1 | #[cfg (all(feature = "serde" , feature = "alloc" ))] |
2 | #[allow (unused_imports)] |
3 | use alloc::string::ToString; |
4 | #[cfg (feature = "bytemuck" )] |
5 | use bytemuck::{Pod, Zeroable}; |
6 | use core::{ |
7 | cmp::Ordering, |
8 | iter::{Product, Sum}, |
9 | num::FpCategory, |
10 | ops::{Add, AddAssign, Div, DivAssign, Mul, MulAssign, Neg, Rem, RemAssign, Sub, SubAssign}, |
11 | }; |
12 | #[cfg (not(target_arch = "spirv" ))] |
13 | use core::{ |
14 | fmt::{ |
15 | Binary, Debug, Display, Error, Formatter, LowerExp, LowerHex, Octal, UpperExp, UpperHex, |
16 | }, |
17 | num::ParseFloatError, |
18 | str::FromStr, |
19 | }; |
20 | #[cfg (feature = "serde" )] |
21 | use serde::{Deserialize, Serialize}; |
22 | #[cfg (feature = "zerocopy" )] |
23 | use zerocopy::{AsBytes, FromBytes}; |
24 | |
25 | pub(crate) mod arch; |
26 | |
27 | /// A 16-bit floating point type implementing the IEEE 754-2008 standard [`binary16`] a.k.a "half" |
28 | /// format. |
29 | /// |
30 | /// This 16-bit floating point type is intended for efficient storage where the full range and |
31 | /// precision of a larger floating point value is not required. |
32 | /// |
33 | /// [`binary16`]: https://en.wikipedia.org/wiki/Half-precision_floating-point_format |
34 | #[allow (non_camel_case_types)] |
35 | #[derive (Clone, Copy, Default)] |
36 | #[repr (transparent)] |
37 | #[cfg_attr (feature = "serde" , derive(Serialize))] |
38 | #[cfg_attr ( |
39 | feature = "rkyv" , |
40 | derive(rkyv::Archive, rkyv::Serialize, rkyv::Deserialize) |
41 | )] |
42 | #[cfg_attr (feature = "rkyv" , archive(resolver = "F16Resolver" ))] |
43 | #[cfg_attr (feature = "bytemuck" , derive(Zeroable, Pod))] |
44 | #[cfg_attr (feature = "zerocopy" , derive(AsBytes, FromBytes))] |
45 | #[cfg_attr (kani, derive(kani::Arbitrary))] |
46 | pub struct f16(u16); |
47 | |
48 | impl f16 { |
49 | /// Constructs a 16-bit floating point value from the raw bits. |
50 | #[inline ] |
51 | #[must_use ] |
52 | pub const fn from_bits(bits: u16) -> f16 { |
53 | f16(bits) |
54 | } |
55 | |
56 | /// Constructs a 16-bit floating point value from a 32-bit floating point value. |
57 | /// |
58 | /// This operation is lossy. If the 32-bit value is to large to fit in 16-bits, ±∞ will result. |
59 | /// NaN values are preserved. 32-bit subnormal values are too tiny to be represented in 16-bits |
60 | /// and result in ±0. Exponents that underflow the minimum 16-bit exponent will result in 16-bit |
61 | /// subnormals or ±0. All other values are truncated and rounded to the nearest representable |
62 | /// 16-bit value. |
63 | #[inline ] |
64 | #[must_use ] |
65 | pub fn from_f32(value: f32) -> f16 { |
66 | f16(arch::f32_to_f16(value)) |
67 | } |
68 | |
69 | /// Constructs a 16-bit floating point value from a 32-bit floating point value. |
70 | /// |
71 | /// This function is identical to [`from_f32`][Self::from_f32] except it never uses hardware |
72 | /// intrinsics, which allows it to be `const`. [`from_f32`][Self::from_f32] should be preferred |
73 | /// in any non-`const` context. |
74 | /// |
75 | /// This operation is lossy. If the 32-bit value is to large to fit in 16-bits, ±∞ will result. |
76 | /// NaN values are preserved. 32-bit subnormal values are too tiny to be represented in 16-bits |
77 | /// and result in ±0. Exponents that underflow the minimum 16-bit exponent will result in 16-bit |
78 | /// subnormals or ±0. All other values are truncated and rounded to the nearest representable |
79 | /// 16-bit value. |
80 | #[inline ] |
81 | #[must_use ] |
82 | pub const fn from_f32_const(value: f32) -> f16 { |
83 | f16(arch::f32_to_f16_fallback(value)) |
84 | } |
85 | |
86 | /// Constructs a 16-bit floating point value from a 64-bit floating point value. |
87 | /// |
88 | /// This operation is lossy. If the 64-bit value is to large to fit in 16-bits, ±∞ will result. |
89 | /// NaN values are preserved. 64-bit subnormal values are too tiny to be represented in 16-bits |
90 | /// and result in ±0. Exponents that underflow the minimum 16-bit exponent will result in 16-bit |
91 | /// subnormals or ±0. All other values are truncated and rounded to the nearest representable |
92 | /// 16-bit value. |
93 | #[inline ] |
94 | #[must_use ] |
95 | pub fn from_f64(value: f64) -> f16 { |
96 | f16(arch::f64_to_f16(value)) |
97 | } |
98 | |
99 | /// Constructs a 16-bit floating point value from a 64-bit floating point value. |
100 | /// |
101 | /// This function is identical to [`from_f64`][Self::from_f64] except it never uses hardware |
102 | /// intrinsics, which allows it to be `const`. [`from_f64`][Self::from_f64] should be preferred |
103 | /// in any non-`const` context. |
104 | /// |
105 | /// This operation is lossy. If the 64-bit value is to large to fit in 16-bits, ±∞ will result. |
106 | /// NaN values are preserved. 64-bit subnormal values are too tiny to be represented in 16-bits |
107 | /// and result in ±0. Exponents that underflow the minimum 16-bit exponent will result in 16-bit |
108 | /// subnormals or ±0. All other values are truncated and rounded to the nearest representable |
109 | /// 16-bit value. |
110 | #[inline ] |
111 | #[must_use ] |
112 | pub const fn from_f64_const(value: f64) -> f16 { |
113 | f16(arch::f64_to_f16_fallback(value)) |
114 | } |
115 | |
116 | /// Converts a [`f16`] into the underlying bit representation. |
117 | #[inline ] |
118 | #[must_use ] |
119 | pub const fn to_bits(self) -> u16 { |
120 | self.0 |
121 | } |
122 | |
123 | /// Returns the memory representation of the underlying bit representation as a byte array in |
124 | /// little-endian byte order. |
125 | /// |
126 | /// # Examples |
127 | /// |
128 | /// ```rust |
129 | /// # use half::prelude::*; |
130 | /// let bytes = f16::from_f32(12.5).to_le_bytes(); |
131 | /// assert_eq!(bytes, [0x40, 0x4A]); |
132 | /// ``` |
133 | #[inline ] |
134 | #[must_use ] |
135 | pub const fn to_le_bytes(self) -> [u8; 2] { |
136 | self.0.to_le_bytes() |
137 | } |
138 | |
139 | /// Returns the memory representation of the underlying bit representation as a byte array in |
140 | /// big-endian (network) byte order. |
141 | /// |
142 | /// # Examples |
143 | /// |
144 | /// ```rust |
145 | /// # use half::prelude::*; |
146 | /// let bytes = f16::from_f32(12.5).to_be_bytes(); |
147 | /// assert_eq!(bytes, [0x4A, 0x40]); |
148 | /// ``` |
149 | #[inline ] |
150 | #[must_use ] |
151 | pub const fn to_be_bytes(self) -> [u8; 2] { |
152 | self.0.to_be_bytes() |
153 | } |
154 | |
155 | /// Returns the memory representation of the underlying bit representation as a byte array in |
156 | /// native byte order. |
157 | /// |
158 | /// As the target platform's native endianness is used, portable code should use |
159 | /// [`to_be_bytes`][Self::to_be_bytes] or [`to_le_bytes`][Self::to_le_bytes], as appropriate, |
160 | /// instead. |
161 | /// |
162 | /// # Examples |
163 | /// |
164 | /// ```rust |
165 | /// # use half::prelude::*; |
166 | /// let bytes = f16::from_f32(12.5).to_ne_bytes(); |
167 | /// assert_eq!(bytes, if cfg!(target_endian = "big" ) { |
168 | /// [0x4A, 0x40] |
169 | /// } else { |
170 | /// [0x40, 0x4A] |
171 | /// }); |
172 | /// ``` |
173 | #[inline ] |
174 | #[must_use ] |
175 | pub const fn to_ne_bytes(self) -> [u8; 2] { |
176 | self.0.to_ne_bytes() |
177 | } |
178 | |
179 | /// Creates a floating point value from its representation as a byte array in little endian. |
180 | /// |
181 | /// # Examples |
182 | /// |
183 | /// ```rust |
184 | /// # use half::prelude::*; |
185 | /// let value = f16::from_le_bytes([0x40, 0x4A]); |
186 | /// assert_eq!(value, f16::from_f32(12.5)); |
187 | /// ``` |
188 | #[inline ] |
189 | #[must_use ] |
190 | pub const fn from_le_bytes(bytes: [u8; 2]) -> f16 { |
191 | f16::from_bits(u16::from_le_bytes(bytes)) |
192 | } |
193 | |
194 | /// Creates a floating point value from its representation as a byte array in big endian. |
195 | /// |
196 | /// # Examples |
197 | /// |
198 | /// ```rust |
199 | /// # use half::prelude::*; |
200 | /// let value = f16::from_be_bytes([0x4A, 0x40]); |
201 | /// assert_eq!(value, f16::from_f32(12.5)); |
202 | /// ``` |
203 | #[inline ] |
204 | #[must_use ] |
205 | pub const fn from_be_bytes(bytes: [u8; 2]) -> f16 { |
206 | f16::from_bits(u16::from_be_bytes(bytes)) |
207 | } |
208 | |
209 | /// Creates a floating point value from its representation as a byte array in native endian. |
210 | /// |
211 | /// As the target platform's native endianness is used, portable code likely wants to use |
212 | /// [`from_be_bytes`][Self::from_be_bytes] or [`from_le_bytes`][Self::from_le_bytes], as |
213 | /// appropriate instead. |
214 | /// |
215 | /// # Examples |
216 | /// |
217 | /// ```rust |
218 | /// # use half::prelude::*; |
219 | /// let value = f16::from_ne_bytes(if cfg!(target_endian = "big" ) { |
220 | /// [0x4A, 0x40] |
221 | /// } else { |
222 | /// [0x40, 0x4A] |
223 | /// }); |
224 | /// assert_eq!(value, f16::from_f32(12.5)); |
225 | /// ``` |
226 | #[inline ] |
227 | #[must_use ] |
228 | pub const fn from_ne_bytes(bytes: [u8; 2]) -> f16 { |
229 | f16::from_bits(u16::from_ne_bytes(bytes)) |
230 | } |
231 | |
232 | /// Converts a [`f16`] value into a `f32` value. |
233 | /// |
234 | /// This conversion is lossless as all 16-bit floating point values can be represented exactly |
235 | /// in 32-bit floating point. |
236 | #[inline ] |
237 | #[must_use ] |
238 | pub fn to_f32(self) -> f32 { |
239 | arch::f16_to_f32(self.0) |
240 | } |
241 | |
242 | /// Converts a [`f16`] value into a `f32` value. |
243 | /// |
244 | /// This function is identical to [`to_f32`][Self::to_f32] except it never uses hardware |
245 | /// intrinsics, which allows it to be `const`. [`to_f32`][Self::to_f32] should be preferred |
246 | /// in any non-`const` context. |
247 | /// |
248 | /// This conversion is lossless as all 16-bit floating point values can be represented exactly |
249 | /// in 32-bit floating point. |
250 | #[inline ] |
251 | #[must_use ] |
252 | pub const fn to_f32_const(self) -> f32 { |
253 | arch::f16_to_f32_fallback(self.0) |
254 | } |
255 | |
256 | /// Converts a [`f16`] value into a `f64` value. |
257 | /// |
258 | /// This conversion is lossless as all 16-bit floating point values can be represented exactly |
259 | /// in 64-bit floating point. |
260 | #[inline ] |
261 | #[must_use ] |
262 | pub fn to_f64(self) -> f64 { |
263 | arch::f16_to_f64(self.0) |
264 | } |
265 | |
266 | /// Converts a [`f16`] value into a `f64` value. |
267 | /// |
268 | /// This function is identical to [`to_f64`][Self::to_f64] except it never uses hardware |
269 | /// intrinsics, which allows it to be `const`. [`to_f64`][Self::to_f64] should be preferred |
270 | /// in any non-`const` context. |
271 | /// |
272 | /// This conversion is lossless as all 16-bit floating point values can be represented exactly |
273 | /// in 64-bit floating point. |
274 | #[inline ] |
275 | #[must_use ] |
276 | pub const fn to_f64_const(self) -> f64 { |
277 | arch::f16_to_f64_fallback(self.0) |
278 | } |
279 | |
280 | /// Returns `true` if this value is `NaN` and `false` otherwise. |
281 | /// |
282 | /// # Examples |
283 | /// |
284 | /// ```rust |
285 | /// # use half::prelude::*; |
286 | /// |
287 | /// let nan = f16::NAN; |
288 | /// let f = f16::from_f32(7.0_f32); |
289 | /// |
290 | /// assert!(nan.is_nan()); |
291 | /// assert!(!f.is_nan()); |
292 | /// ``` |
293 | #[inline ] |
294 | #[must_use ] |
295 | pub const fn is_nan(self) -> bool { |
296 | self.0 & 0x7FFFu16 > 0x7C00u16 |
297 | } |
298 | |
299 | /// Returns `true` if this value is ±∞ and `false`. |
300 | /// otherwise. |
301 | /// |
302 | /// # Examples |
303 | /// |
304 | /// ```rust |
305 | /// # use half::prelude::*; |
306 | /// |
307 | /// let f = f16::from_f32(7.0f32); |
308 | /// let inf = f16::INFINITY; |
309 | /// let neg_inf = f16::NEG_INFINITY; |
310 | /// let nan = f16::NAN; |
311 | /// |
312 | /// assert!(!f.is_infinite()); |
313 | /// assert!(!nan.is_infinite()); |
314 | /// |
315 | /// assert!(inf.is_infinite()); |
316 | /// assert!(neg_inf.is_infinite()); |
317 | /// ``` |
318 | #[inline ] |
319 | #[must_use ] |
320 | pub const fn is_infinite(self) -> bool { |
321 | self.0 & 0x7FFFu16 == 0x7C00u16 |
322 | } |
323 | |
324 | /// Returns `true` if this number is neither infinite nor `NaN`. |
325 | /// |
326 | /// # Examples |
327 | /// |
328 | /// ```rust |
329 | /// # use half::prelude::*; |
330 | /// |
331 | /// let f = f16::from_f32(7.0f32); |
332 | /// let inf = f16::INFINITY; |
333 | /// let neg_inf = f16::NEG_INFINITY; |
334 | /// let nan = f16::NAN; |
335 | /// |
336 | /// assert!(f.is_finite()); |
337 | /// |
338 | /// assert!(!nan.is_finite()); |
339 | /// assert!(!inf.is_finite()); |
340 | /// assert!(!neg_inf.is_finite()); |
341 | /// ``` |
342 | #[inline ] |
343 | #[must_use ] |
344 | pub const fn is_finite(self) -> bool { |
345 | self.0 & 0x7C00u16 != 0x7C00u16 |
346 | } |
347 | |
348 | /// Returns `true` if the number is neither zero, infinite, subnormal, or `NaN`. |
349 | /// |
350 | /// # Examples |
351 | /// |
352 | /// ```rust |
353 | /// # use half::prelude::*; |
354 | /// |
355 | /// let min = f16::MIN_POSITIVE; |
356 | /// let max = f16::MAX; |
357 | /// let lower_than_min = f16::from_f32(1.0e-10_f32); |
358 | /// let zero = f16::from_f32(0.0_f32); |
359 | /// |
360 | /// assert!(min.is_normal()); |
361 | /// assert!(max.is_normal()); |
362 | /// |
363 | /// assert!(!zero.is_normal()); |
364 | /// assert!(!f16::NAN.is_normal()); |
365 | /// assert!(!f16::INFINITY.is_normal()); |
366 | /// // Values between `0` and `min` are Subnormal. |
367 | /// assert!(!lower_than_min.is_normal()); |
368 | /// ``` |
369 | #[inline ] |
370 | #[must_use ] |
371 | pub const fn is_normal(self) -> bool { |
372 | let exp = self.0 & 0x7C00u16; |
373 | exp != 0x7C00u16 && exp != 0 |
374 | } |
375 | |
376 | /// Returns the floating point category of the number. |
377 | /// |
378 | /// If only one property is going to be tested, it is generally faster to use the specific |
379 | /// predicate instead. |
380 | /// |
381 | /// # Examples |
382 | /// |
383 | /// ```rust |
384 | /// use std::num::FpCategory; |
385 | /// # use half::prelude::*; |
386 | /// |
387 | /// let num = f16::from_f32(12.4_f32); |
388 | /// let inf = f16::INFINITY; |
389 | /// |
390 | /// assert_eq!(num.classify(), FpCategory::Normal); |
391 | /// assert_eq!(inf.classify(), FpCategory::Infinite); |
392 | /// ``` |
393 | #[must_use ] |
394 | pub const fn classify(self) -> FpCategory { |
395 | let exp = self.0 & 0x7C00u16; |
396 | let man = self.0 & 0x03FFu16; |
397 | match (exp, man) { |
398 | (0, 0) => FpCategory::Zero, |
399 | (0, _) => FpCategory::Subnormal, |
400 | (0x7C00u16, 0) => FpCategory::Infinite, |
401 | (0x7C00u16, _) => FpCategory::Nan, |
402 | _ => FpCategory::Normal, |
403 | } |
404 | } |
405 | |
406 | /// Returns a number that represents the sign of `self`. |
407 | /// |
408 | /// * `1.0` if the number is positive, `+0.0` or [`INFINITY`][f16::INFINITY] |
409 | /// * `-1.0` if the number is negative, `-0.0` or [`NEG_INFINITY`][f16::NEG_INFINITY] |
410 | /// * [`NAN`][f16::NAN] if the number is `NaN` |
411 | /// |
412 | /// # Examples |
413 | /// |
414 | /// ```rust |
415 | /// # use half::prelude::*; |
416 | /// |
417 | /// let f = f16::from_f32(3.5_f32); |
418 | /// |
419 | /// assert_eq!(f.signum(), f16::from_f32(1.0)); |
420 | /// assert_eq!(f16::NEG_INFINITY.signum(), f16::from_f32(-1.0)); |
421 | /// |
422 | /// assert!(f16::NAN.signum().is_nan()); |
423 | /// ``` |
424 | #[must_use ] |
425 | pub const fn signum(self) -> f16 { |
426 | if self.is_nan() { |
427 | self |
428 | } else if self.0 & 0x8000u16 != 0 { |
429 | Self::NEG_ONE |
430 | } else { |
431 | Self::ONE |
432 | } |
433 | } |
434 | |
435 | /// Returns `true` if and only if `self` has a positive sign, including `+0.0`, `NaNs` with a |
436 | /// positive sign bit and +∞. |
437 | /// |
438 | /// # Examples |
439 | /// |
440 | /// ```rust |
441 | /// # use half::prelude::*; |
442 | /// |
443 | /// let nan = f16::NAN; |
444 | /// let f = f16::from_f32(7.0_f32); |
445 | /// let g = f16::from_f32(-7.0_f32); |
446 | /// |
447 | /// assert!(f.is_sign_positive()); |
448 | /// assert!(!g.is_sign_positive()); |
449 | /// // `NaN` can be either positive or negative |
450 | /// assert!(nan.is_sign_positive() != nan.is_sign_negative()); |
451 | /// ``` |
452 | #[inline ] |
453 | #[must_use ] |
454 | pub const fn is_sign_positive(self) -> bool { |
455 | self.0 & 0x8000u16 == 0 |
456 | } |
457 | |
458 | /// Returns `true` if and only if `self` has a negative sign, including `-0.0`, `NaNs` with a |
459 | /// negative sign bit and −∞. |
460 | /// |
461 | /// # Examples |
462 | /// |
463 | /// ```rust |
464 | /// # use half::prelude::*; |
465 | /// |
466 | /// let nan = f16::NAN; |
467 | /// let f = f16::from_f32(7.0f32); |
468 | /// let g = f16::from_f32(-7.0f32); |
469 | /// |
470 | /// assert!(!f.is_sign_negative()); |
471 | /// assert!(g.is_sign_negative()); |
472 | /// // `NaN` can be either positive or negative |
473 | /// assert!(nan.is_sign_positive() != nan.is_sign_negative()); |
474 | /// ``` |
475 | #[inline ] |
476 | #[must_use ] |
477 | pub const fn is_sign_negative(self) -> bool { |
478 | self.0 & 0x8000u16 != 0 |
479 | } |
480 | |
481 | /// Returns a number composed of the magnitude of `self` and the sign of `sign`. |
482 | /// |
483 | /// Equal to `self` if the sign of `self` and `sign` are the same, otherwise equal to `-self`. |
484 | /// If `self` is NaN, then NaN with the sign of `sign` is returned. |
485 | /// |
486 | /// # Examples |
487 | /// |
488 | /// ``` |
489 | /// # use half::prelude::*; |
490 | /// let f = f16::from_f32(3.5); |
491 | /// |
492 | /// assert_eq!(f.copysign(f16::from_f32(0.42)), f16::from_f32(3.5)); |
493 | /// assert_eq!(f.copysign(f16::from_f32(-0.42)), f16::from_f32(-3.5)); |
494 | /// assert_eq!((-f).copysign(f16::from_f32(0.42)), f16::from_f32(3.5)); |
495 | /// assert_eq!((-f).copysign(f16::from_f32(-0.42)), f16::from_f32(-3.5)); |
496 | /// |
497 | /// assert!(f16::NAN.copysign(f16::from_f32(1.0)).is_nan()); |
498 | /// ``` |
499 | #[inline ] |
500 | #[must_use ] |
501 | pub const fn copysign(self, sign: f16) -> f16 { |
502 | f16((sign.0 & 0x8000u16) | (self.0 & 0x7FFFu16)) |
503 | } |
504 | |
505 | /// Returns the maximum of the two numbers. |
506 | /// |
507 | /// If one of the arguments is NaN, then the other argument is returned. |
508 | /// |
509 | /// # Examples |
510 | /// |
511 | /// ``` |
512 | /// # use half::prelude::*; |
513 | /// let x = f16::from_f32(1.0); |
514 | /// let y = f16::from_f32(2.0); |
515 | /// |
516 | /// assert_eq!(x.max(y), y); |
517 | /// ``` |
518 | #[inline ] |
519 | #[must_use ] |
520 | pub fn max(self, other: f16) -> f16 { |
521 | if other > self && !other.is_nan() { |
522 | other |
523 | } else { |
524 | self |
525 | } |
526 | } |
527 | |
528 | /// Returns the minimum of the two numbers. |
529 | /// |
530 | /// If one of the arguments is NaN, then the other argument is returned. |
531 | /// |
532 | /// # Examples |
533 | /// |
534 | /// ``` |
535 | /// # use half::prelude::*; |
536 | /// let x = f16::from_f32(1.0); |
537 | /// let y = f16::from_f32(2.0); |
538 | /// |
539 | /// assert_eq!(x.min(y), x); |
540 | /// ``` |
541 | #[inline ] |
542 | #[must_use ] |
543 | pub fn min(self, other: f16) -> f16 { |
544 | if other < self && !other.is_nan() { |
545 | other |
546 | } else { |
547 | self |
548 | } |
549 | } |
550 | |
551 | /// Restrict a value to a certain interval unless it is NaN. |
552 | /// |
553 | /// Returns `max` if `self` is greater than `max`, and `min` if `self` is less than `min`. |
554 | /// Otherwise this returns `self`. |
555 | /// |
556 | /// Note that this function returns NaN if the initial value was NaN as well. |
557 | /// |
558 | /// # Panics |
559 | /// Panics if `min > max`, `min` is NaN, or `max` is NaN. |
560 | /// |
561 | /// # Examples |
562 | /// |
563 | /// ``` |
564 | /// # use half::prelude::*; |
565 | /// assert!(f16::from_f32(-3.0).clamp(f16::from_f32(-2.0), f16::from_f32(1.0)) == f16::from_f32(-2.0)); |
566 | /// assert!(f16::from_f32(0.0).clamp(f16::from_f32(-2.0), f16::from_f32(1.0)) == f16::from_f32(0.0)); |
567 | /// assert!(f16::from_f32(2.0).clamp(f16::from_f32(-2.0), f16::from_f32(1.0)) == f16::from_f32(1.0)); |
568 | /// assert!(f16::NAN.clamp(f16::from_f32(-2.0), f16::from_f32(1.0)).is_nan()); |
569 | /// ``` |
570 | #[inline ] |
571 | #[must_use ] |
572 | pub fn clamp(self, min: f16, max: f16) -> f16 { |
573 | assert!(min <= max); |
574 | let mut x = self; |
575 | if x < min { |
576 | x = min; |
577 | } |
578 | if x > max { |
579 | x = max; |
580 | } |
581 | x |
582 | } |
583 | |
584 | /// Returns the ordering between `self` and `other`. |
585 | /// |
586 | /// Unlike the standard partial comparison between floating point numbers, |
587 | /// this comparison always produces an ordering in accordance to |
588 | /// the `totalOrder` predicate as defined in the IEEE 754 (2008 revision) |
589 | /// floating point standard. The values are ordered in the following sequence: |
590 | /// |
591 | /// - negative quiet NaN |
592 | /// - negative signaling NaN |
593 | /// - negative infinity |
594 | /// - negative numbers |
595 | /// - negative subnormal numbers |
596 | /// - negative zero |
597 | /// - positive zero |
598 | /// - positive subnormal numbers |
599 | /// - positive numbers |
600 | /// - positive infinity |
601 | /// - positive signaling NaN |
602 | /// - positive quiet NaN. |
603 | /// |
604 | /// The ordering established by this function does not always agree with the |
605 | /// [`PartialOrd`] and [`PartialEq`] implementations of `f16`. For example, |
606 | /// they consider negative and positive zero equal, while `total_cmp` |
607 | /// doesn't. |
608 | /// |
609 | /// The interpretation of the signaling NaN bit follows the definition in |
610 | /// the IEEE 754 standard, which may not match the interpretation by some of |
611 | /// the older, non-conformant (e.g. MIPS) hardware implementations. |
612 | /// |
613 | /// # Examples |
614 | /// ``` |
615 | /// # use half::f16; |
616 | /// let mut v: Vec<f16> = vec![]; |
617 | /// v.push(f16::ONE); |
618 | /// v.push(f16::INFINITY); |
619 | /// v.push(f16::NEG_INFINITY); |
620 | /// v.push(f16::NAN); |
621 | /// v.push(f16::MAX_SUBNORMAL); |
622 | /// v.push(-f16::MAX_SUBNORMAL); |
623 | /// v.push(f16::ZERO); |
624 | /// v.push(f16::NEG_ZERO); |
625 | /// v.push(f16::NEG_ONE); |
626 | /// v.push(f16::MIN_POSITIVE); |
627 | /// |
628 | /// v.sort_by(|a, b| a.total_cmp(&b)); |
629 | /// |
630 | /// assert!(v |
631 | /// .into_iter() |
632 | /// .zip( |
633 | /// [ |
634 | /// f16::NEG_INFINITY, |
635 | /// f16::NEG_ONE, |
636 | /// -f16::MAX_SUBNORMAL, |
637 | /// f16::NEG_ZERO, |
638 | /// f16::ZERO, |
639 | /// f16::MAX_SUBNORMAL, |
640 | /// f16::MIN_POSITIVE, |
641 | /// f16::ONE, |
642 | /// f16::INFINITY, |
643 | /// f16::NAN |
644 | /// ] |
645 | /// .iter() |
646 | /// ) |
647 | /// .all(|(a, b)| a.to_bits() == b.to_bits())); |
648 | /// ``` |
649 | // Implementation based on: https://doc.rust-lang.org/std/primitive.f32.html#method.total_cmp |
650 | #[inline ] |
651 | #[must_use ] |
652 | pub fn total_cmp(&self, other: &Self) -> Ordering { |
653 | let mut left = self.to_bits() as i16; |
654 | let mut right = other.to_bits() as i16; |
655 | left ^= (((left >> 15) as u16) >> 1) as i16; |
656 | right ^= (((right >> 15) as u16) >> 1) as i16; |
657 | left.cmp(&right) |
658 | } |
659 | |
660 | /// Alternate serialize adapter for serializing as a float. |
661 | /// |
662 | /// By default, [`f16`] serializes as a newtype of [`u16`]. This is an alternate serialize |
663 | /// implementation that serializes as an [`f32`] value. It is designed for use with |
664 | /// `serialize_with` serde attributes. Deserialization from `f32` values is already supported by |
665 | /// the default deserialize implementation. |
666 | /// |
667 | /// # Examples |
668 | /// |
669 | /// A demonstration on how to use this adapater: |
670 | /// |
671 | /// ``` |
672 | /// use serde::{Serialize, Deserialize}; |
673 | /// use half::f16; |
674 | /// |
675 | /// #[derive(Serialize, Deserialize)] |
676 | /// struct MyStruct { |
677 | /// #[serde(serialize_with = "f16::serialize_as_f32")] |
678 | /// value: f16 // Will be serialized as f32 instead of u16 |
679 | /// } |
680 | /// ``` |
681 | #[cfg (feature = "serde" )] |
682 | pub fn serialize_as_f32<S: serde::Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> { |
683 | serializer.serialize_f32(self.to_f32()) |
684 | } |
685 | |
686 | /// Alternate serialize adapter for serializing as a string. |
687 | /// |
688 | /// By default, [`f16`] serializes as a newtype of [`u16`]. This is an alternate serialize |
689 | /// implementation that serializes as a string value. It is designed for use with |
690 | /// `serialize_with` serde attributes. Deserialization from string values is already supported |
691 | /// by the default deserialize implementation. |
692 | /// |
693 | /// # Examples |
694 | /// |
695 | /// A demonstration on how to use this adapater: |
696 | /// |
697 | /// ``` |
698 | /// use serde::{Serialize, Deserialize}; |
699 | /// use half::f16; |
700 | /// |
701 | /// #[derive(Serialize, Deserialize)] |
702 | /// struct MyStruct { |
703 | /// #[serde(serialize_with = "f16::serialize_as_string")] |
704 | /// value: f16 // Will be serialized as a string instead of u16 |
705 | /// } |
706 | /// ``` |
707 | #[cfg (all(feature = "serde" , feature = "alloc" ))] |
708 | pub fn serialize_as_string<S: serde::Serializer>( |
709 | &self, |
710 | serializer: S, |
711 | ) -> Result<S::Ok, S::Error> { |
712 | serializer.serialize_str(&self.to_string()) |
713 | } |
714 | |
715 | /// Approximate number of [`f16`] significant digits in base 10 |
716 | pub const DIGITS: u32 = 3; |
717 | /// [`f16`] |
718 | /// [machine epsilon](https://en.wikipedia.org/wiki/Machine_epsilon) value |
719 | /// |
720 | /// This is the difference between 1.0 and the next largest representable number. |
721 | pub const EPSILON: f16 = f16(0x1400u16); |
722 | /// [`f16`] positive Infinity (+∞) |
723 | pub const INFINITY: f16 = f16(0x7C00u16); |
724 | /// Number of [`f16`] significant digits in base 2 |
725 | pub const MANTISSA_DIGITS: u32 = 11; |
726 | /// Largest finite [`f16`] value |
727 | pub const MAX: f16 = f16(0x7BFF); |
728 | /// Maximum possible [`f16`] power of 10 exponent |
729 | pub const MAX_10_EXP: i32 = 4; |
730 | /// Maximum possible [`f16`] power of 2 exponent |
731 | pub const MAX_EXP: i32 = 16; |
732 | /// Smallest finite [`f16`] value |
733 | pub const MIN: f16 = f16(0xFBFF); |
734 | /// Minimum possible normal [`f16`] power of 10 exponent |
735 | pub const MIN_10_EXP: i32 = -4; |
736 | /// One greater than the minimum possible normal [`f16`] power of 2 exponent |
737 | pub const MIN_EXP: i32 = -13; |
738 | /// Smallest positive normal [`f16`] value |
739 | pub const MIN_POSITIVE: f16 = f16(0x0400u16); |
740 | /// [`f16`] Not a Number (NaN) |
741 | pub const NAN: f16 = f16(0x7E00u16); |
742 | /// [`f16`] negative infinity (-∞) |
743 | pub const NEG_INFINITY: f16 = f16(0xFC00u16); |
744 | /// The radix or base of the internal representation of [`f16`] |
745 | pub const RADIX: u32 = 2; |
746 | |
747 | /// Minimum positive subnormal [`f16`] value |
748 | pub const MIN_POSITIVE_SUBNORMAL: f16 = f16(0x0001u16); |
749 | /// Maximum subnormal [`f16`] value |
750 | pub const MAX_SUBNORMAL: f16 = f16(0x03FFu16); |
751 | |
752 | /// [`f16`] 1 |
753 | pub const ONE: f16 = f16(0x3C00u16); |
754 | /// [`f16`] 0 |
755 | pub const ZERO: f16 = f16(0x0000u16); |
756 | /// [`f16`] -0 |
757 | pub const NEG_ZERO: f16 = f16(0x8000u16); |
758 | /// [`f16`] -1 |
759 | pub const NEG_ONE: f16 = f16(0xBC00u16); |
760 | |
761 | /// [`f16`] Euler's number (ℯ) |
762 | pub const E: f16 = f16(0x4170u16); |
763 | /// [`f16`] Archimedes' constant (π) |
764 | pub const PI: f16 = f16(0x4248u16); |
765 | /// [`f16`] 1/π |
766 | pub const FRAC_1_PI: f16 = f16(0x3518u16); |
767 | /// [`f16`] 1/√2 |
768 | pub const FRAC_1_SQRT_2: f16 = f16(0x39A8u16); |
769 | /// [`f16`] 2/π |
770 | pub const FRAC_2_PI: f16 = f16(0x3918u16); |
771 | /// [`f16`] 2/√π |
772 | pub const FRAC_2_SQRT_PI: f16 = f16(0x3C83u16); |
773 | /// [`f16`] π/2 |
774 | pub const FRAC_PI_2: f16 = f16(0x3E48u16); |
775 | /// [`f16`] π/3 |
776 | pub const FRAC_PI_3: f16 = f16(0x3C30u16); |
777 | /// [`f16`] π/4 |
778 | pub const FRAC_PI_4: f16 = f16(0x3A48u16); |
779 | /// [`f16`] π/6 |
780 | pub const FRAC_PI_6: f16 = f16(0x3830u16); |
781 | /// [`f16`] π/8 |
782 | pub const FRAC_PI_8: f16 = f16(0x3648u16); |
783 | /// [`f16`] 𝗅𝗇 10 |
784 | pub const LN_10: f16 = f16(0x409Bu16); |
785 | /// [`f16`] 𝗅𝗇 2 |
786 | pub const LN_2: f16 = f16(0x398Cu16); |
787 | /// [`f16`] 𝗅𝗈𝗀₁₀ℯ |
788 | pub const LOG10_E: f16 = f16(0x36F3u16); |
789 | /// [`f16`] 𝗅𝗈𝗀₁₀2 |
790 | pub const LOG10_2: f16 = f16(0x34D1u16); |
791 | /// [`f16`] 𝗅𝗈𝗀₂ℯ |
792 | pub const LOG2_E: f16 = f16(0x3DC5u16); |
793 | /// [`f16`] 𝗅𝗈𝗀₂10 |
794 | pub const LOG2_10: f16 = f16(0x42A5u16); |
795 | /// [`f16`] √2 |
796 | pub const SQRT_2: f16 = f16(0x3DA8u16); |
797 | } |
798 | |
799 | impl From<f16> for f32 { |
800 | #[inline ] |
801 | fn from(x: f16) -> f32 { |
802 | x.to_f32() |
803 | } |
804 | } |
805 | |
806 | impl From<f16> for f64 { |
807 | #[inline ] |
808 | fn from(x: f16) -> f64 { |
809 | x.to_f64() |
810 | } |
811 | } |
812 | |
813 | impl From<i8> for f16 { |
814 | #[inline ] |
815 | fn from(x: i8) -> f16 { |
816 | // Convert to f32, then to f16 |
817 | f16::from_f32(f32::from(x)) |
818 | } |
819 | } |
820 | |
821 | impl From<u8> for f16 { |
822 | #[inline ] |
823 | fn from(x: u8) -> f16 { |
824 | // Convert to f32, then to f16 |
825 | f16::from_f32(f32::from(x)) |
826 | } |
827 | } |
828 | |
829 | impl PartialEq for f16 { |
830 | fn eq(&self, other: &f16) -> bool { |
831 | if self.is_nan() || other.is_nan() { |
832 | false |
833 | } else { |
834 | (self.0 == other.0) || ((self.0 | other.0) & 0x7FFFu16 == 0) |
835 | } |
836 | } |
837 | } |
838 | |
839 | impl PartialOrd for f16 { |
840 | fn partial_cmp(&self, other: &f16) -> Option<Ordering> { |
841 | if self.is_nan() || other.is_nan() { |
842 | None |
843 | } else { |
844 | let neg = self.0 & 0x8000u16 != 0; |
845 | let other_neg = other.0 & 0x8000u16 != 0; |
846 | match (neg, other_neg) { |
847 | (false, false) => Some(self.0.cmp(&other.0)), |
848 | (false, true) => { |
849 | if (self.0 | other.0) & 0x7FFFu16 == 0 { |
850 | Some(Ordering::Equal) |
851 | } else { |
852 | Some(Ordering::Greater) |
853 | } |
854 | } |
855 | (true, false) => { |
856 | if (self.0 | other.0) & 0x7FFFu16 == 0 { |
857 | Some(Ordering::Equal) |
858 | } else { |
859 | Some(Ordering::Less) |
860 | } |
861 | } |
862 | (true, true) => Some(other.0.cmp(&self.0)), |
863 | } |
864 | } |
865 | } |
866 | |
867 | fn lt(&self, other: &f16) -> bool { |
868 | if self.is_nan() || other.is_nan() { |
869 | false |
870 | } else { |
871 | let neg = self.0 & 0x8000u16 != 0; |
872 | let other_neg = other.0 & 0x8000u16 != 0; |
873 | match (neg, other_neg) { |
874 | (false, false) => self.0 < other.0, |
875 | (false, true) => false, |
876 | (true, false) => (self.0 | other.0) & 0x7FFFu16 != 0, |
877 | (true, true) => self.0 > other.0, |
878 | } |
879 | } |
880 | } |
881 | |
882 | fn le(&self, other: &f16) -> bool { |
883 | if self.is_nan() || other.is_nan() { |
884 | false |
885 | } else { |
886 | let neg = self.0 & 0x8000u16 != 0; |
887 | let other_neg = other.0 & 0x8000u16 != 0; |
888 | match (neg, other_neg) { |
889 | (false, false) => self.0 <= other.0, |
890 | (false, true) => (self.0 | other.0) & 0x7FFFu16 == 0, |
891 | (true, false) => true, |
892 | (true, true) => self.0 >= other.0, |
893 | } |
894 | } |
895 | } |
896 | |
897 | fn gt(&self, other: &f16) -> bool { |
898 | if self.is_nan() || other.is_nan() { |
899 | false |
900 | } else { |
901 | let neg = self.0 & 0x8000u16 != 0; |
902 | let other_neg = other.0 & 0x8000u16 != 0; |
903 | match (neg, other_neg) { |
904 | (false, false) => self.0 > other.0, |
905 | (false, true) => (self.0 | other.0) & 0x7FFFu16 != 0, |
906 | (true, false) => false, |
907 | (true, true) => self.0 < other.0, |
908 | } |
909 | } |
910 | } |
911 | |
912 | fn ge(&self, other: &f16) -> bool { |
913 | if self.is_nan() || other.is_nan() { |
914 | false |
915 | } else { |
916 | let neg = self.0 & 0x8000u16 != 0; |
917 | let other_neg = other.0 & 0x8000u16 != 0; |
918 | match (neg, other_neg) { |
919 | (false, false) => self.0 >= other.0, |
920 | (false, true) => true, |
921 | (true, false) => (self.0 | other.0) & 0x7FFFu16 == 0, |
922 | (true, true) => self.0 <= other.0, |
923 | } |
924 | } |
925 | } |
926 | } |
927 | |
928 | #[cfg (not(target_arch = "spirv" ))] |
929 | impl FromStr for f16 { |
930 | type Err = ParseFloatError; |
931 | fn from_str(src: &str) -> Result<f16, ParseFloatError> { |
932 | f32::from_str(src).map(op:f16::from_f32) |
933 | } |
934 | } |
935 | |
936 | #[cfg (not(target_arch = "spirv" ))] |
937 | impl Debug for f16 { |
938 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
939 | Debug::fmt(&self.to_f32(), f) |
940 | } |
941 | } |
942 | |
943 | #[cfg (not(target_arch = "spirv" ))] |
944 | impl Display for f16 { |
945 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
946 | Display::fmt(&self.to_f32(), f) |
947 | } |
948 | } |
949 | |
950 | #[cfg (not(target_arch = "spirv" ))] |
951 | impl LowerExp for f16 { |
952 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
953 | write!(f, " {:e}" , self.to_f32()) |
954 | } |
955 | } |
956 | |
957 | #[cfg (not(target_arch = "spirv" ))] |
958 | impl UpperExp for f16 { |
959 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
960 | write!(f, " {:E}" , self.to_f32()) |
961 | } |
962 | } |
963 | |
964 | #[cfg (not(target_arch = "spirv" ))] |
965 | impl Binary for f16 { |
966 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
967 | write!(f, " {:b}" , self.0) |
968 | } |
969 | } |
970 | |
971 | #[cfg (not(target_arch = "spirv" ))] |
972 | impl Octal for f16 { |
973 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
974 | write!(f, " {:o}" , self.0) |
975 | } |
976 | } |
977 | |
978 | #[cfg (not(target_arch = "spirv" ))] |
979 | impl LowerHex for f16 { |
980 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
981 | write!(f, " {:x}" , self.0) |
982 | } |
983 | } |
984 | |
985 | #[cfg (not(target_arch = "spirv" ))] |
986 | impl UpperHex for f16 { |
987 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
988 | write!(f, " {:X}" , self.0) |
989 | } |
990 | } |
991 | |
992 | impl Neg for f16 { |
993 | type Output = Self; |
994 | |
995 | #[inline ] |
996 | fn neg(self) -> Self::Output { |
997 | Self(self.0 ^ 0x8000) |
998 | } |
999 | } |
1000 | |
1001 | impl Neg for &f16 { |
1002 | type Output = <f16 as Neg>::Output; |
1003 | |
1004 | #[inline ] |
1005 | fn neg(self) -> Self::Output { |
1006 | Neg::neg(*self) |
1007 | } |
1008 | } |
1009 | |
1010 | impl Add for f16 { |
1011 | type Output = Self; |
1012 | |
1013 | #[inline ] |
1014 | fn add(self, rhs: Self) -> Self::Output { |
1015 | f16(arch::add_f16(self.0, b:rhs.0)) |
1016 | } |
1017 | } |
1018 | |
1019 | impl Add<&f16> for f16 { |
1020 | type Output = <f16 as Add<f16>>::Output; |
1021 | |
1022 | #[inline ] |
1023 | fn add(self, rhs: &f16) -> Self::Output { |
1024 | self.add(*rhs) |
1025 | } |
1026 | } |
1027 | |
1028 | impl Add<&f16> for &f16 { |
1029 | type Output = <f16 as Add<f16>>::Output; |
1030 | |
1031 | #[inline ] |
1032 | fn add(self, rhs: &f16) -> Self::Output { |
1033 | (*self).add(*rhs) |
1034 | } |
1035 | } |
1036 | |
1037 | impl Add<f16> for &f16 { |
1038 | type Output = <f16 as Add<f16>>::Output; |
1039 | |
1040 | #[inline ] |
1041 | fn add(self, rhs: f16) -> Self::Output { |
1042 | (*self).add(rhs) |
1043 | } |
1044 | } |
1045 | |
1046 | impl AddAssign for f16 { |
1047 | #[inline ] |
1048 | fn add_assign(&mut self, rhs: Self) { |
1049 | *self = (*self).add(rhs); |
1050 | } |
1051 | } |
1052 | |
1053 | impl AddAssign<&f16> for f16 { |
1054 | #[inline ] |
1055 | fn add_assign(&mut self, rhs: &f16) { |
1056 | *self = (*self).add(rhs); |
1057 | } |
1058 | } |
1059 | |
1060 | impl Sub for f16 { |
1061 | type Output = Self; |
1062 | |
1063 | #[inline ] |
1064 | fn sub(self, rhs: Self) -> Self::Output { |
1065 | f16(arch::subtract_f16(self.0, b:rhs.0)) |
1066 | } |
1067 | } |
1068 | |
1069 | impl Sub<&f16> for f16 { |
1070 | type Output = <f16 as Sub<f16>>::Output; |
1071 | |
1072 | #[inline ] |
1073 | fn sub(self, rhs: &f16) -> Self::Output { |
1074 | self.sub(*rhs) |
1075 | } |
1076 | } |
1077 | |
1078 | impl Sub<&f16> for &f16 { |
1079 | type Output = <f16 as Sub<f16>>::Output; |
1080 | |
1081 | #[inline ] |
1082 | fn sub(self, rhs: &f16) -> Self::Output { |
1083 | (*self).sub(*rhs) |
1084 | } |
1085 | } |
1086 | |
1087 | impl Sub<f16> for &f16 { |
1088 | type Output = <f16 as Sub<f16>>::Output; |
1089 | |
1090 | #[inline ] |
1091 | fn sub(self, rhs: f16) -> Self::Output { |
1092 | (*self).sub(rhs) |
1093 | } |
1094 | } |
1095 | |
1096 | impl SubAssign for f16 { |
1097 | #[inline ] |
1098 | fn sub_assign(&mut self, rhs: Self) { |
1099 | *self = (*self).sub(rhs); |
1100 | } |
1101 | } |
1102 | |
1103 | impl SubAssign<&f16> for f16 { |
1104 | #[inline ] |
1105 | fn sub_assign(&mut self, rhs: &f16) { |
1106 | *self = (*self).sub(rhs); |
1107 | } |
1108 | } |
1109 | |
1110 | impl Mul for f16 { |
1111 | type Output = Self; |
1112 | |
1113 | #[inline ] |
1114 | fn mul(self, rhs: Self) -> Self::Output { |
1115 | f16(arch::multiply_f16(self.0, b:rhs.0)) |
1116 | } |
1117 | } |
1118 | |
1119 | impl Mul<&f16> for f16 { |
1120 | type Output = <f16 as Mul<f16>>::Output; |
1121 | |
1122 | #[inline ] |
1123 | fn mul(self, rhs: &f16) -> Self::Output { |
1124 | self.mul(*rhs) |
1125 | } |
1126 | } |
1127 | |
1128 | impl Mul<&f16> for &f16 { |
1129 | type Output = <f16 as Mul<f16>>::Output; |
1130 | |
1131 | #[inline ] |
1132 | fn mul(self, rhs: &f16) -> Self::Output { |
1133 | (*self).mul(*rhs) |
1134 | } |
1135 | } |
1136 | |
1137 | impl Mul<f16> for &f16 { |
1138 | type Output = <f16 as Mul<f16>>::Output; |
1139 | |
1140 | #[inline ] |
1141 | fn mul(self, rhs: f16) -> Self::Output { |
1142 | (*self).mul(rhs) |
1143 | } |
1144 | } |
1145 | |
1146 | impl MulAssign for f16 { |
1147 | #[inline ] |
1148 | fn mul_assign(&mut self, rhs: Self) { |
1149 | *self = (*self).mul(rhs); |
1150 | } |
1151 | } |
1152 | |
1153 | impl MulAssign<&f16> for f16 { |
1154 | #[inline ] |
1155 | fn mul_assign(&mut self, rhs: &f16) { |
1156 | *self = (*self).mul(rhs); |
1157 | } |
1158 | } |
1159 | |
1160 | impl Div for f16 { |
1161 | type Output = Self; |
1162 | |
1163 | #[inline ] |
1164 | fn div(self, rhs: Self) -> Self::Output { |
1165 | f16(arch::divide_f16(self.0, b:rhs.0)) |
1166 | } |
1167 | } |
1168 | |
1169 | impl Div<&f16> for f16 { |
1170 | type Output = <f16 as Div<f16>>::Output; |
1171 | |
1172 | #[inline ] |
1173 | fn div(self, rhs: &f16) -> Self::Output { |
1174 | self.div(*rhs) |
1175 | } |
1176 | } |
1177 | |
1178 | impl Div<&f16> for &f16 { |
1179 | type Output = <f16 as Div<f16>>::Output; |
1180 | |
1181 | #[inline ] |
1182 | fn div(self, rhs: &f16) -> Self::Output { |
1183 | (*self).div(*rhs) |
1184 | } |
1185 | } |
1186 | |
1187 | impl Div<f16> for &f16 { |
1188 | type Output = <f16 as Div<f16>>::Output; |
1189 | |
1190 | #[inline ] |
1191 | fn div(self, rhs: f16) -> Self::Output { |
1192 | (*self).div(rhs) |
1193 | } |
1194 | } |
1195 | |
1196 | impl DivAssign for f16 { |
1197 | #[inline ] |
1198 | fn div_assign(&mut self, rhs: Self) { |
1199 | *self = (*self).div(rhs); |
1200 | } |
1201 | } |
1202 | |
1203 | impl DivAssign<&f16> for f16 { |
1204 | #[inline ] |
1205 | fn div_assign(&mut self, rhs: &f16) { |
1206 | *self = (*self).div(rhs); |
1207 | } |
1208 | } |
1209 | |
1210 | impl Rem for f16 { |
1211 | type Output = Self; |
1212 | |
1213 | #[inline ] |
1214 | fn rem(self, rhs: Self) -> Self::Output { |
1215 | f16(arch::remainder_f16(self.0, b:rhs.0)) |
1216 | } |
1217 | } |
1218 | |
1219 | impl Rem<&f16> for f16 { |
1220 | type Output = <f16 as Rem<f16>>::Output; |
1221 | |
1222 | #[inline ] |
1223 | fn rem(self, rhs: &f16) -> Self::Output { |
1224 | self.rem(*rhs) |
1225 | } |
1226 | } |
1227 | |
1228 | impl Rem<&f16> for &f16 { |
1229 | type Output = <f16 as Rem<f16>>::Output; |
1230 | |
1231 | #[inline ] |
1232 | fn rem(self, rhs: &f16) -> Self::Output { |
1233 | (*self).rem(*rhs) |
1234 | } |
1235 | } |
1236 | |
1237 | impl Rem<f16> for &f16 { |
1238 | type Output = <f16 as Rem<f16>>::Output; |
1239 | |
1240 | #[inline ] |
1241 | fn rem(self, rhs: f16) -> Self::Output { |
1242 | (*self).rem(rhs) |
1243 | } |
1244 | } |
1245 | |
1246 | impl RemAssign for f16 { |
1247 | #[inline ] |
1248 | fn rem_assign(&mut self, rhs: Self) { |
1249 | *self = (*self).rem(rhs); |
1250 | } |
1251 | } |
1252 | |
1253 | impl RemAssign<&f16> for f16 { |
1254 | #[inline ] |
1255 | fn rem_assign(&mut self, rhs: &f16) { |
1256 | *self = (*self).rem(rhs); |
1257 | } |
1258 | } |
1259 | |
1260 | impl Product for f16 { |
1261 | #[inline ] |
1262 | fn product<I: Iterator<Item = Self>>(iter: I) -> Self { |
1263 | f16(arch::product_f16(iter:iter.map(|f: f16| f.to_bits()))) |
1264 | } |
1265 | } |
1266 | |
1267 | impl<'a> Product<&'a f16> for f16 { |
1268 | #[inline ] |
1269 | fn product<I: Iterator<Item = &'a f16>>(iter: I) -> Self { |
1270 | f16(arch::product_f16(iter:iter.map(|f: &f16| f.to_bits()))) |
1271 | } |
1272 | } |
1273 | |
1274 | impl Sum for f16 { |
1275 | #[inline ] |
1276 | fn sum<I: Iterator<Item = Self>>(iter: I) -> Self { |
1277 | f16(arch::sum_f16(iter:iter.map(|f: f16| f.to_bits()))) |
1278 | } |
1279 | } |
1280 | |
1281 | impl<'a> Sum<&'a f16> for f16 { |
1282 | #[inline ] |
1283 | fn sum<I: Iterator<Item = &'a f16>>(iter: I) -> Self { |
1284 | f16(arch::sum_f16(iter:iter.map(|f: &f16| f.to_bits()))) |
1285 | } |
1286 | } |
1287 | |
1288 | #[cfg (feature = "serde" )] |
1289 | struct Visitor; |
1290 | |
1291 | #[cfg (feature = "serde" )] |
1292 | impl<'de> Deserialize<'de> for f16 { |
1293 | fn deserialize<D>(deserializer: D) -> Result<f16, D::Error> |
1294 | where |
1295 | D: serde::de::Deserializer<'de>, |
1296 | { |
1297 | deserializer.deserialize_newtype_struct("f16" , Visitor) |
1298 | } |
1299 | } |
1300 | |
1301 | #[cfg (feature = "serde" )] |
1302 | impl<'de> serde::de::Visitor<'de> for Visitor { |
1303 | type Value = f16; |
1304 | |
1305 | fn expecting(&self, formatter: &mut core::fmt::Formatter) -> core::fmt::Result { |
1306 | write!(formatter, "tuple struct f16" ) |
1307 | } |
1308 | |
1309 | fn visit_newtype_struct<D>(self, deserializer: D) -> Result<Self::Value, D::Error> |
1310 | where |
1311 | D: serde::Deserializer<'de>, |
1312 | { |
1313 | Ok(f16(<u16 as Deserialize>::deserialize(deserializer)?)) |
1314 | } |
1315 | |
1316 | fn visit_str<E>(self, v: &str) -> Result<Self::Value, E> |
1317 | where |
1318 | E: serde::de::Error, |
1319 | { |
1320 | v.parse().map_err(|_| { |
1321 | serde::de::Error::invalid_value(serde::de::Unexpected::Str(v), &"a float string" ) |
1322 | }) |
1323 | } |
1324 | |
1325 | fn visit_f32<E>(self, v: f32) -> Result<Self::Value, E> |
1326 | where |
1327 | E: serde::de::Error, |
1328 | { |
1329 | Ok(f16::from_f32(v)) |
1330 | } |
1331 | |
1332 | fn visit_f64<E>(self, v: f64) -> Result<Self::Value, E> |
1333 | where |
1334 | E: serde::de::Error, |
1335 | { |
1336 | Ok(f16::from_f64(v)) |
1337 | } |
1338 | } |
1339 | |
1340 | #[allow ( |
1341 | clippy::cognitive_complexity, |
1342 | clippy::float_cmp, |
1343 | clippy::neg_cmp_op_on_partial_ord |
1344 | )] |
1345 | #[cfg (test)] |
1346 | mod test { |
1347 | use super::*; |
1348 | #[allow (unused_imports)] |
1349 | use core::cmp::Ordering; |
1350 | #[cfg (feature = "num-traits" )] |
1351 | use num_traits::{AsPrimitive, FromPrimitive, ToPrimitive}; |
1352 | use quickcheck_macros::quickcheck; |
1353 | |
1354 | #[cfg (feature = "num-traits" )] |
1355 | #[test ] |
1356 | fn as_primitive() { |
1357 | let two = f16::from_f32(2.0); |
1358 | assert_eq!(<i32 as AsPrimitive<f16>>::as_(2), two); |
1359 | assert_eq!(<f16 as AsPrimitive<i32>>::as_(two), 2); |
1360 | |
1361 | assert_eq!(<f32 as AsPrimitive<f16>>::as_(2.0), two); |
1362 | assert_eq!(<f16 as AsPrimitive<f32>>::as_(two), 2.0); |
1363 | |
1364 | assert_eq!(<f64 as AsPrimitive<f16>>::as_(2.0), two); |
1365 | assert_eq!(<f16 as AsPrimitive<f64>>::as_(two), 2.0); |
1366 | } |
1367 | |
1368 | #[cfg (feature = "num-traits" )] |
1369 | #[test ] |
1370 | fn to_primitive() { |
1371 | let two = f16::from_f32(2.0); |
1372 | assert_eq!(ToPrimitive::to_i32(&two).unwrap(), 2i32); |
1373 | assert_eq!(ToPrimitive::to_f32(&two).unwrap(), 2.0f32); |
1374 | assert_eq!(ToPrimitive::to_f64(&two).unwrap(), 2.0f64); |
1375 | } |
1376 | |
1377 | #[cfg (feature = "num-traits" )] |
1378 | #[test ] |
1379 | fn from_primitive() { |
1380 | let two = f16::from_f32(2.0); |
1381 | assert_eq!(<f16 as FromPrimitive>::from_i32(2).unwrap(), two); |
1382 | assert_eq!(<f16 as FromPrimitive>::from_f32(2.0).unwrap(), two); |
1383 | assert_eq!(<f16 as FromPrimitive>::from_f64(2.0).unwrap(), two); |
1384 | } |
1385 | |
1386 | #[test ] |
1387 | fn test_f16_consts() { |
1388 | // DIGITS |
1389 | let digits = ((f16::MANTISSA_DIGITS as f32 - 1.0) * 2f32.log10()).floor() as u32; |
1390 | assert_eq!(f16::DIGITS, digits); |
1391 | // sanity check to show test is good |
1392 | let digits32 = ((core::f32::MANTISSA_DIGITS as f32 - 1.0) * 2f32.log10()).floor() as u32; |
1393 | assert_eq!(core::f32::DIGITS, digits32); |
1394 | |
1395 | // EPSILON |
1396 | let one = f16::from_f32(1.0); |
1397 | let one_plus_epsilon = f16::from_bits(one.to_bits() + 1); |
1398 | let epsilon = f16::from_f32(one_plus_epsilon.to_f32() - 1.0); |
1399 | assert_eq!(f16::EPSILON, epsilon); |
1400 | // sanity check to show test is good |
1401 | let one_plus_epsilon32 = f32::from_bits(1.0f32.to_bits() + 1); |
1402 | let epsilon32 = one_plus_epsilon32 - 1f32; |
1403 | assert_eq!(core::f32::EPSILON, epsilon32); |
1404 | |
1405 | // MAX, MIN and MIN_POSITIVE |
1406 | let max = f16::from_bits(f16::INFINITY.to_bits() - 1); |
1407 | let min = f16::from_bits(f16::NEG_INFINITY.to_bits() - 1); |
1408 | let min_pos = f16::from_f32(2f32.powi(f16::MIN_EXP - 1)); |
1409 | assert_eq!(f16::MAX, max); |
1410 | assert_eq!(f16::MIN, min); |
1411 | assert_eq!(f16::MIN_POSITIVE, min_pos); |
1412 | // sanity check to show test is good |
1413 | let max32 = f32::from_bits(core::f32::INFINITY.to_bits() - 1); |
1414 | let min32 = f32::from_bits(core::f32::NEG_INFINITY.to_bits() - 1); |
1415 | let min_pos32 = 2f32.powi(core::f32::MIN_EXP - 1); |
1416 | assert_eq!(core::f32::MAX, max32); |
1417 | assert_eq!(core::f32::MIN, min32); |
1418 | assert_eq!(core::f32::MIN_POSITIVE, min_pos32); |
1419 | |
1420 | // MIN_10_EXP and MAX_10_EXP |
1421 | let ten_to_min = 10f32.powi(f16::MIN_10_EXP); |
1422 | assert!(ten_to_min / 10.0 < f16::MIN_POSITIVE.to_f32()); |
1423 | assert!(ten_to_min > f16::MIN_POSITIVE.to_f32()); |
1424 | let ten_to_max = 10f32.powi(f16::MAX_10_EXP); |
1425 | assert!(ten_to_max < f16::MAX.to_f32()); |
1426 | assert!(ten_to_max * 10.0 > f16::MAX.to_f32()); |
1427 | // sanity check to show test is good |
1428 | let ten_to_min32 = 10f64.powi(core::f32::MIN_10_EXP); |
1429 | assert!(ten_to_min32 / 10.0 < f64::from(core::f32::MIN_POSITIVE)); |
1430 | assert!(ten_to_min32 > f64::from(core::f32::MIN_POSITIVE)); |
1431 | let ten_to_max32 = 10f64.powi(core::f32::MAX_10_EXP); |
1432 | assert!(ten_to_max32 < f64::from(core::f32::MAX)); |
1433 | assert!(ten_to_max32 * 10.0 > f64::from(core::f32::MAX)); |
1434 | } |
1435 | |
1436 | #[test ] |
1437 | fn test_f16_consts_from_f32() { |
1438 | let one = f16::from_f32(1.0); |
1439 | let zero = f16::from_f32(0.0); |
1440 | let neg_zero = f16::from_f32(-0.0); |
1441 | let neg_one = f16::from_f32(-1.0); |
1442 | let inf = f16::from_f32(core::f32::INFINITY); |
1443 | let neg_inf = f16::from_f32(core::f32::NEG_INFINITY); |
1444 | let nan = f16::from_f32(core::f32::NAN); |
1445 | |
1446 | assert_eq!(f16::ONE, one); |
1447 | assert_eq!(f16::ZERO, zero); |
1448 | assert!(zero.is_sign_positive()); |
1449 | assert_eq!(f16::NEG_ZERO, neg_zero); |
1450 | assert!(neg_zero.is_sign_negative()); |
1451 | assert_eq!(f16::NEG_ONE, neg_one); |
1452 | assert!(neg_one.is_sign_negative()); |
1453 | assert_eq!(f16::INFINITY, inf); |
1454 | assert_eq!(f16::NEG_INFINITY, neg_inf); |
1455 | assert!(nan.is_nan()); |
1456 | assert!(f16::NAN.is_nan()); |
1457 | |
1458 | let e = f16::from_f32(core::f32::consts::E); |
1459 | let pi = f16::from_f32(core::f32::consts::PI); |
1460 | let frac_1_pi = f16::from_f32(core::f32::consts::FRAC_1_PI); |
1461 | let frac_1_sqrt_2 = f16::from_f32(core::f32::consts::FRAC_1_SQRT_2); |
1462 | let frac_2_pi = f16::from_f32(core::f32::consts::FRAC_2_PI); |
1463 | let frac_2_sqrt_pi = f16::from_f32(core::f32::consts::FRAC_2_SQRT_PI); |
1464 | let frac_pi_2 = f16::from_f32(core::f32::consts::FRAC_PI_2); |
1465 | let frac_pi_3 = f16::from_f32(core::f32::consts::FRAC_PI_3); |
1466 | let frac_pi_4 = f16::from_f32(core::f32::consts::FRAC_PI_4); |
1467 | let frac_pi_6 = f16::from_f32(core::f32::consts::FRAC_PI_6); |
1468 | let frac_pi_8 = f16::from_f32(core::f32::consts::FRAC_PI_8); |
1469 | let ln_10 = f16::from_f32(core::f32::consts::LN_10); |
1470 | let ln_2 = f16::from_f32(core::f32::consts::LN_2); |
1471 | let log10_e = f16::from_f32(core::f32::consts::LOG10_E); |
1472 | // core::f32::consts::LOG10_2 requires rustc 1.43.0 |
1473 | let log10_2 = f16::from_f32(2f32.log10()); |
1474 | let log2_e = f16::from_f32(core::f32::consts::LOG2_E); |
1475 | // core::f32::consts::LOG2_10 requires rustc 1.43.0 |
1476 | let log2_10 = f16::from_f32(10f32.log2()); |
1477 | let sqrt_2 = f16::from_f32(core::f32::consts::SQRT_2); |
1478 | |
1479 | assert_eq!(f16::E, e); |
1480 | assert_eq!(f16::PI, pi); |
1481 | assert_eq!(f16::FRAC_1_PI, frac_1_pi); |
1482 | assert_eq!(f16::FRAC_1_SQRT_2, frac_1_sqrt_2); |
1483 | assert_eq!(f16::FRAC_2_PI, frac_2_pi); |
1484 | assert_eq!(f16::FRAC_2_SQRT_PI, frac_2_sqrt_pi); |
1485 | assert_eq!(f16::FRAC_PI_2, frac_pi_2); |
1486 | assert_eq!(f16::FRAC_PI_3, frac_pi_3); |
1487 | assert_eq!(f16::FRAC_PI_4, frac_pi_4); |
1488 | assert_eq!(f16::FRAC_PI_6, frac_pi_6); |
1489 | assert_eq!(f16::FRAC_PI_8, frac_pi_8); |
1490 | assert_eq!(f16::LN_10, ln_10); |
1491 | assert_eq!(f16::LN_2, ln_2); |
1492 | assert_eq!(f16::LOG10_E, log10_e); |
1493 | assert_eq!(f16::LOG10_2, log10_2); |
1494 | assert_eq!(f16::LOG2_E, log2_e); |
1495 | assert_eq!(f16::LOG2_10, log2_10); |
1496 | assert_eq!(f16::SQRT_2, sqrt_2); |
1497 | } |
1498 | |
1499 | #[test ] |
1500 | fn test_f16_consts_from_f64() { |
1501 | let one = f16::from_f64(1.0); |
1502 | let zero = f16::from_f64(0.0); |
1503 | let neg_zero = f16::from_f64(-0.0); |
1504 | let inf = f16::from_f64(core::f64::INFINITY); |
1505 | let neg_inf = f16::from_f64(core::f64::NEG_INFINITY); |
1506 | let nan = f16::from_f64(core::f64::NAN); |
1507 | |
1508 | assert_eq!(f16::ONE, one); |
1509 | assert_eq!(f16::ZERO, zero); |
1510 | assert!(zero.is_sign_positive()); |
1511 | assert_eq!(f16::NEG_ZERO, neg_zero); |
1512 | assert!(neg_zero.is_sign_negative()); |
1513 | assert_eq!(f16::INFINITY, inf); |
1514 | assert_eq!(f16::NEG_INFINITY, neg_inf); |
1515 | assert!(nan.is_nan()); |
1516 | assert!(f16::NAN.is_nan()); |
1517 | |
1518 | let e = f16::from_f64(core::f64::consts::E); |
1519 | let pi = f16::from_f64(core::f64::consts::PI); |
1520 | let frac_1_pi = f16::from_f64(core::f64::consts::FRAC_1_PI); |
1521 | let frac_1_sqrt_2 = f16::from_f64(core::f64::consts::FRAC_1_SQRT_2); |
1522 | let frac_2_pi = f16::from_f64(core::f64::consts::FRAC_2_PI); |
1523 | let frac_2_sqrt_pi = f16::from_f64(core::f64::consts::FRAC_2_SQRT_PI); |
1524 | let frac_pi_2 = f16::from_f64(core::f64::consts::FRAC_PI_2); |
1525 | let frac_pi_3 = f16::from_f64(core::f64::consts::FRAC_PI_3); |
1526 | let frac_pi_4 = f16::from_f64(core::f64::consts::FRAC_PI_4); |
1527 | let frac_pi_6 = f16::from_f64(core::f64::consts::FRAC_PI_6); |
1528 | let frac_pi_8 = f16::from_f64(core::f64::consts::FRAC_PI_8); |
1529 | let ln_10 = f16::from_f64(core::f64::consts::LN_10); |
1530 | let ln_2 = f16::from_f64(core::f64::consts::LN_2); |
1531 | let log10_e = f16::from_f64(core::f64::consts::LOG10_E); |
1532 | // core::f64::consts::LOG10_2 requires rustc 1.43.0 |
1533 | let log10_2 = f16::from_f64(2f64.log10()); |
1534 | let log2_e = f16::from_f64(core::f64::consts::LOG2_E); |
1535 | // core::f64::consts::LOG2_10 requires rustc 1.43.0 |
1536 | let log2_10 = f16::from_f64(10f64.log2()); |
1537 | let sqrt_2 = f16::from_f64(core::f64::consts::SQRT_2); |
1538 | |
1539 | assert_eq!(f16::E, e); |
1540 | assert_eq!(f16::PI, pi); |
1541 | assert_eq!(f16::FRAC_1_PI, frac_1_pi); |
1542 | assert_eq!(f16::FRAC_1_SQRT_2, frac_1_sqrt_2); |
1543 | assert_eq!(f16::FRAC_2_PI, frac_2_pi); |
1544 | assert_eq!(f16::FRAC_2_SQRT_PI, frac_2_sqrt_pi); |
1545 | assert_eq!(f16::FRAC_PI_2, frac_pi_2); |
1546 | assert_eq!(f16::FRAC_PI_3, frac_pi_3); |
1547 | assert_eq!(f16::FRAC_PI_4, frac_pi_4); |
1548 | assert_eq!(f16::FRAC_PI_6, frac_pi_6); |
1549 | assert_eq!(f16::FRAC_PI_8, frac_pi_8); |
1550 | assert_eq!(f16::LN_10, ln_10); |
1551 | assert_eq!(f16::LN_2, ln_2); |
1552 | assert_eq!(f16::LOG10_E, log10_e); |
1553 | assert_eq!(f16::LOG10_2, log10_2); |
1554 | assert_eq!(f16::LOG2_E, log2_e); |
1555 | assert_eq!(f16::LOG2_10, log2_10); |
1556 | assert_eq!(f16::SQRT_2, sqrt_2); |
1557 | } |
1558 | |
1559 | #[test ] |
1560 | fn test_nan_conversion_to_smaller() { |
1561 | let nan64 = f64::from_bits(0x7FF0_0000_0000_0001u64); |
1562 | let neg_nan64 = f64::from_bits(0xFFF0_0000_0000_0001u64); |
1563 | let nan32 = f32::from_bits(0x7F80_0001u32); |
1564 | let neg_nan32 = f32::from_bits(0xFF80_0001u32); |
1565 | let nan32_from_64 = nan64 as f32; |
1566 | let neg_nan32_from_64 = neg_nan64 as f32; |
1567 | let nan16_from_64 = f16::from_f64(nan64); |
1568 | let neg_nan16_from_64 = f16::from_f64(neg_nan64); |
1569 | let nan16_from_32 = f16::from_f32(nan32); |
1570 | let neg_nan16_from_32 = f16::from_f32(neg_nan32); |
1571 | |
1572 | assert!(nan64.is_nan() && nan64.is_sign_positive()); |
1573 | assert!(neg_nan64.is_nan() && neg_nan64.is_sign_negative()); |
1574 | assert!(nan32.is_nan() && nan32.is_sign_positive()); |
1575 | assert!(neg_nan32.is_nan() && neg_nan32.is_sign_negative()); |
1576 | |
1577 | // f32/f64 NaN conversion sign is non-deterministic: https://github.com/starkat99/half-rs/issues/103 |
1578 | assert!(nan32_from_64.is_nan()); |
1579 | assert!(neg_nan32_from_64.is_nan()); |
1580 | assert!(nan16_from_64.is_nan()); |
1581 | assert!(neg_nan16_from_64.is_nan()); |
1582 | assert!(nan16_from_32.is_nan()); |
1583 | assert!(neg_nan16_from_32.is_nan()); |
1584 | } |
1585 | |
1586 | #[test ] |
1587 | fn test_nan_conversion_to_larger() { |
1588 | let nan16 = f16::from_bits(0x7C01u16); |
1589 | let neg_nan16 = f16::from_bits(0xFC01u16); |
1590 | let nan32 = f32::from_bits(0x7F80_0001u32); |
1591 | let neg_nan32 = f32::from_bits(0xFF80_0001u32); |
1592 | let nan32_from_16 = f32::from(nan16); |
1593 | let neg_nan32_from_16 = f32::from(neg_nan16); |
1594 | let nan64_from_16 = f64::from(nan16); |
1595 | let neg_nan64_from_16 = f64::from(neg_nan16); |
1596 | let nan64_from_32 = f64::from(nan32); |
1597 | let neg_nan64_from_32 = f64::from(neg_nan32); |
1598 | |
1599 | assert!(nan16.is_nan() && nan16.is_sign_positive()); |
1600 | assert!(neg_nan16.is_nan() && neg_nan16.is_sign_negative()); |
1601 | assert!(nan32.is_nan() && nan32.is_sign_positive()); |
1602 | assert!(neg_nan32.is_nan() && neg_nan32.is_sign_negative()); |
1603 | |
1604 | // f32/f64 NaN conversion sign is non-deterministic: https://github.com/starkat99/half-rs/issues/103 |
1605 | assert!(nan32_from_16.is_nan()); |
1606 | assert!(neg_nan32_from_16.is_nan()); |
1607 | assert!(nan64_from_16.is_nan()); |
1608 | assert!(neg_nan64_from_16.is_nan()); |
1609 | assert!(nan64_from_32.is_nan()); |
1610 | assert!(neg_nan64_from_32.is_nan()); |
1611 | } |
1612 | |
1613 | #[test ] |
1614 | fn test_f16_to_f32() { |
1615 | let f = f16::from_f32(7.0); |
1616 | assert_eq!(f.to_f32(), 7.0f32); |
1617 | |
1618 | // 7.1 is NOT exactly representable in 16-bit, it's rounded |
1619 | let f = f16::from_f32(7.1); |
1620 | let diff = (f.to_f32() - 7.1f32).abs(); |
1621 | // diff must be <= 4 * EPSILON, as 7 has two more significant bits than 1 |
1622 | assert!(diff <= 4.0 * f16::EPSILON.to_f32()); |
1623 | |
1624 | assert_eq!(f16::from_bits(0x0000_0001).to_f32(), 2.0f32.powi(-24)); |
1625 | assert_eq!(f16::from_bits(0x0000_0005).to_f32(), 5.0 * 2.0f32.powi(-24)); |
1626 | |
1627 | assert_eq!(f16::from_bits(0x0000_0001), f16::from_f32(2.0f32.powi(-24))); |
1628 | assert_eq!( |
1629 | f16::from_bits(0x0000_0005), |
1630 | f16::from_f32(5.0 * 2.0f32.powi(-24)) |
1631 | ); |
1632 | } |
1633 | |
1634 | #[test ] |
1635 | fn test_f16_to_f64() { |
1636 | let f = f16::from_f64(7.0); |
1637 | assert_eq!(f.to_f64(), 7.0f64); |
1638 | |
1639 | // 7.1 is NOT exactly representable in 16-bit, it's rounded |
1640 | let f = f16::from_f64(7.1); |
1641 | let diff = (f.to_f64() - 7.1f64).abs(); |
1642 | // diff must be <= 4 * EPSILON, as 7 has two more significant bits than 1 |
1643 | assert!(diff <= 4.0 * f16::EPSILON.to_f64()); |
1644 | |
1645 | assert_eq!(f16::from_bits(0x0000_0001).to_f64(), 2.0f64.powi(-24)); |
1646 | assert_eq!(f16::from_bits(0x0000_0005).to_f64(), 5.0 * 2.0f64.powi(-24)); |
1647 | |
1648 | assert_eq!(f16::from_bits(0x0000_0001), f16::from_f64(2.0f64.powi(-24))); |
1649 | assert_eq!( |
1650 | f16::from_bits(0x0000_0005), |
1651 | f16::from_f64(5.0 * 2.0f64.powi(-24)) |
1652 | ); |
1653 | } |
1654 | |
1655 | #[test ] |
1656 | fn test_comparisons() { |
1657 | let zero = f16::from_f64(0.0); |
1658 | let one = f16::from_f64(1.0); |
1659 | let neg_zero = f16::from_f64(-0.0); |
1660 | let neg_one = f16::from_f64(-1.0); |
1661 | |
1662 | assert_eq!(zero.partial_cmp(&neg_zero), Some(Ordering::Equal)); |
1663 | assert_eq!(neg_zero.partial_cmp(&zero), Some(Ordering::Equal)); |
1664 | assert!(zero == neg_zero); |
1665 | assert!(neg_zero == zero); |
1666 | assert!(!(zero != neg_zero)); |
1667 | assert!(!(neg_zero != zero)); |
1668 | assert!(!(zero < neg_zero)); |
1669 | assert!(!(neg_zero < zero)); |
1670 | assert!(zero <= neg_zero); |
1671 | assert!(neg_zero <= zero); |
1672 | assert!(!(zero > neg_zero)); |
1673 | assert!(!(neg_zero > zero)); |
1674 | assert!(zero >= neg_zero); |
1675 | assert!(neg_zero >= zero); |
1676 | |
1677 | assert_eq!(one.partial_cmp(&neg_zero), Some(Ordering::Greater)); |
1678 | assert_eq!(neg_zero.partial_cmp(&one), Some(Ordering::Less)); |
1679 | assert!(!(one == neg_zero)); |
1680 | assert!(!(neg_zero == one)); |
1681 | assert!(one != neg_zero); |
1682 | assert!(neg_zero != one); |
1683 | assert!(!(one < neg_zero)); |
1684 | assert!(neg_zero < one); |
1685 | assert!(!(one <= neg_zero)); |
1686 | assert!(neg_zero <= one); |
1687 | assert!(one > neg_zero); |
1688 | assert!(!(neg_zero > one)); |
1689 | assert!(one >= neg_zero); |
1690 | assert!(!(neg_zero >= one)); |
1691 | |
1692 | assert_eq!(one.partial_cmp(&neg_one), Some(Ordering::Greater)); |
1693 | assert_eq!(neg_one.partial_cmp(&one), Some(Ordering::Less)); |
1694 | assert!(!(one == neg_one)); |
1695 | assert!(!(neg_one == one)); |
1696 | assert!(one != neg_one); |
1697 | assert!(neg_one != one); |
1698 | assert!(!(one < neg_one)); |
1699 | assert!(neg_one < one); |
1700 | assert!(!(one <= neg_one)); |
1701 | assert!(neg_one <= one); |
1702 | assert!(one > neg_one); |
1703 | assert!(!(neg_one > one)); |
1704 | assert!(one >= neg_one); |
1705 | assert!(!(neg_one >= one)); |
1706 | } |
1707 | |
1708 | #[test ] |
1709 | #[allow (clippy::erasing_op, clippy::identity_op)] |
1710 | fn round_to_even_f32() { |
1711 | // smallest positive subnormal = 0b0.0000_0000_01 * 2^-14 = 2^-24 |
1712 | let min_sub = f16::from_bits(1); |
1713 | let min_sub_f = (-24f32).exp2(); |
1714 | assert_eq!(f16::from_f32(min_sub_f).to_bits(), min_sub.to_bits()); |
1715 | assert_eq!(f32::from(min_sub).to_bits(), min_sub_f.to_bits()); |
1716 | |
1717 | // 0.0000000000_011111 rounded to 0.0000000000 (< tie, no rounding) |
1718 | // 0.0000000000_100000 rounded to 0.0000000000 (tie and even, remains at even) |
1719 | // 0.0000000000_100001 rounded to 0.0000000001 (> tie, rounds up) |
1720 | assert_eq!( |
1721 | f16::from_f32(min_sub_f * 0.49).to_bits(), |
1722 | min_sub.to_bits() * 0 |
1723 | ); |
1724 | assert_eq!( |
1725 | f16::from_f32(min_sub_f * 0.50).to_bits(), |
1726 | min_sub.to_bits() * 0 |
1727 | ); |
1728 | assert_eq!( |
1729 | f16::from_f32(min_sub_f * 0.51).to_bits(), |
1730 | min_sub.to_bits() * 1 |
1731 | ); |
1732 | |
1733 | // 0.0000000001_011111 rounded to 0.0000000001 (< tie, no rounding) |
1734 | // 0.0000000001_100000 rounded to 0.0000000010 (tie and odd, rounds up to even) |
1735 | // 0.0000000001_100001 rounded to 0.0000000010 (> tie, rounds up) |
1736 | assert_eq!( |
1737 | f16::from_f32(min_sub_f * 1.49).to_bits(), |
1738 | min_sub.to_bits() * 1 |
1739 | ); |
1740 | assert_eq!( |
1741 | f16::from_f32(min_sub_f * 1.50).to_bits(), |
1742 | min_sub.to_bits() * 2 |
1743 | ); |
1744 | assert_eq!( |
1745 | f16::from_f32(min_sub_f * 1.51).to_bits(), |
1746 | min_sub.to_bits() * 2 |
1747 | ); |
1748 | |
1749 | // 0.0000000010_011111 rounded to 0.0000000010 (< tie, no rounding) |
1750 | // 0.0000000010_100000 rounded to 0.0000000010 (tie and even, remains at even) |
1751 | // 0.0000000010_100001 rounded to 0.0000000011 (> tie, rounds up) |
1752 | assert_eq!( |
1753 | f16::from_f32(min_sub_f * 2.49).to_bits(), |
1754 | min_sub.to_bits() * 2 |
1755 | ); |
1756 | assert_eq!( |
1757 | f16::from_f32(min_sub_f * 2.50).to_bits(), |
1758 | min_sub.to_bits() * 2 |
1759 | ); |
1760 | assert_eq!( |
1761 | f16::from_f32(min_sub_f * 2.51).to_bits(), |
1762 | min_sub.to_bits() * 3 |
1763 | ); |
1764 | |
1765 | assert_eq!( |
1766 | f16::from_f32(2000.49f32).to_bits(), |
1767 | f16::from_f32(2000.0).to_bits() |
1768 | ); |
1769 | assert_eq!( |
1770 | f16::from_f32(2000.50f32).to_bits(), |
1771 | f16::from_f32(2000.0).to_bits() |
1772 | ); |
1773 | assert_eq!( |
1774 | f16::from_f32(2000.51f32).to_bits(), |
1775 | f16::from_f32(2001.0).to_bits() |
1776 | ); |
1777 | assert_eq!( |
1778 | f16::from_f32(2001.49f32).to_bits(), |
1779 | f16::from_f32(2001.0).to_bits() |
1780 | ); |
1781 | assert_eq!( |
1782 | f16::from_f32(2001.50f32).to_bits(), |
1783 | f16::from_f32(2002.0).to_bits() |
1784 | ); |
1785 | assert_eq!( |
1786 | f16::from_f32(2001.51f32).to_bits(), |
1787 | f16::from_f32(2002.0).to_bits() |
1788 | ); |
1789 | assert_eq!( |
1790 | f16::from_f32(2002.49f32).to_bits(), |
1791 | f16::from_f32(2002.0).to_bits() |
1792 | ); |
1793 | assert_eq!( |
1794 | f16::from_f32(2002.50f32).to_bits(), |
1795 | f16::from_f32(2002.0).to_bits() |
1796 | ); |
1797 | assert_eq!( |
1798 | f16::from_f32(2002.51f32).to_bits(), |
1799 | f16::from_f32(2003.0).to_bits() |
1800 | ); |
1801 | } |
1802 | |
1803 | #[test ] |
1804 | #[allow (clippy::erasing_op, clippy::identity_op)] |
1805 | fn round_to_even_f64() { |
1806 | // smallest positive subnormal = 0b0.0000_0000_01 * 2^-14 = 2^-24 |
1807 | let min_sub = f16::from_bits(1); |
1808 | let min_sub_f = (-24f64).exp2(); |
1809 | assert_eq!(f16::from_f64(min_sub_f).to_bits(), min_sub.to_bits()); |
1810 | assert_eq!(f64::from(min_sub).to_bits(), min_sub_f.to_bits()); |
1811 | |
1812 | // 0.0000000000_011111 rounded to 0.0000000000 (< tie, no rounding) |
1813 | // 0.0000000000_100000 rounded to 0.0000000000 (tie and even, remains at even) |
1814 | // 0.0000000000_100001 rounded to 0.0000000001 (> tie, rounds up) |
1815 | assert_eq!( |
1816 | f16::from_f64(min_sub_f * 0.49).to_bits(), |
1817 | min_sub.to_bits() * 0 |
1818 | ); |
1819 | assert_eq!( |
1820 | f16::from_f64(min_sub_f * 0.50).to_bits(), |
1821 | min_sub.to_bits() * 0 |
1822 | ); |
1823 | assert_eq!( |
1824 | f16::from_f64(min_sub_f * 0.51).to_bits(), |
1825 | min_sub.to_bits() * 1 |
1826 | ); |
1827 | |
1828 | // 0.0000000001_011111 rounded to 0.0000000001 (< tie, no rounding) |
1829 | // 0.0000000001_100000 rounded to 0.0000000010 (tie and odd, rounds up to even) |
1830 | // 0.0000000001_100001 rounded to 0.0000000010 (> tie, rounds up) |
1831 | assert_eq!( |
1832 | f16::from_f64(min_sub_f * 1.49).to_bits(), |
1833 | min_sub.to_bits() * 1 |
1834 | ); |
1835 | assert_eq!( |
1836 | f16::from_f64(min_sub_f * 1.50).to_bits(), |
1837 | min_sub.to_bits() * 2 |
1838 | ); |
1839 | assert_eq!( |
1840 | f16::from_f64(min_sub_f * 1.51).to_bits(), |
1841 | min_sub.to_bits() * 2 |
1842 | ); |
1843 | |
1844 | // 0.0000000010_011111 rounded to 0.0000000010 (< tie, no rounding) |
1845 | // 0.0000000010_100000 rounded to 0.0000000010 (tie and even, remains at even) |
1846 | // 0.0000000010_100001 rounded to 0.0000000011 (> tie, rounds up) |
1847 | assert_eq!( |
1848 | f16::from_f64(min_sub_f * 2.49).to_bits(), |
1849 | min_sub.to_bits() * 2 |
1850 | ); |
1851 | assert_eq!( |
1852 | f16::from_f64(min_sub_f * 2.50).to_bits(), |
1853 | min_sub.to_bits() * 2 |
1854 | ); |
1855 | assert_eq!( |
1856 | f16::from_f64(min_sub_f * 2.51).to_bits(), |
1857 | min_sub.to_bits() * 3 |
1858 | ); |
1859 | |
1860 | assert_eq!( |
1861 | f16::from_f64(2000.49f64).to_bits(), |
1862 | f16::from_f64(2000.0).to_bits() |
1863 | ); |
1864 | assert_eq!( |
1865 | f16::from_f64(2000.50f64).to_bits(), |
1866 | f16::from_f64(2000.0).to_bits() |
1867 | ); |
1868 | assert_eq!( |
1869 | f16::from_f64(2000.51f64).to_bits(), |
1870 | f16::from_f64(2001.0).to_bits() |
1871 | ); |
1872 | assert_eq!( |
1873 | f16::from_f64(2001.49f64).to_bits(), |
1874 | f16::from_f64(2001.0).to_bits() |
1875 | ); |
1876 | assert_eq!( |
1877 | f16::from_f64(2001.50f64).to_bits(), |
1878 | f16::from_f64(2002.0).to_bits() |
1879 | ); |
1880 | assert_eq!( |
1881 | f16::from_f64(2001.51f64).to_bits(), |
1882 | f16::from_f64(2002.0).to_bits() |
1883 | ); |
1884 | assert_eq!( |
1885 | f16::from_f64(2002.49f64).to_bits(), |
1886 | f16::from_f64(2002.0).to_bits() |
1887 | ); |
1888 | assert_eq!( |
1889 | f16::from_f64(2002.50f64).to_bits(), |
1890 | f16::from_f64(2002.0).to_bits() |
1891 | ); |
1892 | assert_eq!( |
1893 | f16::from_f64(2002.51f64).to_bits(), |
1894 | f16::from_f64(2003.0).to_bits() |
1895 | ); |
1896 | } |
1897 | |
1898 | #[test ] |
1899 | fn arithmetic() { |
1900 | assert_eq!(f16::ONE + f16::ONE, f16::from_f32(2.)); |
1901 | assert_eq!(f16::ONE - f16::ONE, f16::ZERO); |
1902 | assert_eq!(f16::ONE * f16::ONE, f16::ONE); |
1903 | assert_eq!(f16::from_f32(2.) * f16::from_f32(2.), f16::from_f32(4.)); |
1904 | assert_eq!(f16::ONE / f16::ONE, f16::ONE); |
1905 | assert_eq!(f16::from_f32(4.) / f16::from_f32(2.), f16::from_f32(2.)); |
1906 | assert_eq!(f16::from_f32(4.) % f16::from_f32(3.), f16::from_f32(1.)); |
1907 | } |
1908 | |
1909 | #[cfg (feature = "std" )] |
1910 | #[test ] |
1911 | fn formatting() { |
1912 | let f = f16::from_f32(0.1152344); |
1913 | |
1914 | assert_eq!(format!(" {:.3}" , f), "0.115" ); |
1915 | assert_eq!(format!(" {:.4}" , f), "0.1152" ); |
1916 | assert_eq!(format!(" {:+.4}" , f), "+0.1152" ); |
1917 | assert_eq!(format!(" {:>+10.4}" , f), " +0.1152" ); |
1918 | |
1919 | assert_eq!(format!(" {:.3?}" , f), "0.115" ); |
1920 | assert_eq!(format!(" {:.4?}" , f), "0.1152" ); |
1921 | assert_eq!(format!(" {:+.4?}" , f), "+0.1152" ); |
1922 | assert_eq!(format!(" {:>+10.4?}" , f), " +0.1152" ); |
1923 | } |
1924 | |
1925 | impl quickcheck::Arbitrary for f16 { |
1926 | fn arbitrary(g: &mut quickcheck::Gen) -> Self { |
1927 | f16(u16::arbitrary(g)) |
1928 | } |
1929 | } |
1930 | |
1931 | #[quickcheck] |
1932 | fn qc_roundtrip_f16_f32_is_identity(f: f16) -> bool { |
1933 | let roundtrip = f16::from_f32(f.to_f32()); |
1934 | if f.is_nan() { |
1935 | roundtrip.is_nan() && f.is_sign_negative() == roundtrip.is_sign_negative() |
1936 | } else { |
1937 | f.0 == roundtrip.0 |
1938 | } |
1939 | } |
1940 | |
1941 | #[quickcheck] |
1942 | fn qc_roundtrip_f16_f64_is_identity(f: f16) -> bool { |
1943 | let roundtrip = f16::from_f64(f.to_f64()); |
1944 | if f.is_nan() { |
1945 | roundtrip.is_nan() && f.is_sign_negative() == roundtrip.is_sign_negative() |
1946 | } else { |
1947 | f.0 == roundtrip.0 |
1948 | } |
1949 | } |
1950 | } |
1951 | |