[−][src]Crate float_eq
Explicitly bounded comparison of floating point numbers.
Comparing floating point values for equality is really hard. To get it right requires careful thought and iteration based on the needs of each specific algorithm's inputs and error margins. This API provides a toolbox of components to make your options clear and your choices explicit to future maintainers.
Table of Contents
Background
Given how widely algorithmic requirements can vary, float_eq
explores the
idea that there are no generally sensible default margins for comparisons.
This is in contrast to the approach taken by many other crates, which often
provide default epsilon values in checks or implicitly favour particular
algorithms. The author's hope is that by exposing the inherent complexity in
a uniform way, programmers will find it easier to develop an intuition for how
to write effective comparisons. The trade-off is that each individual
comparison requires more iteration time and thought.
And yes, this is yet another crate built on the principles described in that Random ASCII floating point comparison article, which is highly recommended background reading 🙂.
Making comparisons
The float_eq!
and float_ne!
macros compare two floating point
expressions for equality based on the result of one or more different kinds
of check. Each check is invoked by name and an upper boundary, so for example
rel <= 0.1
, should be read as "a relative epsilon comparison
with a maximum difference of less than or equal to 0.1
". If multiple checks
are provided then they are executed in order from left to right, shortcutting
to return early if one passes. The corresponding assert_float_eq!
and
assert_float_ne!
use the same interface:
use float_eq::{assert_float_eq, assert_float_ne, float_eq, float_ne}; assert!(float_eq!(1000.0f32, 1000.0002, ulps <= 4)); const ROUNDING_ERROR: f32 = 0.000_345_266_98; // f32::EPSILON.sqrt() assert!(float_ne!(4.0f32, 4.1, rel <= ROUNDING_ERROR)); const RECIP_REL_EPSILON: f32 = 0.000_366_210_94; // 1.5 * 2f32.powi(-12) assert_float_eq!(0.1f32.recip(), 10.0, rel <= RECIP_REL_EPSILON); assert_float_ne!(0.0f32, 0.000_1, abs <= 0.000_05, ulps <= 4);
The ideal choice of comparison will vary on a case by case basis, and depends
on the input range and error margins of the expressions to be compared. For
example, a test of the result of finite difference approximation of
derivatives might use a relative epsilon check with a max_diff
of the sqrt
of machine epsilon, whereas a test of the SSE _mm_rcp_ps
operation could
instead opt for a maximum relative error of 1.5 * 2^(-12)
based on the
available documentation. Algorithm stability can play a big part in the size
of these margins, and it can be worth seeing if code might be rearranged to
reduce loss of precision if you find yourself using large bounds.
Relative comparisons (ulps
and rel
) are usually a good choice for comparing
normal floats (e.g. when f32::is_normal
is true). However, they become
far too strict for comparisons of very small numbers with zero, where the
relative differences are very large but the absolute difference is tiny. This
is where you might choose to use an absolute epsilon (abs
) comparison instead.
There are also potential performance implications based on the target hardware.
Be prepared to research, test, benchmark and iterate on your comparisons. The floating point comparison article which informed this crate's implementation is a good place to start.
Absolute epsilon comparison
A check to see how far apart two expressions are by comparing the absolute
difference between them to an absolute, unscaled epsilon. Equivalent to, using
f32
as an example:
fn float_eq_abs(a: f32, b: f32, max_diff: f32) -> bool { // the PartialEq check covers equality of infinities a == b || (a - b).abs() <= max_diff }
Absolute epsilon tests do not work well for general floating point comparison,
because they do not take into account that floating point values' precision
changes with their magnitude. Thus max_diff
must be very specific and
dependent on the exact values being compared:
let a = 1.0f32; let b = 1.000_000_1f32; // the next representable value above 1.0 assert_float_eq!(a, b, abs <= 0.000_000_2); // equal assert_float_ne!(a * 4.0, b * 4.0, abs <= 0.000_000_2); // not equal assert_float_eq!(a * 4.0, b * 4.0, abs <= 0.000_000_5); // equal
Whereas a relative epsilon comparison could cope with this since it scales by the size of the largest input parameter:
assert_float_eq!(a, b, rel <= 0.000_000_2); assert_float_eq!(a * 4.0, b * 4.0, rel <= 0.000_000_2);
However, absolute epsilon comparison is often the best choice when comparing values directly against zero, especially when those values have undergone catastrophic cancellation, like the subtractions below. In this case, the relative comparison methods break down due to the relative ratio between values being so high compared to their absolute difference:
assert_float_eq!(1.0f32 - 1.000_000_1, 0.0, abs <= 0.000_000_2); // equal assert_float_ne!(1.0f32 - 1.000_000_1, 0.0, rel <= 0.000_000_2); // not equal assert_float_ne!(1.0f32 - 1.000_000_1, 0.0, ulps <= 1); // not equal
Absolute epsilon comparisons:
- Are useful for checking if a float is equal to zero, especially if it has undergone an operation that suffers from catastrophic cancellation or is a subnormal value.
- Are almost certainly not what you want to use when testing normal floats
for equality.
rel
andulps
checks can be easier to parameterise and reason about. - Can be useful for testing equality of infinities.
Relative epsilon comparison
A check to see how far apart two expressions are by comparing the absolute
difference between them to an epsilon that is scaled to the precision of the
larger input. Equivalent to, using f32
as an example:
fn float_eq_rel(a: f32, b: f32, max_diff: f32) -> bool { // the PartialEq check covers equality of infinities a == b || { let largest = a.abs().max(b.abs()); (a - b).abs() <= (largest * max_diff) } }
This makes it suitable for general comparison of values where the ratio between those values is relatively stable (e.g. normal floats, excluding infinity):
let a: f32 = 1.0; let b: f32 = 1.000_000_1; // the next representable value above 1.0 assert_float_eq!(a, b, rel <= 0.000_000_2); assert_float_eq!(a * 4.0, b * 4.0, rel <= 0.000_000_2);
However, relative epsilon comparison becomes far too strict when the numbers being checked are too close to zero, since the relative ratio between the values can be huge whilst the absolute difference remains tiny. In these circumstances, it is usually better to make an absolute epsilon check instead, especially if your algorithm contains some form of catastrophic cancellation, like these subtractions:
assert_float_ne!(1.0f32 - 1.000_000_1, 0.0, rel <= 0.000_000_2); // not equal assert_float_eq!(1.0f32 - 1.000_000_1, 0.0, abs <= 0.000_000_2); // equal
Relative epsilon comparisons:
- Are useful for checking if two normal floats are equal.
- Aren't a good choice when checking values against zero, where
abs
is often far better. - Have slightly counterintuitive results around powers of two values, where the relative precision ratio changes due to way the floating point exponent works.
- Are not useful at infinity, where any comparison using a non-zero margin will compare true.
Units in the Last Place (ULPs) comparison
A check to see how far apart two expressions are by comparing the number of
discrete values that can be expressed between them. This works by interpreting
the bitwise representation of the input values as integers and comparing the
absolute difference between those. Equivalent to, using f32
as an example:
fn float_eq_ulps(a: f32, b: f32, max_diff: u32) -> bool { if a.is_nan() || b.is_nan() { false // NaNs are never equal } else if a.is_sign_positive() != b.is_sign_positive() { a == b // values of different signs are only equal if both are zero. } else { let a_bits = a.to_bits(); let b_bits = b.to_bits(); let max = a_bits.max(b_bits); let min = a_bits.min(b_bits); (max - min) <= max_diff } }
Thanks to a deliberate quirk in the way the underlying format of IEEE floats was designed, this is a good measure of how near two values are that scales with their relative precision:
assert_float_eq!(1.0f32, 1.000_000_1, ulps <= 1); assert_float_eq!(4.0f32, 4.000_000_5, ulps <= 1); assert_float_eq!(-1_000_000.0f32, -1_000_000.06, ulps <= 1);
However, it becames far too strict when both expressions are close to zero, since the relative difference between them can be very large, whilst the absolute difference remains small. In these circumstances, it is usually better to make an absolute epsilon check instead, especially if your algorithm contains some form of catastrophic cancellation, like these subtractions:
assert_float_ne!(1.0f32 - 1.000_000_1, 0.0, ulps <= 1); // not equal assert_float_eq!(1.0f32 - 1.000_000_1, 0.0, abs <= 0.000_000_2); // equal
ULPs based comparisons:
- Are useful for checking if two normal floats are equal.
- Aren't a good choice when checking values against zero, where
abs
is often a better choice. - Provide a way to precisely tweak
max_diff
margins, since they have a 1-to-1 correlation with the underlying representation. - Have slightly counterintuitive results around powers of two values, where the relative precision ratio changes due to way the floating point exponent works.
- Do not work at all if the two values being checked have different signs.
- Whilst slightly counterintuitive at infinity (
MAX
is one ULP away fromINFINITY
), are more useful thanrel
checks for this.
Comparing composite types
When comparing composite values, it can be helpful to specify thresholds
separately for each individual field. The abs
, rel
and ulps
checks
expect this behaviour. Conversely, the abs_all
, rel_all
and ulps_all
checks accept a single epsilon that is then used to compare across all fields.
For example, arrays may be compared using an epsilon that covers each index
separately:
let a = [1.0, -2.0, 3.0]; let b = [-1.0, 2.0, 3.5]; assert_float_eq!(a, b, abs <= [2.0, 4.0, 0.5]);
Or with the same threshold across all values:
assert_float_eq!(a, b, abs_all <= 4.0);
Similarly, if FloatEq
and FloatEqAll
have been implemented for a
struct type:
let a = Complex32 { re: 2.0, im: 4.000_002 }; let b = Complex32 { re: 2.000_000_5, im: 4.0 }; assert_float_eq!(a, b, rel <= Complex32 { re: 0.000_000_25, im: 0.000_000_5 }); assert_float_eq!(a, b, rel_all <= 0.000_000_5); assert_float_eq!(a, b, ulps <= Complex32Ulps { re: 2, im: 4 }); assert_float_eq!(a, b, ulps_all <= 4);
Error messages
Assertion failure messages provide context information that hopefully helps
in determining how a check failed. The absolute difference (abs_diff
) and
ULPs difference (ulps_diff
) between the values are always provided, and
then the epsilon values used in the check are listed afterwards. For example,
this line:
assert_float_eq!(4.0f32, 4.000_008, rel <= 0.000_001);
Panics with this error message, where the relative epsilon, [rel] ε
, has
been scaled based on the size of the inputs (ε is the greek letter epsilon):
thread 'test' panicked at 'assertion failed: `float_eq!(left, right, rel <= ε)`
left: `4.0`,
right: `4.000008`,
abs_diff: `0.000008106232`,
ulps_diff: `Some(17)`,
[rel] ε: `0.000004000008`', assert_failure.rs:15:5
Comparing custom types
Comparison of new types is supported by implementing FloatEq
and FloatEqAll
.
If assert support is required, then FloatDiff
and FloatEqDebug
/FloatEqAllDebug
should also be implemented, as they provide important context information on
failure.
Macros
assert_float_eq | Asserts that two floating point expressions are equal to each other. |
assert_float_ne | Asserts that two floating point expressions are not equal to each other. |
debug_assert_float_eq | Asserts that two floating point expressions are equal to each other. |
debug_assert_float_ne | Asserts that two floating point expressions are not equal to each other. |
float_eq | Checks if two floating point expressions are equal to each other. |
float_ne | Checks if two floating point expressions are not equal to each other. |
Structs
ComplexUlps | The absolute difference between two floating point |
Traits
FloatDiff | Compute the difference between IEEE floating point values. |
FloatEq | Compare IEEE floating point values for equality using per-field thresholds. |
FloatEqAll | Compare IEEE floating point values for equality using a uniform threshold. |
FloatEqAllDebug | Debug context for when an assert using |
FloatEqDebug | Debug context for when an assert using |
Type Definitions
ComplexUlps32 |
|
ComplexUlps64 |
|