1 | use core::{fmt, mem, ops}; |
2 | |
3 | use super::int_traits::{CastFrom, Int, MinInt}; |
4 | |
5 | /// Trait for some basic operations on floats |
6 | // #[allow(dead_code)] |
7 | pub trait Float: |
8 | Copy |
9 | + fmt::Debug |
10 | + PartialEq |
11 | + PartialOrd |
12 | + ops::AddAssign |
13 | + ops::MulAssign |
14 | + ops::Add<Output = Self> |
15 | + ops::Sub<Output = Self> |
16 | + ops::Mul<Output = Self> |
17 | + ops::Div<Output = Self> |
18 | + ops::Rem<Output = Self> |
19 | + ops::Neg<Output = Self> |
20 | + 'static |
21 | { |
22 | /// A uint of the same width as the float |
23 | type Int: Int<OtherSign = Self::SignedInt, Unsigned = Self::Int>; |
24 | |
25 | /// A int of the same width as the float |
26 | type SignedInt: Int |
27 | + MinInt<OtherSign = Self::Int, Unsigned = Self::Int> |
28 | + ops::Neg<Output = Self::SignedInt>; |
29 | |
30 | const ZERO: Self; |
31 | const NEG_ZERO: Self; |
32 | const ONE: Self; |
33 | const NEG_ONE: Self; |
34 | const INFINITY: Self; |
35 | const NEG_INFINITY: Self; |
36 | const NAN: Self; |
37 | const NEG_NAN: Self; |
38 | const MAX: Self; |
39 | const MIN: Self; |
40 | const EPSILON: Self; |
41 | const PI: Self; |
42 | const NEG_PI: Self; |
43 | const FRAC_PI_2: Self; |
44 | |
45 | const MIN_POSITIVE_NORMAL: Self; |
46 | |
47 | /// The bitwidth of the float type |
48 | const BITS: u32; |
49 | |
50 | /// The bitwidth of the significand |
51 | const SIG_BITS: u32; |
52 | |
53 | /// The bitwidth of the exponent |
54 | const EXP_BITS: u32 = Self::BITS - Self::SIG_BITS - 1; |
55 | |
56 | /// The saturated (maximum bitpattern) value of the exponent, i.e. the infinite |
57 | /// representation. |
58 | /// |
59 | /// This shifted fully right, use `EXP_MASK` for the shifted value. |
60 | const EXP_SAT: u32 = (1 << Self::EXP_BITS) - 1; |
61 | |
62 | /// The exponent bias value |
63 | const EXP_BIAS: u32 = Self::EXP_SAT >> 1; |
64 | |
65 | /// Maximum unbiased exponent value. |
66 | const EXP_MAX: i32 = Self::EXP_BIAS as i32; |
67 | |
68 | /// Minimum *NORMAL* unbiased exponent value. |
69 | const EXP_MIN: i32 = -(Self::EXP_MAX - 1); |
70 | |
71 | /// Minimum subnormal exponent value. |
72 | const EXP_MIN_SUBNORM: i32 = Self::EXP_MIN - Self::SIG_BITS as i32; |
73 | |
74 | /// A mask for the sign bit |
75 | const SIGN_MASK: Self::Int; |
76 | |
77 | /// A mask for the significand |
78 | const SIG_MASK: Self::Int; |
79 | |
80 | /// A mask for the exponent |
81 | const EXP_MASK: Self::Int; |
82 | |
83 | /// The implicit bit of the float format |
84 | const IMPLICIT_BIT: Self::Int; |
85 | |
86 | /// Returns `self` transmuted to `Self::Int` |
87 | fn to_bits(self) -> Self::Int; |
88 | |
89 | /// Returns `self` transmuted to `Self::SignedInt` |
90 | #[allow (dead_code)] |
91 | fn to_bits_signed(self) -> Self::SignedInt { |
92 | self.to_bits().signed() |
93 | } |
94 | |
95 | /// Check bitwise equality. |
96 | #[allow (dead_code)] |
97 | fn biteq(self, rhs: Self) -> bool { |
98 | self.to_bits() == rhs.to_bits() |
99 | } |
100 | |
101 | /// Checks if two floats have the same bit representation. *Except* for NaNs! NaN can be |
102 | /// represented in multiple different ways. |
103 | /// |
104 | /// This method returns `true` if two NaNs are compared. Use [`biteq`](Self::biteq) instead |
105 | /// if `NaN` should not be treated separately. |
106 | #[allow (dead_code)] |
107 | fn eq_repr(self, rhs: Self) -> bool { |
108 | if self.is_nan() && rhs.is_nan() { |
109 | true |
110 | } else { |
111 | self.biteq(rhs) |
112 | } |
113 | } |
114 | |
115 | /// Returns true if the value is NaN. |
116 | fn is_nan(self) -> bool; |
117 | |
118 | /// Returns true if the value is +inf or -inf. |
119 | fn is_infinite(self) -> bool; |
120 | |
121 | /// Returns true if the sign is negative. Extracts the sign bit regardless of zero or NaN. |
122 | fn is_sign_negative(self) -> bool; |
123 | |
124 | /// Returns true if the sign is positive. Extracts the sign bit regardless of zero or NaN. |
125 | fn is_sign_positive(self) -> bool { |
126 | !self.is_sign_negative() |
127 | } |
128 | |
129 | /// Returns if `self` is subnormal. |
130 | #[allow (dead_code)] |
131 | fn is_subnormal(self) -> bool { |
132 | (self.to_bits() & Self::EXP_MASK) == Self::Int::ZERO |
133 | } |
134 | |
135 | /// Returns the exponent, not adjusting for bias, not accounting for subnormals or zero. |
136 | fn ex(self) -> u32 { |
137 | u32::cast_from(self.to_bits() >> Self::SIG_BITS) & Self::EXP_SAT |
138 | } |
139 | |
140 | /// Extract the exponent and adjust it for bias, not accounting for subnormals or zero. |
141 | fn exp_unbiased(self) -> i32 { |
142 | self.ex().signed() - (Self::EXP_BIAS as i32) |
143 | } |
144 | |
145 | /// Returns the significand with no implicit bit (or the "fractional" part) |
146 | #[allow (dead_code)] |
147 | fn frac(self) -> Self::Int { |
148 | self.to_bits() & Self::SIG_MASK |
149 | } |
150 | |
151 | /// Returns a `Self::Int` transmuted back to `Self` |
152 | fn from_bits(a: Self::Int) -> Self; |
153 | |
154 | /// Constructs a `Self` from its parts. Inputs are treated as bits and shifted into position. |
155 | fn from_parts(negative: bool, exponent: u32, significand: Self::Int) -> Self { |
156 | let sign = if negative { |
157 | Self::Int::ONE |
158 | } else { |
159 | Self::Int::ZERO |
160 | }; |
161 | Self::from_bits( |
162 | (sign << (Self::BITS - 1)) |
163 | | (Self::Int::cast_from(exponent & Self::EXP_SAT) << Self::SIG_BITS) |
164 | | (significand & Self::SIG_MASK), |
165 | ) |
166 | } |
167 | |
168 | #[allow (dead_code)] |
169 | fn abs(self) -> Self; |
170 | |
171 | /// Returns a number composed of the magnitude of self and the sign of sign. |
172 | fn copysign(self, other: Self) -> Self; |
173 | |
174 | /// Fused multiply add, rounding once. |
175 | fn fma(self, y: Self, z: Self) -> Self; |
176 | |
177 | /// Returns (normalized exponent, normalized significand) |
178 | #[allow (dead_code)] |
179 | fn normalize(significand: Self::Int) -> (i32, Self::Int); |
180 | |
181 | /// Returns a number that represents the sign of self. |
182 | #[allow (dead_code)] |
183 | fn signum(self) -> Self { |
184 | if self.is_nan() { |
185 | self |
186 | } else { |
187 | Self::ONE.copysign(self) |
188 | } |
189 | } |
190 | } |
191 | |
192 | /// Access the associated `Int` type from a float (helper to avoid ambiguous associated types). |
193 | pub type IntTy<F> = <F as Float>::Int; |
194 | |
195 | macro_rules! float_impl { |
196 | ( |
197 | $ty:ident, |
198 | $ity:ident, |
199 | $sity:ident, |
200 | $bits:expr, |
201 | $significand_bits:expr, |
202 | $from_bits:path, |
203 | $to_bits:path, |
204 | $fma_fn:ident, |
205 | $fma_intrinsic:ident |
206 | ) => { |
207 | impl Float for $ty { |
208 | type Int = $ity; |
209 | type SignedInt = $sity; |
210 | |
211 | const ZERO: Self = 0.0; |
212 | const NEG_ZERO: Self = -0.0; |
213 | const ONE: Self = 1.0; |
214 | const NEG_ONE: Self = -1.0; |
215 | const INFINITY: Self = Self::INFINITY; |
216 | const NEG_INFINITY: Self = Self::NEG_INFINITY; |
217 | const NAN: Self = Self::NAN; |
218 | // NAN isn't guaranteed to be positive but it usually is. We only use this for |
219 | // tests. |
220 | const NEG_NAN: Self = $from_bits($to_bits(Self::NAN) | Self::SIGN_MASK); |
221 | const MAX: Self = -Self::MIN; |
222 | // Sign bit set, saturated mantissa, saturated exponent with last bit zeroed |
223 | const MIN: Self = $from_bits(Self::Int::MAX & !(1 << Self::SIG_BITS)); |
224 | const EPSILON: Self = <$ty>::EPSILON; |
225 | |
226 | // Exponent is a 1 in the LSB |
227 | const MIN_POSITIVE_NORMAL: Self = $from_bits(1 << Self::SIG_BITS); |
228 | |
229 | const PI: Self = core::$ty::consts::PI; |
230 | const NEG_PI: Self = -Self::PI; |
231 | const FRAC_PI_2: Self = core::$ty::consts::FRAC_PI_2; |
232 | |
233 | const BITS: u32 = $bits; |
234 | const SIG_BITS: u32 = $significand_bits; |
235 | |
236 | const SIGN_MASK: Self::Int = 1 << (Self::BITS - 1); |
237 | const SIG_MASK: Self::Int = (1 << Self::SIG_BITS) - 1; |
238 | const EXP_MASK: Self::Int = !(Self::SIGN_MASK | Self::SIG_MASK); |
239 | const IMPLICIT_BIT: Self::Int = 1 << Self::SIG_BITS; |
240 | |
241 | fn to_bits(self) -> Self::Int { |
242 | self.to_bits() |
243 | } |
244 | fn is_nan(self) -> bool { |
245 | self.is_nan() |
246 | } |
247 | fn is_infinite(self) -> bool { |
248 | self.is_infinite() |
249 | } |
250 | fn is_sign_negative(self) -> bool { |
251 | self.is_sign_negative() |
252 | } |
253 | fn from_bits(a: Self::Int) -> Self { |
254 | Self::from_bits(a) |
255 | } |
256 | fn abs(self) -> Self { |
257 | cfg_if! { |
258 | // FIXME(msrv): `abs` is available in `core` starting with 1.85. |
259 | if #[cfg(intrinsics_enabled)] { |
260 | self.abs() |
261 | } else { |
262 | super::super::generic::fabs(self) |
263 | } |
264 | } |
265 | } |
266 | fn copysign(self, other: Self) -> Self { |
267 | cfg_if! { |
268 | // FIXME(msrv): `copysign` is available in `core` starting with 1.85. |
269 | if #[cfg(intrinsics_enabled)] { |
270 | self.copysign(other) |
271 | } else { |
272 | super::super::generic::copysign(self, other) |
273 | } |
274 | } |
275 | } |
276 | fn fma(self, y: Self, z: Self) -> Self { |
277 | cfg_if! { |
278 | // fma is not yet available in `core` |
279 | if #[cfg(intrinsics_enabled)] { |
280 | unsafe{ core::intrinsics::$fma_intrinsic(self, y, z) } |
281 | } else { |
282 | super::super::$fma_fn(self, y, z) |
283 | } |
284 | } |
285 | } |
286 | fn normalize(significand: Self::Int) -> (i32, Self::Int) { |
287 | let shift = significand.leading_zeros().wrapping_sub(Self::EXP_BITS); |
288 | ( |
289 | 1i32.wrapping_sub(shift as i32), |
290 | significand << shift as Self::Int, |
291 | ) |
292 | } |
293 | } |
294 | }; |
295 | } |
296 | |
297 | #[cfg (f16_enabled)] |
298 | float_impl!( |
299 | f16, |
300 | u16, |
301 | i16, |
302 | 16, |
303 | 10, |
304 | f16::from_bits, |
305 | f16::to_bits, |
306 | fmaf16, |
307 | fmaf16 |
308 | ); |
309 | float_impl!( |
310 | f32, |
311 | u32, |
312 | i32, |
313 | 32, |
314 | 23, |
315 | f32_from_bits, |
316 | f32_to_bits, |
317 | fmaf, |
318 | fmaf32 |
319 | ); |
320 | float_impl!( |
321 | f64, |
322 | u64, |
323 | i64, |
324 | 64, |
325 | 52, |
326 | f64_from_bits, |
327 | f64_to_bits, |
328 | fma, |
329 | fmaf64 |
330 | ); |
331 | #[cfg (f128_enabled)] |
332 | float_impl!( |
333 | f128, |
334 | u128, |
335 | i128, |
336 | 128, |
337 | 112, |
338 | f128::from_bits, |
339 | f128::to_bits, |
340 | fmaf128, |
341 | fmaf128 |
342 | ); |
343 | |
344 | /* FIXME(msrv): vendor some things that are not const stable at our MSRV */ |
345 | |
346 | /// `f32::from_bits` |
347 | pub const fn f32_from_bits(bits: u32) -> f32 { |
348 | // SAFETY: POD cast with no preconditions |
349 | unsafe { mem::transmute::<u32, f32>(src:bits) } |
350 | } |
351 | |
352 | /// `f32::to_bits` |
353 | pub const fn f32_to_bits(x: f32) -> u32 { |
354 | // SAFETY: POD cast with no preconditions |
355 | unsafe { mem::transmute::<f32, u32>(src:x) } |
356 | } |
357 | |
358 | /// `f64::from_bits` |
359 | pub const fn f64_from_bits(bits: u64) -> f64 { |
360 | // SAFETY: POD cast with no preconditions |
361 | unsafe { mem::transmute::<u64, f64>(src:bits) } |
362 | } |
363 | |
364 | /// `f64::to_bits` |
365 | pub const fn f64_to_bits(x: f64) -> u64 { |
366 | // SAFETY: POD cast with no preconditions |
367 | unsafe { mem::transmute::<f64, u64>(src:x) } |
368 | } |
369 | |
370 | /// Trait for floats twice the bit width of another integer. |
371 | pub trait DFloat: Float { |
372 | /// Float that is half the bit width of the floatthis trait is implemented for. |
373 | type H: HFloat<D = Self>; |
374 | |
375 | /// Narrow the float type. |
376 | fn narrow(self) -> Self::H; |
377 | } |
378 | |
379 | /// Trait for floats half the bit width of another float. |
380 | pub trait HFloat: Float { |
381 | /// Float that is double the bit width of the float this trait is implemented for. |
382 | type D: DFloat<H = Self>; |
383 | |
384 | /// Widen the float type. |
385 | fn widen(self) -> Self::D; |
386 | } |
387 | |
388 | macro_rules! impl_d_float { |
389 | ($($X:ident $D:ident),*) => { |
390 | $( |
391 | impl DFloat for $D { |
392 | type H = $X; |
393 | |
394 | fn narrow(self) -> Self::H { |
395 | self as $X |
396 | } |
397 | } |
398 | )* |
399 | }; |
400 | } |
401 | |
402 | macro_rules! impl_h_float { |
403 | ($($H:ident $X:ident),*) => { |
404 | $( |
405 | impl HFloat for $H { |
406 | type D = $X; |
407 | |
408 | fn widen(self) -> Self::D { |
409 | self as $X |
410 | } |
411 | } |
412 | )* |
413 | }; |
414 | } |
415 | |
416 | impl_d_float!(f32 f64); |
417 | #[cfg (f16_enabled)] |
418 | impl_d_float!(f16 f32); |
419 | #[cfg (f128_enabled)] |
420 | impl_d_float!(f64 f128); |
421 | |
422 | impl_h_float!(f32 f64); |
423 | #[cfg (f16_enabled)] |
424 | impl_h_float!(f16 f32); |
425 | #[cfg (f128_enabled)] |
426 | impl_h_float!(f64 f128); |
427 | |
428 | #[cfg (test)] |
429 | mod tests { |
430 | use super::*; |
431 | |
432 | #[test ] |
433 | #[cfg (f16_enabled)] |
434 | fn check_f16() { |
435 | // Constants |
436 | assert_eq!(f16::EXP_SAT, 0b11111); |
437 | assert_eq!(f16::EXP_BIAS, 15); |
438 | assert_eq!(f16::EXP_MAX, 15); |
439 | assert_eq!(f16::EXP_MIN, -14); |
440 | assert_eq!(f16::EXP_MIN_SUBNORM, -24); |
441 | |
442 | // `exp_unbiased` |
443 | assert_eq!(f16::FRAC_PI_2.exp_unbiased(), 0); |
444 | assert_eq!((1.0f16 / 2.0).exp_unbiased(), -1); |
445 | assert_eq!(f16::MAX.exp_unbiased(), 15); |
446 | assert_eq!(f16::MIN.exp_unbiased(), 15); |
447 | assert_eq!(f16::MIN_POSITIVE.exp_unbiased(), -14); |
448 | // This is a convenience method and not ldexp, `exp_unbiased` does not return correct |
449 | // results for zero and subnormals. |
450 | assert_eq!(f16::ZERO.exp_unbiased(), -15); |
451 | assert_eq!(f16::from_bits(0x1).exp_unbiased(), -15); |
452 | assert_eq!(f16::MIN_POSITIVE, f16::MIN_POSITIVE_NORMAL); |
453 | |
454 | // `from_parts` |
455 | assert_biteq!(f16::from_parts(true, f16::EXP_BIAS, 0), -1.0f16); |
456 | assert_biteq!(f16::from_parts(false, 0, 1), f16::from_bits(0x1)); |
457 | } |
458 | |
459 | #[test ] |
460 | fn check_f32() { |
461 | // Constants |
462 | assert_eq!(f32::EXP_SAT, 0b11111111); |
463 | assert_eq!(f32::EXP_BIAS, 127); |
464 | assert_eq!(f32::EXP_MAX, 127); |
465 | assert_eq!(f32::EXP_MIN, -126); |
466 | assert_eq!(f32::EXP_MIN_SUBNORM, -149); |
467 | |
468 | // `exp_unbiased` |
469 | assert_eq!(f32::FRAC_PI_2.exp_unbiased(), 0); |
470 | assert_eq!((1.0f32 / 2.0).exp_unbiased(), -1); |
471 | assert_eq!(f32::MAX.exp_unbiased(), 127); |
472 | assert_eq!(f32::MIN.exp_unbiased(), 127); |
473 | assert_eq!(f32::MIN_POSITIVE.exp_unbiased(), -126); |
474 | // This is a convenience method and not ldexp, `exp_unbiased` does not return correct |
475 | // results for zero and subnormals. |
476 | assert_eq!(f32::ZERO.exp_unbiased(), -127); |
477 | assert_eq!(f32::from_bits(0x1).exp_unbiased(), -127); |
478 | assert_eq!(f32::MIN_POSITIVE, f32::MIN_POSITIVE_NORMAL); |
479 | |
480 | // `from_parts` |
481 | assert_biteq!(f32::from_parts(true, f32::EXP_BIAS, 0), -1.0f32); |
482 | assert_biteq!( |
483 | f32::from_parts(false, 10 + f32::EXP_BIAS, 0), |
484 | hf32!("0x1p10" ) |
485 | ); |
486 | assert_biteq!(f32::from_parts(false, 0, 1), f32::from_bits(0x1)); |
487 | } |
488 | |
489 | #[test ] |
490 | fn check_f64() { |
491 | // Constants |
492 | assert_eq!(f64::EXP_SAT, 0b11111111111); |
493 | assert_eq!(f64::EXP_BIAS, 1023); |
494 | assert_eq!(f64::EXP_MAX, 1023); |
495 | assert_eq!(f64::EXP_MIN, -1022); |
496 | assert_eq!(f64::EXP_MIN_SUBNORM, -1074); |
497 | |
498 | // `exp_unbiased` |
499 | assert_eq!(f64::FRAC_PI_2.exp_unbiased(), 0); |
500 | assert_eq!((1.0f64 / 2.0).exp_unbiased(), -1); |
501 | assert_eq!(f64::MAX.exp_unbiased(), 1023); |
502 | assert_eq!(f64::MIN.exp_unbiased(), 1023); |
503 | assert_eq!(f64::MIN_POSITIVE.exp_unbiased(), -1022); |
504 | // This is a convenience method and not ldexp, `exp_unbiased` does not return correct |
505 | // results for zero and subnormals. |
506 | assert_eq!(f64::ZERO.exp_unbiased(), -1023); |
507 | assert_eq!(f64::from_bits(0x1).exp_unbiased(), -1023); |
508 | assert_eq!(f64::MIN_POSITIVE, f64::MIN_POSITIVE_NORMAL); |
509 | |
510 | // `from_parts` |
511 | assert_biteq!(f64::from_parts(true, f64::EXP_BIAS, 0), -1.0f64); |
512 | assert_biteq!( |
513 | f64::from_parts(false, 10 + f64::EXP_BIAS, 0), |
514 | hf64!("0x1p10" ) |
515 | ); |
516 | assert_biteq!(f64::from_parts(false, 0, 1), f64::from_bits(0x1)); |
517 | } |
518 | |
519 | #[test ] |
520 | #[cfg (f128_enabled)] |
521 | fn check_f128() { |
522 | // Constants |
523 | assert_eq!(f128::EXP_SAT, 0b111111111111111); |
524 | assert_eq!(f128::EXP_BIAS, 16383); |
525 | assert_eq!(f128::EXP_MAX, 16383); |
526 | assert_eq!(f128::EXP_MIN, -16382); |
527 | assert_eq!(f128::EXP_MIN_SUBNORM, -16494); |
528 | |
529 | // `exp_unbiased` |
530 | assert_eq!(f128::FRAC_PI_2.exp_unbiased(), 0); |
531 | assert_eq!((1.0f128 / 2.0).exp_unbiased(), -1); |
532 | assert_eq!(f128::MAX.exp_unbiased(), 16383); |
533 | assert_eq!(f128::MIN.exp_unbiased(), 16383); |
534 | assert_eq!(f128::MIN_POSITIVE.exp_unbiased(), -16382); |
535 | // This is a convenience method and not ldexp, `exp_unbiased` does not return correct |
536 | // results for zero and subnormals. |
537 | assert_eq!(f128::ZERO.exp_unbiased(), -16383); |
538 | assert_eq!(f128::from_bits(0x1).exp_unbiased(), -16383); |
539 | assert_eq!(f128::MIN_POSITIVE, f128::MIN_POSITIVE_NORMAL); |
540 | |
541 | // `from_parts` |
542 | assert_biteq!(f128::from_parts(true, f128::EXP_BIAS, 0), -1.0f128); |
543 | assert_biteq!(f128::from_parts(false, 0, 1), f128::from_bits(0x1)); |
544 | } |
545 | } |
546 | |