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 convert; |
26 | |
27 | /// A 16-bit floating point type implementing the [`bfloat16`] format. |
28 | /// |
29 | /// The [`bfloat16`] floating point format is a truncated 16-bit version of the IEEE 754 standard |
30 | /// `binary32`, a.k.a [`f32`]. [`bf16`] has approximately the same dynamic range as [`f32`] by |
31 | /// having a lower precision than [`f16`][crate::f16]. While [`f16`][crate::f16] has a precision of |
32 | /// 11 bits, [`bf16`] has a precision of only 8 bits. |
33 | /// |
34 | /// [`bfloat16`]: https://en.wikipedia.org/wiki/Bfloat16_floating-point_format |
35 | #[allow (non_camel_case_types)] |
36 | #[derive (Clone, Copy, Default)] |
37 | #[repr (transparent)] |
38 | #[cfg_attr (feature = "serde" , derive(Serialize))] |
39 | #[cfg_attr ( |
40 | feature = "rkyv" , |
41 | derive(rkyv::Archive, rkyv::Serialize, rkyv::Deserialize) |
42 | )] |
43 | #[cfg_attr (feature = "rkyv" , archive(resolver = "Bf16Resolver" ))] |
44 | #[cfg_attr (feature = "bytemuck" , derive(Zeroable, Pod))] |
45 | #[cfg_attr (feature = "zerocopy" , derive(AsBytes, FromBytes))] |
46 | #[cfg_attr (kani, derive(kani::Arbitrary))] |
47 | pub struct bf16(u16); |
48 | |
49 | impl bf16 { |
50 | /// Constructs a [`bf16`] value from the raw bits. |
51 | #[inline ] |
52 | #[must_use ] |
53 | pub const fn from_bits(bits: u16) -> bf16 { |
54 | bf16(bits) |
55 | } |
56 | |
57 | /// Constructs a [`bf16`] value from a 32-bit floating point value. |
58 | /// |
59 | /// This operation is lossy. If the 32-bit value is too large to fit, ±∞ will result. NaN values |
60 | /// are preserved. Subnormal values that are too tiny to be represented will result in ±0. All |
61 | /// other values are truncated and rounded to the nearest representable value. |
62 | #[inline ] |
63 | #[must_use ] |
64 | pub fn from_f32(value: f32) -> bf16 { |
65 | Self::from_f32_const(value) |
66 | } |
67 | |
68 | /// Constructs a [`bf16`] value from a 32-bit floating point value. |
69 | /// |
70 | /// This function is identical to [`from_f32`][Self::from_f32] except it never uses hardware |
71 | /// intrinsics, which allows it to be `const`. [`from_f32`][Self::from_f32] should be preferred |
72 | /// in any non-`const` context. |
73 | /// |
74 | /// This operation is lossy. If the 32-bit value is too large to fit, ±∞ will result. NaN values |
75 | /// are preserved. Subnormal values that are too tiny to be represented will result in ±0. All |
76 | /// other values are truncated and rounded to the nearest representable value. |
77 | #[inline ] |
78 | #[must_use ] |
79 | pub const fn from_f32_const(value: f32) -> bf16 { |
80 | bf16(convert::f32_to_bf16(value)) |
81 | } |
82 | |
83 | /// Constructs a [`bf16`] value from a 64-bit floating point value. |
84 | /// |
85 | /// This operation is lossy. If the 64-bit value is to large to fit, ±∞ will result. NaN values |
86 | /// are preserved. 64-bit subnormal values are too tiny to be represented and result in ±0. |
87 | /// Exponents that underflow the minimum exponent will result in subnormals or ±0. All other |
88 | /// values are truncated and rounded to the nearest representable value. |
89 | #[inline ] |
90 | #[must_use ] |
91 | pub fn from_f64(value: f64) -> bf16 { |
92 | Self::from_f64_const(value) |
93 | } |
94 | |
95 | /// Constructs a [`bf16`] value from a 64-bit floating point value. |
96 | /// |
97 | /// This function is identical to [`from_f64`][Self::from_f64] except it never uses hardware |
98 | /// intrinsics, which allows it to be `const`. [`from_f64`][Self::from_f64] should be preferred |
99 | /// in any non-`const` context. |
100 | /// |
101 | /// This operation is lossy. If the 64-bit value is to large to fit, ±∞ will result. NaN values |
102 | /// are preserved. 64-bit subnormal values are too tiny to be represented and result in ±0. |
103 | /// Exponents that underflow the minimum exponent will result in subnormals or ±0. All other |
104 | /// values are truncated and rounded to the nearest representable value. |
105 | #[inline ] |
106 | #[must_use ] |
107 | pub const fn from_f64_const(value: f64) -> bf16 { |
108 | bf16(convert::f64_to_bf16(value)) |
109 | } |
110 | |
111 | /// Converts a [`bf16`] into the underlying bit representation. |
112 | #[inline ] |
113 | #[must_use ] |
114 | pub const fn to_bits(self) -> u16 { |
115 | self.0 |
116 | } |
117 | |
118 | /// Returns the memory representation of the underlying bit representation as a byte array in |
119 | /// little-endian byte order. |
120 | /// |
121 | /// # Examples |
122 | /// |
123 | /// ```rust |
124 | /// # use half::prelude::*; |
125 | /// let bytes = bf16::from_f32(12.5).to_le_bytes(); |
126 | /// assert_eq!(bytes, [0x48, 0x41]); |
127 | /// ``` |
128 | #[inline ] |
129 | #[must_use ] |
130 | pub const fn to_le_bytes(self) -> [u8; 2] { |
131 | self.0.to_le_bytes() |
132 | } |
133 | |
134 | /// Returns the memory representation of the underlying bit representation as a byte array in |
135 | /// big-endian (network) byte order. |
136 | /// |
137 | /// # Examples |
138 | /// |
139 | /// ```rust |
140 | /// # use half::prelude::*; |
141 | /// let bytes = bf16::from_f32(12.5).to_be_bytes(); |
142 | /// assert_eq!(bytes, [0x41, 0x48]); |
143 | /// ``` |
144 | #[inline ] |
145 | #[must_use ] |
146 | pub const fn to_be_bytes(self) -> [u8; 2] { |
147 | self.0.to_be_bytes() |
148 | } |
149 | |
150 | /// Returns the memory representation of the underlying bit representation as a byte array in |
151 | /// native byte order. |
152 | /// |
153 | /// As the target platform's native endianness is used, portable code should use |
154 | /// [`to_be_bytes`][bf16::to_be_bytes] or [`to_le_bytes`][bf16::to_le_bytes], as appropriate, |
155 | /// instead. |
156 | /// |
157 | /// # Examples |
158 | /// |
159 | /// ```rust |
160 | /// # use half::prelude::*; |
161 | /// let bytes = bf16::from_f32(12.5).to_ne_bytes(); |
162 | /// assert_eq!(bytes, if cfg!(target_endian = "big" ) { |
163 | /// [0x41, 0x48] |
164 | /// } else { |
165 | /// [0x48, 0x41] |
166 | /// }); |
167 | /// ``` |
168 | #[inline ] |
169 | #[must_use ] |
170 | pub const fn to_ne_bytes(self) -> [u8; 2] { |
171 | self.0.to_ne_bytes() |
172 | } |
173 | |
174 | /// Creates a floating point value from its representation as a byte array in little endian. |
175 | /// |
176 | /// # Examples |
177 | /// |
178 | /// ```rust |
179 | /// # use half::prelude::*; |
180 | /// let value = bf16::from_le_bytes([0x48, 0x41]); |
181 | /// assert_eq!(value, bf16::from_f32(12.5)); |
182 | /// ``` |
183 | #[inline ] |
184 | #[must_use ] |
185 | pub const fn from_le_bytes(bytes: [u8; 2]) -> bf16 { |
186 | bf16::from_bits(u16::from_le_bytes(bytes)) |
187 | } |
188 | |
189 | /// Creates a floating point value from its representation as a byte array in big endian. |
190 | /// |
191 | /// # Examples |
192 | /// |
193 | /// ```rust |
194 | /// # use half::prelude::*; |
195 | /// let value = bf16::from_be_bytes([0x41, 0x48]); |
196 | /// assert_eq!(value, bf16::from_f32(12.5)); |
197 | /// ``` |
198 | #[inline ] |
199 | #[must_use ] |
200 | pub const fn from_be_bytes(bytes: [u8; 2]) -> bf16 { |
201 | bf16::from_bits(u16::from_be_bytes(bytes)) |
202 | } |
203 | |
204 | /// Creates a floating point value from its representation as a byte array in native endian. |
205 | /// |
206 | /// As the target platform's native endianness is used, portable code likely wants to use |
207 | /// [`from_be_bytes`][bf16::from_be_bytes] or [`from_le_bytes`][bf16::from_le_bytes], as |
208 | /// appropriate instead. |
209 | /// |
210 | /// # Examples |
211 | /// |
212 | /// ```rust |
213 | /// # use half::prelude::*; |
214 | /// let value = bf16::from_ne_bytes(if cfg!(target_endian = "big" ) { |
215 | /// [0x41, 0x48] |
216 | /// } else { |
217 | /// [0x48, 0x41] |
218 | /// }); |
219 | /// assert_eq!(value, bf16::from_f32(12.5)); |
220 | /// ``` |
221 | #[inline ] |
222 | #[must_use ] |
223 | pub const fn from_ne_bytes(bytes: [u8; 2]) -> bf16 { |
224 | bf16::from_bits(u16::from_ne_bytes(bytes)) |
225 | } |
226 | |
227 | /// Converts a [`bf16`] value into an [`f32`] value. |
228 | /// |
229 | /// This conversion is lossless as all values can be represented exactly in [`f32`]. |
230 | #[inline ] |
231 | #[must_use ] |
232 | pub fn to_f32(self) -> f32 { |
233 | self.to_f32_const() |
234 | } |
235 | |
236 | /// Converts a [`bf16`] value into an [`f32`] value. |
237 | /// |
238 | /// This function is identical to [`to_f32`][Self::to_f32] except it never uses hardware |
239 | /// intrinsics, which allows it to be `const`. [`to_f32`][Self::to_f32] should be preferred |
240 | /// in any non-`const` context. |
241 | /// |
242 | /// This conversion is lossless as all values can be represented exactly in [`f32`]. |
243 | #[inline ] |
244 | #[must_use ] |
245 | pub const fn to_f32_const(self) -> f32 { |
246 | convert::bf16_to_f32(self.0) |
247 | } |
248 | |
249 | /// Converts a [`bf16`] value into an [`f64`] value. |
250 | /// |
251 | /// This conversion is lossless as all values can be represented exactly in [`f64`]. |
252 | #[inline ] |
253 | #[must_use ] |
254 | pub fn to_f64(self) -> f64 { |
255 | self.to_f64_const() |
256 | } |
257 | |
258 | /// Converts a [`bf16`] value into an [`f64`] value. |
259 | /// |
260 | /// This function is identical to [`to_f64`][Self::to_f64] except it never uses hardware |
261 | /// intrinsics, which allows it to be `const`. [`to_f64`][Self::to_f64] should be preferred |
262 | /// in any non-`const` context. |
263 | /// |
264 | /// This conversion is lossless as all values can be represented exactly in [`f64`]. |
265 | #[inline ] |
266 | #[must_use ] |
267 | pub const fn to_f64_const(self) -> f64 { |
268 | convert::bf16_to_f64(self.0) |
269 | } |
270 | |
271 | /// Returns `true` if this value is NaN and `false` otherwise. |
272 | /// |
273 | /// # Examples |
274 | /// |
275 | /// ```rust |
276 | /// # use half::prelude::*; |
277 | /// |
278 | /// let nan = bf16::NAN; |
279 | /// let f = bf16::from_f32(7.0_f32); |
280 | /// |
281 | /// assert!(nan.is_nan()); |
282 | /// assert!(!f.is_nan()); |
283 | /// ``` |
284 | #[inline ] |
285 | #[must_use ] |
286 | pub const fn is_nan(self) -> bool { |
287 | self.0 & 0x7FFFu16 > 0x7F80u16 |
288 | } |
289 | |
290 | /// Returns `true` if this value is ±∞ and `false` otherwise. |
291 | /// |
292 | /// # Examples |
293 | /// |
294 | /// ```rust |
295 | /// # use half::prelude::*; |
296 | /// |
297 | /// let f = bf16::from_f32(7.0f32); |
298 | /// let inf = bf16::INFINITY; |
299 | /// let neg_inf = bf16::NEG_INFINITY; |
300 | /// let nan = bf16::NAN; |
301 | /// |
302 | /// assert!(!f.is_infinite()); |
303 | /// assert!(!nan.is_infinite()); |
304 | /// |
305 | /// assert!(inf.is_infinite()); |
306 | /// assert!(neg_inf.is_infinite()); |
307 | /// ``` |
308 | #[inline ] |
309 | #[must_use ] |
310 | pub const fn is_infinite(self) -> bool { |
311 | self.0 & 0x7FFFu16 == 0x7F80u16 |
312 | } |
313 | |
314 | /// Returns `true` if this number is neither infinite nor NaN. |
315 | /// |
316 | /// # Examples |
317 | /// |
318 | /// ```rust |
319 | /// # use half::prelude::*; |
320 | /// |
321 | /// let f = bf16::from_f32(7.0f32); |
322 | /// let inf = bf16::INFINITY; |
323 | /// let neg_inf = bf16::NEG_INFINITY; |
324 | /// let nan = bf16::NAN; |
325 | /// |
326 | /// assert!(f.is_finite()); |
327 | /// |
328 | /// assert!(!nan.is_finite()); |
329 | /// assert!(!inf.is_finite()); |
330 | /// assert!(!neg_inf.is_finite()); |
331 | /// ``` |
332 | #[inline ] |
333 | #[must_use ] |
334 | pub const fn is_finite(self) -> bool { |
335 | self.0 & 0x7F80u16 != 0x7F80u16 |
336 | } |
337 | |
338 | /// Returns `true` if the number is neither zero, infinite, subnormal, or NaN. |
339 | /// |
340 | /// # Examples |
341 | /// |
342 | /// ```rust |
343 | /// # use half::prelude::*; |
344 | /// |
345 | /// let min = bf16::MIN_POSITIVE; |
346 | /// let max = bf16::MAX; |
347 | /// let lower_than_min = bf16::from_f32(1.0e-39_f32); |
348 | /// let zero = bf16::from_f32(0.0_f32); |
349 | /// |
350 | /// assert!(min.is_normal()); |
351 | /// assert!(max.is_normal()); |
352 | /// |
353 | /// assert!(!zero.is_normal()); |
354 | /// assert!(!bf16::NAN.is_normal()); |
355 | /// assert!(!bf16::INFINITY.is_normal()); |
356 | /// // Values between 0 and `min` are subnormal. |
357 | /// assert!(!lower_than_min.is_normal()); |
358 | /// ``` |
359 | #[inline ] |
360 | #[must_use ] |
361 | pub const fn is_normal(self) -> bool { |
362 | let exp = self.0 & 0x7F80u16; |
363 | exp != 0x7F80u16 && exp != 0 |
364 | } |
365 | |
366 | /// Returns the floating point category of the number. |
367 | /// |
368 | /// If only one property is going to be tested, it is generally faster to use the specific |
369 | /// predicate instead. |
370 | /// |
371 | /// # Examples |
372 | /// |
373 | /// ```rust |
374 | /// use std::num::FpCategory; |
375 | /// # use half::prelude::*; |
376 | /// |
377 | /// let num = bf16::from_f32(12.4_f32); |
378 | /// let inf = bf16::INFINITY; |
379 | /// |
380 | /// assert_eq!(num.classify(), FpCategory::Normal); |
381 | /// assert_eq!(inf.classify(), FpCategory::Infinite); |
382 | /// ``` |
383 | #[must_use ] |
384 | pub const fn classify(self) -> FpCategory { |
385 | let exp = self.0 & 0x7F80u16; |
386 | let man = self.0 & 0x007Fu16; |
387 | match (exp, man) { |
388 | (0, 0) => FpCategory::Zero, |
389 | (0, _) => FpCategory::Subnormal, |
390 | (0x7F80u16, 0) => FpCategory::Infinite, |
391 | (0x7F80u16, _) => FpCategory::Nan, |
392 | _ => FpCategory::Normal, |
393 | } |
394 | } |
395 | |
396 | /// Returns a number that represents the sign of `self`. |
397 | /// |
398 | /// * 1.0 if the number is positive, +0.0 or [`INFINITY`][bf16::INFINITY] |
399 | /// * −1.0 if the number is negative, −0.0` or [`NEG_INFINITY`][bf16::NEG_INFINITY] |
400 | /// * [`NAN`][bf16::NAN] if the number is NaN |
401 | /// |
402 | /// # Examples |
403 | /// |
404 | /// ```rust |
405 | /// # use half::prelude::*; |
406 | /// |
407 | /// let f = bf16::from_f32(3.5_f32); |
408 | /// |
409 | /// assert_eq!(f.signum(), bf16::from_f32(1.0)); |
410 | /// assert_eq!(bf16::NEG_INFINITY.signum(), bf16::from_f32(-1.0)); |
411 | /// |
412 | /// assert!(bf16::NAN.signum().is_nan()); |
413 | /// ``` |
414 | #[must_use ] |
415 | pub const fn signum(self) -> bf16 { |
416 | if self.is_nan() { |
417 | self |
418 | } else if self.0 & 0x8000u16 != 0 { |
419 | Self::NEG_ONE |
420 | } else { |
421 | Self::ONE |
422 | } |
423 | } |
424 | |
425 | /// Returns `true` if and only if `self` has a positive sign, including +0.0, NaNs with a |
426 | /// positive sign bit and +∞. |
427 | /// |
428 | /// # Examples |
429 | /// |
430 | /// ```rust |
431 | /// # use half::prelude::*; |
432 | /// |
433 | /// let nan = bf16::NAN; |
434 | /// let f = bf16::from_f32(7.0_f32); |
435 | /// let g = bf16::from_f32(-7.0_f32); |
436 | /// |
437 | /// assert!(f.is_sign_positive()); |
438 | /// assert!(!g.is_sign_positive()); |
439 | /// // NaN can be either positive or negative |
440 | /// assert!(nan.is_sign_positive() != nan.is_sign_negative()); |
441 | /// ``` |
442 | #[inline ] |
443 | #[must_use ] |
444 | pub const fn is_sign_positive(self) -> bool { |
445 | self.0 & 0x8000u16 == 0 |
446 | } |
447 | |
448 | /// Returns `true` if and only if `self` has a negative sign, including −0.0, NaNs with a |
449 | /// negative sign bit and −∞. |
450 | /// |
451 | /// # Examples |
452 | /// |
453 | /// ```rust |
454 | /// # use half::prelude::*; |
455 | /// |
456 | /// let nan = bf16::NAN; |
457 | /// let f = bf16::from_f32(7.0f32); |
458 | /// let g = bf16::from_f32(-7.0f32); |
459 | /// |
460 | /// assert!(!f.is_sign_negative()); |
461 | /// assert!(g.is_sign_negative()); |
462 | /// // NaN can be either positive or negative |
463 | /// assert!(nan.is_sign_positive() != nan.is_sign_negative()); |
464 | /// ``` |
465 | #[inline ] |
466 | #[must_use ] |
467 | pub const fn is_sign_negative(self) -> bool { |
468 | self.0 & 0x8000u16 != 0 |
469 | } |
470 | |
471 | /// Returns a number composed of the magnitude of `self` and the sign of `sign`. |
472 | /// |
473 | /// Equal to `self` if the sign of `self` and `sign` are the same, otherwise equal to `-self`. |
474 | /// If `self` is NaN, then NaN with the sign of `sign` is returned. |
475 | /// |
476 | /// # Examples |
477 | /// |
478 | /// ``` |
479 | /// # use half::prelude::*; |
480 | /// let f = bf16::from_f32(3.5); |
481 | /// |
482 | /// assert_eq!(f.copysign(bf16::from_f32(0.42)), bf16::from_f32(3.5)); |
483 | /// assert_eq!(f.copysign(bf16::from_f32(-0.42)), bf16::from_f32(-3.5)); |
484 | /// assert_eq!((-f).copysign(bf16::from_f32(0.42)), bf16::from_f32(3.5)); |
485 | /// assert_eq!((-f).copysign(bf16::from_f32(-0.42)), bf16::from_f32(-3.5)); |
486 | /// |
487 | /// assert!(bf16::NAN.copysign(bf16::from_f32(1.0)).is_nan()); |
488 | /// ``` |
489 | #[inline ] |
490 | #[must_use ] |
491 | pub const fn copysign(self, sign: bf16) -> bf16 { |
492 | bf16((sign.0 & 0x8000u16) | (self.0 & 0x7FFFu16)) |
493 | } |
494 | |
495 | /// Returns the maximum of the two numbers. |
496 | /// |
497 | /// If one of the arguments is NaN, then the other argument is returned. |
498 | /// |
499 | /// # Examples |
500 | /// |
501 | /// ``` |
502 | /// # use half::prelude::*; |
503 | /// let x = bf16::from_f32(1.0); |
504 | /// let y = bf16::from_f32(2.0); |
505 | /// |
506 | /// assert_eq!(x.max(y), y); |
507 | /// ``` |
508 | #[inline ] |
509 | #[must_use ] |
510 | pub fn max(self, other: bf16) -> bf16 { |
511 | if other > self && !other.is_nan() { |
512 | other |
513 | } else { |
514 | self |
515 | } |
516 | } |
517 | |
518 | /// Returns the minimum of the two numbers. |
519 | /// |
520 | /// If one of the arguments is NaN, then the other argument is returned. |
521 | /// |
522 | /// # Examples |
523 | /// |
524 | /// ``` |
525 | /// # use half::prelude::*; |
526 | /// let x = bf16::from_f32(1.0); |
527 | /// let y = bf16::from_f32(2.0); |
528 | /// |
529 | /// assert_eq!(x.min(y), x); |
530 | /// ``` |
531 | #[inline ] |
532 | #[must_use ] |
533 | pub fn min(self, other: bf16) -> bf16 { |
534 | if other < self && !other.is_nan() { |
535 | other |
536 | } else { |
537 | self |
538 | } |
539 | } |
540 | |
541 | /// Restrict a value to a certain interval unless it is NaN. |
542 | /// |
543 | /// Returns `max` if `self` is greater than `max`, and `min` if `self` is less than `min`. |
544 | /// Otherwise this returns `self`. |
545 | /// |
546 | /// Note that this function returns NaN if the initial value was NaN as well. |
547 | /// |
548 | /// # Panics |
549 | /// Panics if `min > max`, `min` is NaN, or `max` is NaN. |
550 | /// |
551 | /// # Examples |
552 | /// |
553 | /// ``` |
554 | /// # use half::prelude::*; |
555 | /// assert!(bf16::from_f32(-3.0).clamp(bf16::from_f32(-2.0), bf16::from_f32(1.0)) == bf16::from_f32(-2.0)); |
556 | /// assert!(bf16::from_f32(0.0).clamp(bf16::from_f32(-2.0), bf16::from_f32(1.0)) == bf16::from_f32(0.0)); |
557 | /// assert!(bf16::from_f32(2.0).clamp(bf16::from_f32(-2.0), bf16::from_f32(1.0)) == bf16::from_f32(1.0)); |
558 | /// assert!(bf16::NAN.clamp(bf16::from_f32(-2.0), bf16::from_f32(1.0)).is_nan()); |
559 | /// ``` |
560 | #[inline ] |
561 | #[must_use ] |
562 | pub fn clamp(self, min: bf16, max: bf16) -> bf16 { |
563 | assert!(min <= max); |
564 | let mut x = self; |
565 | if x < min { |
566 | x = min; |
567 | } |
568 | if x > max { |
569 | x = max; |
570 | } |
571 | x |
572 | } |
573 | |
574 | /// Returns the ordering between `self` and `other`. |
575 | /// |
576 | /// Unlike the standard partial comparison between floating point numbers, |
577 | /// this comparison always produces an ordering in accordance to |
578 | /// the `totalOrder` predicate as defined in the IEEE 754 (2008 revision) |
579 | /// floating point standard. The values are ordered in the following sequence: |
580 | /// |
581 | /// - negative quiet NaN |
582 | /// - negative signaling NaN |
583 | /// - negative infinity |
584 | /// - negative numbers |
585 | /// - negative subnormal numbers |
586 | /// - negative zero |
587 | /// - positive zero |
588 | /// - positive subnormal numbers |
589 | /// - positive numbers |
590 | /// - positive infinity |
591 | /// - positive signaling NaN |
592 | /// - positive quiet NaN. |
593 | /// |
594 | /// The ordering established by this function does not always agree with the |
595 | /// [`PartialOrd`] and [`PartialEq`] implementations of `bf16`. For example, |
596 | /// they consider negative and positive zero equal, while `total_cmp` |
597 | /// doesn't. |
598 | /// |
599 | /// The interpretation of the signaling NaN bit follows the definition in |
600 | /// the IEEE 754 standard, which may not match the interpretation by some of |
601 | /// the older, non-conformant (e.g. MIPS) hardware implementations. |
602 | /// |
603 | /// # Examples |
604 | /// ``` |
605 | /// # use half::bf16; |
606 | /// let mut v: Vec<bf16> = vec![]; |
607 | /// v.push(bf16::ONE); |
608 | /// v.push(bf16::INFINITY); |
609 | /// v.push(bf16::NEG_INFINITY); |
610 | /// v.push(bf16::NAN); |
611 | /// v.push(bf16::MAX_SUBNORMAL); |
612 | /// v.push(-bf16::MAX_SUBNORMAL); |
613 | /// v.push(bf16::ZERO); |
614 | /// v.push(bf16::NEG_ZERO); |
615 | /// v.push(bf16::NEG_ONE); |
616 | /// v.push(bf16::MIN_POSITIVE); |
617 | /// |
618 | /// v.sort_by(|a, b| a.total_cmp(&b)); |
619 | /// |
620 | /// assert!(v |
621 | /// .into_iter() |
622 | /// .zip( |
623 | /// [ |
624 | /// bf16::NEG_INFINITY, |
625 | /// bf16::NEG_ONE, |
626 | /// -bf16::MAX_SUBNORMAL, |
627 | /// bf16::NEG_ZERO, |
628 | /// bf16::ZERO, |
629 | /// bf16::MAX_SUBNORMAL, |
630 | /// bf16::MIN_POSITIVE, |
631 | /// bf16::ONE, |
632 | /// bf16::INFINITY, |
633 | /// bf16::NAN |
634 | /// ] |
635 | /// .iter() |
636 | /// ) |
637 | /// .all(|(a, b)| a.to_bits() == b.to_bits())); |
638 | /// ``` |
639 | // Implementation based on: https://doc.rust-lang.org/std/primitive.f32.html#method.total_cmp |
640 | #[inline ] |
641 | #[must_use ] |
642 | pub fn total_cmp(&self, other: &Self) -> Ordering { |
643 | let mut left = self.to_bits() as i16; |
644 | let mut right = other.to_bits() as i16; |
645 | left ^= (((left >> 15) as u16) >> 1) as i16; |
646 | right ^= (((right >> 15) as u16) >> 1) as i16; |
647 | left.cmp(&right) |
648 | } |
649 | |
650 | /// Alternate serialize adapter for serializing as a float. |
651 | /// |
652 | /// By default, [`bf16`] serializes as a newtype of [`u16`]. This is an alternate serialize |
653 | /// implementation that serializes as an [`f32`] value. It is designed for use with |
654 | /// `serialize_with` serde attributes. Deserialization from `f32` values is already supported by |
655 | /// the default deserialize implementation. |
656 | /// |
657 | /// # Examples |
658 | /// |
659 | /// A demonstration on how to use this adapater: |
660 | /// |
661 | /// ``` |
662 | /// use serde::{Serialize, Deserialize}; |
663 | /// use half::bf16; |
664 | /// |
665 | /// #[derive(Serialize, Deserialize)] |
666 | /// struct MyStruct { |
667 | /// #[serde(serialize_with = "bf16::serialize_as_f32")] |
668 | /// value: bf16 // Will be serialized as f32 instead of u16 |
669 | /// } |
670 | /// ``` |
671 | #[cfg (feature = "serde" )] |
672 | pub fn serialize_as_f32<S: serde::Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> { |
673 | serializer.serialize_f32(self.to_f32()) |
674 | } |
675 | |
676 | /// Alternate serialize adapter for serializing as a string. |
677 | /// |
678 | /// By default, [`bf16`] serializes as a newtype of [`u16`]. This is an alternate serialize |
679 | /// implementation that serializes as a string value. It is designed for use with |
680 | /// `serialize_with` serde attributes. Deserialization from string values is already supported |
681 | /// by the default deserialize implementation. |
682 | /// |
683 | /// # Examples |
684 | /// |
685 | /// A demonstration on how to use this adapater: |
686 | /// |
687 | /// ``` |
688 | /// use serde::{Serialize, Deserialize}; |
689 | /// use half::bf16; |
690 | /// |
691 | /// #[derive(Serialize, Deserialize)] |
692 | /// struct MyStruct { |
693 | /// #[serde(serialize_with = "bf16::serialize_as_string")] |
694 | /// value: bf16 // Will be serialized as a string instead of u16 |
695 | /// } |
696 | /// ``` |
697 | #[cfg (all(feature = "serde" , feature = "alloc" ))] |
698 | pub fn serialize_as_string<S: serde::Serializer>( |
699 | &self, |
700 | serializer: S, |
701 | ) -> Result<S::Ok, S::Error> { |
702 | serializer.serialize_str(&self.to_string()) |
703 | } |
704 | |
705 | /// Approximate number of [`bf16`] significant digits in base 10 |
706 | pub const DIGITS: u32 = 2; |
707 | /// [`bf16`] |
708 | /// [machine epsilon](https://en.wikipedia.org/wiki/Machine_epsilon) value |
709 | /// |
710 | /// This is the difference between 1.0 and the next largest representable number. |
711 | pub const EPSILON: bf16 = bf16(0x3C00u16); |
712 | /// [`bf16`] positive Infinity (+∞) |
713 | pub const INFINITY: bf16 = bf16(0x7F80u16); |
714 | /// Number of [`bf16`] significant digits in base 2 |
715 | pub const MANTISSA_DIGITS: u32 = 8; |
716 | /// Largest finite [`bf16`] value |
717 | pub const MAX: bf16 = bf16(0x7F7F); |
718 | /// Maximum possible [`bf16`] power of 10 exponent |
719 | pub const MAX_10_EXP: i32 = 38; |
720 | /// Maximum possible [`bf16`] power of 2 exponent |
721 | pub const MAX_EXP: i32 = 128; |
722 | /// Smallest finite [`bf16`] value |
723 | pub const MIN: bf16 = bf16(0xFF7F); |
724 | /// Minimum possible normal [`bf16`] power of 10 exponent |
725 | pub const MIN_10_EXP: i32 = -37; |
726 | /// One greater than the minimum possible normal [`bf16`] power of 2 exponent |
727 | pub const MIN_EXP: i32 = -125; |
728 | /// Smallest positive normal [`bf16`] value |
729 | pub const MIN_POSITIVE: bf16 = bf16(0x0080u16); |
730 | /// [`bf16`] Not a Number (NaN) |
731 | pub const NAN: bf16 = bf16(0x7FC0u16); |
732 | /// [`bf16`] negative infinity (-∞). |
733 | pub const NEG_INFINITY: bf16 = bf16(0xFF80u16); |
734 | /// The radix or base of the internal representation of [`bf16`] |
735 | pub const RADIX: u32 = 2; |
736 | |
737 | /// Minimum positive subnormal [`bf16`] value |
738 | pub const MIN_POSITIVE_SUBNORMAL: bf16 = bf16(0x0001u16); |
739 | /// Maximum subnormal [`bf16`] value |
740 | pub const MAX_SUBNORMAL: bf16 = bf16(0x007Fu16); |
741 | |
742 | /// [`bf16`] 1 |
743 | pub const ONE: bf16 = bf16(0x3F80u16); |
744 | /// [`bf16`] 0 |
745 | pub const ZERO: bf16 = bf16(0x0000u16); |
746 | /// [`bf16`] -0 |
747 | pub const NEG_ZERO: bf16 = bf16(0x8000u16); |
748 | /// [`bf16`] -1 |
749 | pub const NEG_ONE: bf16 = bf16(0xBF80u16); |
750 | |
751 | /// [`bf16`] Euler's number (ℯ) |
752 | pub const E: bf16 = bf16(0x402Eu16); |
753 | /// [`bf16`] Archimedes' constant (π) |
754 | pub const PI: bf16 = bf16(0x4049u16); |
755 | /// [`bf16`] 1/π |
756 | pub const FRAC_1_PI: bf16 = bf16(0x3EA3u16); |
757 | /// [`bf16`] 1/√2 |
758 | pub const FRAC_1_SQRT_2: bf16 = bf16(0x3F35u16); |
759 | /// [`bf16`] 2/π |
760 | pub const FRAC_2_PI: bf16 = bf16(0x3F23u16); |
761 | /// [`bf16`] 2/√π |
762 | pub const FRAC_2_SQRT_PI: bf16 = bf16(0x3F90u16); |
763 | /// [`bf16`] π/2 |
764 | pub const FRAC_PI_2: bf16 = bf16(0x3FC9u16); |
765 | /// [`bf16`] π/3 |
766 | pub const FRAC_PI_3: bf16 = bf16(0x3F86u16); |
767 | /// [`bf16`] π/4 |
768 | pub const FRAC_PI_4: bf16 = bf16(0x3F49u16); |
769 | /// [`bf16`] π/6 |
770 | pub const FRAC_PI_6: bf16 = bf16(0x3F06u16); |
771 | /// [`bf16`] π/8 |
772 | pub const FRAC_PI_8: bf16 = bf16(0x3EC9u16); |
773 | /// [`bf16`] 𝗅𝗇 10 |
774 | pub const LN_10: bf16 = bf16(0x4013u16); |
775 | /// [`bf16`] 𝗅𝗇 2 |
776 | pub const LN_2: bf16 = bf16(0x3F31u16); |
777 | /// [`bf16`] 𝗅𝗈𝗀₁₀ℯ |
778 | pub const LOG10_E: bf16 = bf16(0x3EDEu16); |
779 | /// [`bf16`] 𝗅𝗈𝗀₁₀2 |
780 | pub const LOG10_2: bf16 = bf16(0x3E9Au16); |
781 | /// [`bf16`] 𝗅𝗈𝗀₂ℯ |
782 | pub const LOG2_E: bf16 = bf16(0x3FB9u16); |
783 | /// [`bf16`] 𝗅𝗈𝗀₂10 |
784 | pub const LOG2_10: bf16 = bf16(0x4055u16); |
785 | /// [`bf16`] √2 |
786 | pub const SQRT_2: bf16 = bf16(0x3FB5u16); |
787 | } |
788 | |
789 | impl From<bf16> for f32 { |
790 | #[inline ] |
791 | fn from(x: bf16) -> f32 { |
792 | x.to_f32() |
793 | } |
794 | } |
795 | |
796 | impl From<bf16> for f64 { |
797 | #[inline ] |
798 | fn from(x: bf16) -> f64 { |
799 | x.to_f64() |
800 | } |
801 | } |
802 | |
803 | impl From<i8> for bf16 { |
804 | #[inline ] |
805 | fn from(x: i8) -> bf16 { |
806 | // Convert to f32, then to bf16 |
807 | bf16::from_f32(f32::from(x)) |
808 | } |
809 | } |
810 | |
811 | impl From<u8> for bf16 { |
812 | #[inline ] |
813 | fn from(x: u8) -> bf16 { |
814 | // Convert to f32, then to f16 |
815 | bf16::from_f32(f32::from(x)) |
816 | } |
817 | } |
818 | |
819 | impl PartialEq for bf16 { |
820 | fn eq(&self, other: &bf16) -> bool { |
821 | if self.is_nan() || other.is_nan() { |
822 | false |
823 | } else { |
824 | (self.0 == other.0) || ((self.0 | other.0) & 0x7FFFu16 == 0) |
825 | } |
826 | } |
827 | } |
828 | |
829 | impl PartialOrd for bf16 { |
830 | fn partial_cmp(&self, other: &bf16) -> Option<Ordering> { |
831 | if self.is_nan() || other.is_nan() { |
832 | None |
833 | } else { |
834 | let neg = self.0 & 0x8000u16 != 0; |
835 | let other_neg = other.0 & 0x8000u16 != 0; |
836 | match (neg, other_neg) { |
837 | (false, false) => Some(self.0.cmp(&other.0)), |
838 | (false, true) => { |
839 | if (self.0 | other.0) & 0x7FFFu16 == 0 { |
840 | Some(Ordering::Equal) |
841 | } else { |
842 | Some(Ordering::Greater) |
843 | } |
844 | } |
845 | (true, false) => { |
846 | if (self.0 | other.0) & 0x7FFFu16 == 0 { |
847 | Some(Ordering::Equal) |
848 | } else { |
849 | Some(Ordering::Less) |
850 | } |
851 | } |
852 | (true, true) => Some(other.0.cmp(&self.0)), |
853 | } |
854 | } |
855 | } |
856 | |
857 | fn lt(&self, other: &bf16) -> bool { |
858 | if self.is_nan() || other.is_nan() { |
859 | false |
860 | } else { |
861 | let neg = self.0 & 0x8000u16 != 0; |
862 | let other_neg = other.0 & 0x8000u16 != 0; |
863 | match (neg, other_neg) { |
864 | (false, false) => self.0 < other.0, |
865 | (false, true) => false, |
866 | (true, false) => (self.0 | other.0) & 0x7FFFu16 != 0, |
867 | (true, true) => self.0 > other.0, |
868 | } |
869 | } |
870 | } |
871 | |
872 | fn le(&self, other: &bf16) -> bool { |
873 | if self.is_nan() || other.is_nan() { |
874 | false |
875 | } else { |
876 | let neg = self.0 & 0x8000u16 != 0; |
877 | let other_neg = other.0 & 0x8000u16 != 0; |
878 | match (neg, other_neg) { |
879 | (false, false) => self.0 <= other.0, |
880 | (false, true) => (self.0 | other.0) & 0x7FFFu16 == 0, |
881 | (true, false) => true, |
882 | (true, true) => self.0 >= other.0, |
883 | } |
884 | } |
885 | } |
886 | |
887 | fn gt(&self, other: &bf16) -> bool { |
888 | if self.is_nan() || other.is_nan() { |
889 | false |
890 | } else { |
891 | let neg = self.0 & 0x8000u16 != 0; |
892 | let other_neg = other.0 & 0x8000u16 != 0; |
893 | match (neg, other_neg) { |
894 | (false, false) => self.0 > other.0, |
895 | (false, true) => (self.0 | other.0) & 0x7FFFu16 != 0, |
896 | (true, false) => false, |
897 | (true, true) => self.0 < other.0, |
898 | } |
899 | } |
900 | } |
901 | |
902 | fn ge(&self, other: &bf16) -> bool { |
903 | if self.is_nan() || other.is_nan() { |
904 | false |
905 | } else { |
906 | let neg = self.0 & 0x8000u16 != 0; |
907 | let other_neg = other.0 & 0x8000u16 != 0; |
908 | match (neg, other_neg) { |
909 | (false, false) => self.0 >= other.0, |
910 | (false, true) => true, |
911 | (true, false) => (self.0 | other.0) & 0x7FFFu16 == 0, |
912 | (true, true) => self.0 <= other.0, |
913 | } |
914 | } |
915 | } |
916 | } |
917 | |
918 | #[cfg (not(target_arch = "spirv" ))] |
919 | impl FromStr for bf16 { |
920 | type Err = ParseFloatError; |
921 | fn from_str(src: &str) -> Result<bf16, ParseFloatError> { |
922 | f32::from_str(src).map(op:bf16::from_f32) |
923 | } |
924 | } |
925 | |
926 | #[cfg (not(target_arch = "spirv" ))] |
927 | impl Debug for bf16 { |
928 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
929 | Debug::fmt(&self.to_f32(), f) |
930 | } |
931 | } |
932 | |
933 | #[cfg (not(target_arch = "spirv" ))] |
934 | impl Display for bf16 { |
935 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
936 | Display::fmt(&self.to_f32(), f) |
937 | } |
938 | } |
939 | |
940 | #[cfg (not(target_arch = "spirv" ))] |
941 | impl LowerExp for bf16 { |
942 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
943 | write!(f, " {:e}" , self.to_f32()) |
944 | } |
945 | } |
946 | |
947 | #[cfg (not(target_arch = "spirv" ))] |
948 | impl UpperExp for bf16 { |
949 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
950 | write!(f, " {:E}" , self.to_f32()) |
951 | } |
952 | } |
953 | |
954 | #[cfg (not(target_arch = "spirv" ))] |
955 | impl Binary for bf16 { |
956 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
957 | write!(f, " {:b}" , self.0) |
958 | } |
959 | } |
960 | |
961 | #[cfg (not(target_arch = "spirv" ))] |
962 | impl Octal for bf16 { |
963 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
964 | write!(f, " {:o}" , self.0) |
965 | } |
966 | } |
967 | |
968 | #[cfg (not(target_arch = "spirv" ))] |
969 | impl LowerHex for bf16 { |
970 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
971 | write!(f, " {:x}" , self.0) |
972 | } |
973 | } |
974 | |
975 | #[cfg (not(target_arch = "spirv" ))] |
976 | impl UpperHex for bf16 { |
977 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> { |
978 | write!(f, " {:X}" , self.0) |
979 | } |
980 | } |
981 | |
982 | impl Neg for bf16 { |
983 | type Output = Self; |
984 | |
985 | fn neg(self) -> Self::Output { |
986 | Self(self.0 ^ 0x8000) |
987 | } |
988 | } |
989 | |
990 | impl Neg for &bf16 { |
991 | type Output = <bf16 as Neg>::Output; |
992 | |
993 | #[inline ] |
994 | fn neg(self) -> Self::Output { |
995 | Neg::neg(*self) |
996 | } |
997 | } |
998 | |
999 | impl Add for bf16 { |
1000 | type Output = Self; |
1001 | |
1002 | fn add(self, rhs: Self) -> Self::Output { |
1003 | Self::from_f32(Self::to_f32(self) + Self::to_f32(self:rhs)) |
1004 | } |
1005 | } |
1006 | |
1007 | impl Add<&bf16> for bf16 { |
1008 | type Output = <bf16 as Add<bf16>>::Output; |
1009 | |
1010 | #[inline ] |
1011 | fn add(self, rhs: &bf16) -> Self::Output { |
1012 | self.add(*rhs) |
1013 | } |
1014 | } |
1015 | |
1016 | impl Add<&bf16> for &bf16 { |
1017 | type Output = <bf16 as Add<bf16>>::Output; |
1018 | |
1019 | #[inline ] |
1020 | fn add(self, rhs: &bf16) -> Self::Output { |
1021 | (*self).add(*rhs) |
1022 | } |
1023 | } |
1024 | |
1025 | impl Add<bf16> for &bf16 { |
1026 | type Output = <bf16 as Add<bf16>>::Output; |
1027 | |
1028 | #[inline ] |
1029 | fn add(self, rhs: bf16) -> Self::Output { |
1030 | (*self).add(rhs) |
1031 | } |
1032 | } |
1033 | |
1034 | impl AddAssign for bf16 { |
1035 | #[inline ] |
1036 | fn add_assign(&mut self, rhs: Self) { |
1037 | *self = (*self).add(rhs); |
1038 | } |
1039 | } |
1040 | |
1041 | impl AddAssign<&bf16> for bf16 { |
1042 | #[inline ] |
1043 | fn add_assign(&mut self, rhs: &bf16) { |
1044 | *self = (*self).add(rhs); |
1045 | } |
1046 | } |
1047 | |
1048 | impl Sub for bf16 { |
1049 | type Output = Self; |
1050 | |
1051 | fn sub(self, rhs: Self) -> Self::Output { |
1052 | Self::from_f32(Self::to_f32(self) - Self::to_f32(self:rhs)) |
1053 | } |
1054 | } |
1055 | |
1056 | impl Sub<&bf16> for bf16 { |
1057 | type Output = <bf16 as Sub<bf16>>::Output; |
1058 | |
1059 | #[inline ] |
1060 | fn sub(self, rhs: &bf16) -> Self::Output { |
1061 | self.sub(*rhs) |
1062 | } |
1063 | } |
1064 | |
1065 | impl Sub<&bf16> for &bf16 { |
1066 | type Output = <bf16 as Sub<bf16>>::Output; |
1067 | |
1068 | #[inline ] |
1069 | fn sub(self, rhs: &bf16) -> Self::Output { |
1070 | (*self).sub(*rhs) |
1071 | } |
1072 | } |
1073 | |
1074 | impl Sub<bf16> for &bf16 { |
1075 | type Output = <bf16 as Sub<bf16>>::Output; |
1076 | |
1077 | #[inline ] |
1078 | fn sub(self, rhs: bf16) -> Self::Output { |
1079 | (*self).sub(rhs) |
1080 | } |
1081 | } |
1082 | |
1083 | impl SubAssign for bf16 { |
1084 | #[inline ] |
1085 | fn sub_assign(&mut self, rhs: Self) { |
1086 | *self = (*self).sub(rhs); |
1087 | } |
1088 | } |
1089 | |
1090 | impl SubAssign<&bf16> for bf16 { |
1091 | #[inline ] |
1092 | fn sub_assign(&mut self, rhs: &bf16) { |
1093 | *self = (*self).sub(rhs); |
1094 | } |
1095 | } |
1096 | |
1097 | impl Mul for bf16 { |
1098 | type Output = Self; |
1099 | |
1100 | fn mul(self, rhs: Self) -> Self::Output { |
1101 | Self::from_f32(Self::to_f32(self) * Self::to_f32(self:rhs)) |
1102 | } |
1103 | } |
1104 | |
1105 | impl Mul<&bf16> for bf16 { |
1106 | type Output = <bf16 as Mul<bf16>>::Output; |
1107 | |
1108 | #[inline ] |
1109 | fn mul(self, rhs: &bf16) -> Self::Output { |
1110 | self.mul(*rhs) |
1111 | } |
1112 | } |
1113 | |
1114 | impl Mul<&bf16> for &bf16 { |
1115 | type Output = <bf16 as Mul<bf16>>::Output; |
1116 | |
1117 | #[inline ] |
1118 | fn mul(self, rhs: &bf16) -> Self::Output { |
1119 | (*self).mul(*rhs) |
1120 | } |
1121 | } |
1122 | |
1123 | impl Mul<bf16> for &bf16 { |
1124 | type Output = <bf16 as Mul<bf16>>::Output; |
1125 | |
1126 | #[inline ] |
1127 | fn mul(self, rhs: bf16) -> Self::Output { |
1128 | (*self).mul(rhs) |
1129 | } |
1130 | } |
1131 | |
1132 | impl MulAssign for bf16 { |
1133 | #[inline ] |
1134 | fn mul_assign(&mut self, rhs: Self) { |
1135 | *self = (*self).mul(rhs); |
1136 | } |
1137 | } |
1138 | |
1139 | impl MulAssign<&bf16> for bf16 { |
1140 | #[inline ] |
1141 | fn mul_assign(&mut self, rhs: &bf16) { |
1142 | *self = (*self).mul(rhs); |
1143 | } |
1144 | } |
1145 | |
1146 | impl Div for bf16 { |
1147 | type Output = Self; |
1148 | |
1149 | fn div(self, rhs: Self) -> Self::Output { |
1150 | Self::from_f32(Self::to_f32(self) / Self::to_f32(self:rhs)) |
1151 | } |
1152 | } |
1153 | |
1154 | impl Div<&bf16> for bf16 { |
1155 | type Output = <bf16 as Div<bf16>>::Output; |
1156 | |
1157 | #[inline ] |
1158 | fn div(self, rhs: &bf16) -> Self::Output { |
1159 | self.div(*rhs) |
1160 | } |
1161 | } |
1162 | |
1163 | impl Div<&bf16> for &bf16 { |
1164 | type Output = <bf16 as Div<bf16>>::Output; |
1165 | |
1166 | #[inline ] |
1167 | fn div(self, rhs: &bf16) -> Self::Output { |
1168 | (*self).div(*rhs) |
1169 | } |
1170 | } |
1171 | |
1172 | impl Div<bf16> for &bf16 { |
1173 | type Output = <bf16 as Div<bf16>>::Output; |
1174 | |
1175 | #[inline ] |
1176 | fn div(self, rhs: bf16) -> Self::Output { |
1177 | (*self).div(rhs) |
1178 | } |
1179 | } |
1180 | |
1181 | impl DivAssign for bf16 { |
1182 | #[inline ] |
1183 | fn div_assign(&mut self, rhs: Self) { |
1184 | *self = (*self).div(rhs); |
1185 | } |
1186 | } |
1187 | |
1188 | impl DivAssign<&bf16> for bf16 { |
1189 | #[inline ] |
1190 | fn div_assign(&mut self, rhs: &bf16) { |
1191 | *self = (*self).div(rhs); |
1192 | } |
1193 | } |
1194 | |
1195 | impl Rem for bf16 { |
1196 | type Output = Self; |
1197 | |
1198 | fn rem(self, rhs: Self) -> Self::Output { |
1199 | Self::from_f32(Self::to_f32(self) % Self::to_f32(self:rhs)) |
1200 | } |
1201 | } |
1202 | |
1203 | impl Rem<&bf16> for bf16 { |
1204 | type Output = <bf16 as Rem<bf16>>::Output; |
1205 | |
1206 | #[inline ] |
1207 | fn rem(self, rhs: &bf16) -> Self::Output { |
1208 | self.rem(*rhs) |
1209 | } |
1210 | } |
1211 | |
1212 | impl Rem<&bf16> for &bf16 { |
1213 | type Output = <bf16 as Rem<bf16>>::Output; |
1214 | |
1215 | #[inline ] |
1216 | fn rem(self, rhs: &bf16) -> Self::Output { |
1217 | (*self).rem(*rhs) |
1218 | } |
1219 | } |
1220 | |
1221 | impl Rem<bf16> for &bf16 { |
1222 | type Output = <bf16 as Rem<bf16>>::Output; |
1223 | |
1224 | #[inline ] |
1225 | fn rem(self, rhs: bf16) -> Self::Output { |
1226 | (*self).rem(rhs) |
1227 | } |
1228 | } |
1229 | |
1230 | impl RemAssign for bf16 { |
1231 | #[inline ] |
1232 | fn rem_assign(&mut self, rhs: Self) { |
1233 | *self = (*self).rem(rhs); |
1234 | } |
1235 | } |
1236 | |
1237 | impl RemAssign<&bf16> for bf16 { |
1238 | #[inline ] |
1239 | fn rem_assign(&mut self, rhs: &bf16) { |
1240 | *self = (*self).rem(rhs); |
1241 | } |
1242 | } |
1243 | |
1244 | impl Product for bf16 { |
1245 | #[inline ] |
1246 | fn product<I: Iterator<Item = Self>>(iter: I) -> Self { |
1247 | bf16::from_f32(iter.map(|f: bf16| f.to_f32()).product()) |
1248 | } |
1249 | } |
1250 | |
1251 | impl<'a> Product<&'a bf16> for bf16 { |
1252 | #[inline ] |
1253 | fn product<I: Iterator<Item = &'a bf16>>(iter: I) -> Self { |
1254 | bf16::from_f32(iter.map(|f: &bf16| f.to_f32()).product()) |
1255 | } |
1256 | } |
1257 | |
1258 | impl Sum for bf16 { |
1259 | #[inline ] |
1260 | fn sum<I: Iterator<Item = Self>>(iter: I) -> Self { |
1261 | bf16::from_f32(iter.map(|f: bf16| f.to_f32()).sum()) |
1262 | } |
1263 | } |
1264 | |
1265 | impl<'a> Sum<&'a bf16> for bf16 { |
1266 | #[inline ] |
1267 | fn sum<I: Iterator<Item = &'a bf16>>(iter: I) -> Self { |
1268 | bf16::from_f32(iter.map(|f: &bf16| f.to_f32()).sum()) |
1269 | } |
1270 | } |
1271 | |
1272 | #[cfg (feature = "serde" )] |
1273 | struct Visitor; |
1274 | |
1275 | #[cfg (feature = "serde" )] |
1276 | impl<'de> Deserialize<'de> for bf16 { |
1277 | fn deserialize<D>(deserializer: D) -> Result<bf16, D::Error> |
1278 | where |
1279 | D: serde::de::Deserializer<'de>, |
1280 | { |
1281 | deserializer.deserialize_newtype_struct("bf16" , Visitor) |
1282 | } |
1283 | } |
1284 | |
1285 | #[cfg (feature = "serde" )] |
1286 | impl<'de> serde::de::Visitor<'de> for Visitor { |
1287 | type Value = bf16; |
1288 | |
1289 | fn expecting(&self, formatter: &mut core::fmt::Formatter) -> core::fmt::Result { |
1290 | write!(formatter, "tuple struct bf16" ) |
1291 | } |
1292 | |
1293 | fn visit_newtype_struct<D>(self, deserializer: D) -> Result<Self::Value, D::Error> |
1294 | where |
1295 | D: serde::Deserializer<'de>, |
1296 | { |
1297 | Ok(bf16(<u16 as Deserialize>::deserialize(deserializer)?)) |
1298 | } |
1299 | |
1300 | fn visit_str<E>(self, v: &str) -> Result<Self::Value, E> |
1301 | where |
1302 | E: serde::de::Error, |
1303 | { |
1304 | v.parse().map_err(|_| { |
1305 | serde::de::Error::invalid_value(serde::de::Unexpected::Str(v), &"a float string" ) |
1306 | }) |
1307 | } |
1308 | |
1309 | fn visit_f32<E>(self, v: f32) -> Result<Self::Value, E> |
1310 | where |
1311 | E: serde::de::Error, |
1312 | { |
1313 | Ok(bf16::from_f32(v)) |
1314 | } |
1315 | |
1316 | fn visit_f64<E>(self, v: f64) -> Result<Self::Value, E> |
1317 | where |
1318 | E: serde::de::Error, |
1319 | { |
1320 | Ok(bf16::from_f64(v)) |
1321 | } |
1322 | } |
1323 | |
1324 | #[allow ( |
1325 | clippy::cognitive_complexity, |
1326 | clippy::float_cmp, |
1327 | clippy::neg_cmp_op_on_partial_ord |
1328 | )] |
1329 | #[cfg (test)] |
1330 | mod test { |
1331 | use super::*; |
1332 | #[allow (unused_imports)] |
1333 | use core::cmp::Ordering; |
1334 | #[cfg (feature = "num-traits" )] |
1335 | use num_traits::{AsPrimitive, FromPrimitive, ToPrimitive}; |
1336 | use quickcheck_macros::quickcheck; |
1337 | |
1338 | #[cfg (feature = "num-traits" )] |
1339 | #[test ] |
1340 | fn as_primitive() { |
1341 | let two = bf16::from_f32(2.0); |
1342 | assert_eq!(<i32 as AsPrimitive<bf16>>::as_(2), two); |
1343 | assert_eq!(<bf16 as AsPrimitive<i32>>::as_(two), 2); |
1344 | |
1345 | assert_eq!(<f32 as AsPrimitive<bf16>>::as_(2.0), two); |
1346 | assert_eq!(<bf16 as AsPrimitive<f32>>::as_(two), 2.0); |
1347 | |
1348 | assert_eq!(<f64 as AsPrimitive<bf16>>::as_(2.0), two); |
1349 | assert_eq!(<bf16 as AsPrimitive<f64>>::as_(two), 2.0); |
1350 | } |
1351 | |
1352 | #[cfg (feature = "num-traits" )] |
1353 | #[test ] |
1354 | fn to_primitive() { |
1355 | let two = bf16::from_f32(2.0); |
1356 | assert_eq!(ToPrimitive::to_i32(&two).unwrap(), 2i32); |
1357 | assert_eq!(ToPrimitive::to_f32(&two).unwrap(), 2.0f32); |
1358 | assert_eq!(ToPrimitive::to_f64(&two).unwrap(), 2.0f64); |
1359 | } |
1360 | |
1361 | #[cfg (feature = "num-traits" )] |
1362 | #[test ] |
1363 | fn from_primitive() { |
1364 | let two = bf16::from_f32(2.0); |
1365 | assert_eq!(<bf16 as FromPrimitive>::from_i32(2).unwrap(), two); |
1366 | assert_eq!(<bf16 as FromPrimitive>::from_f32(2.0).unwrap(), two); |
1367 | assert_eq!(<bf16 as FromPrimitive>::from_f64(2.0).unwrap(), two); |
1368 | } |
1369 | |
1370 | #[test ] |
1371 | fn test_bf16_consts_from_f32() { |
1372 | let one = bf16::from_f32(1.0); |
1373 | let zero = bf16::from_f32(0.0); |
1374 | let neg_zero = bf16::from_f32(-0.0); |
1375 | let neg_one = bf16::from_f32(-1.0); |
1376 | let inf = bf16::from_f32(core::f32::INFINITY); |
1377 | let neg_inf = bf16::from_f32(core::f32::NEG_INFINITY); |
1378 | let nan = bf16::from_f32(core::f32::NAN); |
1379 | |
1380 | assert_eq!(bf16::ONE, one); |
1381 | assert_eq!(bf16::ZERO, zero); |
1382 | assert!(zero.is_sign_positive()); |
1383 | assert_eq!(bf16::NEG_ZERO, neg_zero); |
1384 | assert!(neg_zero.is_sign_negative()); |
1385 | assert_eq!(bf16::NEG_ONE, neg_one); |
1386 | assert!(neg_one.is_sign_negative()); |
1387 | assert_eq!(bf16::INFINITY, inf); |
1388 | assert_eq!(bf16::NEG_INFINITY, neg_inf); |
1389 | assert!(nan.is_nan()); |
1390 | assert!(bf16::NAN.is_nan()); |
1391 | |
1392 | let e = bf16::from_f32(core::f32::consts::E); |
1393 | let pi = bf16::from_f32(core::f32::consts::PI); |
1394 | let frac_1_pi = bf16::from_f32(core::f32::consts::FRAC_1_PI); |
1395 | let frac_1_sqrt_2 = bf16::from_f32(core::f32::consts::FRAC_1_SQRT_2); |
1396 | let frac_2_pi = bf16::from_f32(core::f32::consts::FRAC_2_PI); |
1397 | let frac_2_sqrt_pi = bf16::from_f32(core::f32::consts::FRAC_2_SQRT_PI); |
1398 | let frac_pi_2 = bf16::from_f32(core::f32::consts::FRAC_PI_2); |
1399 | let frac_pi_3 = bf16::from_f32(core::f32::consts::FRAC_PI_3); |
1400 | let frac_pi_4 = bf16::from_f32(core::f32::consts::FRAC_PI_4); |
1401 | let frac_pi_6 = bf16::from_f32(core::f32::consts::FRAC_PI_6); |
1402 | let frac_pi_8 = bf16::from_f32(core::f32::consts::FRAC_PI_8); |
1403 | let ln_10 = bf16::from_f32(core::f32::consts::LN_10); |
1404 | let ln_2 = bf16::from_f32(core::f32::consts::LN_2); |
1405 | let log10_e = bf16::from_f32(core::f32::consts::LOG10_E); |
1406 | // core::f32::consts::LOG10_2 requires rustc 1.43.0 |
1407 | let log10_2 = bf16::from_f32(2f32.log10()); |
1408 | let log2_e = bf16::from_f32(core::f32::consts::LOG2_E); |
1409 | // core::f32::consts::LOG2_10 requires rustc 1.43.0 |
1410 | let log2_10 = bf16::from_f32(10f32.log2()); |
1411 | let sqrt_2 = bf16::from_f32(core::f32::consts::SQRT_2); |
1412 | |
1413 | assert_eq!(bf16::E, e); |
1414 | assert_eq!(bf16::PI, pi); |
1415 | assert_eq!(bf16::FRAC_1_PI, frac_1_pi); |
1416 | assert_eq!(bf16::FRAC_1_SQRT_2, frac_1_sqrt_2); |
1417 | assert_eq!(bf16::FRAC_2_PI, frac_2_pi); |
1418 | assert_eq!(bf16::FRAC_2_SQRT_PI, frac_2_sqrt_pi); |
1419 | assert_eq!(bf16::FRAC_PI_2, frac_pi_2); |
1420 | assert_eq!(bf16::FRAC_PI_3, frac_pi_3); |
1421 | assert_eq!(bf16::FRAC_PI_4, frac_pi_4); |
1422 | assert_eq!(bf16::FRAC_PI_6, frac_pi_6); |
1423 | assert_eq!(bf16::FRAC_PI_8, frac_pi_8); |
1424 | assert_eq!(bf16::LN_10, ln_10); |
1425 | assert_eq!(bf16::LN_2, ln_2); |
1426 | assert_eq!(bf16::LOG10_E, log10_e); |
1427 | assert_eq!(bf16::LOG10_2, log10_2); |
1428 | assert_eq!(bf16::LOG2_E, log2_e); |
1429 | assert_eq!(bf16::LOG2_10, log2_10); |
1430 | assert_eq!(bf16::SQRT_2, sqrt_2); |
1431 | } |
1432 | |
1433 | #[test ] |
1434 | fn test_bf16_consts_from_f64() { |
1435 | let one = bf16::from_f64(1.0); |
1436 | let zero = bf16::from_f64(0.0); |
1437 | let neg_zero = bf16::from_f64(-0.0); |
1438 | let inf = bf16::from_f64(core::f64::INFINITY); |
1439 | let neg_inf = bf16::from_f64(core::f64::NEG_INFINITY); |
1440 | let nan = bf16::from_f64(core::f64::NAN); |
1441 | |
1442 | assert_eq!(bf16::ONE, one); |
1443 | assert_eq!(bf16::ZERO, zero); |
1444 | assert_eq!(bf16::NEG_ZERO, neg_zero); |
1445 | assert_eq!(bf16::INFINITY, inf); |
1446 | assert_eq!(bf16::NEG_INFINITY, neg_inf); |
1447 | assert!(nan.is_nan()); |
1448 | assert!(bf16::NAN.is_nan()); |
1449 | |
1450 | let e = bf16::from_f64(core::f64::consts::E); |
1451 | let pi = bf16::from_f64(core::f64::consts::PI); |
1452 | let frac_1_pi = bf16::from_f64(core::f64::consts::FRAC_1_PI); |
1453 | let frac_1_sqrt_2 = bf16::from_f64(core::f64::consts::FRAC_1_SQRT_2); |
1454 | let frac_2_pi = bf16::from_f64(core::f64::consts::FRAC_2_PI); |
1455 | let frac_2_sqrt_pi = bf16::from_f64(core::f64::consts::FRAC_2_SQRT_PI); |
1456 | let frac_pi_2 = bf16::from_f64(core::f64::consts::FRAC_PI_2); |
1457 | let frac_pi_3 = bf16::from_f64(core::f64::consts::FRAC_PI_3); |
1458 | let frac_pi_4 = bf16::from_f64(core::f64::consts::FRAC_PI_4); |
1459 | let frac_pi_6 = bf16::from_f64(core::f64::consts::FRAC_PI_6); |
1460 | let frac_pi_8 = bf16::from_f64(core::f64::consts::FRAC_PI_8); |
1461 | let ln_10 = bf16::from_f64(core::f64::consts::LN_10); |
1462 | let ln_2 = bf16::from_f64(core::f64::consts::LN_2); |
1463 | let log10_e = bf16::from_f64(core::f64::consts::LOG10_E); |
1464 | // core::f64::consts::LOG10_2 requires rustc 1.43.0 |
1465 | let log10_2 = bf16::from_f64(2f64.log10()); |
1466 | let log2_e = bf16::from_f64(core::f64::consts::LOG2_E); |
1467 | // core::f64::consts::LOG2_10 requires rustc 1.43.0 |
1468 | let log2_10 = bf16::from_f64(10f64.log2()); |
1469 | let sqrt_2 = bf16::from_f64(core::f64::consts::SQRT_2); |
1470 | |
1471 | assert_eq!(bf16::E, e); |
1472 | assert_eq!(bf16::PI, pi); |
1473 | assert_eq!(bf16::FRAC_1_PI, frac_1_pi); |
1474 | assert_eq!(bf16::FRAC_1_SQRT_2, frac_1_sqrt_2); |
1475 | assert_eq!(bf16::FRAC_2_PI, frac_2_pi); |
1476 | assert_eq!(bf16::FRAC_2_SQRT_PI, frac_2_sqrt_pi); |
1477 | assert_eq!(bf16::FRAC_PI_2, frac_pi_2); |
1478 | assert_eq!(bf16::FRAC_PI_3, frac_pi_3); |
1479 | assert_eq!(bf16::FRAC_PI_4, frac_pi_4); |
1480 | assert_eq!(bf16::FRAC_PI_6, frac_pi_6); |
1481 | assert_eq!(bf16::FRAC_PI_8, frac_pi_8); |
1482 | assert_eq!(bf16::LN_10, ln_10); |
1483 | assert_eq!(bf16::LN_2, ln_2); |
1484 | assert_eq!(bf16::LOG10_E, log10_e); |
1485 | assert_eq!(bf16::LOG10_2, log10_2); |
1486 | assert_eq!(bf16::LOG2_E, log2_e); |
1487 | assert_eq!(bf16::LOG2_10, log2_10); |
1488 | assert_eq!(bf16::SQRT_2, sqrt_2); |
1489 | } |
1490 | |
1491 | #[test ] |
1492 | fn test_nan_conversion_to_smaller() { |
1493 | let nan64 = f64::from_bits(0x7FF0_0000_0000_0001u64); |
1494 | let neg_nan64 = f64::from_bits(0xFFF0_0000_0000_0001u64); |
1495 | let nan32 = f32::from_bits(0x7F80_0001u32); |
1496 | let neg_nan32 = f32::from_bits(0xFF80_0001u32); |
1497 | let nan32_from_64 = nan64 as f32; |
1498 | let neg_nan32_from_64 = neg_nan64 as f32; |
1499 | let nan16_from_64 = bf16::from_f64(nan64); |
1500 | let neg_nan16_from_64 = bf16::from_f64(neg_nan64); |
1501 | let nan16_from_32 = bf16::from_f32(nan32); |
1502 | let neg_nan16_from_32 = bf16::from_f32(neg_nan32); |
1503 | |
1504 | assert!(nan64.is_nan() && nan64.is_sign_positive()); |
1505 | assert!(neg_nan64.is_nan() && neg_nan64.is_sign_negative()); |
1506 | assert!(nan32.is_nan() && nan32.is_sign_positive()); |
1507 | assert!(neg_nan32.is_nan() && neg_nan32.is_sign_negative()); |
1508 | |
1509 | // f32/f64 NaN conversion sign is non-deterministic: https://github.com/starkat99/half-rs/issues/103 |
1510 | assert!(neg_nan32_from_64.is_nan()); |
1511 | assert!(nan32_from_64.is_nan()); |
1512 | assert!(nan16_from_64.is_nan()); |
1513 | assert!(neg_nan16_from_64.is_nan()); |
1514 | assert!(nan16_from_32.is_nan()); |
1515 | assert!(neg_nan16_from_32.is_nan()); |
1516 | } |
1517 | |
1518 | #[test ] |
1519 | fn test_nan_conversion_to_larger() { |
1520 | let nan16 = bf16::from_bits(0x7F81u16); |
1521 | let neg_nan16 = bf16::from_bits(0xFF81u16); |
1522 | let nan32 = f32::from_bits(0x7F80_0001u32); |
1523 | let neg_nan32 = f32::from_bits(0xFF80_0001u32); |
1524 | let nan32_from_16 = f32::from(nan16); |
1525 | let neg_nan32_from_16 = f32::from(neg_nan16); |
1526 | let nan64_from_16 = f64::from(nan16); |
1527 | let neg_nan64_from_16 = f64::from(neg_nan16); |
1528 | let nan64_from_32 = f64::from(nan32); |
1529 | let neg_nan64_from_32 = f64::from(neg_nan32); |
1530 | |
1531 | assert!(nan16.is_nan() && nan16.is_sign_positive()); |
1532 | assert!(neg_nan16.is_nan() && neg_nan16.is_sign_negative()); |
1533 | assert!(nan32.is_nan() && nan32.is_sign_positive()); |
1534 | assert!(neg_nan32.is_nan() && neg_nan32.is_sign_negative()); |
1535 | |
1536 | // // f32/f64 NaN conversion sign is non-deterministic: https://github.com/starkat99/half-rs/issues/103 |
1537 | assert!(nan32_from_16.is_nan()); |
1538 | assert!(neg_nan32_from_16.is_nan()); |
1539 | assert!(nan64_from_16.is_nan()); |
1540 | assert!(neg_nan64_from_16.is_nan()); |
1541 | assert!(nan64_from_32.is_nan()); |
1542 | assert!(neg_nan64_from_32.is_nan()); |
1543 | } |
1544 | |
1545 | #[test ] |
1546 | fn test_bf16_to_f32() { |
1547 | let f = bf16::from_f32(7.0); |
1548 | assert_eq!(f.to_f32(), 7.0f32); |
1549 | |
1550 | // 7.1 is NOT exactly representable in 16-bit, it's rounded |
1551 | let f = bf16::from_f32(7.1); |
1552 | let diff = (f.to_f32() - 7.1f32).abs(); |
1553 | // diff must be <= 4 * EPSILON, as 7 has two more significant bits than 1 |
1554 | assert!(diff <= 4.0 * bf16::EPSILON.to_f32()); |
1555 | |
1556 | let tiny32 = f32::from_bits(0x0001_0000u32); |
1557 | assert_eq!(bf16::from_bits(0x0001).to_f32(), tiny32); |
1558 | assert_eq!(bf16::from_bits(0x0005).to_f32(), 5.0 * tiny32); |
1559 | |
1560 | assert_eq!(bf16::from_bits(0x0001), bf16::from_f32(tiny32)); |
1561 | assert_eq!(bf16::from_bits(0x0005), bf16::from_f32(5.0 * tiny32)); |
1562 | } |
1563 | |
1564 | #[test ] |
1565 | fn test_bf16_to_f64() { |
1566 | let f = bf16::from_f64(7.0); |
1567 | assert_eq!(f.to_f64(), 7.0f64); |
1568 | |
1569 | // 7.1 is NOT exactly representable in 16-bit, it's rounded |
1570 | let f = bf16::from_f64(7.1); |
1571 | let diff = (f.to_f64() - 7.1f64).abs(); |
1572 | // diff must be <= 4 * EPSILON, as 7 has two more significant bits than 1 |
1573 | assert!(diff <= 4.0 * bf16::EPSILON.to_f64()); |
1574 | |
1575 | let tiny64 = 2.0f64.powi(-133); |
1576 | assert_eq!(bf16::from_bits(0x0001).to_f64(), tiny64); |
1577 | assert_eq!(bf16::from_bits(0x0005).to_f64(), 5.0 * tiny64); |
1578 | |
1579 | assert_eq!(bf16::from_bits(0x0001), bf16::from_f64(tiny64)); |
1580 | assert_eq!(bf16::from_bits(0x0005), bf16::from_f64(5.0 * tiny64)); |
1581 | } |
1582 | |
1583 | #[test ] |
1584 | fn test_comparisons() { |
1585 | let zero = bf16::from_f64(0.0); |
1586 | let one = bf16::from_f64(1.0); |
1587 | let neg_zero = bf16::from_f64(-0.0); |
1588 | let neg_one = bf16::from_f64(-1.0); |
1589 | |
1590 | assert_eq!(zero.partial_cmp(&neg_zero), Some(Ordering::Equal)); |
1591 | assert_eq!(neg_zero.partial_cmp(&zero), Some(Ordering::Equal)); |
1592 | assert!(zero == neg_zero); |
1593 | assert!(neg_zero == zero); |
1594 | assert!(!(zero != neg_zero)); |
1595 | assert!(!(neg_zero != zero)); |
1596 | assert!(!(zero < neg_zero)); |
1597 | assert!(!(neg_zero < zero)); |
1598 | assert!(zero <= neg_zero); |
1599 | assert!(neg_zero <= zero); |
1600 | assert!(!(zero > neg_zero)); |
1601 | assert!(!(neg_zero > zero)); |
1602 | assert!(zero >= neg_zero); |
1603 | assert!(neg_zero >= zero); |
1604 | |
1605 | assert_eq!(one.partial_cmp(&neg_zero), Some(Ordering::Greater)); |
1606 | assert_eq!(neg_zero.partial_cmp(&one), Some(Ordering::Less)); |
1607 | assert!(!(one == neg_zero)); |
1608 | assert!(!(neg_zero == one)); |
1609 | assert!(one != neg_zero); |
1610 | assert!(neg_zero != one); |
1611 | assert!(!(one < neg_zero)); |
1612 | assert!(neg_zero < one); |
1613 | assert!(!(one <= neg_zero)); |
1614 | assert!(neg_zero <= one); |
1615 | assert!(one > neg_zero); |
1616 | assert!(!(neg_zero > one)); |
1617 | assert!(one >= neg_zero); |
1618 | assert!(!(neg_zero >= one)); |
1619 | |
1620 | assert_eq!(one.partial_cmp(&neg_one), Some(Ordering::Greater)); |
1621 | assert_eq!(neg_one.partial_cmp(&one), Some(Ordering::Less)); |
1622 | assert!(!(one == neg_one)); |
1623 | assert!(!(neg_one == one)); |
1624 | assert!(one != neg_one); |
1625 | assert!(neg_one != one); |
1626 | assert!(!(one < neg_one)); |
1627 | assert!(neg_one < one); |
1628 | assert!(!(one <= neg_one)); |
1629 | assert!(neg_one <= one); |
1630 | assert!(one > neg_one); |
1631 | assert!(!(neg_one > one)); |
1632 | assert!(one >= neg_one); |
1633 | assert!(!(neg_one >= one)); |
1634 | } |
1635 | |
1636 | #[test ] |
1637 | #[allow (clippy::erasing_op, clippy::identity_op)] |
1638 | fn round_to_even_f32() { |
1639 | // smallest positive subnormal = 0b0.0000_001 * 2^-126 = 2^-133 |
1640 | let min_sub = bf16::from_bits(1); |
1641 | let min_sub_f = (-133f32).exp2(); |
1642 | assert_eq!(bf16::from_f32(min_sub_f).to_bits(), min_sub.to_bits()); |
1643 | assert_eq!(f32::from(min_sub).to_bits(), min_sub_f.to_bits()); |
1644 | |
1645 | // 0.0000000_011111 rounded to 0.0000000 (< tie, no rounding) |
1646 | // 0.0000000_100000 rounded to 0.0000000 (tie and even, remains at even) |
1647 | // 0.0000000_100001 rounded to 0.0000001 (> tie, rounds up) |
1648 | assert_eq!( |
1649 | bf16::from_f32(min_sub_f * 0.49).to_bits(), |
1650 | min_sub.to_bits() * 0 |
1651 | ); |
1652 | assert_eq!( |
1653 | bf16::from_f32(min_sub_f * 0.50).to_bits(), |
1654 | min_sub.to_bits() * 0 |
1655 | ); |
1656 | assert_eq!( |
1657 | bf16::from_f32(min_sub_f * 0.51).to_bits(), |
1658 | min_sub.to_bits() * 1 |
1659 | ); |
1660 | |
1661 | // 0.0000001_011111 rounded to 0.0000001 (< tie, no rounding) |
1662 | // 0.0000001_100000 rounded to 0.0000010 (tie and odd, rounds up to even) |
1663 | // 0.0000001_100001 rounded to 0.0000010 (> tie, rounds up) |
1664 | assert_eq!( |
1665 | bf16::from_f32(min_sub_f * 1.49).to_bits(), |
1666 | min_sub.to_bits() * 1 |
1667 | ); |
1668 | assert_eq!( |
1669 | bf16::from_f32(min_sub_f * 1.50).to_bits(), |
1670 | min_sub.to_bits() * 2 |
1671 | ); |
1672 | assert_eq!( |
1673 | bf16::from_f32(min_sub_f * 1.51).to_bits(), |
1674 | min_sub.to_bits() * 2 |
1675 | ); |
1676 | |
1677 | // 0.0000010_011111 rounded to 0.0000010 (< tie, no rounding) |
1678 | // 0.0000010_100000 rounded to 0.0000010 (tie and even, remains at even) |
1679 | // 0.0000010_100001 rounded to 0.0000011 (> tie, rounds up) |
1680 | assert_eq!( |
1681 | bf16::from_f32(min_sub_f * 2.49).to_bits(), |
1682 | min_sub.to_bits() * 2 |
1683 | ); |
1684 | assert_eq!( |
1685 | bf16::from_f32(min_sub_f * 2.50).to_bits(), |
1686 | min_sub.to_bits() * 2 |
1687 | ); |
1688 | assert_eq!( |
1689 | bf16::from_f32(min_sub_f * 2.51).to_bits(), |
1690 | min_sub.to_bits() * 3 |
1691 | ); |
1692 | |
1693 | assert_eq!( |
1694 | bf16::from_f32(250.49f32).to_bits(), |
1695 | bf16::from_f32(250.0).to_bits() |
1696 | ); |
1697 | assert_eq!( |
1698 | bf16::from_f32(250.50f32).to_bits(), |
1699 | bf16::from_f32(250.0).to_bits() |
1700 | ); |
1701 | assert_eq!( |
1702 | bf16::from_f32(250.51f32).to_bits(), |
1703 | bf16::from_f32(251.0).to_bits() |
1704 | ); |
1705 | assert_eq!( |
1706 | bf16::from_f32(251.49f32).to_bits(), |
1707 | bf16::from_f32(251.0).to_bits() |
1708 | ); |
1709 | assert_eq!( |
1710 | bf16::from_f32(251.50f32).to_bits(), |
1711 | bf16::from_f32(252.0).to_bits() |
1712 | ); |
1713 | assert_eq!( |
1714 | bf16::from_f32(251.51f32).to_bits(), |
1715 | bf16::from_f32(252.0).to_bits() |
1716 | ); |
1717 | assert_eq!( |
1718 | bf16::from_f32(252.49f32).to_bits(), |
1719 | bf16::from_f32(252.0).to_bits() |
1720 | ); |
1721 | assert_eq!( |
1722 | bf16::from_f32(252.50f32).to_bits(), |
1723 | bf16::from_f32(252.0).to_bits() |
1724 | ); |
1725 | assert_eq!( |
1726 | bf16::from_f32(252.51f32).to_bits(), |
1727 | bf16::from_f32(253.0).to_bits() |
1728 | ); |
1729 | } |
1730 | |
1731 | #[test ] |
1732 | #[allow (clippy::erasing_op, clippy::identity_op)] |
1733 | fn round_to_even_f64() { |
1734 | // smallest positive subnormal = 0b0.0000_001 * 2^-126 = 2^-133 |
1735 | let min_sub = bf16::from_bits(1); |
1736 | let min_sub_f = (-133f64).exp2(); |
1737 | assert_eq!(bf16::from_f64(min_sub_f).to_bits(), min_sub.to_bits()); |
1738 | assert_eq!(f64::from(min_sub).to_bits(), min_sub_f.to_bits()); |
1739 | |
1740 | // 0.0000000_011111 rounded to 0.0000000 (< tie, no rounding) |
1741 | // 0.0000000_100000 rounded to 0.0000000 (tie and even, remains at even) |
1742 | // 0.0000000_100001 rounded to 0.0000001 (> tie, rounds up) |
1743 | assert_eq!( |
1744 | bf16::from_f64(min_sub_f * 0.49).to_bits(), |
1745 | min_sub.to_bits() * 0 |
1746 | ); |
1747 | assert_eq!( |
1748 | bf16::from_f64(min_sub_f * 0.50).to_bits(), |
1749 | min_sub.to_bits() * 0 |
1750 | ); |
1751 | assert_eq!( |
1752 | bf16::from_f64(min_sub_f * 0.51).to_bits(), |
1753 | min_sub.to_bits() * 1 |
1754 | ); |
1755 | |
1756 | // 0.0000001_011111 rounded to 0.0000001 (< tie, no rounding) |
1757 | // 0.0000001_100000 rounded to 0.0000010 (tie and odd, rounds up to even) |
1758 | // 0.0000001_100001 rounded to 0.0000010 (> tie, rounds up) |
1759 | assert_eq!( |
1760 | bf16::from_f64(min_sub_f * 1.49).to_bits(), |
1761 | min_sub.to_bits() * 1 |
1762 | ); |
1763 | assert_eq!( |
1764 | bf16::from_f64(min_sub_f * 1.50).to_bits(), |
1765 | min_sub.to_bits() * 2 |
1766 | ); |
1767 | assert_eq!( |
1768 | bf16::from_f64(min_sub_f * 1.51).to_bits(), |
1769 | min_sub.to_bits() * 2 |
1770 | ); |
1771 | |
1772 | // 0.0000010_011111 rounded to 0.0000010 (< tie, no rounding) |
1773 | // 0.0000010_100000 rounded to 0.0000010 (tie and even, remains at even) |
1774 | // 0.0000010_100001 rounded to 0.0000011 (> tie, rounds up) |
1775 | assert_eq!( |
1776 | bf16::from_f64(min_sub_f * 2.49).to_bits(), |
1777 | min_sub.to_bits() * 2 |
1778 | ); |
1779 | assert_eq!( |
1780 | bf16::from_f64(min_sub_f * 2.50).to_bits(), |
1781 | min_sub.to_bits() * 2 |
1782 | ); |
1783 | assert_eq!( |
1784 | bf16::from_f64(min_sub_f * 2.51).to_bits(), |
1785 | min_sub.to_bits() * 3 |
1786 | ); |
1787 | |
1788 | assert_eq!( |
1789 | bf16::from_f64(250.49f64).to_bits(), |
1790 | bf16::from_f64(250.0).to_bits() |
1791 | ); |
1792 | assert_eq!( |
1793 | bf16::from_f64(250.50f64).to_bits(), |
1794 | bf16::from_f64(250.0).to_bits() |
1795 | ); |
1796 | assert_eq!( |
1797 | bf16::from_f64(250.51f64).to_bits(), |
1798 | bf16::from_f64(251.0).to_bits() |
1799 | ); |
1800 | assert_eq!( |
1801 | bf16::from_f64(251.49f64).to_bits(), |
1802 | bf16::from_f64(251.0).to_bits() |
1803 | ); |
1804 | assert_eq!( |
1805 | bf16::from_f64(251.50f64).to_bits(), |
1806 | bf16::from_f64(252.0).to_bits() |
1807 | ); |
1808 | assert_eq!( |
1809 | bf16::from_f64(251.51f64).to_bits(), |
1810 | bf16::from_f64(252.0).to_bits() |
1811 | ); |
1812 | assert_eq!( |
1813 | bf16::from_f64(252.49f64).to_bits(), |
1814 | bf16::from_f64(252.0).to_bits() |
1815 | ); |
1816 | assert_eq!( |
1817 | bf16::from_f64(252.50f64).to_bits(), |
1818 | bf16::from_f64(252.0).to_bits() |
1819 | ); |
1820 | assert_eq!( |
1821 | bf16::from_f64(252.51f64).to_bits(), |
1822 | bf16::from_f64(253.0).to_bits() |
1823 | ); |
1824 | } |
1825 | |
1826 | #[cfg (feature = "std" )] |
1827 | #[test ] |
1828 | fn formatting() { |
1829 | let f = bf16::from_f32(0.1152344); |
1830 | |
1831 | assert_eq!(format!(" {:.3}" , f), "0.115" ); |
1832 | assert_eq!(format!(" {:.4}" , f), "0.1152" ); |
1833 | assert_eq!(format!(" {:+.4}" , f), "+0.1152" ); |
1834 | assert_eq!(format!(" {:>+10.4}" , f), " +0.1152" ); |
1835 | |
1836 | assert_eq!(format!(" {:.3?}" , f), "0.115" ); |
1837 | assert_eq!(format!(" {:.4?}" , f), "0.1152" ); |
1838 | assert_eq!(format!(" {:+.4?}" , f), "+0.1152" ); |
1839 | assert_eq!(format!(" {:>+10.4?}" , f), " +0.1152" ); |
1840 | } |
1841 | |
1842 | impl quickcheck::Arbitrary for bf16 { |
1843 | fn arbitrary(g: &mut quickcheck::Gen) -> Self { |
1844 | bf16(u16::arbitrary(g)) |
1845 | } |
1846 | } |
1847 | |
1848 | #[quickcheck] |
1849 | fn qc_roundtrip_bf16_f32_is_identity(f: bf16) -> bool { |
1850 | let roundtrip = bf16::from_f32(f.to_f32()); |
1851 | if f.is_nan() { |
1852 | roundtrip.is_nan() && f.is_sign_negative() == roundtrip.is_sign_negative() |
1853 | } else { |
1854 | f.0 == roundtrip.0 |
1855 | } |
1856 | } |
1857 | |
1858 | #[quickcheck] |
1859 | fn qc_roundtrip_bf16_f64_is_identity(f: bf16) -> bool { |
1860 | let roundtrip = bf16::from_f64(f.to_f64()); |
1861 | if f.is_nan() { |
1862 | roundtrip.is_nan() && f.is_sign_negative() == roundtrip.is_sign_negative() |
1863 | } else { |
1864 | f.0 == roundtrip.0 |
1865 | } |
1866 | } |
1867 | } |
1868 | |