src/f16.rs

   1 use crate::{
   2     prim::PrimFloat,
   3     scalar::Value,
   4     traits::{ConvertFrom, ConvertTo, Float},
   5 };
   6 use core::{
   7     fmt,
   8     ops::{Add, AddAssign, Div, DivAssign, Mul, MulAssign, Neg, Rem, RemAssign, Sub, SubAssign},
   9 };
  10
  11 #[cfg(feature = "f16")]
  12 use half::f16 as F16Impl;
  13
  14 #[cfg(not(feature = "f16"))]
  15 type F16Impl = u16;
  16
  17 #[derive(Clone, Copy, PartialEq, PartialOrd)]
  18 #[repr(transparent)]
  19 pub struct F16(F16Impl);
  20
  21 #[cfg(not(feature = "f16"))]
  22 #[track_caller]
  23 pub(crate) fn panic_f16_feature_disabled() -> ! {
  24     panic!("f16 feature is not enabled")
  25 }
  26
  27 #[cfg(feature = "f16")]
  28 macro_rules! f16_impl {
  29     ($v:expr, [$($vars:ident),*]) => {
  30         $v
  31     };
  32 }
  33
  34 #[cfg(not(feature = "f16"))]
  35 macro_rules! f16_impl {
  36     ($v:expr, [$($vars:ident),*]) => {
  37         {
  38             $(let _ = $vars;)*
  39             panic_f16_feature_disabled()
  40         }
  41     };
  42 }
  43
  44 impl fmt::Display for F16 {
  45     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
  46         f16_impl!(self.0.fmt(f), [f])
  47     }
  48 }
  49
  50 impl fmt::Debug for F16 {
  51     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
  52         f16_impl!(self.0.fmt(f), [f])
  53     }
  54 }
  55
  56 impl Default for F16 {
  57     fn default() -> Self {
  58         f16_impl!(F16(F16Impl::default()), [])
  59     }
  60 }
  61
  62 impl From<F16Impl> for F16 {
  63     fn from(v: F16Impl) -> Self {
  64         F16(v)
  65     }
  66 }
  67
  68 impl From<F16> for F16Impl {
  69     fn from(v: F16) -> Self {
  70         v.0
  71     }
  72 }
  73
  74 macro_rules! impl_f16_from {
  75     ($($ty:ident,)*) => {
  76         $(
  77             impl From<$ty> for F16 {
  78                 fn from(v: $ty) -> Self {
  79                     f16_impl!(F16(F16Impl::from(v)), [v])
  80                 }
  81             }
  82
  83             impl ConvertFrom<$ty> for F16 {
  84                 fn cvt_from(v: $ty) -> F16 {
  85                     v.into()
  86                 }
  87             }
  88         )*
  89     };
  90 }
  91
  92 macro_rules! impl_from_f16 {
  93     ($($ty:ident,)*) => {
  94         $(
  95             impl From<F16> for $ty {
  96                 fn from(v: F16) -> Self {
  97                     f16_impl!(v.0.into(), [v])
  98                 }
  99             }
 100
 101             impl ConvertFrom<F16> for $ty {
 102                 fn cvt_from(v: F16) -> Self {
 103                     v.into()
 104                 }
 105             }
 106         )*
 107     };
 108 }
 109
 110 impl_f16_from![i8, u8,];
 111
 112 impl_from_f16![f32, f64,];
 113
 114 macro_rules! impl_int_to_f16 {
 115     ($($int:ident),*) => {
 116         $(
 117             impl ConvertFrom<$int> for F16 {
 118                 fn cvt_from(v: $int) -> Self {
 119                     // f32 has enough mantissa bits such that f16 overflows to
 120                     // infinity before f32 stops being able to properly
 121                     // represent integer values, making the below conversion correct.
 122                     F16::cvt_from(v as f32)
 123                 }
 124             }
 125         )*
 126     };
 127 }
 128
 129 macro_rules! impl_f16_to_int {
 130     ($($int:ident),*) => {
 131         $(
 132             impl ConvertFrom<F16> for $int {
 133                 fn cvt_from(v: F16) -> Self {
 134                     f32::from(v) as $int
 135                 }
 136             }
 137         )*
 138     };
 139 }
 140
 141 impl_int_to_f16![i16, u16, i32, u32, i64, u64, i128, u128];
 142 impl_f16_to_int![i8, u8, i16, u16, i32, u32, i64, u64, i128, u128];
 143
 144 impl ConvertFrom<f32> for F16 {
 145     fn cvt_from(v: f32) -> Self {
 146         f16_impl!(F16(F16Impl::from_f32(v)), [v])
 147     }
 148 }
 149
 150 impl ConvertFrom<f64> for F16 {
 151     fn cvt_from(v: f64) -> Self {
 152         f16_impl!(F16(F16Impl::from_f64(v)), [v])
 153     }
 154 }
 155
 156 impl Neg for F16 {
 157     type Output = Self;
 158
 159     fn neg(self) -> Self::Output {
 160         f16_impl!(Self::from_bits(self.to_bits() ^ 0x8000), [])
 161     }
 162 }
 163
 164 macro_rules! impl_bin_op_using_f32 {
 165     ($($op:ident, $op_fn:ident, $op_assign:ident, $op_assign_fn:ident;)*) => {
 166         $(
 167             impl $op for F16 {
 168                 type Output = Self;
 169
 170                 fn $op_fn(self, rhs: Self) -> Self::Output {
 171                     f32::from(self).$op_fn(f32::from(rhs)).to()
 172                 }
 173             }
 174
 175             impl $op_assign for F16 {
 176                 fn $op_assign_fn(&mut self, rhs: Self) {
 177                     *self = (*self).$op_fn(rhs);
 178                 }
 179             }
 180         )*
 181     };
 182 }
 183
 184 impl_bin_op_using_f32! {
 185     Add, add, AddAssign, add_assign;
 186     Sub, sub, SubAssign, sub_assign;
 187     Mul, mul, MulAssign, mul_assign;
 188     Div, div, DivAssign, div_assign;
 189     Rem, rem, RemAssign, rem_assign;
 190 }
 191
 192 impl F16 {
 193     pub fn from_bits(v: u16) -> Self {
 194         #[cfg(feature = "f16")]
 195         return F16(F16Impl::from_bits(v));
 196         #[cfg(not(feature = "f16"))]
 197         return F16(v);
 198     }
 199     pub fn to_bits(self) -> u16 {
 200         #[cfg(feature = "f16")]
 201         return self.0.to_bits();
 202         #[cfg(not(feature = "f16"))]
 203         return self.0;
 204     }
 205     pub fn abs(self) -> Self {
 206         f16_impl!(Self::from_bits(self.to_bits() & 0x7FFF), [])
 207     }
 208     pub fn copysign(self, sign: Self) -> Self {
 209         f16_impl!(
 210             Self::from_bits((self.to_bits() & 0x7FFF) | (sign.to_bits() & 0x8000)),
 211             [sign]
 212         )
 213     }
 214     pub fn trunc(self) -> Self {
 215         return PrimFloat::trunc(f32::from(self)).to();
 216     }
 217     pub fn ceil(self) -> Self {
 218         return PrimFloat::ceil(f32::from(self)).to();
 219     }
 220     pub fn floor(self) -> Self {
 221         return PrimFloat::floor(f32::from(self)).to();
 222     }
 223     /// round to nearest, ties to unspecified
 224     pub fn round(self) -> Self {
 225         return PrimFloat::round(f32::from(self)).to();
 226     }
 227     #[cfg(feature = "fma")]
 228     pub fn fma(self, a: Self, b: Self) -> Self {
 229         (f64::from(self) * f64::from(a) + f64::from(b)).to()
 230     }
 231
 232     pub fn is_nan(self) -> bool {
 233         f16_impl!(self.0.is_nan(), [])
 234     }
 235
 236     pub fn is_infinite(self) -> bool {
 237         f16_impl!(self.0.is_infinite(), [])
 238     }
 239
 240     pub fn is_finite(self) -> bool {
 241         f16_impl!(self.0.is_finite(), [])
 242     }
 243 }
 244
 245 impl Float for Value<F16> {
 246     type PrimFloat = F16;
 247     type BitsType = Value<u16>;
 248     type SignedBitsType = Value<i16>;
 249
 250     fn abs(self) -> Self {
 251         Value(self.0.abs())
 252     }
 253
 254     fn trunc(self) -> Self {
 255         Value(self.0.trunc())
 256     }
 257
 258     fn ceil(self) -> Self {
 259         Value(self.0.ceil())
 260     }
 261
 262     fn floor(self) -> Self {
 263         Value(self.0.floor())
 264     }
 265
 266     fn round(self) -> Self {
 267         Value(self.0.round())
 268     }
 269
 270     #[cfg(feature = "fma")]
 271     fn fma(self, a: Self, b: Self) -> Self {
 272         Value(self.0.fma(a.0, b.0))
 273     }
 274
 275     fn is_nan(self) -> Self::Bool {
 276         Value(self.0.is_nan())
 277     }
 278
 279     fn is_infinite(self) -> Self::Bool {
 280         Value(self.0.is_infinite())
 281     }
 282
 283     fn is_finite(self) -> Self::Bool {
 284         Value(self.0.is_finite())
 285     }
 286
 287     fn from_bits(v: Self::BitsType) -> Self {
 288         Value(F16::from_bits(v.0))
 289     }
 290
 291     fn to_bits(self) -> Self::BitsType {
 292         Value(self.0.to_bits())
 293     }
 294 }
 295
 296 #[cfg(test)]
 297 mod tests {
 298     use super::*;
 299     use core::cmp::Ordering;
 300
 301     #[test]
 302     #[cfg_attr(
 303         not(feature = "f16"),
 304         should_panic(expected = "f16 feature is not enabled")
 305     )]
 306     fn test_abs() {
 307         assert_eq!(F16::from_bits(0x8000).abs().to_bits(), 0);
 308         assert_eq!(F16::from_bits(0).abs().to_bits(), 0);
 309         assert_eq!(F16::from_bits(0x8ABC).abs().to_bits(), 0xABC);
 310         assert_eq!(F16::from_bits(0xFE00).abs().to_bits(), 0x7E00);
 311         assert_eq!(F16::from_bits(0x7E00).abs().to_bits(), 0x7E00);
 312     }
 313
 314     #[test]
 315     #[cfg_attr(
 316         not(feature = "f16"),
 317         should_panic(expected = "f16 feature is not enabled")
 318     )]
 319     fn test_neg() {
 320         assert_eq!(F16::from_bits(0x8000).neg().to_bits(), 0);
 321         assert_eq!(F16::from_bits(0).neg().to_bits(), 0x8000);
 322         assert_eq!(F16::from_bits(0x8ABC).neg().to_bits(), 0xABC);
 323         assert_eq!(F16::from_bits(0xFE00).neg().to_bits(), 0x7E00);
 324         assert_eq!(F16::from_bits(0x7E00).neg().to_bits(), 0xFE00);
 325     }
 326
 327     #[test]
 328     #[cfg_attr(
 329         not(feature = "f16"),
 330         should_panic(expected = "f16 feature is not enabled")
 331     )]
 332     fn test_int_to_f16() {
 333         assert_eq!(F16::to_bits(0u32.to()), 0);
 334         for v in 1..0x20000u32 {
 335             let leading_zeros = u32::leading_zeros(v);
 336             let shifted_v = v << leading_zeros;
 337             // round to nearest, ties to even
 338             let round_up = match (shifted_v & 0x1FFFFF).cmp(&0x100000) {
 339                 Ordering::Less => false,
 340                 Ordering::Equal => (shifted_v & 0x200000) != 0,
 341                 Ordering::Greater => true,
 342             };
 343             let (rounded, carry) =
 344                 (shifted_v & !0x1FFFFF).overflowing_add(round_up.then(|| 0x200000).unwrap_or(0));
 345             let mantissa;
 346             if carry {
 347                 mantissa = (rounded >> 22) as u16 + 0x400;
 348             } else {
 349                 mantissa = (rounded >> 21) as u16;
 350             }
 351             assert_eq!((mantissa & !0x3FF), 0x400);
 352             let exponent = 31 - leading_zeros as u16 + 15 + carry as u16;
 353             let expected = if exponent < 0x1F {
 354                 (mantissa & 0x3FF) + (exponent << 10)
 355             } else {
 356                 0x7C00
 357             };
 358             let actual = F16::to_bits(v.to());
 359             assert_eq!(
 360                 actual, expected,
 361                 "actual = {:#X}, expected = {:#X}, v = {:#X}",
 362                 actual, expected, v
 363             );
 364         }
 365     }
 366 }