#[cfg(feature = "bytemuck")]
use bytemuck::{Pod, Zeroable};
use core::{
cmp::Ordering,
fmt::{
Binary, Debug, Display, Error, Formatter, LowerExp, LowerHex, Octal, UpperExp, UpperHex,
},
iter::{Product, Sum},
num::{FpCategory, ParseFloatError},
ops::{Add, AddAssign, Div, DivAssign, Mul, MulAssign, Neg, Rem, RemAssign, Sub, SubAssign},
str::FromStr,
};
#[cfg(feature = "serde")]
use serde::{Deserialize, Serialize};
#[cfg(feature = "zerocopy")]
use zerocopy::{AsBytes, FromBytes};
pub(crate) mod convert;
#[allow(non_camel_case_types)]
#[derive(Clone, Copy, Default)]
#[repr(transparent)]
#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
#[cfg_attr(feature = "bytemuck", derive(Zeroable, Pod))]
#[cfg_attr(feature = "zerocopy", derive(AsBytes, FromBytes))]
pub struct f16(u16);
#[doc(hidden)]
#[deprecated(
since = "1.4.0",
note = "all constants moved to associated constants of `f16`"
)]
pub mod consts {
use super::f16;
#[deprecated(since = "1.4.0", note = "moved to `f16::DIGITS`")]
pub const DIGITS: u32 = f16::DIGITS;
#[deprecated(since = "1.4.0", note = "moved to `f16::EPSILON`")]
pub const EPSILON: f16 = f16::EPSILON;
#[deprecated(since = "1.4.0", note = "moved to `f16::INFINITY`")]
pub const INFINITY: f16 = f16::INFINITY;
#[deprecated(since = "1.4.0", note = "moved to `f16::MANTISSA_DIGITS`")]
pub const MANTISSA_DIGITS: u32 = f16::MANTISSA_DIGITS;
#[deprecated(since = "1.4.0", note = "moved to `f16::MAX`")]
pub const MAX: f16 = f16::MAX;
#[deprecated(since = "1.4.0", note = "moved to `f16::MAX_10_EXP`")]
pub const MAX_10_EXP: i32 = f16::MAX_10_EXP;
#[deprecated(since = "1.4.0", note = "moved to `f16::MAX_EXP`")]
pub const MAX_EXP: i32 = f16::MAX_EXP;
#[deprecated(since = "1.4.0", note = "moved to `f16::MIN`")]
pub const MIN: f16 = f16::MIN;
#[deprecated(since = "1.4.0", note = "moved to `f16::MIN_10_EXP`")]
pub const MIN_10_EXP: i32 = f16::MIN_10_EXP;
#[deprecated(since = "1.4.0", note = "moved to `f16::MIN_EXP`")]
pub const MIN_EXP: i32 = f16::MIN_EXP;
#[deprecated(since = "1.4.0", note = "moved to `f16::MIN_POSITIVE`")]
pub const MIN_POSITIVE: f16 = f16::MIN_POSITIVE;
#[deprecated(since = "1.4.0", note = "moved to `f16::NAN`")]
pub const NAN: f16 = f16::NAN;
#[deprecated(since = "1.4.0", note = "moved to `f16::NEG_INFINITY`")]
pub const NEG_INFINITY: f16 = f16::NEG_INFINITY;
#[deprecated(since = "1.4.0", note = "moved to `f16::RADIX`")]
pub const RADIX: u32 = f16::RADIX;
#[deprecated(since = "1.4.0", note = "moved to `f16::MIN_POSITIVE_SUBNORMAL`")]
pub const MIN_POSITIVE_SUBNORMAL: f16 = f16::MIN_POSITIVE_SUBNORMAL;
#[deprecated(since = "1.4.0", note = "moved to `f16::MAX_SUBNORMAL`")]
pub const MAX_SUBNORMAL: f16 = f16::MAX_SUBNORMAL;
#[deprecated(since = "1.4.0", note = "moved to `f16::ONE`")]
pub const ONE: f16 = f16::ONE;
#[deprecated(since = "1.4.0", note = "moved to `f16::ZERO`")]
pub const ZERO: f16 = f16::ZERO;
#[deprecated(since = "1.4.0", note = "moved to `f16::NEG_ZERO`")]
pub const NEG_ZERO: f16 = f16::NEG_ZERO;
#[deprecated(since = "1.4.0", note = "moved to `f16::E`")]
pub const E: f16 = f16::E;
#[deprecated(since = "1.4.0", note = "moved to `f16::PI`")]
pub const PI: f16 = f16::PI;
#[deprecated(since = "1.4.0", note = "moved to `f16::FRAC_1_PI`")]
pub const FRAC_1_PI: f16 = f16::FRAC_1_PI;
#[deprecated(since = "1.4.0", note = "moved to `f16::FRAC_1_SQRT_2`")]
pub const FRAC_1_SQRT_2: f16 = f16::FRAC_1_SQRT_2;
#[deprecated(since = "1.4.0", note = "moved to `f16::FRAC_2_PI`")]
pub const FRAC_2_PI: f16 = f16::FRAC_2_PI;
#[deprecated(since = "1.4.0", note = "moved to `f16::FRAC_2_SQRT_PI`")]
pub const FRAC_2_SQRT_PI: f16 = f16::FRAC_2_SQRT_PI;
#[deprecated(since = "1.4.0", note = "moved to `f16::FRAC_PI_2`")]
pub const FRAC_PI_2: f16 = f16::FRAC_PI_2;
#[deprecated(since = "1.4.0", note = "moved to `f16::FRAC_PI_3`")]
pub const FRAC_PI_3: f16 = f16::FRAC_PI_3;
#[deprecated(since = "1.4.0", note = "moved to `f16::FRAC_PI_4`")]
pub const FRAC_PI_4: f16 = f16::FRAC_PI_4;
#[deprecated(since = "1.4.0", note = "moved to `f16::FRAC_PI_6`")]
pub const FRAC_PI_6: f16 = f16::FRAC_PI_6;
#[deprecated(since = "1.4.0", note = "moved to `f16::FRAC_PI_8`")]
pub const FRAC_PI_8: f16 = f16::FRAC_PI_8;
#[deprecated(since = "1.4.0", note = "moved to `f16::LN_10`")]
pub const LN_10: f16 = f16::LN_10;
#[deprecated(since = "1.4.0", note = "moved to `f16::LN_2`")]
pub const LN_2: f16 = f16::LN_2;
#[deprecated(since = "1.4.0", note = "moved to `f16::LOG10_E`")]
pub const LOG10_E: f16 = f16::LOG10_E;
#[deprecated(since = "1.4.0", note = "moved to `f16::LOG2_E`")]
pub const LOG2_E: f16 = f16::LOG2_E;
#[deprecated(since = "1.4.0", note = "moved to `f16::SQRT_2`")]
pub const SQRT_2: f16 = f16::SQRT_2;
}
impl f16 {
#[inline]
pub const fn from_bits(bits: u16) -> f16 {
f16(bits)
}
#[inline]
pub fn from_f32(value: f32) -> f16 {
f16(convert::f32_to_f16(value))
}
#[inline]
pub fn from_f64(value: f64) -> f16 {
f16(convert::f64_to_f16(value))
}
#[inline]
pub const fn to_bits(self) -> u16 {
self.0
}
#[inline]
pub const fn to_le_bytes(self) -> [u8; 2] {
self.0.to_le_bytes()
}
#[inline]
pub const fn to_be_bytes(self) -> [u8; 2] {
self.0.to_be_bytes()
}
#[inline]
pub const fn to_ne_bytes(self) -> [u8; 2] {
self.0.to_ne_bytes()
}
#[inline]
pub const fn from_le_bytes(bytes: [u8; 2]) -> f16 {
f16::from_bits(u16::from_le_bytes(bytes))
}
#[inline]
pub const fn from_be_bytes(bytes: [u8; 2]) -> f16 {
f16::from_bits(u16::from_be_bytes(bytes))
}
#[inline]
pub const fn from_ne_bytes(bytes: [u8; 2]) -> f16 {
f16::from_bits(u16::from_ne_bytes(bytes))
}
#[doc(hidden)]
#[deprecated(since = "1.2.0", note = "renamed to `to_bits`")]
#[inline]
pub fn as_bits(self) -> u16 {
self.to_bits()
}
#[inline]
pub fn to_f32(self) -> f32 {
convert::f16_to_f32(self.0)
}
#[inline]
pub fn to_f64(self) -> f64 {
convert::f16_to_f64(self.0)
}
#[inline]
pub const fn is_nan(self) -> bool {
self.0 & 0x7FFFu16 > 0x7C00u16
}
#[inline]
pub const fn is_infinite(self) -> bool {
self.0 & 0x7FFFu16 == 0x7C00u16
}
#[inline]
pub const fn is_finite(self) -> bool {
self.0 & 0x7C00u16 != 0x7C00u16
}
#[inline]
pub const fn is_normal(self) -> bool {
let exp = self.0 & 0x7C00u16;
exp != 0x7C00u16 && exp != 0
}
pub const fn classify(self) -> FpCategory {
let exp = self.0 & 0x7C00u16;
let man = self.0 & 0x03FFu16;
match (exp, man) {
(0, 0) => FpCategory::Zero,
(0, _) => FpCategory::Subnormal,
(0x7C00u16, 0) => FpCategory::Infinite,
(0x7C00u16, _) => FpCategory::Nan,
_ => FpCategory::Normal,
}
}
pub const fn signum(self) -> f16 {
if self.is_nan() {
self
} else if self.0 & 0x8000u16 != 0 {
Self::NEG_ONE
} else {
Self::ONE
}
}
#[inline]
pub const fn is_sign_positive(self) -> bool {
self.0 & 0x8000u16 == 0
}
#[inline]
pub const fn is_sign_negative(self) -> bool {
self.0 & 0x8000u16 != 0
}
#[inline]
pub const fn copysign(self, sign: f16) -> f16 {
f16((sign.0 & 0x8000u16) | (self.0 & 0x7FFFu16))
}
#[inline]
pub fn max(self, other: f16) -> f16 {
if other > self && !other.is_nan() {
other
} else {
self
}
}
#[inline]
pub fn min(self, other: f16) -> f16 {
if other < self && !other.is_nan() {
other
} else {
self
}
}
#[inline]
pub fn clamp(self, min: f16, max: f16) -> f16 {
assert!(min <= max);
let mut x = self;
if x < min {
x = min;
}
if x > max {
x = max;
}
x
}
pub const DIGITS: u32 = 3;
pub const EPSILON: f16 = f16(0x1400u16);
pub const INFINITY: f16 = f16(0x7C00u16);
pub const MANTISSA_DIGITS: u32 = 11;
pub const MAX: f16 = f16(0x7BFF);
pub const MAX_10_EXP: i32 = 4;
pub const MAX_EXP: i32 = 16;
pub const MIN: f16 = f16(0xFBFF);
pub const MIN_10_EXP: i32 = -4;
pub const MIN_EXP: i32 = -13;
pub const MIN_POSITIVE: f16 = f16(0x0400u16);
pub const NAN: f16 = f16(0x7E00u16);
pub const NEG_INFINITY: f16 = f16(0xFC00u16);
pub const RADIX: u32 = 2;
pub const MIN_POSITIVE_SUBNORMAL: f16 = f16(0x0001u16);
pub const MAX_SUBNORMAL: f16 = f16(0x03FFu16);
pub const ONE: f16 = f16(0x3C00u16);
pub const ZERO: f16 = f16(0x0000u16);
pub const NEG_ZERO: f16 = f16(0x8000u16);
pub const NEG_ONE: f16 = f16(0xBC00u16);
pub const E: f16 = f16(0x4170u16);
pub const PI: f16 = f16(0x4248u16);
pub const FRAC_1_PI: f16 = f16(0x3518u16);
pub const FRAC_1_SQRT_2: f16 = f16(0x39A8u16);
pub const FRAC_2_PI: f16 = f16(0x3918u16);
pub const FRAC_2_SQRT_PI: f16 = f16(0x3C83u16);
pub const FRAC_PI_2: f16 = f16(0x3E48u16);
pub const FRAC_PI_3: f16 = f16(0x3C30u16);
pub const FRAC_PI_4: f16 = f16(0x3A48u16);
pub const FRAC_PI_6: f16 = f16(0x3830u16);
pub const FRAC_PI_8: f16 = f16(0x3648u16);
pub const LN_10: f16 = f16(0x409Bu16);
pub const LN_2: f16 = f16(0x398Cu16);
pub const LOG10_E: f16 = f16(0x36F3u16);
pub const LOG10_2: f16 = f16(0x34D1u16);
pub const LOG2_E: f16 = f16(0x3DC5u16);
pub const LOG2_10: f16 = f16(0x42A5u16);
pub const SQRT_2: f16 = f16(0x3DA8u16);
}
impl From<f16> for f32 {
#[inline]
fn from(x: f16) -> f32 {
x.to_f32()
}
}
impl From<f16> for f64 {
#[inline]
fn from(x: f16) -> f64 {
x.to_f64()
}
}
impl From<i8> for f16 {
#[inline]
fn from(x: i8) -> f16 {
f16::from_f32(f32::from(x))
}
}
impl From<u8> for f16 {
#[inline]
fn from(x: u8) -> f16 {
f16::from_f32(f32::from(x))
}
}
impl PartialEq for f16 {
fn eq(&self, other: &f16) -> bool {
if self.is_nan() || other.is_nan() {
false
} else {
(self.0 == other.0) || ((self.0 | other.0) & 0x7FFFu16 == 0)
}
}
}
impl PartialOrd for f16 {
fn partial_cmp(&self, other: &f16) -> Option<Ordering> {
if self.is_nan() || other.is_nan() {
None
} else {
let neg = self.0 & 0x8000u16 != 0;
let other_neg = other.0 & 0x8000u16 != 0;
match (neg, other_neg) {
(false, false) => Some(self.0.cmp(&other.0)),
(false, true) => {
if (self.0 | other.0) & 0x7FFFu16 == 0 {
Some(Ordering::Equal)
} else {
Some(Ordering::Greater)
}
}
(true, false) => {
if (self.0 | other.0) & 0x7FFFu16 == 0 {
Some(Ordering::Equal)
} else {
Some(Ordering::Less)
}
}
(true, true) => Some(other.0.cmp(&self.0)),
}
}
}
fn lt(&self, other: &f16) -> bool {
if self.is_nan() || other.is_nan() {
false
} else {
let neg = self.0 & 0x8000u16 != 0;
let other_neg = other.0 & 0x8000u16 != 0;
match (neg, other_neg) {
(false, false) => self.0 < other.0,
(false, true) => false,
(true, false) => (self.0 | other.0) & 0x7FFFu16 != 0,
(true, true) => self.0 > other.0,
}
}
}
fn le(&self, other: &f16) -> bool {
if self.is_nan() || other.is_nan() {
false
} else {
let neg = self.0 & 0x8000u16 != 0;
let other_neg = other.0 & 0x8000u16 != 0;
match (neg, other_neg) {
(false, false) => self.0 <= other.0,
(false, true) => (self.0 | other.0) & 0x7FFFu16 == 0,
(true, false) => true,
(true, true) => self.0 >= other.0,
}
}
}
fn gt(&self, other: &f16) -> bool {
if self.is_nan() || other.is_nan() {
false
} else {
let neg = self.0 & 0x8000u16 != 0;
let other_neg = other.0 & 0x8000u16 != 0;
match (neg, other_neg) {
(false, false) => self.0 > other.0,
(false, true) => (self.0 | other.0) & 0x7FFFu16 != 0,
(true, false) => false,
(true, true) => self.0 < other.0,
}
}
}
fn ge(&self, other: &f16) -> bool {
if self.is_nan() || other.is_nan() {
false
} else {
let neg = self.0 & 0x8000u16 != 0;
let other_neg = other.0 & 0x8000u16 != 0;
match (neg, other_neg) {
(false, false) => self.0 >= other.0,
(false, true) => true,
(true, false) => (self.0 | other.0) & 0x7FFFu16 == 0,
(true, true) => self.0 <= other.0,
}
}
}
}
impl FromStr for f16 {
type Err = ParseFloatError;
fn from_str(src: &str) -> Result<f16, ParseFloatError> {
f32::from_str(src).map(f16::from_f32)
}
}
impl Debug for f16 {
fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
write!(f, "{:?}", self.to_f32())
}
}
impl Display for f16 {
fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
write!(f, "{}", self.to_f32())
}
}
impl LowerExp for f16 {
fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
write!(f, "{:e}", self.to_f32())
}
}
impl UpperExp for f16 {
fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
write!(f, "{:E}", self.to_f32())
}
}
impl Binary for f16 {
fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
write!(f, "{:b}", self.0)
}
}
impl Octal for f16 {
fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
write!(f, "{:o}", self.0)
}
}
impl LowerHex for f16 {
fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
write!(f, "{:x}", self.0)
}
}
impl UpperHex for f16 {
fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error> {
write!(f, "{:X}", self.0)
}
}
impl Neg for f16 {
type Output = Self;
#[inline]
fn neg(self) -> Self::Output {
Self(self.0 ^ 0x8000)
}
}
impl Add for f16 {
type Output = Self;
#[inline]
fn add(self, rhs: Self) -> Self::Output {
Self::from_f32(Self::to_f32(self) + Self::to_f32(rhs))
}
}
impl Add<&f16> for f16 {
type Output = <f16 as Add<f16>>::Output;
#[inline]
fn add(self, rhs: &f16) -> Self::Output {
self.add(*rhs)
}
}
impl Add<&f16> for &f16 {
type Output = <f16 as Add<f16>>::Output;
#[inline]
fn add(self, rhs: &f16) -> Self::Output {
(*self).add(*rhs)
}
}
impl Add<f16> for &f16 {
type Output = <f16 as Add<f16>>::Output;
#[inline]
fn add(self, rhs: f16) -> Self::Output {
(*self).add(rhs)
}
}
impl AddAssign for f16 {
#[inline]
fn add_assign(&mut self, rhs: Self) {
*self = (*self).add(rhs);
}
}
impl AddAssign<&f16> for f16 {
#[inline]
fn add_assign(&mut self, rhs: &f16) {
*self = (*self).add(rhs);
}
}
impl Sub for f16 {
type Output = Self;
#[inline]
fn sub(self, rhs: Self) -> Self::Output {
Self::from_f32(Self::to_f32(self) - Self::to_f32(rhs))
}
}
impl Sub<&f16> for f16 {
type Output = <f16 as Sub<f16>>::Output;
#[inline]
fn sub(self, rhs: &f16) -> Self::Output {
self.sub(*rhs)
}
}
impl Sub<&f16> for &f16 {
type Output = <f16 as Sub<f16>>::Output;
#[inline]
fn sub(self, rhs: &f16) -> Self::Output {
(*self).sub(*rhs)
}
}
impl Sub<f16> for &f16 {
type Output = <f16 as Sub<f16>>::Output;
#[inline]
fn sub(self, rhs: f16) -> Self::Output {
(*self).sub(rhs)
}
}
impl SubAssign for f16 {
#[inline]
fn sub_assign(&mut self, rhs: Self) {
*self = (*self).sub(rhs);
}
}
impl SubAssign<&f16> for f16 {
#[inline]
fn sub_assign(&mut self, rhs: &f16) {
*self = (*self).sub(rhs);
}
}
impl Mul for f16 {
type Output = Self;
#[inline]
fn mul(self, rhs: Self) -> Self::Output {
Self::from_f32(Self::to_f32(self) * Self::to_f32(rhs))
}
}
impl Mul<&f16> for f16 {
type Output = <f16 as Mul<f16>>::Output;
#[inline]
fn mul(self, rhs: &f16) -> Self::Output {
self.mul(*rhs)
}
}
impl Mul<&f16> for &f16 {
type Output = <f16 as Mul<f16>>::Output;
#[inline]
fn mul(self, rhs: &f16) -> Self::Output {
(*self).mul(*rhs)
}
}
impl Mul<f16> for &f16 {
type Output = <f16 as Mul<f16>>::Output;
#[inline]
fn mul(self, rhs: f16) -> Self::Output {
(*self).mul(rhs)
}
}
impl MulAssign for f16 {
#[inline]
fn mul_assign(&mut self, rhs: Self) {
*self = (*self).mul(rhs);
}
}
impl MulAssign<&f16> for f16 {
#[inline]
fn mul_assign(&mut self, rhs: &f16) {
*self = (*self).mul(rhs);
}
}
impl Div for f16 {
type Output = Self;
#[inline]
fn div(self, rhs: Self) -> Self::Output {
Self::from_f32(Self::to_f32(self) / Self::to_f32(rhs))
}
}
impl Div<&f16> for f16 {
type Output = <f16 as Div<f16>>::Output;
#[inline]
fn div(self, rhs: &f16) -> Self::Output {
self.div(*rhs)
}
}
impl Div<&f16> for &f16 {
type Output = <f16 as Div<f16>>::Output;
#[inline]
fn div(self, rhs: &f16) -> Self::Output {
(*self).div(*rhs)
}
}
impl Div<f16> for &f16 {
type Output = <f16 as Div<f16>>::Output;
#[inline]
fn div(self, rhs: f16) -> Self::Output {
(*self).div(rhs)
}
}
impl DivAssign for f16 {
#[inline]
fn div_assign(&mut self, rhs: Self) {
*self = (*self).div(rhs);
}
}
impl DivAssign<&f16> for f16 {
#[inline]
fn div_assign(&mut self, rhs: &f16) {
*self = (*self).div(rhs);
}
}
impl Rem for f16 {
type Output = Self;
#[inline]
fn rem(self, rhs: Self) -> Self::Output {
Self::from_f32(Self::to_f32(self) % Self::to_f32(rhs))
}
}
impl Rem<&f16> for f16 {
type Output = <f16 as Rem<f16>>::Output;
#[inline]
fn rem(self, rhs: &f16) -> Self::Output {
self.rem(*rhs)
}
}
impl Rem<&f16> for &f16 {
type Output = <f16 as Rem<f16>>::Output;
#[inline]
fn rem(self, rhs: &f16) -> Self::Output {
(*self).rem(*rhs)
}
}
impl Rem<f16> for &f16 {
type Output = <f16 as Rem<f16>>::Output;
#[inline]
fn rem(self, rhs: f16) -> Self::Output {
(*self).rem(rhs)
}
}
impl RemAssign for f16 {
#[inline]
fn rem_assign(&mut self, rhs: Self) {
*self = (*self).rem(rhs);
}
}
impl RemAssign<&f16> for f16 {
#[inline]
fn rem_assign(&mut self, rhs: &f16) {
*self = (*self).rem(rhs);
}
}
impl Product for f16 {
#[inline]
fn product<I: Iterator<Item = Self>>(iter: I) -> Self {
f16::from_f32(iter.map(|f| f.to_f32()).product())
}
}
impl<'a> Product<&'a f16> for f16 {
#[inline]
fn product<I: Iterator<Item = &'a f16>>(iter: I) -> Self {
f16::from_f32(iter.map(|f| f.to_f32()).product())
}
}
impl Sum for f16 {
#[inline]
fn sum<I: Iterator<Item = Self>>(iter: I) -> Self {
f16::from_f32(iter.map(|f| f.to_f32()).sum())
}
}
impl<'a> Sum<&'a f16> for f16 {
#[inline]
fn sum<I: Iterator<Item = &'a f16>>(iter: I) -> Self {
f16::from_f32(iter.map(|f| f.to_f32()).product())
}
}
#[allow(
clippy::cognitive_complexity,
clippy::float_cmp,
clippy::neg_cmp_op_on_partial_ord
)]
#[cfg(test)]
mod test {
use super::*;
use core::cmp::Ordering;
#[cfg(feature = "num-traits")]
use num_traits::{AsPrimitive, FromPrimitive, ToPrimitive};
use quickcheck_macros::quickcheck;
#[cfg(feature = "num-traits")]
#[test]
fn as_primitive() {
let two = f16::from_f32(2.0);
assert_eq!(<i32 as AsPrimitive<f16>>::as_(2), two);
assert_eq!(<f16 as AsPrimitive<i32>>::as_(two), 2);
assert_eq!(<f32 as AsPrimitive<f16>>::as_(2.0), two);
assert_eq!(<f16 as AsPrimitive<f32>>::as_(two), 2.0);
assert_eq!(<f64 as AsPrimitive<f16>>::as_(2.0), two);
assert_eq!(<f16 as AsPrimitive<f64>>::as_(two), 2.0);
}
#[cfg(feature = "num-traits")]
#[test]
fn to_primitive() {
let two = f16::from_f32(2.0);
assert_eq!(ToPrimitive::to_i32(&two).unwrap(), 2i32);
assert_eq!(ToPrimitive::to_f32(&two).unwrap(), 2.0f32);
assert_eq!(ToPrimitive::to_f64(&two).unwrap(), 2.0f64);
}
#[cfg(feature = "num-traits")]
#[test]
fn from_primitive() {
let two = f16::from_f32(2.0);
assert_eq!(<f16 as FromPrimitive>::from_i32(2).unwrap(), two);
assert_eq!(<f16 as FromPrimitive>::from_f32(2.0).unwrap(), two);
assert_eq!(<f16 as FromPrimitive>::from_f64(2.0).unwrap(), two);
}
#[test]
fn test_f16_consts() {
let digits = ((f16::MANTISSA_DIGITS as f32 - 1.0) * 2f32.log10()).floor() as u32;
assert_eq!(f16::DIGITS, digits);
let digits32 = ((core::f32::MANTISSA_DIGITS as f32 - 1.0) * 2f32.log10()).floor() as u32;
assert_eq!(core::f32::DIGITS, digits32);
let one = f16::from_f32(1.0);
let one_plus_epsilon = f16::from_bits(one.to_bits() + 1);
let epsilon = f16::from_f32(one_plus_epsilon.to_f32() - 1.0);
assert_eq!(f16::EPSILON, epsilon);
let one_plus_epsilon32 = f32::from_bits(1.0f32.to_bits() + 1);
let epsilon32 = one_plus_epsilon32 - 1f32;
assert_eq!(core::f32::EPSILON, epsilon32);
let max = f16::from_bits(f16::INFINITY.to_bits() - 1);
let min = f16::from_bits(f16::NEG_INFINITY.to_bits() - 1);
let min_pos = f16::from_f32(2f32.powi(f16::MIN_EXP - 1));
assert_eq!(f16::MAX, max);
assert_eq!(f16::MIN, min);
assert_eq!(f16::MIN_POSITIVE, min_pos);
let max32 = f32::from_bits(core::f32::INFINITY.to_bits() - 1);
let min32 = f32::from_bits(core::f32::NEG_INFINITY.to_bits() - 1);
let min_pos32 = 2f32.powi(core::f32::MIN_EXP - 1);
assert_eq!(core::f32::MAX, max32);
assert_eq!(core::f32::MIN, min32);
assert_eq!(core::f32::MIN_POSITIVE, min_pos32);
let ten_to_min = 10f32.powi(f16::MIN_10_EXP);
assert!(ten_to_min / 10.0 < f16::MIN_POSITIVE.to_f32());
assert!(ten_to_min > f16::MIN_POSITIVE.to_f32());
let ten_to_max = 10f32.powi(f16::MAX_10_EXP);
assert!(ten_to_max < f16::MAX.to_f32());
assert!(ten_to_max * 10.0 > f16::MAX.to_f32());
let ten_to_min32 = 10f64.powi(core::f32::MIN_10_EXP);
assert!(ten_to_min32 / 10.0 < f64::from(core::f32::MIN_POSITIVE));
assert!(ten_to_min32 > f64::from(core::f32::MIN_POSITIVE));
let ten_to_max32 = 10f64.powi(core::f32::MAX_10_EXP);
assert!(ten_to_max32 < f64::from(core::f32::MAX));
assert!(ten_to_max32 * 10.0 > f64::from(core::f32::MAX));
}
#[test]
fn test_f16_consts_from_f32() {
let one = f16::from_f32(1.0);
let zero = f16::from_f32(0.0);
let neg_zero = f16::from_f32(-0.0);
let neg_one = f16::from_f32(-1.0);
let inf = f16::from_f32(core::f32::INFINITY);
let neg_inf = f16::from_f32(core::f32::NEG_INFINITY);
let nan = f16::from_f32(core::f32::NAN);
assert_eq!(f16::ONE, one);
assert_eq!(f16::ZERO, zero);
assert!(zero.is_sign_positive());
assert_eq!(f16::NEG_ZERO, neg_zero);
assert!(neg_zero.is_sign_negative());
assert_eq!(f16::NEG_ONE, neg_one);
assert!(neg_one.is_sign_negative());
assert_eq!(f16::INFINITY, inf);
assert_eq!(f16::NEG_INFINITY, neg_inf);
assert!(nan.is_nan());
assert!(f16::NAN.is_nan());
let e = f16::from_f32(core::f32::consts::E);
let pi = f16::from_f32(core::f32::consts::PI);
let frac_1_pi = f16::from_f32(core::f32::consts::FRAC_1_PI);
let frac_1_sqrt_2 = f16::from_f32(core::f32::consts::FRAC_1_SQRT_2);
let frac_2_pi = f16::from_f32(core::f32::consts::FRAC_2_PI);
let frac_2_sqrt_pi = f16::from_f32(core::f32::consts::FRAC_2_SQRT_PI);
let frac_pi_2 = f16::from_f32(core::f32::consts::FRAC_PI_2);
let frac_pi_3 = f16::from_f32(core::f32::consts::FRAC_PI_3);
let frac_pi_4 = f16::from_f32(core::f32::consts::FRAC_PI_4);
let frac_pi_6 = f16::from_f32(core::f32::consts::FRAC_PI_6);
let frac_pi_8 = f16::from_f32(core::f32::consts::FRAC_PI_8);
let ln_10 = f16::from_f32(core::f32::consts::LN_10);
let ln_2 = f16::from_f32(core::f32::consts::LN_2);
let log10_e = f16::from_f32(core::f32::consts::LOG10_E);
let log10_2 = f16::from_f32(2f32.log10());
let log2_e = f16::from_f32(core::f32::consts::LOG2_E);
let log2_10 = f16::from_f32(10f32.log2());
let sqrt_2 = f16::from_f32(core::f32::consts::SQRT_2);
assert_eq!(f16::E, e);
assert_eq!(f16::PI, pi);
assert_eq!(f16::FRAC_1_PI, frac_1_pi);
assert_eq!(f16::FRAC_1_SQRT_2, frac_1_sqrt_2);
assert_eq!(f16::FRAC_2_PI, frac_2_pi);
assert_eq!(f16::FRAC_2_SQRT_PI, frac_2_sqrt_pi);
assert_eq!(f16::FRAC_PI_2, frac_pi_2);
assert_eq!(f16::FRAC_PI_3, frac_pi_3);
assert_eq!(f16::FRAC_PI_4, frac_pi_4);
assert_eq!(f16::FRAC_PI_6, frac_pi_6);
assert_eq!(f16::FRAC_PI_8, frac_pi_8);
assert_eq!(f16::LN_10, ln_10);
assert_eq!(f16::LN_2, ln_2);
assert_eq!(f16::LOG10_E, log10_e);
assert_eq!(f16::LOG10_2, log10_2);
assert_eq!(f16::LOG2_E, log2_e);
assert_eq!(f16::LOG2_10, log2_10);
assert_eq!(f16::SQRT_2, sqrt_2);
}
#[test]
fn test_f16_consts_from_f64() {
let one = f16::from_f64(1.0);
let zero = f16::from_f64(0.0);
let neg_zero = f16::from_f64(-0.0);
let inf = f16::from_f64(core::f64::INFINITY);
let neg_inf = f16::from_f64(core::f64::NEG_INFINITY);
let nan = f16::from_f64(core::f64::NAN);
assert_eq!(f16::ONE, one);
assert_eq!(f16::ZERO, zero);
assert!(zero.is_sign_positive());
assert_eq!(f16::NEG_ZERO, neg_zero);
assert!(neg_zero.is_sign_negative());
assert_eq!(f16::INFINITY, inf);
assert_eq!(f16::NEG_INFINITY, neg_inf);
assert!(nan.is_nan());
assert!(f16::NAN.is_nan());
let e = f16::from_f64(core::f64::consts::E);
let pi = f16::from_f64(core::f64::consts::PI);
let frac_1_pi = f16::from_f64(core::f64::consts::FRAC_1_PI);
let frac_1_sqrt_2 = f16::from_f64(core::f64::consts::FRAC_1_SQRT_2);
let frac_2_pi = f16::from_f64(core::f64::consts::FRAC_2_PI);
let frac_2_sqrt_pi = f16::from_f64(core::f64::consts::FRAC_2_SQRT_PI);
let frac_pi_2 = f16::from_f64(core::f64::consts::FRAC_PI_2);
let frac_pi_3 = f16::from_f64(core::f64::consts::FRAC_PI_3);
let frac_pi_4 = f16::from_f64(core::f64::consts::FRAC_PI_4);
let frac_pi_6 = f16::from_f64(core::f64::consts::FRAC_PI_6);
let frac_pi_8 = f16::from_f64(core::f64::consts::FRAC_PI_8);
let ln_10 = f16::from_f64(core::f64::consts::LN_10);
let ln_2 = f16::from_f64(core::f64::consts::LN_2);
let log10_e = f16::from_f64(core::f64::consts::LOG10_E);
let log10_2 = f16::from_f64(2f64.log10());
let log2_e = f16::from_f64(core::f64::consts::LOG2_E);
let log2_10 = f16::from_f64(10f64.log2());
let sqrt_2 = f16::from_f64(core::f64::consts::SQRT_2);
assert_eq!(f16::E, e);
assert_eq!(f16::PI, pi);
assert_eq!(f16::FRAC_1_PI, frac_1_pi);
assert_eq!(f16::FRAC_1_SQRT_2, frac_1_sqrt_2);
assert_eq!(f16::FRAC_2_PI, frac_2_pi);
assert_eq!(f16::FRAC_2_SQRT_PI, frac_2_sqrt_pi);
assert_eq!(f16::FRAC_PI_2, frac_pi_2);
assert_eq!(f16::FRAC_PI_3, frac_pi_3);
assert_eq!(f16::FRAC_PI_4, frac_pi_4);
assert_eq!(f16::FRAC_PI_6, frac_pi_6);
assert_eq!(f16::FRAC_PI_8, frac_pi_8);
assert_eq!(f16::LN_10, ln_10);
assert_eq!(f16::LN_2, ln_2);
assert_eq!(f16::LOG10_E, log10_e);
assert_eq!(f16::LOG10_2, log10_2);
assert_eq!(f16::LOG2_E, log2_e);
assert_eq!(f16::LOG2_10, log2_10);
assert_eq!(f16::SQRT_2, sqrt_2);
}
#[test]
fn test_nan_conversion_to_smaller() {
let nan64 = f64::from_bits(0x7FF0_0000_0000_0001u64);
let neg_nan64 = f64::from_bits(0xFFF0_0000_0000_0001u64);
let nan32 = f32::from_bits(0x7F80_0001u32);
let neg_nan32 = f32::from_bits(0xFF80_0001u32);
let nan32_from_64 = nan64 as f32;
let neg_nan32_from_64 = neg_nan64 as f32;
let nan16_from_64 = f16::from_f64(nan64);
let neg_nan16_from_64 = f16::from_f64(neg_nan64);
let nan16_from_32 = f16::from_f32(nan32);
let neg_nan16_from_32 = f16::from_f32(neg_nan32);
assert!(nan64.is_nan() && nan64.is_sign_positive());
assert!(neg_nan64.is_nan() && neg_nan64.is_sign_negative());
assert!(nan32.is_nan() && nan32.is_sign_positive());
assert!(neg_nan32.is_nan() && neg_nan32.is_sign_negative());
assert!(nan32_from_64.is_nan() && nan32_from_64.is_sign_positive());
assert!(neg_nan32_from_64.is_nan() && neg_nan32_from_64.is_sign_negative());
assert!(nan16_from_64.is_nan() && nan16_from_64.is_sign_positive());
assert!(neg_nan16_from_64.is_nan() && neg_nan16_from_64.is_sign_negative());
assert!(nan16_from_32.is_nan() && nan16_from_32.is_sign_positive());
assert!(neg_nan16_from_32.is_nan() && neg_nan16_from_32.is_sign_negative());
}
#[test]
fn test_nan_conversion_to_larger() {
let nan16 = f16::from_bits(0x7C01u16);
let neg_nan16 = f16::from_bits(0xFC01u16);
let nan32 = f32::from_bits(0x7F80_0001u32);
let neg_nan32 = f32::from_bits(0xFF80_0001u32);
let nan32_from_16 = f32::from(nan16);
let neg_nan32_from_16 = f32::from(neg_nan16);
let nan64_from_16 = f64::from(nan16);
let neg_nan64_from_16 = f64::from(neg_nan16);
let nan64_from_32 = f64::from(nan32);
let neg_nan64_from_32 = f64::from(neg_nan32);
assert!(nan16.is_nan() && nan16.is_sign_positive());
assert!(neg_nan16.is_nan() && neg_nan16.is_sign_negative());
assert!(nan32.is_nan() && nan32.is_sign_positive());
assert!(neg_nan32.is_nan() && neg_nan32.is_sign_negative());
assert!(nan32_from_16.is_nan() && nan32_from_16.is_sign_positive());
assert!(neg_nan32_from_16.is_nan() && neg_nan32_from_16.is_sign_negative());
assert!(nan64_from_16.is_nan() && nan64_from_16.is_sign_positive());
assert!(neg_nan64_from_16.is_nan() && neg_nan64_from_16.is_sign_negative());
assert!(nan64_from_32.is_nan() && nan64_from_32.is_sign_positive());
assert!(neg_nan64_from_32.is_nan() && neg_nan64_from_32.is_sign_negative());
}
#[test]
fn test_f16_to_f32() {
let f = f16::from_f32(7.0);
assert_eq!(f.to_f32(), 7.0f32);
let f = f16::from_f32(7.1);
let diff = (f.to_f32() - 7.1f32).abs();
assert!(diff <= 4.0 * f16::EPSILON.to_f32());
assert_eq!(f16::from_bits(0x0000_0001).to_f32(), 2.0f32.powi(-24));
assert_eq!(f16::from_bits(0x0000_0005).to_f32(), 5.0 * 2.0f32.powi(-24));
assert_eq!(f16::from_bits(0x0000_0001), f16::from_f32(2.0f32.powi(-24)));
assert_eq!(
f16::from_bits(0x0000_0005),
f16::from_f32(5.0 * 2.0f32.powi(-24))
);
}
#[test]
fn test_f16_to_f64() {
let f = f16::from_f64(7.0);
assert_eq!(f.to_f64(), 7.0f64);
let f = f16::from_f64(7.1);
let diff = (f.to_f64() - 7.1f64).abs();
assert!(diff <= 4.0 * f16::EPSILON.to_f64());
assert_eq!(f16::from_bits(0x0000_0001).to_f64(), 2.0f64.powi(-24));
assert_eq!(f16::from_bits(0x0000_0005).to_f64(), 5.0 * 2.0f64.powi(-24));
assert_eq!(f16::from_bits(0x0000_0001), f16::from_f64(2.0f64.powi(-24)));
assert_eq!(
f16::from_bits(0x0000_0005),
f16::from_f64(5.0 * 2.0f64.powi(-24))
);
}
#[test]
fn test_comparisons() {
let zero = f16::from_f64(0.0);
let one = f16::from_f64(1.0);
let neg_zero = f16::from_f64(-0.0);
let neg_one = f16::from_f64(-1.0);
assert_eq!(zero.partial_cmp(&neg_zero), Some(Ordering::Equal));
assert_eq!(neg_zero.partial_cmp(&zero), Some(Ordering::Equal));
assert!(zero == neg_zero);
assert!(neg_zero == zero);
assert!(!(zero != neg_zero));
assert!(!(neg_zero != zero));
assert!(!(zero < neg_zero));
assert!(!(neg_zero < zero));
assert!(zero <= neg_zero);
assert!(neg_zero <= zero);
assert!(!(zero > neg_zero));
assert!(!(neg_zero > zero));
assert!(zero >= neg_zero);
assert!(neg_zero >= zero);
assert_eq!(one.partial_cmp(&neg_zero), Some(Ordering::Greater));
assert_eq!(neg_zero.partial_cmp(&one), Some(Ordering::Less));
assert!(!(one == neg_zero));
assert!(!(neg_zero == one));
assert!(one != neg_zero);
assert!(neg_zero != one);
assert!(!(one < neg_zero));
assert!(neg_zero < one);
assert!(!(one <= neg_zero));
assert!(neg_zero <= one);
assert!(one > neg_zero);
assert!(!(neg_zero > one));
assert!(one >= neg_zero);
assert!(!(neg_zero >= one));
assert_eq!(one.partial_cmp(&neg_one), Some(Ordering::Greater));
assert_eq!(neg_one.partial_cmp(&one), Some(Ordering::Less));
assert!(!(one == neg_one));
assert!(!(neg_one == one));
assert!(one != neg_one);
assert!(neg_one != one);
assert!(!(one < neg_one));
assert!(neg_one < one);
assert!(!(one <= neg_one));
assert!(neg_one <= one);
assert!(one > neg_one);
assert!(!(neg_one > one));
assert!(one >= neg_one);
assert!(!(neg_one >= one));
}
#[test]
#[allow(clippy::erasing_op, clippy::identity_op)]
fn round_to_even_f32() {
let min_sub = f16::from_bits(1);
let min_sub_f = (-24f32).exp2();
assert_eq!(f16::from_f32(min_sub_f).to_bits(), min_sub.to_bits());
assert_eq!(f32::from(min_sub).to_bits(), min_sub_f.to_bits());
assert_eq!(
f16::from_f32(min_sub_f * 0.49).to_bits(),
min_sub.to_bits() * 0
);
assert_eq!(
f16::from_f32(min_sub_f * 0.50).to_bits(),
min_sub.to_bits() * 0
);
assert_eq!(
f16::from_f32(min_sub_f * 0.51).to_bits(),
min_sub.to_bits() * 1
);
assert_eq!(
f16::from_f32(min_sub_f * 1.49).to_bits(),
min_sub.to_bits() * 1
);
assert_eq!(
f16::from_f32(min_sub_f * 1.50).to_bits(),
min_sub.to_bits() * 2
);
assert_eq!(
f16::from_f32(min_sub_f * 1.51).to_bits(),
min_sub.to_bits() * 2
);
assert_eq!(
f16::from_f32(min_sub_f * 2.49).to_bits(),
min_sub.to_bits() * 2
);
assert_eq!(
f16::from_f32(min_sub_f * 2.50).to_bits(),
min_sub.to_bits() * 2
);
assert_eq!(
f16::from_f32(min_sub_f * 2.51).to_bits(),
min_sub.to_bits() * 3
);
assert_eq!(
f16::from_f32(2000.49f32).to_bits(),
f16::from_f32(2000.0).to_bits()
);
assert_eq!(
f16::from_f32(2000.50f32).to_bits(),
f16::from_f32(2000.0).to_bits()
);
assert_eq!(
f16::from_f32(2000.51f32).to_bits(),
f16::from_f32(2001.0).to_bits()
);
assert_eq!(
f16::from_f32(2001.49f32).to_bits(),
f16::from_f32(2001.0).to_bits()
);
assert_eq!(
f16::from_f32(2001.50f32).to_bits(),
f16::from_f32(2002.0).to_bits()
);
assert_eq!(
f16::from_f32(2001.51f32).to_bits(),
f16::from_f32(2002.0).to_bits()
);
assert_eq!(
f16::from_f32(2002.49f32).to_bits(),
f16::from_f32(2002.0).to_bits()
);
assert_eq!(
f16::from_f32(2002.50f32).to_bits(),
f16::from_f32(2002.0).to_bits()
);
assert_eq!(
f16::from_f32(2002.51f32).to_bits(),
f16::from_f32(2003.0).to_bits()
);
}
#[test]
#[allow(clippy::erasing_op, clippy::identity_op)]
fn round_to_even_f64() {
let min_sub = f16::from_bits(1);
let min_sub_f = (-24f64).exp2();
assert_eq!(f16::from_f64(min_sub_f).to_bits(), min_sub.to_bits());
assert_eq!(f64::from(min_sub).to_bits(), min_sub_f.to_bits());
assert_eq!(
f16::from_f64(min_sub_f * 0.49).to_bits(),
min_sub.to_bits() * 0
);
assert_eq!(
f16::from_f64(min_sub_f * 0.50).to_bits(),
min_sub.to_bits() * 0
);
assert_eq!(
f16::from_f64(min_sub_f * 0.51).to_bits(),
min_sub.to_bits() * 1
);
assert_eq!(
f16::from_f64(min_sub_f * 1.49).to_bits(),
min_sub.to_bits() * 1
);
assert_eq!(
f16::from_f64(min_sub_f * 1.50).to_bits(),
min_sub.to_bits() * 2
);
assert_eq!(
f16::from_f64(min_sub_f * 1.51).to_bits(),
min_sub.to_bits() * 2
);
assert_eq!(
f16::from_f64(min_sub_f * 2.49).to_bits(),
min_sub.to_bits() * 2
);
assert_eq!(
f16::from_f64(min_sub_f * 2.50).to_bits(),
min_sub.to_bits() * 2
);
assert_eq!(
f16::from_f64(min_sub_f * 2.51).to_bits(),
min_sub.to_bits() * 3
);
assert_eq!(
f16::from_f64(2000.49f64).to_bits(),
f16::from_f64(2000.0).to_bits()
);
assert_eq!(
f16::from_f64(2000.50f64).to_bits(),
f16::from_f64(2000.0).to_bits()
);
assert_eq!(
f16::from_f64(2000.51f64).to_bits(),
f16::from_f64(2001.0).to_bits()
);
assert_eq!(
f16::from_f64(2001.49f64).to_bits(),
f16::from_f64(2001.0).to_bits()
);
assert_eq!(
f16::from_f64(2001.50f64).to_bits(),
f16::from_f64(2002.0).to_bits()
);
assert_eq!(
f16::from_f64(2001.51f64).to_bits(),
f16::from_f64(2002.0).to_bits()
);
assert_eq!(
f16::from_f64(2002.49f64).to_bits(),
f16::from_f64(2002.0).to_bits()
);
assert_eq!(
f16::from_f64(2002.50f64).to_bits(),
f16::from_f64(2002.0).to_bits()
);
assert_eq!(
f16::from_f64(2002.51f64).to_bits(),
f16::from_f64(2003.0).to_bits()
);
}
impl quickcheck::Arbitrary for f16 {
fn arbitrary(g: &mut quickcheck::Gen) -> Self {
f16(u16::arbitrary(g))
}
}
#[quickcheck]
fn qc_roundtrip_f16_f32_is_identity(f: f16) -> bool {
let roundtrip = f16::from_f32(f.to_f32());
if f.is_nan() {
roundtrip.is_nan() && f.is_sign_negative() == roundtrip.is_sign_negative()
} else {
f.0 == roundtrip.0
}
}
#[quickcheck]
fn qc_roundtrip_f16_f64_is_identity(f: f16) -> bool {
let roundtrip = f16::from_f64(f.to_f64());
if f.is_nan() {
roundtrip.is_nan() && f.is_sign_negative() == roundtrip.is_sign_negative()
} else {
f.0 == roundtrip.0
}
}
}