Crate noisy_float[][src]

This crate contains floating point types that panic if they are set to an illegal value, such as NaN.

The name “Noisy Float” comes from the terms “quiet NaN” and “signaling NaN”; “signaling” was too long to put in a struct/crate name, so “noisy” is used instead, being the opposite of “quiet.”

The standard types defined in noisy_float::types follow the principle demonstrated by Rust’s handling of integer overflow: a bad arithmetic operation is considered an error, but it is too costly to check everywhere in optimized builds. For each floating point number that is created, a debug_assert! invocation is used to check if it is valid or not. This way, there are guarantees when developing code that floating point numbers have valid values, but during a release run there is no overhead for using these floating point types compared to using f32 or f64 directly.

This crate makes use of the num, bounded, signed and floating point traits in the popular num_traits crate. This crate can be compiled with no_std.


An example using the R64 type, which corresponds to finite f64 values.

use noisy_float::prelude::*;

fn geometric_mean(a: R64, b: R64) -> R64 {
    (a * b).sqrt() //used just like regular floating point numbers

fn mean(a: R64, b: R64) -> R64 {
    (a + b) * 0.5 //the RHS of ops can be the underlying float type

    "geometric_mean(10.0, 20.0) = {}",
    geometric_mean(r64(10.0), r64(20.0))
//prints 14.142...
assert!(mean(r64(10.0), r64(20.0)) == 15.0);

An example using the N32 type, which corresponds to non-NaN f32 values. The float types in this crate are able to implement Eq and Ord properly, since NaN is not allowed.

use noisy_float::prelude::*;

let values = vec![n32(3.0), n32(-1.5), n32(71.3), N32::infinity()];
assert!(values.iter().cloned().min() == Some(n32(-1.5)));
assert!(values.iter().cloned().max() == Some(N32::infinity()));

An example converting from R64 to primitive types.

use noisy_float::prelude::*;
use num_traits::cast::ToPrimitive;

let value_r64: R64 = r64(1.0);
let value_f64_a: f64 = value_r64.into();
let value_f64_b: f64 = value_r64.raw();
let value_u64: u64 = value_r64.to_u64().unwrap();

assert!(value_f64_a == value_f64_b);
assert!(value_f64_a as u64 == value_u64);


This crate has the following cargo features:

  • serde: Enable serialization for all NoisyFloats using serde 1.0 and will transparently serialize then as floats
  • approx: Adds implementations to use NoisyFloat with the approx crate



Standard implementations of FloatChecker.


Prelude for the noisy_float crate.


Standard definitions of NoisyFloat.



A floating point number with a restricted set of legal values.



Trait for checking whether a floating point number is valid.