1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
//! A crate to test if floats are almost equal.
//!
//! ```
//! # let (x, y, z) = (0.5, 1.0, 2.0);
//! // Compare two variables.
//! if almost::equal(x, y) {
//!    println!("They're almost equal!");
//! }
//! // Or, if you need need to compare with a constant zero:
//! if almost::zero(z) {
//!    println!("It's almost zero!");
//! }
//! ```
//!
//! # Why another crate?
//!
//! There are a lot of crates for doing this already.
//!
//! The author crate has fairly strong opinions on how this should be done, and
//! thinks most of the similar crates in the wild make dubious choices, make it
//! easy for the user to misuse, or follow poor numerical robustness practices.
//!
//! Specific differences / benefits compared to other crates:
//!
//! 1. Better choice of default tolerances for unknown inputs. Often the value
//!    of the `EPSILON` is used as a value for tolerance, or its use is
//!    encouraged by the API).
//!
//!    This is the wrong choice more often than it is right. The machine epsilon
//!    is a quite strict bound for comparison, and after just a few arithmetic
//!    operations you will no longer be within it.
//!
//!    This library chooses a default tolerance value that is much more
//!    forgiving while still tight enough for it to be unlikely to cause false
//!    positives (specifically, it assumes roughly half of the bits have been
//!    lost to rounding, e.g. the *square root* of the machine epsilon).
//!
//! 2. Relative comparison by default. Most of the crates in the wild seem to
//!    use a hybrid between relative and absolute comparison. This is bad for
//!    arbitrary numbers which may have any scale, and gives up a number of
//!    desirable properties of the floating point number system.
//!
//! 3. Absolute comparison with zero. The only downside to using relative
//!    comparison by default is that it is essentially never useful to use
//!    relative comparison where one of the values is known in advance to be
//!    zero.
//!
//!    As a result, this library provides `almost::zero(v)` as well, which uses
//!    absolute comparison.
//!
//! 4. Properly handling both overflow and underflow.
//!
//!    Because this library uses relative comparison, denormal numbers behave
//!    properly, as well as comparisons where one of the values has overflowed
//!    to infinity. The second might sound impossible, but we can just rescale
//!    both values, and compare with the same tolerance.
//!
//! 5. Simple API. We don't expose other ways of comparing numbers, most of
//!    which are either dubious choices for non-niche use cases.
//!
//! That said, there's no one size fits all here. Numerical robustness is full
//! of tradeoffs, and while I believe the ones made by this library are good for
//! most cases, they do not and cannot satisfy every possible case.
#![no_std]

pub(crate) mod imp;

/// Returns `true` if `lhs` and `rhs` are almost equal.
///
/// ```
/// assert!(almost::equal(0.1 + 0.2, 0.3));
/// ```
///
/// Do not use this to compare a value with a constant zero. Instead, for this
/// you should use [`almost::zero`](zero).
///
/// This uses a relative comparison with a threshold chosen to be good for
/// values produced by arbitrary computation, while still tight enough for it to
/// be unlikely to cause false positives in practice This is a good default, but
/// you have a tighter bound you need to use,
/// [`almost::equal_with_tolerance`](equal_with_tolerance) is also available.
///
/// Note that this returns false in the case that both values are NaN.
#[inline]
pub fn equal<T: AlmostEqual>(lhs: T, rhs: T) -> bool {
    lhs.almost_equals(rhs)
}

/// Returns `true` if `a` is almost zero.
///
/// ```
/// # use core as std;
/// assert!(almost::zero(std::f32::EPSILON));
/// ```
///
/// This is the correct function to use when comparing to see if a value is
/// almost zero.
///
/// Testing if a value is zero is a unique case where a relative comparison
/// becomes almost useless, and an absolute comparison becomes the obviously
/// correct choice.
///
/// As such, this performs comparison with a *absolute* threshold. The threshold
/// used assumes half of the bits have been lost to rounding, which is a good
/// default for user code that does not keep track of this, while still tight
/// enough for it to be unlikely to cause false positives in practice. However,
/// if you need a tighter bound, the function
/// [`almost::zero_with_tolerance`](zero_with_tolerance) can be used.
#[inline]
pub fn zero<T: AlmostEqual>(a: T) -> bool {
    a.almost_zero()
}

/// Returns `true` if `a` is almost zero, using the specified absolute
/// tolerance.
///
/// ```
/// # use core as std;
/// assert!(!almost::zero_with_tolerance(std::f32::EPSILON, std::f32::EPSILON));
/// ```
///
/// This is a version of [`almost::zero`](zero) which does not define a
/// tolerance value for you.
///
/// The correct choice of tolerance value is tricky and differs depending on:
///
/// - The type of numerical operations you're doing.
/// - The scale of the values used.
/// - The number of operations performed.
///
/// However, for comparison with zero it is possible to give broad guidelines
/// (note that these do *not* apply to
/// [`almost::equal_with_tolerance`](equal_with_tolerance), which for which the
/// correct decision is more challenging).
///
/// # Panics
/// This function panics in debug mode if `tolerance` is not greater than zero,
/// as the results are unlikely to be sensible.
///
/// In release builds it should never panic.
#[inline]
pub fn zero_with_tolerance<T: AlmostEqual>(v: T, tolerance: T) -> bool{
    v.almost_zero_with_tolerance(tolerance)
}

/// Returns `true` if `lhs` and `rhs` are almost equal using the provided
/// relative tolerance.
///
/// ```
/// const MY_TOLERANCE: f32 = almost::F32_TOLERANCE / 2.0;
/// assert!(almost::equal_with_tolerance(0.1 + 0.2, 0.3, MY_TOLERANCE));
/// ```
///
/// Do not use this to compare a value with a constant zero. Instead, for this
/// you should use [`almost::zero_with_tolerance`](zero_with_tolerance).
///
/// # Panics
/// This function panics in debug mode if `tolerance` is less than `T::EPSILON`
/// or greater than 1.0, as the results are unlikely to be sensible.
///
/// In release builds it should never panic.
#[inline]
pub fn equal_with_tolerance<T: AlmostEqual>(lhs: T, rhs: T, tolerance: T) -> bool {
    lhs.almost_equals_with_tolerance(rhs, tolerance)
}

/// A trait for comparing floating point numbers. Not broadly intended to be
/// used by most code (instead, use the functions at the crate root), however it
/// could be useful for generic code too.
pub trait AlmostEqual {
    /// The default tolerance value for this type. Typically equivalent to
    /// `T::EPSILON.sqrt()`, as we assume that around half of the precision bits
    /// of any arbitrary computation have been rounded away.
    const DEFAULT_TOLERANCE: Self;

    /// The machine epsilon for this type. This generally should not be used as
    /// a tolerance value (it's frequently too strict), however it can be useful
    /// when computing tolerances.
    const MACHINE_EPSILON: Self;

    /// Equivalent to [`almost::zero`](zero).
    /// ```
    /// # let v = 0.000001f32;
    /// # use almost::AlmostEqual;
    /// assert!(v.almost_zero());
    /// ```
    #[inline]
    fn almost_zero(self) -> bool where Self: Sized {
        self.almost_zero_with_tolerance(Self::DEFAULT_TOLERANCE)
    }

    /// Equivalent to [`almost::equal`](equal).
    /// ```
    /// # let (a, b) = (0.5, 0.5);
    /// # use almost::AlmostEqual;
    /// assert!(a.almost_equals(b));
    /// ```
    #[inline]
    fn almost_equals(self, rhs: Self) -> bool where Self: Sized {
        self.almost_equals_with_tolerance(rhs, Self::DEFAULT_TOLERANCE)
    }

    /// Equivalent to [`almost::equal_with_tolerance`](equal_with_tolerance).
    /// ```
    /// # let (a, b) = (0.5, 0.5);
    /// # use almost::AlmostEqual;
    /// const MY_TOLERANCE: f32 = almost::F32_TOLERANCE / 2.0;
    /// assert!(a.almost_equals_with_tolerance(b, MY_TOLERANCE));
    /// ```
    fn almost_equals_with_tolerance(self, rhs: Self, tol: Self) -> bool;

    /// Equivalent to [`almost::zero_with_tolerance`](zero_with_tolerance).
    /// ```
    /// # use almost::AlmostEqual;
    /// assert!(0.01.almost_zero_with_tolerance(0.05));
    /// ```
    fn almost_zero_with_tolerance(self, tol: Self) -> bool;
}

/// The default tolerance used for `f64`. Equivalent to `f64::EPSILON.sqrt()`
/// (or `0.000000014901161193847656_f64`), as we assume that around half of the
/// precision bits of any arbitrary value have been rounded away.
pub const F64_TOLERANCE: f64 = 0.000000014901161193847656_f64;
/// The default tolerance used for `f32`. Equivalent to `f32::EPSILON.sqrt()`
/// (or `0.00034526698_f32`), as we assume that around half of the precision
/// bits of any arbitrary value have been rounded away.
pub const F32_TOLERANCE: f32 = 0.00034526698_f32;

impl AlmostEqual for f64 {

    const MACHINE_EPSILON: Self = core::f64::EPSILON;

    const DEFAULT_TOLERANCE: Self = F64_TOLERANCE;

    fn almost_equals_with_tolerance(self, rhs: Self, tol: Self) -> bool {
        debug_assert!(tol < 1.0, "Tolerance should not be greater than 1.0");
        debug_assert!(tol >= Self::MACHINE_EPSILON, "Tolerance should not be smaller than the machine epsilon");
        crate::imp::f64::eq_with_tol_impl(self, rhs, tol)
    }

    fn almost_zero_with_tolerance(self, tol: Self) -> bool {
        debug_assert!(tol > 0.0);
        crate::imp::f64::abs(self) < tol
    }
}


impl AlmostEqual for f32 {

    const MACHINE_EPSILON: Self = core::f32::EPSILON;

    const DEFAULT_TOLERANCE: Self = F32_TOLERANCE;

    fn almost_equals_with_tolerance(self, rhs: Self, tol: Self) -> bool {
        debug_assert!(tol < 1.0, "Tolerance should not be greater than 1.0");
        debug_assert!(tol >= Self::MACHINE_EPSILON, "Tolerance should not be smaller than the machine epsilon");
        crate::imp::f32::eq_with_tol_impl(self, rhs, tol)
    }

    fn almost_zero_with_tolerance(self, tol: Self) -> bool {
        debug_assert!(tol > 0.0);
        crate::imp::f32::abs(self) < tol
    }
}