20F12
. Aliases are provided for all
combinations of integer and fractional bits adding up to a total of eight, 16,
32, 64 or 128 bits.
use ;
// -8 ≤ I4F4 < 8 with steps of 1/16 (~0.06)
let a = I4F4 from_num;
// multiplication and division by integers are possible
let ans1 = a / 5 * 17;
// 1 / 5 × 17 = 3 2/5 (3.4), but we get 3 3/16 (~3.2)
assert_eq!;
assert_eq!;
// -8 ≤ I4F12 < 8 with steps of 1/4096 (~0.0002)
let wider_a = I4F12 from;
let wider_ans = wider_a / 5 * 17;
let ans2 = I4F4 from_num;
// now the answer is the much closer 3 6/16 (~3.4)
assert_eq!;
assert_eq!;
The second example shows some precision and conversion issues. The low precision
of a
means that a / 5
is 3⁄16 instead of 1⁄5, leading to an inaccurate
result ans1
= 3 3⁄16 (~3.2). With a higher precision, we get wider_a / 5
equal to 819⁄4096, leading to a more accurate intermediate result wider_ans
=
3 1635⁄4096. When we convert back to four fractional bits, we get ans2
= 3
6⁄16 (~3.4).
Note that we can convert from I4F4
to I4F12
using From
, as the
target type has the same number of integer bits and a larger number of
fractional bits. Converting from I4F12
to I4F4
cannot use From
as we
have less fractional bits, so we use from_num
instead.
Writing fixed-point constants and values literally
The lit
method, which is available as a const
function, can be used to
parse literals. It supports
- underscores as separators;
- prefixes “
0b
”, “0o
” and “0x
” for binary, octal and hexadecimal numbers; - an optional decimal exponent with separator “
e
” or “E
” for decimal, binary and octal numbers, or with separator “@
” for all supported radices including hexadecimal.
use I16F16;
// 0.1275e2 is 12.75
const TWELVE_POINT_75: I16F16 = I16F16 lit;
// 1.8 hexadecimal is 1.5 decimal, and 18@-1 is 1.8
const ONE_POINT_5: I16F16 = I16F16 lit;
// 12.75 + 1.5 = 14.25
let sum = TWELVE_POINT_75 + ONE_POINT_5;
assert_eq!;
Using the fixed crate
The fixed crate is available on crates.io. To use it in your crate, add it as a dependency inside Cargo.toml:
[]
= "1.28"
The fixed crate requires rustc version 1.79.0 or later.
Optional features
The fixed crate has these optional feature:
arbitrary
, disabled by default. This provides the generation of arbitrary fixed-point numbers from raw, unstructured data. This feature requires the arbitrary crate.borsh
, disabled by default. This implements serialization and deserialization using the borsh crate.serde
, disabled by default. This provides serialization support for the fixed-point types. This feature requires the serde crate.std
, disabled by default. This is for features that are not possible underno_std
: currently the implementation of theError
trait forParseFixedError
.serde-str
, disabled by default. Fixed-point numbers are serialized as strings showing the value when using human-readable formats. This feature requires theserde
and thestd
optional features. Warning: numbers serialized when this feature is enabled cannot be deserialized when this feature is disabled, and vice versa.
To enable features, you can add the dependency like this to Cargo.toml:
[]
= ["serde"]
= "1.28"
Experimental optional features
It is not considered a breaking change if the following experimental features are removed. The removal of experimental features would however require a minor version bump. Similarly, on a minor version bump, optional dependencies can be updated to an incompatible newer version.
num-traits
, disabled by default. This implements some traits from the num-traits crate. (The plan is to promote this to an optional feature once the num-traits crate reaches version 1.0.0.)nightly-float
, disabled by default. This requires the nightly compiler, and implements conversions and comparisons with the experimentalf16
andf128
primitives. (The plan is to always implement the conversions and comparisons and remove this experimental feature once the primitives are stabilized.)
Deprecated optional features
The following optional features are deprecated and will be removed in the next major version of the crate.
az
, has no effect. Previously required for theaz
cast traits. Now these cast traits are always provided.f16
, has no effect. Previously required for conversion to/from half::f16 and half::bf16. Now these conversions are always provided.
License
This crate is free software: you can redistribute it and/or modify it under the terms of either
- the Apache License, Version 2.0 or
- the MIT License
at your option.
Contribution
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache License, Version 2.0, shall be dual licensed as above, without any additional terms or conditions.