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::{FromBytes, Immutable, IntoBytes, KnownLayout};
|
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`]. [`struct@bf16`] has approximately the same dynamic range as [`f32`] by
|
31 | /// having a lower precision than [`struct@f16`][crate::f16]. While [`struct@f16`][crate::f16] has a precision of
|
32 | /// 11 bits, [`struct@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" , rkyv(resolver = Bf16Resolver))]
|
44 | #[cfg_attr (feature = "bytemuck" , derive(Zeroable, Pod))]
|
45 | #[cfg_attr (
|
46 | feature = "zerocopy" ,
|
47 | derive(FromBytes, Immutable, IntoBytes, KnownLayout)
|
48 | )]
|
49 | #[cfg_attr (kani, derive(kani::Arbitrary))]
|
50 | pub struct bf16(u16);
|
51 |
|
52 | impl bf16 {
|
53 | /// Constructs a [`struct@bf16`] value from the raw bits.
|
54 | #[inline ]
|
55 | #[must_use ]
|
56 | pub const fn from_bits(bits: u16) -> bf16 {
|
57 | bf16(bits)
|
58 | }
|
59 |
|
60 | /// Constructs a [`struct@bf16`] value from a 32-bit floating point value.
|
61 | ///
|
62 | /// This operation is lossy. If the 32-bit value is too large to fit, ±∞ will result. NaN values
|
63 | /// are preserved. Subnormal values that are too tiny to be represented will result in ±0. All
|
64 | /// other values are truncated and rounded to the nearest representable value.
|
65 | #[inline ]
|
66 | #[must_use ]
|
67 | pub fn from_f32(value: f32) -> bf16 {
|
68 | Self::from_f32_const(value)
|
69 | }
|
70 |
|
71 | /// Constructs a [`struct@bf16`] value from a 32-bit floating point value.
|
72 | ///
|
73 | /// This function is identical to [`from_f32`][Self::from_f32] except it never uses hardware
|
74 | /// intrinsics, which allows it to be `const`. [`from_f32`][Self::from_f32] should be preferred
|
75 | /// in any non-`const` context.
|
76 | ///
|
77 | /// This operation is lossy. If the 32-bit value is too large to fit, ±∞ will result. NaN values
|
78 | /// are preserved. Subnormal values that are too tiny to be represented will result in ±0. All
|
79 | /// other values are truncated and rounded to the nearest representable value.
|
80 | #[inline ]
|
81 | #[must_use ]
|
82 | pub const fn from_f32_const(value: f32) -> bf16 {
|
83 | bf16(convert::f32_to_bf16(value))
|
84 | }
|
85 |
|
86 | /// Constructs a [`struct@bf16`] value from a 64-bit floating point value.
|
87 | ///
|
88 | /// This operation is lossy. If the 64-bit value is to large to fit, ±∞ will result. NaN values
|
89 | /// are preserved. 64-bit subnormal values are too tiny to be represented and result in ±0.
|
90 | /// Exponents that underflow the minimum exponent will result in subnormals or ±0. All other
|
91 | /// values are truncated and rounded to the nearest representable value.
|
92 | #[inline ]
|
93 | #[must_use ]
|
94 | pub fn from_f64(value: f64) -> bf16 {
|
95 | Self::from_f64_const(value)
|
96 | }
|
97 |
|
98 | /// Constructs a [`struct@bf16`] value from a 64-bit floating point value.
|
99 | ///
|
100 | /// This function is identical to [`from_f64`][Self::from_f64] except it never uses hardware
|
101 | /// intrinsics, which allows it to be `const`. [`from_f64`][Self::from_f64] should be preferred
|
102 | /// in any non-`const` context.
|
103 | ///
|
104 | /// This operation is lossy. If the 64-bit value is to large to fit, ±∞ will result. NaN values
|
105 | /// are preserved. 64-bit subnormal values are too tiny to be represented and result in ±0.
|
106 | /// Exponents that underflow the minimum exponent will result in subnormals or ±0. All other
|
107 | /// values are truncated and rounded to the nearest representable value.
|
108 | #[inline ]
|
109 | #[must_use ]
|
110 | pub const fn from_f64_const(value: f64) -> bf16 {
|
111 | bf16(convert::f64_to_bf16(value))
|
112 | }
|
113 |
|
114 | /// Converts a [`struct@bf16`] into the underlying bit representation.
|
115 | #[inline ]
|
116 | #[must_use ]
|
117 | pub const fn to_bits(self) -> u16 {
|
118 | self.0
|
119 | }
|
120 |
|
121 | /// Returns the memory representation of the underlying bit representation as a byte array in
|
122 | /// little-endian byte order.
|
123 | ///
|
124 | /// # Examples
|
125 | ///
|
126 | /// ```rust
|
127 | /// # use half::prelude::*;
|
128 | /// let bytes = bf16::from_f32(12.5).to_le_bytes();
|
129 | /// assert_eq!(bytes, [0x48, 0x41]);
|
130 | /// ```
|
131 | #[inline ]
|
132 | #[must_use ]
|
133 | pub const fn to_le_bytes(self) -> [u8; 2] {
|
134 | self.0.to_le_bytes()
|
135 | }
|
136 |
|
137 | /// Returns the memory representation of the underlying bit representation as a byte array in
|
138 | /// big-endian (network) byte order.
|
139 | ///
|
140 | /// # Examples
|
141 | ///
|
142 | /// ```rust
|
143 | /// # use half::prelude::*;
|
144 | /// let bytes = bf16::from_f32(12.5).to_be_bytes();
|
145 | /// assert_eq!(bytes, [0x41, 0x48]);
|
146 | /// ```
|
147 | #[inline ]
|
148 | #[must_use ]
|
149 | pub const fn to_be_bytes(self) -> [u8; 2] {
|
150 | self.0.to_be_bytes()
|
151 | }
|
152 |
|
153 | /// Returns the memory representation of the underlying bit representation as a byte array in
|
154 | /// native byte order.
|
155 | ///
|
156 | /// As the target platform's native endianness is used, portable code should use
|
157 | /// [`to_be_bytes`][bf16::to_be_bytes] or [`to_le_bytes`][bf16::to_le_bytes], as appropriate,
|
158 | /// instead.
|
159 | ///
|
160 | /// # Examples
|
161 | ///
|
162 | /// ```rust
|
163 | /// # use half::prelude::*;
|
164 | /// let bytes = bf16::from_f32(12.5).to_ne_bytes();
|
165 | /// assert_eq!(bytes, if cfg!(target_endian = "big" ) {
|
166 | /// [0x41, 0x48]
|
167 | /// } else {
|
168 | /// [0x48, 0x41]
|
169 | /// });
|
170 | /// ```
|
171 | #[inline ]
|
172 | #[must_use ]
|
173 | pub const fn to_ne_bytes(self) -> [u8; 2] {
|
174 | self.0.to_ne_bytes()
|
175 | }
|
176 |
|
177 | /// Creates a floating point value from its representation as a byte array in little endian.
|
178 | ///
|
179 | /// # Examples
|
180 | ///
|
181 | /// ```rust
|
182 | /// # use half::prelude::*;
|
183 | /// let value = bf16::from_le_bytes([0x48, 0x41]);
|
184 | /// assert_eq!(value, bf16::from_f32(12.5));
|
185 | /// ```
|
186 | #[inline ]
|
187 | #[must_use ]
|
188 | pub const fn from_le_bytes(bytes: [u8; 2]) -> bf16 {
|
189 | bf16::from_bits(u16::from_le_bytes(bytes))
|
190 | }
|
191 |
|
192 | /// Creates a floating point value from its representation as a byte array in big endian.
|
193 | ///
|
194 | /// # Examples
|
195 | ///
|
196 | /// ```rust
|
197 | /// # use half::prelude::*;
|
198 | /// let value = bf16::from_be_bytes([0x41, 0x48]);
|
199 | /// assert_eq!(value, bf16::from_f32(12.5));
|
200 | /// ```
|
201 | #[inline ]
|
202 | #[must_use ]
|
203 | pub const fn from_be_bytes(bytes: [u8; 2]) -> bf16 {
|
204 | bf16::from_bits(u16::from_be_bytes(bytes))
|
205 | }
|
206 |
|
207 | /// Creates a floating point value from its representation as a byte array in native endian.
|
208 | ///
|
209 | /// As the target platform's native endianness is used, portable code likely wants to use
|
210 | /// [`from_be_bytes`][bf16::from_be_bytes] or [`from_le_bytes`][bf16::from_le_bytes], as
|
211 | /// appropriate instead.
|
212 | ///
|
213 | /// # Examples
|
214 | ///
|
215 | /// ```rust
|
216 | /// # use half::prelude::*;
|
217 | /// let value = bf16::from_ne_bytes(if cfg!(target_endian = "big" ) {
|
218 | /// [0x41, 0x48]
|
219 | /// } else {
|
220 | /// [0x48, 0x41]
|
221 | /// });
|
222 | /// assert_eq!(value, bf16::from_f32(12.5));
|
223 | /// ```
|
224 | #[inline ]
|
225 | #[must_use ]
|
226 | pub const fn from_ne_bytes(bytes: [u8; 2]) -> bf16 {
|
227 | bf16::from_bits(u16::from_ne_bytes(bytes))
|
228 | }
|
229 |
|
230 | /// Converts a [`struct@bf16`] value into an [`f32`] value.
|
231 | ///
|
232 | /// This conversion is lossless as all values can be represented exactly in [`f32`].
|
233 | #[inline ]
|
234 | #[must_use ]
|
235 | pub fn to_f32(self) -> f32 {
|
236 | self.to_f32_const()
|
237 | }
|
238 |
|
239 | /// Converts a [`struct@bf16`] value into an [`f32`] value.
|
240 | ///
|
241 | /// This function is identical to [`to_f32`][Self::to_f32] except it never uses hardware
|
242 | /// intrinsics, which allows it to be `const`. [`to_f32`][Self::to_f32] should be preferred
|
243 | /// in any non-`const` context.
|
244 | ///
|
245 | /// This conversion is lossless as all values can be represented exactly in [`f32`].
|
246 | #[inline ]
|
247 | #[must_use ]
|
248 | pub const fn to_f32_const(self) -> f32 {
|
249 | convert::bf16_to_f32(self.0)
|
250 | }
|
251 |
|
252 | /// Converts a [`struct@bf16`] value into an [`f64`] value.
|
253 | ///
|
254 | /// This conversion is lossless as all values can be represented exactly in [`f64`].
|
255 | #[inline ]
|
256 | #[must_use ]
|
257 | pub fn to_f64(self) -> f64 {
|
258 | self.to_f64_const()
|
259 | }
|
260 |
|
261 | /// Converts a [`struct@bf16`] value into an [`f64`] value.
|
262 | ///
|
263 | /// This function is identical to [`to_f64`][Self::to_f64] except it never uses hardware
|
264 | /// intrinsics, which allows it to be `const`. [`to_f64`][Self::to_f64] should be preferred
|
265 | /// in any non-`const` context.
|
266 | ///
|
267 | /// This conversion is lossless as all values can be represented exactly in [`f64`].
|
268 | #[inline ]
|
269 | #[must_use ]
|
270 | pub const fn to_f64_const(self) -> f64 {
|
271 | convert::bf16_to_f64(self.0)
|
272 | }
|
273 |
|
274 | /// Returns `true` if this value is NaN and `false` otherwise.
|
275 | ///
|
276 | /// # Examples
|
277 | ///
|
278 | /// ```rust
|
279 | /// # use half::prelude::*;
|
280 | ///
|
281 | /// let nan = bf16::NAN;
|
282 | /// let f = bf16::from_f32(7.0_f32);
|
283 | ///
|
284 | /// assert!(nan.is_nan());
|
285 | /// assert!(!f.is_nan());
|
286 | /// ```
|
287 | #[inline ]
|
288 | #[must_use ]
|
289 | pub const fn is_nan(self) -> bool {
|
290 | self.0 & 0x7FFFu16 > 0x7F80u16
|
291 | }
|
292 |
|
293 | /// Returns `true` if this value is ±∞ and `false` otherwise.
|
294 | ///
|
295 | /// # Examples
|
296 | ///
|
297 | /// ```rust
|
298 | /// # use half::prelude::*;
|
299 | ///
|
300 | /// let f = bf16::from_f32(7.0f32);
|
301 | /// let inf = bf16::INFINITY;
|
302 | /// let neg_inf = bf16::NEG_INFINITY;
|
303 | /// let nan = bf16::NAN;
|
304 | ///
|
305 | /// assert!(!f.is_infinite());
|
306 | /// assert!(!nan.is_infinite());
|
307 | ///
|
308 | /// assert!(inf.is_infinite());
|
309 | /// assert!(neg_inf.is_infinite());
|
310 | /// ```
|
311 | #[inline ]
|
312 | #[must_use ]
|
313 | pub const fn is_infinite(self) -> bool {
|
314 | self.0 & 0x7FFFu16 == 0x7F80u16
|
315 | }
|
316 |
|
317 | /// Returns `true` if this number is neither infinite nor NaN.
|
318 | ///
|
319 | /// # Examples
|
320 | ///
|
321 | /// ```rust
|
322 | /// # use half::prelude::*;
|
323 | ///
|
324 | /// let f = bf16::from_f32(7.0f32);
|
325 | /// let inf = bf16::INFINITY;
|
326 | /// let neg_inf = bf16::NEG_INFINITY;
|
327 | /// let nan = bf16::NAN;
|
328 | ///
|
329 | /// assert!(f.is_finite());
|
330 | ///
|
331 | /// assert!(!nan.is_finite());
|
332 | /// assert!(!inf.is_finite());
|
333 | /// assert!(!neg_inf.is_finite());
|
334 | /// ```
|
335 | #[inline ]
|
336 | #[must_use ]
|
337 | pub const fn is_finite(self) -> bool {
|
338 | self.0 & 0x7F80u16 != 0x7F80u16
|
339 | }
|
340 |
|
341 | /// Returns `true` if the number is neither zero, infinite, subnormal, or NaN.
|
342 | ///
|
343 | /// # Examples
|
344 | ///
|
345 | /// ```rust
|
346 | /// # use half::prelude::*;
|
347 | ///
|
348 | /// let min = bf16::MIN_POSITIVE;
|
349 | /// let max = bf16::MAX;
|
350 | /// let lower_than_min = bf16::from_f32(1.0e-39_f32);
|
351 | /// let zero = bf16::from_f32(0.0_f32);
|
352 | ///
|
353 | /// assert!(min.is_normal());
|
354 | /// assert!(max.is_normal());
|
355 | ///
|
356 | /// assert!(!zero.is_normal());
|
357 | /// assert!(!bf16::NAN.is_normal());
|
358 | /// assert!(!bf16::INFINITY.is_normal());
|
359 | /// // Values between 0 and `min` are subnormal.
|
360 | /// assert!(!lower_than_min.is_normal());
|
361 | /// ```
|
362 | #[inline ]
|
363 | #[must_use ]
|
364 | pub const fn is_normal(self) -> bool {
|
365 | let exp = self.0 & 0x7F80u16;
|
366 | exp != 0x7F80u16 && exp != 0
|
367 | }
|
368 |
|
369 | /// Returns the floating point category of the number.
|
370 | ///
|
371 | /// If only one property is going to be tested, it is generally faster to use the specific
|
372 | /// predicate instead.
|
373 | ///
|
374 | /// # Examples
|
375 | ///
|
376 | /// ```rust
|
377 | /// use std::num::FpCategory;
|
378 | /// # use half::prelude::*;
|
379 | ///
|
380 | /// let num = bf16::from_f32(12.4_f32);
|
381 | /// let inf = bf16::INFINITY;
|
382 | ///
|
383 | /// assert_eq!(num.classify(), FpCategory::Normal);
|
384 | /// assert_eq!(inf.classify(), FpCategory::Infinite);
|
385 | /// ```
|
386 | #[must_use ]
|
387 | pub const fn classify(self) -> FpCategory {
|
388 | let exp = self.0 & 0x7F80u16;
|
389 | let man = self.0 & 0x007Fu16;
|
390 | match (exp, man) {
|
391 | (0, 0) => FpCategory::Zero,
|
392 | (0, _) => FpCategory::Subnormal,
|
393 | (0x7F80u16, 0) => FpCategory::Infinite,
|
394 | (0x7F80u16, _) => FpCategory::Nan,
|
395 | _ => FpCategory::Normal,
|
396 | }
|
397 | }
|
398 |
|
399 | /// Returns a number that represents the sign of `self`.
|
400 | ///
|
401 | /// * 1.0 if the number is positive, +0.0 or [`INFINITY`][bf16::INFINITY]
|
402 | /// * −1.0 if the number is negative, −0.0` or [`NEG_INFINITY`][bf16::NEG_INFINITY]
|
403 | /// * [`NAN`][bf16::NAN] if the number is NaN
|
404 | ///
|
405 | /// # Examples
|
406 | ///
|
407 | /// ```rust
|
408 | /// # use half::prelude::*;
|
409 | ///
|
410 | /// let f = bf16::from_f32(3.5_f32);
|
411 | ///
|
412 | /// assert_eq!(f.signum(), bf16::from_f32(1.0));
|
413 | /// assert_eq!(bf16::NEG_INFINITY.signum(), bf16::from_f32(-1.0));
|
414 | ///
|
415 | /// assert!(bf16::NAN.signum().is_nan());
|
416 | /// ```
|
417 | #[must_use ]
|
418 | pub const fn signum(self) -> bf16 {
|
419 | if self.is_nan() {
|
420 | self
|
421 | } else if self.0 & 0x8000u16 != 0 {
|
422 | Self::NEG_ONE
|
423 | } else {
|
424 | Self::ONE
|
425 | }
|
426 | }
|
427 |
|
428 | /// Returns `true` if and only if `self` has a positive sign, including +0.0, NaNs with a
|
429 | /// positive sign bit and +∞.
|
430 | ///
|
431 | /// # Examples
|
432 | ///
|
433 | /// ```rust
|
434 | /// # use half::prelude::*;
|
435 | ///
|
436 | /// let nan = bf16::NAN;
|
437 | /// let f = bf16::from_f32(7.0_f32);
|
438 | /// let g = bf16::from_f32(-7.0_f32);
|
439 | ///
|
440 | /// assert!(f.is_sign_positive());
|
441 | /// assert!(!g.is_sign_positive());
|
442 | /// // NaN can be either positive or negative
|
443 | /// assert!(nan.is_sign_positive() != nan.is_sign_negative());
|
444 | /// ```
|
445 | #[inline ]
|
446 | #[must_use ]
|
447 | pub const fn is_sign_positive(self) -> bool {
|
448 | self.0 & 0x8000u16 == 0
|
449 | }
|
450 |
|
451 | /// Returns `true` if and only if `self` has a negative sign, including −0.0, NaNs with a
|
452 | /// negative sign bit and −∞.
|
453 | ///
|
454 | /// # Examples
|
455 | ///
|
456 | /// ```rust
|
457 | /// # use half::prelude::*;
|
458 | ///
|
459 | /// let nan = bf16::NAN;
|
460 | /// let f = bf16::from_f32(7.0f32);
|
461 | /// let g = bf16::from_f32(-7.0f32);
|
462 | ///
|
463 | /// assert!(!f.is_sign_negative());
|
464 | /// assert!(g.is_sign_negative());
|
465 | /// // NaN can be either positive or negative
|
466 | /// assert!(nan.is_sign_positive() != nan.is_sign_negative());
|
467 | /// ```
|
468 | #[inline ]
|
469 | #[must_use ]
|
470 | pub const fn is_sign_negative(self) -> bool {
|
471 | self.0 & 0x8000u16 != 0
|
472 | }
|
473 |
|
474 | /// Returns a number composed of the magnitude of `self` and the sign of `sign`.
|
475 | ///
|
476 | /// Equal to `self` if the sign of `self` and `sign` are the same, otherwise equal to `-self`.
|
477 | /// If `self` is NaN, then NaN with the sign of `sign` is returned.
|
478 | ///
|
479 | /// # Examples
|
480 | ///
|
481 | /// ```
|
482 | /// # use half::prelude::*;
|
483 | /// let f = bf16::from_f32(3.5);
|
484 | ///
|
485 | /// assert_eq!(f.copysign(bf16::from_f32(0.42)), bf16::from_f32(3.5));
|
486 | /// assert_eq!(f.copysign(bf16::from_f32(-0.42)), bf16::from_f32(-3.5));
|
487 | /// assert_eq!((-f).copysign(bf16::from_f32(0.42)), bf16::from_f32(3.5));
|
488 | /// assert_eq!((-f).copysign(bf16::from_f32(-0.42)), bf16::from_f32(-3.5));
|
489 | ///
|
490 | /// assert!(bf16::NAN.copysign(bf16::from_f32(1.0)).is_nan());
|
491 | /// ```
|
492 | #[inline ]
|
493 | #[must_use ]
|
494 | pub const fn copysign(self, sign: bf16) -> bf16 {
|
495 | bf16((sign.0 & 0x8000u16) | (self.0 & 0x7FFFu16))
|
496 | }
|
497 |
|
498 | /// Returns the maximum of the two numbers.
|
499 | ///
|
500 | /// If one of the arguments is NaN, then the other argument is returned.
|
501 | ///
|
502 | /// # Examples
|
503 | ///
|
504 | /// ```
|
505 | /// # use half::prelude::*;
|
506 | /// let x = bf16::from_f32(1.0);
|
507 | /// let y = bf16::from_f32(2.0);
|
508 | ///
|
509 | /// assert_eq!(x.max(y), y);
|
510 | /// ```
|
511 | #[inline ]
|
512 | #[must_use ]
|
513 | pub fn max(self, other: bf16) -> bf16 {
|
514 | if other > self && !other.is_nan() {
|
515 | other
|
516 | } else {
|
517 | self
|
518 | }
|
519 | }
|
520 |
|
521 | /// Returns the minimum of the two numbers.
|
522 | ///
|
523 | /// If one of the arguments is NaN, then the other argument is returned.
|
524 | ///
|
525 | /// # Examples
|
526 | ///
|
527 | /// ```
|
528 | /// # use half::prelude::*;
|
529 | /// let x = bf16::from_f32(1.0);
|
530 | /// let y = bf16::from_f32(2.0);
|
531 | ///
|
532 | /// assert_eq!(x.min(y), x);
|
533 | /// ```
|
534 | #[inline ]
|
535 | #[must_use ]
|
536 | pub fn min(self, other: bf16) -> bf16 {
|
537 | if other < self && !other.is_nan() {
|
538 | other
|
539 | } else {
|
540 | self
|
541 | }
|
542 | }
|
543 |
|
544 | /// Restrict a value to a certain interval unless it is NaN.
|
545 | ///
|
546 | /// Returns `max` if `self` is greater than `max`, and `min` if `self` is less than `min`.
|
547 | /// Otherwise this returns `self`.
|
548 | ///
|
549 | /// Note that this function returns NaN if the initial value was NaN as well.
|
550 | ///
|
551 | /// # Panics
|
552 | /// Panics if `min > max`, `min` is NaN, or `max` is NaN.
|
553 | ///
|
554 | /// # Examples
|
555 | ///
|
556 | /// ```
|
557 | /// # use half::prelude::*;
|
558 | /// assert!(bf16::from_f32(-3.0).clamp(bf16::from_f32(-2.0), bf16::from_f32(1.0)) == bf16::from_f32(-2.0));
|
559 | /// assert!(bf16::from_f32(0.0).clamp(bf16::from_f32(-2.0), bf16::from_f32(1.0)) == bf16::from_f32(0.0));
|
560 | /// assert!(bf16::from_f32(2.0).clamp(bf16::from_f32(-2.0), bf16::from_f32(1.0)) == bf16::from_f32(1.0));
|
561 | /// assert!(bf16::NAN.clamp(bf16::from_f32(-2.0), bf16::from_f32(1.0)).is_nan());
|
562 | /// ```
|
563 | #[inline ]
|
564 | #[must_use ]
|
565 | pub fn clamp(self, min: bf16, max: bf16) -> bf16 {
|
566 | assert!(min <= max);
|
567 | let mut x = self;
|
568 | if x < min {
|
569 | x = min;
|
570 | }
|
571 | if x > max {
|
572 | x = max;
|
573 | }
|
574 | x
|
575 | }
|
576 |
|
577 | /// Returns the ordering between `self` and `other`.
|
578 | ///
|
579 | /// Unlike the standard partial comparison between floating point numbers,
|
580 | /// this comparison always produces an ordering in accordance to
|
581 | /// the `totalOrder` predicate as defined in the IEEE 754 (2008 revision)
|
582 | /// floating point standard. The values are ordered in the following sequence:
|
583 | ///
|
584 | /// - negative quiet NaN
|
585 | /// - negative signaling NaN
|
586 | /// - negative infinity
|
587 | /// - negative numbers
|
588 | /// - negative subnormal numbers
|
589 | /// - negative zero
|
590 | /// - positive zero
|
591 | /// - positive subnormal numbers
|
592 | /// - positive numbers
|
593 | /// - positive infinity
|
594 | /// - positive signaling NaN
|
595 | /// - positive quiet NaN.
|
596 | ///
|
597 | /// The ordering established by this function does not always agree with the
|
598 | /// [`PartialOrd`] and [`PartialEq`] implementations of `bf16`. For example,
|
599 | /// they consider negative and positive zero equal, while `total_cmp`
|
600 | /// doesn't.
|
601 | ///
|
602 | /// The interpretation of the signaling NaN bit follows the definition in
|
603 | /// the IEEE 754 standard, which may not match the interpretation by some of
|
604 | /// the older, non-conformant (e.g. MIPS) hardware implementations.
|
605 | ///
|
606 | /// # Examples
|
607 | /// ```
|
608 | /// # use half::bf16;
|
609 | /// let mut v: Vec<bf16> = vec![];
|
610 | /// v.push(bf16::ONE);
|
611 | /// v.push(bf16::INFINITY);
|
612 | /// v.push(bf16::NEG_INFINITY);
|
613 | /// v.push(bf16::NAN);
|
614 | /// v.push(bf16::MAX_SUBNORMAL);
|
615 | /// v.push(-bf16::MAX_SUBNORMAL);
|
616 | /// v.push(bf16::ZERO);
|
617 | /// v.push(bf16::NEG_ZERO);
|
618 | /// v.push(bf16::NEG_ONE);
|
619 | /// v.push(bf16::MIN_POSITIVE);
|
620 | ///
|
621 | /// v.sort_by(|a, b| a.total_cmp(&b));
|
622 | ///
|
623 | /// assert!(v
|
624 | /// .into_iter()
|
625 | /// .zip(
|
626 | /// [
|
627 | /// bf16::NEG_INFINITY,
|
628 | /// bf16::NEG_ONE,
|
629 | /// -bf16::MAX_SUBNORMAL,
|
630 | /// bf16::NEG_ZERO,
|
631 | /// bf16::ZERO,
|
632 | /// bf16::MAX_SUBNORMAL,
|
633 | /// bf16::MIN_POSITIVE,
|
634 | /// bf16::ONE,
|
635 | /// bf16::INFINITY,
|
636 | /// bf16::NAN
|
637 | /// ]
|
638 | /// .iter()
|
639 | /// )
|
640 | /// .all(|(a, b)| a.to_bits() == b.to_bits()));
|
641 | /// ```
|
642 | // Implementation based on: https://doc.rust-lang.org/std/primitive.f32.html#method.total_cmp
|
643 | #[inline ]
|
644 | #[must_use ]
|
645 | pub fn total_cmp(&self, other: &Self) -> Ordering {
|
646 | let mut left = self.to_bits() as i16;
|
647 | let mut right = other.to_bits() as i16;
|
648 | left ^= (((left >> 15) as u16) >> 1) as i16;
|
649 | right ^= (((right >> 15) as u16) >> 1) as i16;
|
650 | left.cmp(&right)
|
651 | }
|
652 |
|
653 | /// Alternate serialize adapter for serializing as a float.
|
654 | ///
|
655 | /// By default, [`struct@bf16`] serializes as a newtype of [`u16`]. This is an alternate serialize
|
656 | /// implementation that serializes as an [`f32`] value. It is designed for use with
|
657 | /// `serialize_with` serde attributes. Deserialization from `f32` values is already supported by
|
658 | /// the default deserialize implementation.
|
659 | ///
|
660 | /// # Examples
|
661 | ///
|
662 | /// A demonstration on how to use this adapater:
|
663 | ///
|
664 | /// ```
|
665 | /// use serde::{Serialize, Deserialize};
|
666 | /// use half::bf16;
|
667 | ///
|
668 | /// #[derive(Serialize, Deserialize)]
|
669 | /// struct MyStruct {
|
670 | /// #[serde(serialize_with = "bf16::serialize_as_f32")]
|
671 | /// value: bf16 // Will be serialized as f32 instead of u16
|
672 | /// }
|
673 | /// ```
|
674 | #[cfg (feature = "serde" )]
|
675 | pub fn serialize_as_f32<S: serde::Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
|
676 | serializer.serialize_f32(self.to_f32())
|
677 | }
|
678 |
|
679 | /// Alternate serialize adapter for serializing as a string.
|
680 | ///
|
681 | /// By default, [`struct@bf16`] serializes as a newtype of [`u16`]. This is an alternate serialize
|
682 | /// implementation that serializes as a string value. It is designed for use with
|
683 | /// `serialize_with` serde attributes. Deserialization from string values is already supported
|
684 | /// by the default deserialize implementation.
|
685 | ///
|
686 | /// # Examples
|
687 | ///
|
688 | /// A demonstration on how to use this adapater:
|
689 | ///
|
690 | /// ```
|
691 | /// use serde::{Serialize, Deserialize};
|
692 | /// use half::bf16;
|
693 | ///
|
694 | /// #[derive(Serialize, Deserialize)]
|
695 | /// struct MyStruct {
|
696 | /// #[serde(serialize_with = "bf16::serialize_as_string")]
|
697 | /// value: bf16 // Will be serialized as a string instead of u16
|
698 | /// }
|
699 | /// ```
|
700 | #[cfg (all(feature = "serde" , feature = "alloc" ))]
|
701 | pub fn serialize_as_string<S: serde::Serializer>(
|
702 | &self,
|
703 | serializer: S,
|
704 | ) -> Result<S::Ok, S::Error> {
|
705 | serializer.serialize_str(&self.to_string())
|
706 | }
|
707 |
|
708 | /// Approximate number of [`struct@bf16`] significant digits in base 10
|
709 | pub const DIGITS: u32 = 2;
|
710 | /// [`struct@bf16`]
|
711 | /// [machine epsilon](https://en.wikipedia.org/wiki/Machine_epsilon) value
|
712 | ///
|
713 | /// This is the difference between 1.0 and the next largest representable number.
|
714 | pub const EPSILON: bf16 = bf16(0x3C00u16);
|
715 | /// [`struct@bf16`] positive Infinity (+∞)
|
716 | pub const INFINITY: bf16 = bf16(0x7F80u16);
|
717 | /// Number of [`struct@bf16`] significant digits in base 2
|
718 | pub const MANTISSA_DIGITS: u32 = 8;
|
719 | /// Largest finite [`struct@bf16`] value
|
720 | pub const MAX: bf16 = bf16(0x7F7F);
|
721 | /// Maximum possible [`struct@bf16`] power of 10 exponent
|
722 | pub const MAX_10_EXP: i32 = 38;
|
723 | /// Maximum possible [`struct@bf16`] power of 2 exponent
|
724 | pub const MAX_EXP: i32 = 128;
|
725 | /// Smallest finite [`struct@bf16`] value
|
726 | pub const MIN: bf16 = bf16(0xFF7F);
|
727 | /// Minimum possible normal [`struct@bf16`] power of 10 exponent
|
728 | pub const MIN_10_EXP: i32 = -37;
|
729 | /// One greater than the minimum possible normal [`struct@bf16`] power of 2 exponent
|
730 | pub const MIN_EXP: i32 = -125;
|
731 | /// Smallest positive normal [`struct@bf16`] value
|
732 | pub const MIN_POSITIVE: bf16 = bf16(0x0080u16);
|
733 | /// [`struct@bf16`] Not a Number (NaN)
|
734 | pub const NAN: bf16 = bf16(0x7FC0u16);
|
735 | /// [`struct@bf16`] negative infinity (-∞).
|
736 | pub const NEG_INFINITY: bf16 = bf16(0xFF80u16);
|
737 | /// The radix or base of the internal representation of [`struct@bf16`]
|
738 | pub const RADIX: u32 = 2;
|
739 |
|
740 | /// Minimum positive subnormal [`struct@bf16`] value
|
741 | pub const MIN_POSITIVE_SUBNORMAL: bf16 = bf16(0x0001u16);
|
742 | /// Maximum subnormal [`struct@bf16`] value
|
743 | pub const MAX_SUBNORMAL: bf16 = bf16(0x007Fu16);
|
744 |
|
745 | /// [`struct@bf16`] 1
|
746 | pub const ONE: bf16 = bf16(0x3F80u16);
|
747 | /// [`struct@bf16`] 0
|
748 | pub const ZERO: bf16 = bf16(0x0000u16);
|
749 | /// [`struct@bf16`] -0
|
750 | pub const NEG_ZERO: bf16 = bf16(0x8000u16);
|
751 | /// [`struct@bf16`] -1
|
752 | pub const NEG_ONE: bf16 = bf16(0xBF80u16);
|
753 |
|
754 | /// [`struct@bf16`] Euler's number (ℯ)
|
755 | pub const E: bf16 = bf16(0x402Eu16);
|
756 | /// [`struct@bf16`] Archimedes' constant (π)
|
757 | pub const PI: bf16 = bf16(0x4049u16);
|
758 | /// [`struct@bf16`] 1/π
|
759 | pub const FRAC_1_PI: bf16 = bf16(0x3EA3u16);
|
760 | /// [`struct@bf16`] 1/√2
|
761 | pub const FRAC_1_SQRT_2: bf16 = bf16(0x3F35u16);
|
762 | /// [`struct@bf16`] 2/π
|
763 | pub const FRAC_2_PI: bf16 = bf16(0x3F23u16);
|
764 | /// [`struct@bf16`] 2/√π
|
765 | pub const FRAC_2_SQRT_PI: bf16 = bf16(0x3F90u16);
|
766 | /// [`struct@bf16`] π/2
|
767 | pub const FRAC_PI_2: bf16 = bf16(0x3FC9u16);
|
768 | /// [`struct@bf16`] π/3
|
769 | pub const FRAC_PI_3: bf16 = bf16(0x3F86u16);
|
770 | /// [`struct@bf16`] π/4
|
771 | pub const FRAC_PI_4: bf16 = bf16(0x3F49u16);
|
772 | /// [`struct@bf16`] π/6
|
773 | pub const FRAC_PI_6: bf16 = bf16(0x3F06u16);
|
774 | /// [`struct@bf16`] π/8
|
775 | pub const FRAC_PI_8: bf16 = bf16(0x3EC9u16);
|
776 | /// [`struct@bf16`] 𝗅𝗇 10
|
777 | pub const LN_10: bf16 = bf16(0x4013u16);
|
778 | /// [`struct@bf16`] 𝗅𝗇 2
|
779 | pub const LN_2: bf16 = bf16(0x3F31u16);
|
780 | /// [`struct@bf16`] 𝗅𝗈𝗀₁₀ℯ
|
781 | pub const LOG10_E: bf16 = bf16(0x3EDEu16);
|
782 | /// [`struct@bf16`] 𝗅𝗈𝗀₁₀2
|
783 | pub const LOG10_2: bf16 = bf16(0x3E9Au16);
|
784 | /// [`struct@bf16`] 𝗅𝗈𝗀₂ℯ
|
785 | pub const LOG2_E: bf16 = bf16(0x3FB9u16);
|
786 | /// [`struct@bf16`] 𝗅𝗈𝗀₂10
|
787 | pub const LOG2_10: bf16 = bf16(0x4055u16);
|
788 | /// [`struct@bf16`] √2
|
789 | pub const SQRT_2: bf16 = bf16(0x3FB5u16);
|
790 | }
|
791 |
|
792 | impl From<bf16> for f32 {
|
793 | #[inline ]
|
794 | fn from(x: bf16) -> f32 {
|
795 | x.to_f32()
|
796 | }
|
797 | }
|
798 |
|
799 | impl From<bf16> for f64 {
|
800 | #[inline ]
|
801 | fn from(x: bf16) -> f64 {
|
802 | x.to_f64()
|
803 | }
|
804 | }
|
805 |
|
806 | impl From<i8> for bf16 {
|
807 | #[inline ]
|
808 | fn from(x: i8) -> bf16 {
|
809 | // Convert to f32, then to bf16
|
810 | bf16::from_f32(f32::from(x))
|
811 | }
|
812 | }
|
813 |
|
814 | impl From<u8> for bf16 {
|
815 | #[inline ]
|
816 | fn from(x: u8) -> bf16 {
|
817 | // Convert to f32, then to f16
|
818 | bf16::from_f32(f32::from(x))
|
819 | }
|
820 | }
|
821 |
|
822 | impl PartialEq for bf16 {
|
823 | fn eq(&self, other: &bf16) -> bool {
|
824 | if self.is_nan() || other.is_nan() {
|
825 | false
|
826 | } else {
|
827 | (self.0 == other.0) || ((self.0 | other.0) & 0x7FFFu16 == 0)
|
828 | }
|
829 | }
|
830 | }
|
831 |
|
832 | impl PartialOrd for bf16 {
|
833 | fn partial_cmp(&self, other: &bf16) -> Option<Ordering> {
|
834 | if self.is_nan() || other.is_nan() {
|
835 | None
|
836 | } else {
|
837 | let neg = self.0 & 0x8000u16 != 0;
|
838 | let other_neg = other.0 & 0x8000u16 != 0;
|
839 | match (neg, other_neg) {
|
840 | (false, false) => Some(self.0.cmp(&other.0)),
|
841 | (false, true) => {
|
842 | if (self.0 | other.0) & 0x7FFFu16 == 0 {
|
843 | Some(Ordering::Equal)
|
844 | } else {
|
845 | Some(Ordering::Greater)
|
846 | }
|
847 | }
|
848 | (true, false) => {
|
849 | if (self.0 | other.0) & 0x7FFFu16 == 0 {
|
850 | Some(Ordering::Equal)
|
851 | } else {
|
852 | Some(Ordering::Less)
|
853 | }
|
854 | }
|
855 | (true, true) => Some(other.0.cmp(&self.0)),
|
856 | }
|
857 | }
|
858 | }
|
859 |
|
860 | fn lt(&self, other: &bf16) -> bool {
|
861 | if self.is_nan() || other.is_nan() {
|
862 | false
|
863 | } else {
|
864 | let neg = self.0 & 0x8000u16 != 0;
|
865 | let other_neg = other.0 & 0x8000u16 != 0;
|
866 | match (neg, other_neg) {
|
867 | (false, false) => self.0 < other.0,
|
868 | (false, true) => false,
|
869 | (true, false) => (self.0 | other.0) & 0x7FFFu16 != 0,
|
870 | (true, true) => self.0 > other.0,
|
871 | }
|
872 | }
|
873 | }
|
874 |
|
875 | fn le(&self, other: &bf16) -> bool {
|
876 | if self.is_nan() || other.is_nan() {
|
877 | false
|
878 | } else {
|
879 | let neg = self.0 & 0x8000u16 != 0;
|
880 | let other_neg = other.0 & 0x8000u16 != 0;
|
881 | match (neg, other_neg) {
|
882 | (false, false) => self.0 <= other.0,
|
883 | (false, true) => (self.0 | other.0) & 0x7FFFu16 == 0,
|
884 | (true, false) => true,
|
885 | (true, true) => self.0 >= other.0,
|
886 | }
|
887 | }
|
888 | }
|
889 |
|
890 | fn gt(&self, other: &bf16) -> bool {
|
891 | if self.is_nan() || other.is_nan() {
|
892 | false
|
893 | } else {
|
894 | let neg = self.0 & 0x8000u16 != 0;
|
895 | let other_neg = other.0 & 0x8000u16 != 0;
|
896 | match (neg, other_neg) {
|
897 | (false, false) => self.0 > other.0,
|
898 | (false, true) => (self.0 | other.0) & 0x7FFFu16 != 0,
|
899 | (true, false) => false,
|
900 | (true, true) => self.0 < other.0,
|
901 | }
|
902 | }
|
903 | }
|
904 |
|
905 | fn ge(&self, other: &bf16) -> bool {
|
906 | if self.is_nan() || other.is_nan() {
|
907 | false
|
908 | } else {
|
909 | let neg = self.0 & 0x8000u16 != 0;
|
910 | let other_neg = other.0 & 0x8000u16 != 0;
|
911 | match (neg, other_neg) {
|
912 | (false, false) => self.0 >= other.0,
|
913 | (false, true) => true,
|
914 | (true, false) => (self.0 | other.0) & 0x7FFFu16 == 0,
|
915 | (true, true) => self.0 <= other.0,
|
916 | }
|
917 | }
|
918 | }
|
919 | }
|
920 |
|
921 | #[cfg (not(target_arch = "spirv" ))]
|
922 | impl FromStr for bf16 {
|
923 | type Err = ParseFloatError;
|
924 | fn from_str(src: &str) -> Result<bf16, ParseFloatError> {
|
925 | f32::from_str(src).map(op:bf16::from_f32)
|
926 | }
|
927 | }
|
928 |
|
929 | #[cfg (not(target_arch = "spirv" ))]
|
930 | impl Debug for bf16 {
|
931 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
|
932 | Debug::fmt(&self.to_f32(), f)
|
933 | }
|
934 | }
|
935 |
|
936 | #[cfg (not(target_arch = "spirv" ))]
|
937 | impl Display for bf16 {
|
938 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
|
939 | Display::fmt(&self.to_f32(), f)
|
940 | }
|
941 | }
|
942 |
|
943 | #[cfg (not(target_arch = "spirv" ))]
|
944 | impl LowerExp for bf16 {
|
945 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
|
946 | write!(f, " {:e}" , self.to_f32())
|
947 | }
|
948 | }
|
949 |
|
950 | #[cfg (not(target_arch = "spirv" ))]
|
951 | impl UpperExp for bf16 {
|
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 Binary for bf16 {
|
959 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
|
960 | write!(f, " {:b}" , self.0)
|
961 | }
|
962 | }
|
963 |
|
964 | #[cfg (not(target_arch = "spirv" ))]
|
965 | impl Octal for bf16 {
|
966 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
|
967 | write!(f, " {:o}" , self.0)
|
968 | }
|
969 | }
|
970 |
|
971 | #[cfg (not(target_arch = "spirv" ))]
|
972 | impl LowerHex for bf16 {
|
973 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
|
974 | write!(f, " {:x}" , self.0)
|
975 | }
|
976 | }
|
977 |
|
978 | #[cfg (not(target_arch = "spirv" ))]
|
979 | impl UpperHex for bf16 {
|
980 | fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
|
981 | write!(f, " {:X}" , self.0)
|
982 | }
|
983 | }
|
984 |
|
985 | impl Neg for bf16 {
|
986 | type Output = Self;
|
987 |
|
988 | fn neg(self) -> Self::Output {
|
989 | Self(self.0 ^ 0x8000)
|
990 | }
|
991 | }
|
992 |
|
993 | impl Neg for &bf16 {
|
994 | type Output = <bf16 as Neg>::Output;
|
995 |
|
996 | #[inline ]
|
997 | fn neg(self) -> Self::Output {
|
998 | Neg::neg(*self)
|
999 | }
|
1000 | }
|
1001 |
|
1002 | impl Add for bf16 {
|
1003 | type Output = Self;
|
1004 |
|
1005 | fn add(self, rhs: Self) -> Self::Output {
|
1006 | Self::from_f32(Self::to_f32(self) + Self::to_f32(self:rhs))
|
1007 | }
|
1008 | }
|
1009 |
|
1010 | impl Add<&bf16> for bf16 {
|
1011 | type Output = <bf16 as Add<bf16>>::Output;
|
1012 |
|
1013 | #[inline ]
|
1014 | fn add(self, rhs: &bf16) -> Self::Output {
|
1015 | self.add(*rhs)
|
1016 | }
|
1017 | }
|
1018 |
|
1019 | impl Add<&bf16> for &bf16 {
|
1020 | type Output = <bf16 as Add<bf16>>::Output;
|
1021 |
|
1022 | #[inline ]
|
1023 | fn add(self, rhs: &bf16) -> Self::Output {
|
1024 | (*self).add(*rhs)
|
1025 | }
|
1026 | }
|
1027 |
|
1028 | impl Add<bf16> for &bf16 {
|
1029 | type Output = <bf16 as Add<bf16>>::Output;
|
1030 |
|
1031 | #[inline ]
|
1032 | fn add(self, rhs: bf16) -> Self::Output {
|
1033 | (*self).add(rhs)
|
1034 | }
|
1035 | }
|
1036 |
|
1037 | impl AddAssign for bf16 {
|
1038 | #[inline ]
|
1039 | fn add_assign(&mut self, rhs: Self) {
|
1040 | *self = (*self).add(rhs);
|
1041 | }
|
1042 | }
|
1043 |
|
1044 | impl AddAssign<&bf16> for bf16 {
|
1045 | #[inline ]
|
1046 | fn add_assign(&mut self, rhs: &bf16) {
|
1047 | *self = (*self).add(rhs);
|
1048 | }
|
1049 | }
|
1050 |
|
1051 | impl Sub for bf16 {
|
1052 | type Output = Self;
|
1053 |
|
1054 | fn sub(self, rhs: Self) -> Self::Output {
|
1055 | Self::from_f32(Self::to_f32(self) - Self::to_f32(self:rhs))
|
1056 | }
|
1057 | }
|
1058 |
|
1059 | impl Sub<&bf16> for bf16 {
|
1060 | type Output = <bf16 as Sub<bf16>>::Output;
|
1061 |
|
1062 | #[inline ]
|
1063 | fn sub(self, rhs: &bf16) -> Self::Output {
|
1064 | self.sub(*rhs)
|
1065 | }
|
1066 | }
|
1067 |
|
1068 | impl Sub<&bf16> for &bf16 {
|
1069 | type Output = <bf16 as Sub<bf16>>::Output;
|
1070 |
|
1071 | #[inline ]
|
1072 | fn sub(self, rhs: &bf16) -> Self::Output {
|
1073 | (*self).sub(*rhs)
|
1074 | }
|
1075 | }
|
1076 |
|
1077 | impl Sub<bf16> for &bf16 {
|
1078 | type Output = <bf16 as Sub<bf16>>::Output;
|
1079 |
|
1080 | #[inline ]
|
1081 | fn sub(self, rhs: bf16) -> Self::Output {
|
1082 | (*self).sub(rhs)
|
1083 | }
|
1084 | }
|
1085 |
|
1086 | impl SubAssign for bf16 {
|
1087 | #[inline ]
|
1088 | fn sub_assign(&mut self, rhs: Self) {
|
1089 | *self = (*self).sub(rhs);
|
1090 | }
|
1091 | }
|
1092 |
|
1093 | impl SubAssign<&bf16> for bf16 {
|
1094 | #[inline ]
|
1095 | fn sub_assign(&mut self, rhs: &bf16) {
|
1096 | *self = (*self).sub(rhs);
|
1097 | }
|
1098 | }
|
1099 |
|
1100 | impl Mul for bf16 {
|
1101 | type Output = Self;
|
1102 |
|
1103 | fn mul(self, rhs: Self) -> Self::Output {
|
1104 | Self::from_f32(Self::to_f32(self) * Self::to_f32(self:rhs))
|
1105 | }
|
1106 | }
|
1107 |
|
1108 | impl Mul<&bf16> for bf16 {
|
1109 | type Output = <bf16 as Mul<bf16>>::Output;
|
1110 |
|
1111 | #[inline ]
|
1112 | fn mul(self, rhs: &bf16) -> Self::Output {
|
1113 | self.mul(*rhs)
|
1114 | }
|
1115 | }
|
1116 |
|
1117 | impl Mul<&bf16> for &bf16 {
|
1118 | type Output = <bf16 as Mul<bf16>>::Output;
|
1119 |
|
1120 | #[inline ]
|
1121 | fn mul(self, rhs: &bf16) -> Self::Output {
|
1122 | (*self).mul(*rhs)
|
1123 | }
|
1124 | }
|
1125 |
|
1126 | impl Mul<bf16> for &bf16 {
|
1127 | type Output = <bf16 as Mul<bf16>>::Output;
|
1128 |
|
1129 | #[inline ]
|
1130 | fn mul(self, rhs: bf16) -> Self::Output {
|
1131 | (*self).mul(rhs)
|
1132 | }
|
1133 | }
|
1134 |
|
1135 | impl MulAssign for bf16 {
|
1136 | #[inline ]
|
1137 | fn mul_assign(&mut self, rhs: Self) {
|
1138 | *self = (*self).mul(rhs);
|
1139 | }
|
1140 | }
|
1141 |
|
1142 | impl MulAssign<&bf16> for bf16 {
|
1143 | #[inline ]
|
1144 | fn mul_assign(&mut self, rhs: &bf16) {
|
1145 | *self = (*self).mul(rhs);
|
1146 | }
|
1147 | }
|
1148 |
|
1149 | impl Div for bf16 {
|
1150 | type Output = Self;
|
1151 |
|
1152 | fn div(self, rhs: Self) -> Self::Output {
|
1153 | Self::from_f32(Self::to_f32(self) / Self::to_f32(self:rhs))
|
1154 | }
|
1155 | }
|
1156 |
|
1157 | impl Div<&bf16> for bf16 {
|
1158 | type Output = <bf16 as Div<bf16>>::Output;
|
1159 |
|
1160 | #[inline ]
|
1161 | fn div(self, rhs: &bf16) -> Self::Output {
|
1162 | self.div(*rhs)
|
1163 | }
|
1164 | }
|
1165 |
|
1166 | impl Div<&bf16> for &bf16 {
|
1167 | type Output = <bf16 as Div<bf16>>::Output;
|
1168 |
|
1169 | #[inline ]
|
1170 | fn div(self, rhs: &bf16) -> Self::Output {
|
1171 | (*self).div(*rhs)
|
1172 | }
|
1173 | }
|
1174 |
|
1175 | impl Div<bf16> for &bf16 {
|
1176 | type Output = <bf16 as Div<bf16>>::Output;
|
1177 |
|
1178 | #[inline ]
|
1179 | fn div(self, rhs: bf16) -> Self::Output {
|
1180 | (*self).div(rhs)
|
1181 | }
|
1182 | }
|
1183 |
|
1184 | impl DivAssign for bf16 {
|
1185 | #[inline ]
|
1186 | fn div_assign(&mut self, rhs: Self) {
|
1187 | *self = (*self).div(rhs);
|
1188 | }
|
1189 | }
|
1190 |
|
1191 | impl DivAssign<&bf16> for bf16 {
|
1192 | #[inline ]
|
1193 | fn div_assign(&mut self, rhs: &bf16) {
|
1194 | *self = (*self).div(rhs);
|
1195 | }
|
1196 | }
|
1197 |
|
1198 | impl Rem for bf16 {
|
1199 | type Output = Self;
|
1200 |
|
1201 | fn rem(self, rhs: Self) -> Self::Output {
|
1202 | Self::from_f32(Self::to_f32(self) % Self::to_f32(self:rhs))
|
1203 | }
|
1204 | }
|
1205 |
|
1206 | impl Rem<&bf16> for bf16 {
|
1207 | type Output = <bf16 as Rem<bf16>>::Output;
|
1208 |
|
1209 | #[inline ]
|
1210 | fn rem(self, rhs: &bf16) -> Self::Output {
|
1211 | self.rem(*rhs)
|
1212 | }
|
1213 | }
|
1214 |
|
1215 | impl Rem<&bf16> for &bf16 {
|
1216 | type Output = <bf16 as Rem<bf16>>::Output;
|
1217 |
|
1218 | #[inline ]
|
1219 | fn rem(self, rhs: &bf16) -> Self::Output {
|
1220 | (*self).rem(*rhs)
|
1221 | }
|
1222 | }
|
1223 |
|
1224 | impl Rem<bf16> for &bf16 {
|
1225 | type Output = <bf16 as Rem<bf16>>::Output;
|
1226 |
|
1227 | #[inline ]
|
1228 | fn rem(self, rhs: bf16) -> Self::Output {
|
1229 | (*self).rem(rhs)
|
1230 | }
|
1231 | }
|
1232 |
|
1233 | impl RemAssign for bf16 {
|
1234 | #[inline ]
|
1235 | fn rem_assign(&mut self, rhs: Self) {
|
1236 | *self = (*self).rem(rhs);
|
1237 | }
|
1238 | }
|
1239 |
|
1240 | impl RemAssign<&bf16> for bf16 {
|
1241 | #[inline ]
|
1242 | fn rem_assign(&mut self, rhs: &bf16) {
|
1243 | *self = (*self).rem(rhs);
|
1244 | }
|
1245 | }
|
1246 |
|
1247 | impl Product for bf16 {
|
1248 | #[inline ]
|
1249 | fn product<I: Iterator<Item = Self>>(iter: I) -> Self {
|
1250 | bf16::from_f32(iter.map(|f: bf16| f.to_f32()).product())
|
1251 | }
|
1252 | }
|
1253 |
|
1254 | impl<'a> Product<&'a bf16> for bf16 {
|
1255 | #[inline ]
|
1256 | fn product<I: Iterator<Item = &'a bf16>>(iter: I) -> Self {
|
1257 | bf16::from_f32(iter.map(|f: &'a bf16| f.to_f32()).product())
|
1258 | }
|
1259 | }
|
1260 |
|
1261 | impl Sum for bf16 {
|
1262 | #[inline ]
|
1263 | fn sum<I: Iterator<Item = Self>>(iter: I) -> Self {
|
1264 | bf16::from_f32(iter.map(|f: bf16| f.to_f32()).sum())
|
1265 | }
|
1266 | }
|
1267 |
|
1268 | impl<'a> Sum<&'a bf16> for bf16 {
|
1269 | #[inline ]
|
1270 | fn sum<I: Iterator<Item = &'a bf16>>(iter: I) -> Self {
|
1271 | bf16::from_f32(iter.map(|f: &'a bf16| f.to_f32()).sum())
|
1272 | }
|
1273 | }
|
1274 |
|
1275 | #[cfg (feature = "serde" )]
|
1276 | struct Visitor;
|
1277 |
|
1278 | #[cfg (feature = "serde" )]
|
1279 | impl<'de> Deserialize<'de> for bf16 {
|
1280 | fn deserialize<D>(deserializer: D) -> Result<bf16, D::Error>
|
1281 | where
|
1282 | D: serde::de::Deserializer<'de>,
|
1283 | {
|
1284 | deserializer.deserialize_newtype_struct("bf16" , Visitor)
|
1285 | }
|
1286 | }
|
1287 |
|
1288 | #[cfg (feature = "serde" )]
|
1289 | impl<'de> serde::de::Visitor<'de> for Visitor {
|
1290 | type Value = bf16;
|
1291 |
|
1292 | fn expecting(&self, formatter: &mut core::fmt::Formatter) -> core::fmt::Result {
|
1293 | write!(formatter, "tuple struct bf16" )
|
1294 | }
|
1295 |
|
1296 | fn visit_newtype_struct<D>(self, deserializer: D) -> Result<Self::Value, D::Error>
|
1297 | where
|
1298 | D: serde::Deserializer<'de>,
|
1299 | {
|
1300 | Ok(bf16(<u16 as Deserialize>::deserialize(deserializer)?))
|
1301 | }
|
1302 |
|
1303 | fn visit_str<E>(self, v: &str) -> Result<Self::Value, E>
|
1304 | where
|
1305 | E: serde::de::Error,
|
1306 | {
|
1307 | v.parse().map_err(|_| {
|
1308 | serde::de::Error::invalid_value(serde::de::Unexpected::Str(v), &"a float string" )
|
1309 | })
|
1310 | }
|
1311 |
|
1312 | fn visit_f32<E>(self, v: f32) -> Result<Self::Value, E>
|
1313 | where
|
1314 | E: serde::de::Error,
|
1315 | {
|
1316 | Ok(bf16::from_f32(v))
|
1317 | }
|
1318 |
|
1319 | fn visit_f64<E>(self, v: f64) -> Result<Self::Value, E>
|
1320 | where
|
1321 | E: serde::de::Error,
|
1322 | {
|
1323 | Ok(bf16::from_f64(v))
|
1324 | }
|
1325 | }
|
1326 |
|
1327 | #[allow (
|
1328 | clippy::cognitive_complexity,
|
1329 | clippy::float_cmp,
|
1330 | clippy::neg_cmp_op_on_partial_ord
|
1331 | )]
|
1332 | #[cfg (test)]
|
1333 | mod test {
|
1334 | use super::*;
|
1335 | #[allow (unused_imports)]
|
1336 | use core::cmp::Ordering;
|
1337 | #[cfg (feature = "num-traits" )]
|
1338 | use num_traits::{AsPrimitive, FromBytes, FromPrimitive, ToBytes, ToPrimitive};
|
1339 | use quickcheck_macros::quickcheck;
|
1340 |
|
1341 | #[cfg (feature = "num-traits" )]
|
1342 | #[test ]
|
1343 | fn as_primitive() {
|
1344 | let two = bf16::from_f32(2.0);
|
1345 | assert_eq!(<i32 as AsPrimitive<bf16>>::as_(2), two);
|
1346 | assert_eq!(<bf16 as AsPrimitive<i32>>::as_(two), 2);
|
1347 |
|
1348 | assert_eq!(<f32 as AsPrimitive<bf16>>::as_(2.0), two);
|
1349 | assert_eq!(<bf16 as AsPrimitive<f32>>::as_(two), 2.0);
|
1350 |
|
1351 | assert_eq!(<f64 as AsPrimitive<bf16>>::as_(2.0), two);
|
1352 | assert_eq!(<bf16 as AsPrimitive<f64>>::as_(two), 2.0);
|
1353 | }
|
1354 |
|
1355 | #[cfg (feature = "num-traits" )]
|
1356 | #[test ]
|
1357 | fn to_primitive() {
|
1358 | let two = bf16::from_f32(2.0);
|
1359 | assert_eq!(ToPrimitive::to_i32(&two).unwrap(), 2i32);
|
1360 | assert_eq!(ToPrimitive::to_f32(&two).unwrap(), 2.0f32);
|
1361 | assert_eq!(ToPrimitive::to_f64(&two).unwrap(), 2.0f64);
|
1362 | }
|
1363 |
|
1364 | #[cfg (feature = "num-traits" )]
|
1365 | #[test ]
|
1366 | fn from_primitive() {
|
1367 | let two = bf16::from_f32(2.0);
|
1368 | assert_eq!(<bf16 as FromPrimitive>::from_i32(2).unwrap(), two);
|
1369 | assert_eq!(<bf16 as FromPrimitive>::from_f32(2.0).unwrap(), two);
|
1370 | assert_eq!(<bf16 as FromPrimitive>::from_f64(2.0).unwrap(), two);
|
1371 | }
|
1372 |
|
1373 | #[cfg (feature = "num-traits" )]
|
1374 | #[test ]
|
1375 | fn to_and_from_bytes() {
|
1376 | let two = bf16::from_f32(2.0);
|
1377 | assert_eq!(<bf16 as ToBytes>::to_le_bytes(&two), [0, 64]);
|
1378 | assert_eq!(<bf16 as FromBytes>::from_le_bytes(&[0, 64]), two);
|
1379 | assert_eq!(<bf16 as ToBytes>::to_be_bytes(&two), [64, 0]);
|
1380 | assert_eq!(<bf16 as FromBytes>::from_be_bytes(&[64, 0]), two);
|
1381 | }
|
1382 |
|
1383 | #[test ]
|
1384 | fn test_bf16_consts_from_f32() {
|
1385 | let one = bf16::from_f32(1.0);
|
1386 | let zero = bf16::from_f32(0.0);
|
1387 | let neg_zero = bf16::from_f32(-0.0);
|
1388 | let neg_one = bf16::from_f32(-1.0);
|
1389 | let inf = bf16::from_f32(core::f32::INFINITY);
|
1390 | let neg_inf = bf16::from_f32(core::f32::NEG_INFINITY);
|
1391 | let nan = bf16::from_f32(core::f32::NAN);
|
1392 |
|
1393 | assert_eq!(bf16::ONE, one);
|
1394 | assert_eq!(bf16::ZERO, zero);
|
1395 | assert!(zero.is_sign_positive());
|
1396 | assert_eq!(bf16::NEG_ZERO, neg_zero);
|
1397 | assert!(neg_zero.is_sign_negative());
|
1398 | assert_eq!(bf16::NEG_ONE, neg_one);
|
1399 | assert!(neg_one.is_sign_negative());
|
1400 | assert_eq!(bf16::INFINITY, inf);
|
1401 | assert_eq!(bf16::NEG_INFINITY, neg_inf);
|
1402 | assert!(nan.is_nan());
|
1403 | assert!(bf16::NAN.is_nan());
|
1404 |
|
1405 | let e = bf16::from_f32(core::f32::consts::E);
|
1406 | let pi = bf16::from_f32(core::f32::consts::PI);
|
1407 | let frac_1_pi = bf16::from_f32(core::f32::consts::FRAC_1_PI);
|
1408 | let frac_1_sqrt_2 = bf16::from_f32(core::f32::consts::FRAC_1_SQRT_2);
|
1409 | let frac_2_pi = bf16::from_f32(core::f32::consts::FRAC_2_PI);
|
1410 | let frac_2_sqrt_pi = bf16::from_f32(core::f32::consts::FRAC_2_SQRT_PI);
|
1411 | let frac_pi_2 = bf16::from_f32(core::f32::consts::FRAC_PI_2);
|
1412 | let frac_pi_3 = bf16::from_f32(core::f32::consts::FRAC_PI_3);
|
1413 | let frac_pi_4 = bf16::from_f32(core::f32::consts::FRAC_PI_4);
|
1414 | let frac_pi_6 = bf16::from_f32(core::f32::consts::FRAC_PI_6);
|
1415 | let frac_pi_8 = bf16::from_f32(core::f32::consts::FRAC_PI_8);
|
1416 | let ln_10 = bf16::from_f32(core::f32::consts::LN_10);
|
1417 | let ln_2 = bf16::from_f32(core::f32::consts::LN_2);
|
1418 | let log10_e = bf16::from_f32(core::f32::consts::LOG10_E);
|
1419 | // core::f32::consts::LOG10_2 requires rustc 1.43.0
|
1420 | let log10_2 = bf16::from_f32(2f32.log10());
|
1421 | let log2_e = bf16::from_f32(core::f32::consts::LOG2_E);
|
1422 | // core::f32::consts::LOG2_10 requires rustc 1.43.0
|
1423 | let log2_10 = bf16::from_f32(10f32.log2());
|
1424 | let sqrt_2 = bf16::from_f32(core::f32::consts::SQRT_2);
|
1425 |
|
1426 | assert_eq!(bf16::E, e);
|
1427 | assert_eq!(bf16::PI, pi);
|
1428 | assert_eq!(bf16::FRAC_1_PI, frac_1_pi);
|
1429 | assert_eq!(bf16::FRAC_1_SQRT_2, frac_1_sqrt_2);
|
1430 | assert_eq!(bf16::FRAC_2_PI, frac_2_pi);
|
1431 | assert_eq!(bf16::FRAC_2_SQRT_PI, frac_2_sqrt_pi);
|
1432 | assert_eq!(bf16::FRAC_PI_2, frac_pi_2);
|
1433 | assert_eq!(bf16::FRAC_PI_3, frac_pi_3);
|
1434 | assert_eq!(bf16::FRAC_PI_4, frac_pi_4);
|
1435 | assert_eq!(bf16::FRAC_PI_6, frac_pi_6);
|
1436 | assert_eq!(bf16::FRAC_PI_8, frac_pi_8);
|
1437 | assert_eq!(bf16::LN_10, ln_10);
|
1438 | assert_eq!(bf16::LN_2, ln_2);
|
1439 | assert_eq!(bf16::LOG10_E, log10_e);
|
1440 | assert_eq!(bf16::LOG10_2, log10_2);
|
1441 | assert_eq!(bf16::LOG2_E, log2_e);
|
1442 | assert_eq!(bf16::LOG2_10, log2_10);
|
1443 | assert_eq!(bf16::SQRT_2, sqrt_2);
|
1444 | }
|
1445 |
|
1446 | #[test ]
|
1447 | fn test_bf16_consts_from_f64() {
|
1448 | let one = bf16::from_f64(1.0);
|
1449 | let zero = bf16::from_f64(0.0);
|
1450 | let neg_zero = bf16::from_f64(-0.0);
|
1451 | let inf = bf16::from_f64(core::f64::INFINITY);
|
1452 | let neg_inf = bf16::from_f64(core::f64::NEG_INFINITY);
|
1453 | let nan = bf16::from_f64(core::f64::NAN);
|
1454 |
|
1455 | assert_eq!(bf16::ONE, one);
|
1456 | assert_eq!(bf16::ZERO, zero);
|
1457 | assert_eq!(bf16::NEG_ZERO, neg_zero);
|
1458 | assert_eq!(bf16::INFINITY, inf);
|
1459 | assert_eq!(bf16::NEG_INFINITY, neg_inf);
|
1460 | assert!(nan.is_nan());
|
1461 | assert!(bf16::NAN.is_nan());
|
1462 |
|
1463 | let e = bf16::from_f64(core::f64::consts::E);
|
1464 | let pi = bf16::from_f64(core::f64::consts::PI);
|
1465 | let frac_1_pi = bf16::from_f64(core::f64::consts::FRAC_1_PI);
|
1466 | let frac_1_sqrt_2 = bf16::from_f64(core::f64::consts::FRAC_1_SQRT_2);
|
1467 | let frac_2_pi = bf16::from_f64(core::f64::consts::FRAC_2_PI);
|
1468 | let frac_2_sqrt_pi = bf16::from_f64(core::f64::consts::FRAC_2_SQRT_PI);
|
1469 | let frac_pi_2 = bf16::from_f64(core::f64::consts::FRAC_PI_2);
|
1470 | let frac_pi_3 = bf16::from_f64(core::f64::consts::FRAC_PI_3);
|
1471 | let frac_pi_4 = bf16::from_f64(core::f64::consts::FRAC_PI_4);
|
1472 | let frac_pi_6 = bf16::from_f64(core::f64::consts::FRAC_PI_6);
|
1473 | let frac_pi_8 = bf16::from_f64(core::f64::consts::FRAC_PI_8);
|
1474 | let ln_10 = bf16::from_f64(core::f64::consts::LN_10);
|
1475 | let ln_2 = bf16::from_f64(core::f64::consts::LN_2);
|
1476 | let log10_e = bf16::from_f64(core::f64::consts::LOG10_E);
|
1477 | // core::f64::consts::LOG10_2 requires rustc 1.43.0
|
1478 | let log10_2 = bf16::from_f64(2f64.log10());
|
1479 | let log2_e = bf16::from_f64(core::f64::consts::LOG2_E);
|
1480 | // core::f64::consts::LOG2_10 requires rustc 1.43.0
|
1481 | let log2_10 = bf16::from_f64(10f64.log2());
|
1482 | let sqrt_2 = bf16::from_f64(core::f64::consts::SQRT_2);
|
1483 |
|
1484 | assert_eq!(bf16::E, e);
|
1485 | assert_eq!(bf16::PI, pi);
|
1486 | assert_eq!(bf16::FRAC_1_PI, frac_1_pi);
|
1487 | assert_eq!(bf16::FRAC_1_SQRT_2, frac_1_sqrt_2);
|
1488 | assert_eq!(bf16::FRAC_2_PI, frac_2_pi);
|
1489 | assert_eq!(bf16::FRAC_2_SQRT_PI, frac_2_sqrt_pi);
|
1490 | assert_eq!(bf16::FRAC_PI_2, frac_pi_2);
|
1491 | assert_eq!(bf16::FRAC_PI_3, frac_pi_3);
|
1492 | assert_eq!(bf16::FRAC_PI_4, frac_pi_4);
|
1493 | assert_eq!(bf16::FRAC_PI_6, frac_pi_6);
|
1494 | assert_eq!(bf16::FRAC_PI_8, frac_pi_8);
|
1495 | assert_eq!(bf16::LN_10, ln_10);
|
1496 | assert_eq!(bf16::LN_2, ln_2);
|
1497 | assert_eq!(bf16::LOG10_E, log10_e);
|
1498 | assert_eq!(bf16::LOG10_2, log10_2);
|
1499 | assert_eq!(bf16::LOG2_E, log2_e);
|
1500 | assert_eq!(bf16::LOG2_10, log2_10);
|
1501 | assert_eq!(bf16::SQRT_2, sqrt_2);
|
1502 | }
|
1503 |
|
1504 | #[test ]
|
1505 | fn test_nan_conversion_to_smaller() {
|
1506 | let nan64 = f64::from_bits(0x7FF0_0000_0000_0001u64);
|
1507 | let neg_nan64 = f64::from_bits(0xFFF0_0000_0000_0001u64);
|
1508 | let nan32 = f32::from_bits(0x7F80_0001u32);
|
1509 | let neg_nan32 = f32::from_bits(0xFF80_0001u32);
|
1510 | let nan32_from_64 = nan64 as f32;
|
1511 | let neg_nan32_from_64 = neg_nan64 as f32;
|
1512 | let nan16_from_64 = bf16::from_f64(nan64);
|
1513 | let neg_nan16_from_64 = bf16::from_f64(neg_nan64);
|
1514 | let nan16_from_32 = bf16::from_f32(nan32);
|
1515 | let neg_nan16_from_32 = bf16::from_f32(neg_nan32);
|
1516 |
|
1517 | assert!(nan64.is_nan() && nan64.is_sign_positive());
|
1518 | assert!(neg_nan64.is_nan() && neg_nan64.is_sign_negative());
|
1519 | assert!(nan32.is_nan() && nan32.is_sign_positive());
|
1520 | assert!(neg_nan32.is_nan() && neg_nan32.is_sign_negative());
|
1521 |
|
1522 | // f32/f64 NaN conversion sign is non-deterministic: https://github.com/starkat99/half-rs/issues/103
|
1523 | assert!(neg_nan32_from_64.is_nan());
|
1524 | assert!(nan32_from_64.is_nan());
|
1525 | assert!(nan16_from_64.is_nan());
|
1526 | assert!(neg_nan16_from_64.is_nan());
|
1527 | assert!(nan16_from_32.is_nan());
|
1528 | assert!(neg_nan16_from_32.is_nan());
|
1529 | }
|
1530 |
|
1531 | #[test ]
|
1532 | fn test_nan_conversion_to_larger() {
|
1533 | let nan16 = bf16::from_bits(0x7F81u16);
|
1534 | let neg_nan16 = bf16::from_bits(0xFF81u16);
|
1535 | let nan32 = f32::from_bits(0x7F80_0001u32);
|
1536 | let neg_nan32 = f32::from_bits(0xFF80_0001u32);
|
1537 | let nan32_from_16 = f32::from(nan16);
|
1538 | let neg_nan32_from_16 = f32::from(neg_nan16);
|
1539 | let nan64_from_16 = f64::from(nan16);
|
1540 | let neg_nan64_from_16 = f64::from(neg_nan16);
|
1541 | let nan64_from_32 = f64::from(nan32);
|
1542 | let neg_nan64_from_32 = f64::from(neg_nan32);
|
1543 |
|
1544 | assert!(nan16.is_nan() && nan16.is_sign_positive());
|
1545 | assert!(neg_nan16.is_nan() && neg_nan16.is_sign_negative());
|
1546 | assert!(nan32.is_nan() && nan32.is_sign_positive());
|
1547 | assert!(neg_nan32.is_nan() && neg_nan32.is_sign_negative());
|
1548 |
|
1549 | // // f32/f64 NaN conversion sign is non-deterministic: https://github.com/starkat99/half-rs/issues/103
|
1550 | assert!(nan32_from_16.is_nan());
|
1551 | assert!(neg_nan32_from_16.is_nan());
|
1552 | assert!(nan64_from_16.is_nan());
|
1553 | assert!(neg_nan64_from_16.is_nan());
|
1554 | assert!(nan64_from_32.is_nan());
|
1555 | assert!(neg_nan64_from_32.is_nan());
|
1556 | }
|
1557 |
|
1558 | #[test ]
|
1559 | fn test_bf16_to_f32() {
|
1560 | let f = bf16::from_f32(7.0);
|
1561 | assert_eq!(f.to_f32(), 7.0f32);
|
1562 |
|
1563 | // 7.1 is NOT exactly representable in 16-bit, it's rounded
|
1564 | let f = bf16::from_f32(7.1);
|
1565 | let diff = (f.to_f32() - 7.1f32).abs();
|
1566 | // diff must be <= 4 * EPSILON, as 7 has two more significant bits than 1
|
1567 | assert!(diff <= 4.0 * bf16::EPSILON.to_f32());
|
1568 |
|
1569 | let tiny32 = f32::from_bits(0x0001_0000u32);
|
1570 | assert_eq!(bf16::from_bits(0x0001).to_f32(), tiny32);
|
1571 | assert_eq!(bf16::from_bits(0x0005).to_f32(), 5.0 * tiny32);
|
1572 |
|
1573 | assert_eq!(bf16::from_bits(0x0001), bf16::from_f32(tiny32));
|
1574 | assert_eq!(bf16::from_bits(0x0005), bf16::from_f32(5.0 * tiny32));
|
1575 | }
|
1576 |
|
1577 | #[test ]
|
1578 | fn test_bf16_to_f64() {
|
1579 | let f = bf16::from_f64(7.0);
|
1580 | assert_eq!(f.to_f64(), 7.0f64);
|
1581 |
|
1582 | // 7.1 is NOT exactly representable in 16-bit, it's rounded
|
1583 | let f = bf16::from_f64(7.1);
|
1584 | let diff = (f.to_f64() - 7.1f64).abs();
|
1585 | // diff must be <= 4 * EPSILON, as 7 has two more significant bits than 1
|
1586 | assert!(diff <= 4.0 * bf16::EPSILON.to_f64());
|
1587 |
|
1588 | let tiny64 = 2.0f64.powi(-133);
|
1589 | assert_eq!(bf16::from_bits(0x0001).to_f64(), tiny64);
|
1590 | assert_eq!(bf16::from_bits(0x0005).to_f64(), 5.0 * tiny64);
|
1591 |
|
1592 | assert_eq!(bf16::from_bits(0x0001), bf16::from_f64(tiny64));
|
1593 | assert_eq!(bf16::from_bits(0x0005), bf16::from_f64(5.0 * tiny64));
|
1594 | }
|
1595 |
|
1596 | #[test ]
|
1597 | fn test_comparisons() {
|
1598 | let zero = bf16::from_f64(0.0);
|
1599 | let one = bf16::from_f64(1.0);
|
1600 | let neg_zero = bf16::from_f64(-0.0);
|
1601 | let neg_one = bf16::from_f64(-1.0);
|
1602 |
|
1603 | assert_eq!(zero.partial_cmp(&neg_zero), Some(Ordering::Equal));
|
1604 | assert_eq!(neg_zero.partial_cmp(&zero), Some(Ordering::Equal));
|
1605 | assert!(zero == neg_zero);
|
1606 | assert!(neg_zero == zero);
|
1607 | assert!(!(zero != neg_zero));
|
1608 | assert!(!(neg_zero != zero));
|
1609 | assert!(!(zero < neg_zero));
|
1610 | assert!(!(neg_zero < zero));
|
1611 | assert!(zero <= neg_zero);
|
1612 | assert!(neg_zero <= zero);
|
1613 | assert!(!(zero > neg_zero));
|
1614 | assert!(!(neg_zero > zero));
|
1615 | assert!(zero >= neg_zero);
|
1616 | assert!(neg_zero >= zero);
|
1617 |
|
1618 | assert_eq!(one.partial_cmp(&neg_zero), Some(Ordering::Greater));
|
1619 | assert_eq!(neg_zero.partial_cmp(&one), Some(Ordering::Less));
|
1620 | assert!(!(one == neg_zero));
|
1621 | assert!(!(neg_zero == one));
|
1622 | assert!(one != neg_zero);
|
1623 | assert!(neg_zero != one);
|
1624 | assert!(!(one < neg_zero));
|
1625 | assert!(neg_zero < one);
|
1626 | assert!(!(one <= neg_zero));
|
1627 | assert!(neg_zero <= one);
|
1628 | assert!(one > neg_zero);
|
1629 | assert!(!(neg_zero > one));
|
1630 | assert!(one >= neg_zero);
|
1631 | assert!(!(neg_zero >= one));
|
1632 |
|
1633 | assert_eq!(one.partial_cmp(&neg_one), Some(Ordering::Greater));
|
1634 | assert_eq!(neg_one.partial_cmp(&one), Some(Ordering::Less));
|
1635 | assert!(!(one == neg_one));
|
1636 | assert!(!(neg_one == one));
|
1637 | assert!(one != neg_one);
|
1638 | assert!(neg_one != one);
|
1639 | assert!(!(one < neg_one));
|
1640 | assert!(neg_one < one);
|
1641 | assert!(!(one <= neg_one));
|
1642 | assert!(neg_one <= one);
|
1643 | assert!(one > neg_one);
|
1644 | assert!(!(neg_one > one));
|
1645 | assert!(one >= neg_one);
|
1646 | assert!(!(neg_one >= one));
|
1647 | }
|
1648 |
|
1649 | #[test ]
|
1650 | #[allow (clippy::erasing_op, clippy::identity_op)]
|
1651 | fn round_to_even_f32() {
|
1652 | // smallest positive subnormal = 0b0.0000_001 * 2^-126 = 2^-133
|
1653 | let min_sub = bf16::from_bits(1);
|
1654 | let min_sub_f = (-133f32).exp2();
|
1655 | assert_eq!(bf16::from_f32(min_sub_f).to_bits(), min_sub.to_bits());
|
1656 | assert_eq!(f32::from(min_sub).to_bits(), min_sub_f.to_bits());
|
1657 |
|
1658 | // 0.0000000_011111 rounded to 0.0000000 (< tie, no rounding)
|
1659 | // 0.0000000_100000 rounded to 0.0000000 (tie and even, remains at even)
|
1660 | // 0.0000000_100001 rounded to 0.0000001 (> tie, rounds up)
|
1661 | assert_eq!(
|
1662 | bf16::from_f32(min_sub_f * 0.49).to_bits(),
|
1663 | min_sub.to_bits() * 0
|
1664 | );
|
1665 | assert_eq!(
|
1666 | bf16::from_f32(min_sub_f * 0.50).to_bits(),
|
1667 | min_sub.to_bits() * 0
|
1668 | );
|
1669 | assert_eq!(
|
1670 | bf16::from_f32(min_sub_f * 0.51).to_bits(),
|
1671 | min_sub.to_bits() * 1
|
1672 | );
|
1673 |
|
1674 | // 0.0000001_011111 rounded to 0.0000001 (< tie, no rounding)
|
1675 | // 0.0000001_100000 rounded to 0.0000010 (tie and odd, rounds up to even)
|
1676 | // 0.0000001_100001 rounded to 0.0000010 (> tie, rounds up)
|
1677 | assert_eq!(
|
1678 | bf16::from_f32(min_sub_f * 1.49).to_bits(),
|
1679 | min_sub.to_bits() * 1
|
1680 | );
|
1681 | assert_eq!(
|
1682 | bf16::from_f32(min_sub_f * 1.50).to_bits(),
|
1683 | min_sub.to_bits() * 2
|
1684 | );
|
1685 | assert_eq!(
|
1686 | bf16::from_f32(min_sub_f * 1.51).to_bits(),
|
1687 | min_sub.to_bits() * 2
|
1688 | );
|
1689 |
|
1690 | // 0.0000010_011111 rounded to 0.0000010 (< tie, no rounding)
|
1691 | // 0.0000010_100000 rounded to 0.0000010 (tie and even, remains at even)
|
1692 | // 0.0000010_100001 rounded to 0.0000011 (> tie, rounds up)
|
1693 | assert_eq!(
|
1694 | bf16::from_f32(min_sub_f * 2.49).to_bits(),
|
1695 | min_sub.to_bits() * 2
|
1696 | );
|
1697 | assert_eq!(
|
1698 | bf16::from_f32(min_sub_f * 2.50).to_bits(),
|
1699 | min_sub.to_bits() * 2
|
1700 | );
|
1701 | assert_eq!(
|
1702 | bf16::from_f32(min_sub_f * 2.51).to_bits(),
|
1703 | min_sub.to_bits() * 3
|
1704 | );
|
1705 |
|
1706 | assert_eq!(
|
1707 | bf16::from_f32(250.49f32).to_bits(),
|
1708 | bf16::from_f32(250.0).to_bits()
|
1709 | );
|
1710 | assert_eq!(
|
1711 | bf16::from_f32(250.50f32).to_bits(),
|
1712 | bf16::from_f32(250.0).to_bits()
|
1713 | );
|
1714 | assert_eq!(
|
1715 | bf16::from_f32(250.51f32).to_bits(),
|
1716 | bf16::from_f32(251.0).to_bits()
|
1717 | );
|
1718 | assert_eq!(
|
1719 | bf16::from_f32(251.49f32).to_bits(),
|
1720 | bf16::from_f32(251.0).to_bits()
|
1721 | );
|
1722 | assert_eq!(
|
1723 | bf16::from_f32(251.50f32).to_bits(),
|
1724 | bf16::from_f32(252.0).to_bits()
|
1725 | );
|
1726 | assert_eq!(
|
1727 | bf16::from_f32(251.51f32).to_bits(),
|
1728 | bf16::from_f32(252.0).to_bits()
|
1729 | );
|
1730 | assert_eq!(
|
1731 | bf16::from_f32(252.49f32).to_bits(),
|
1732 | bf16::from_f32(252.0).to_bits()
|
1733 | );
|
1734 | assert_eq!(
|
1735 | bf16::from_f32(252.50f32).to_bits(),
|
1736 | bf16::from_f32(252.0).to_bits()
|
1737 | );
|
1738 | assert_eq!(
|
1739 | bf16::from_f32(252.51f32).to_bits(),
|
1740 | bf16::from_f32(253.0).to_bits()
|
1741 | );
|
1742 | }
|
1743 |
|
1744 | #[test ]
|
1745 | #[allow (clippy::erasing_op, clippy::identity_op)]
|
1746 | fn round_to_even_f64() {
|
1747 | // smallest positive subnormal = 0b0.0000_001 * 2^-126 = 2^-133
|
1748 | let min_sub = bf16::from_bits(1);
|
1749 | let min_sub_f = (-133f64).exp2();
|
1750 | assert_eq!(bf16::from_f64(min_sub_f).to_bits(), min_sub.to_bits());
|
1751 | assert_eq!(f64::from(min_sub).to_bits(), min_sub_f.to_bits());
|
1752 |
|
1753 | // 0.0000000_011111 rounded to 0.0000000 (< tie, no rounding)
|
1754 | // 0.0000000_100000 rounded to 0.0000000 (tie and even, remains at even)
|
1755 | // 0.0000000_100001 rounded to 0.0000001 (> tie, rounds up)
|
1756 | assert_eq!(
|
1757 | bf16::from_f64(min_sub_f * 0.49).to_bits(),
|
1758 | min_sub.to_bits() * 0
|
1759 | );
|
1760 | assert_eq!(
|
1761 | bf16::from_f64(min_sub_f * 0.50).to_bits(),
|
1762 | min_sub.to_bits() * 0
|
1763 | );
|
1764 | assert_eq!(
|
1765 | bf16::from_f64(min_sub_f * 0.51).to_bits(),
|
1766 | min_sub.to_bits() * 1
|
1767 | );
|
1768 |
|
1769 | // 0.0000001_011111 rounded to 0.0000001 (< tie, no rounding)
|
1770 | // 0.0000001_100000 rounded to 0.0000010 (tie and odd, rounds up to even)
|
1771 | // 0.0000001_100001 rounded to 0.0000010 (> tie, rounds up)
|
1772 | assert_eq!(
|
1773 | bf16::from_f64(min_sub_f * 1.49).to_bits(),
|
1774 | min_sub.to_bits() * 1
|
1775 | );
|
1776 | assert_eq!(
|
1777 | bf16::from_f64(min_sub_f * 1.50).to_bits(),
|
1778 | min_sub.to_bits() * 2
|
1779 | );
|
1780 | assert_eq!(
|
1781 | bf16::from_f64(min_sub_f * 1.51).to_bits(),
|
1782 | min_sub.to_bits() * 2
|
1783 | );
|
1784 |
|
1785 | // 0.0000010_011111 rounded to 0.0000010 (< tie, no rounding)
|
1786 | // 0.0000010_100000 rounded to 0.0000010 (tie and even, remains at even)
|
1787 | // 0.0000010_100001 rounded to 0.0000011 (> tie, rounds up)
|
1788 | assert_eq!(
|
1789 | bf16::from_f64(min_sub_f * 2.49).to_bits(),
|
1790 | min_sub.to_bits() * 2
|
1791 | );
|
1792 | assert_eq!(
|
1793 | bf16::from_f64(min_sub_f * 2.50).to_bits(),
|
1794 | min_sub.to_bits() * 2
|
1795 | );
|
1796 | assert_eq!(
|
1797 | bf16::from_f64(min_sub_f * 2.51).to_bits(),
|
1798 | min_sub.to_bits() * 3
|
1799 | );
|
1800 |
|
1801 | assert_eq!(
|
1802 | bf16::from_f64(250.49f64).to_bits(),
|
1803 | bf16::from_f64(250.0).to_bits()
|
1804 | );
|
1805 | assert_eq!(
|
1806 | bf16::from_f64(250.50f64).to_bits(),
|
1807 | bf16::from_f64(250.0).to_bits()
|
1808 | );
|
1809 | assert_eq!(
|
1810 | bf16::from_f64(250.51f64).to_bits(),
|
1811 | bf16::from_f64(251.0).to_bits()
|
1812 | );
|
1813 | assert_eq!(
|
1814 | bf16::from_f64(251.49f64).to_bits(),
|
1815 | bf16::from_f64(251.0).to_bits()
|
1816 | );
|
1817 | assert_eq!(
|
1818 | bf16::from_f64(251.50f64).to_bits(),
|
1819 | bf16::from_f64(252.0).to_bits()
|
1820 | );
|
1821 | assert_eq!(
|
1822 | bf16::from_f64(251.51f64).to_bits(),
|
1823 | bf16::from_f64(252.0).to_bits()
|
1824 | );
|
1825 | assert_eq!(
|
1826 | bf16::from_f64(252.49f64).to_bits(),
|
1827 | bf16::from_f64(252.0).to_bits()
|
1828 | );
|
1829 | assert_eq!(
|
1830 | bf16::from_f64(252.50f64).to_bits(),
|
1831 | bf16::from_f64(252.0).to_bits()
|
1832 | );
|
1833 | assert_eq!(
|
1834 | bf16::from_f64(252.51f64).to_bits(),
|
1835 | bf16::from_f64(253.0).to_bits()
|
1836 | );
|
1837 | }
|
1838 |
|
1839 | #[cfg (feature = "std" )]
|
1840 | #[test ]
|
1841 | fn formatting() {
|
1842 | let f = bf16::from_f32(0.1152344);
|
1843 |
|
1844 | assert_eq!(format!("{:.3}" , f), "0.115" );
|
1845 | assert_eq!(format!("{:.4}" , f), "0.1152" );
|
1846 | assert_eq!(format!("{:+.4}" , f), "+0.1152" );
|
1847 | assert_eq!(format!("{:>+10.4}" , f), " +0.1152" );
|
1848 |
|
1849 | assert_eq!(format!("{:.3?}" , f), "0.115" );
|
1850 | assert_eq!(format!("{:.4?}" , f), "0.1152" );
|
1851 | assert_eq!(format!("{:+.4?}" , f), "+0.1152" );
|
1852 | assert_eq!(format!("{:>+10.4?}" , f), " +0.1152" );
|
1853 | }
|
1854 |
|
1855 | impl quickcheck::Arbitrary for bf16 {
|
1856 | fn arbitrary(g: &mut quickcheck::Gen) -> Self {
|
1857 | bf16(u16::arbitrary(g))
|
1858 | }
|
1859 | }
|
1860 |
|
1861 | #[quickcheck]
|
1862 | fn qc_roundtrip_bf16_f32_is_identity(f: bf16) -> bool {
|
1863 | let roundtrip = bf16::from_f32(f.to_f32());
|
1864 | if f.is_nan() {
|
1865 | roundtrip.is_nan() && f.is_sign_negative() == roundtrip.is_sign_negative()
|
1866 | } else {
|
1867 | f.0 == roundtrip.0
|
1868 | }
|
1869 | }
|
1870 |
|
1871 | #[quickcheck]
|
1872 | fn qc_roundtrip_bf16_f64_is_identity(f: bf16) -> bool {
|
1873 | let roundtrip = bf16::from_f64(f.to_f64());
|
1874 | if f.is_nan() {
|
1875 | roundtrip.is_nan() && f.is_sign_negative() == roundtrip.is_sign_negative()
|
1876 | } else {
|
1877 | f.0 == roundtrip.0
|
1878 | }
|
1879 | }
|
1880 | }
|
1881 | |