use core::cmp::Ordering;
use core::fmt;
use core::ops::{Add, Div, Mul, Sub};
use crate::{ConversionError, dispatch};
macro_rules! impl_float_format {
($ty:ty, $to_float:ident) => {
impl fmt::Display for $ty {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
fmt::Display::fmt(&self.$to_float(), f)
}
}
impl fmt::LowerExp for $ty {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
fmt::LowerExp::fmt(&self.$to_float(), f)
}
}
impl fmt::UpperExp for $ty {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
fmt::UpperExp::fmt(&self.$to_float(), f)
}
}
};
}
#[derive(Clone, Copy, Default, Hash, PartialEq, Eq)]
#[repr(transparent)]
pub struct Uf8E4M4(u8);
pub type Uf8 = Uf8E4M4;
impl Uf8E4M4 {
pub const EXPONENT_BITS: u32 = 4;
pub const MANTISSA_BITS: u32 = 4;
pub const EXPONENT_BIAS: i32 = 7;
pub const EXPONENT_MASK: u8 = 0xf0;
pub const MANTISSA_MASK: u8 = 0x0f;
pub const ZERO: Self = Self(0);
pub const ONE: Self = Self(0x70);
pub const INFINITY: Self = Self(0xf0);
pub const NAN: Self = Self(0xf8);
pub const MAX: Self = Self(0xef);
pub const MIN_POSITIVE: Self = Self(0x01);
pub const MIN_NORMAL: Self = Self(0x10);
pub const fn from_bits(bits: u8) -> Self {
Self(bits)
}
pub const fn to_bits(self) -> u8 {
self.0
}
pub fn from_f32(value: f32) -> Self {
Self(dispatch::f32_to_uf8(value))
}
pub fn to_f32(self) -> f32 {
dispatch::uf8_to_f32(self.0)
}
pub fn from_f64(value: f64) -> Self {
Self(dispatch::f32_to_uf8(value as f32))
}
pub fn to_f64(self) -> f64 {
self.to_f32() as f64
}
pub fn try_from_f64(value: f64) -> Result<Self, ConversionError> {
crate::convert::check_finite_non_negative(value)?;
let encoded = Self::from_f64(value);
crate::convert::check_encoded(value, encoded.is_zero(), encoded.is_infinite())?;
Ok(encoded)
}
#[cfg(feature = "f16")]
pub fn from_f16(value: f16) -> Self {
Self::from_f32(value as f32)
}
#[cfg(feature = "f16")]
pub fn to_f16(self) -> f16 {
self.to_f32() as f16
}
pub const fn exponent(self) -> u8 {
(self.0 & Self::EXPONENT_MASK) >> Self::MANTISSA_BITS
}
pub const fn mantissa(self) -> u8 {
self.0 & Self::MANTISSA_MASK
}
pub const fn is_zero(self) -> bool {
self.0 == 0
}
pub const fn is_nan(self) -> bool {
self.exponent() == 0x0f && self.mantissa() != 0
}
pub const fn is_infinite(self) -> bool {
self.0 == Self::INFINITY.0
}
pub const fn is_finite(self) -> bool {
self.exponent() != 0x0f
}
pub const fn is_subnormal(self) -> bool {
self.exponent() == 0 && self.mantissa() != 0
}
}
impl From<f32> for Uf8E4M4 {
fn from(value: f32) -> Self {
Self::from_f32(value)
}
}
#[cfg(feature = "f16")]
impl From<f16> for Uf8E4M4 {
fn from(value: f16) -> Self {
Self::from_f16(value)
}
}
impl From<Uf8E4M4> for f32 {
fn from(value: Uf8E4M4) -> Self {
value.to_f32()
}
}
#[cfg(feature = "f16")]
impl From<Uf8E4M4> for f16 {
fn from(value: Uf8E4M4) -> Self {
value.to_f16()
}
}
impl From<Uf8E4M4> for f64 {
fn from(value: Uf8E4M4) -> Self {
value.to_f64()
}
}
impl Ord for Uf8E4M4 {
fn cmp(&self, other: &Self) -> Ordering {
self.0.cmp(&other.0)
}
}
impl PartialOrd for Uf8E4M4 {
fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
Some(self.cmp(other))
}
}
impl Add for Uf8E4M4 {
type Output = Self;
fn add(self, rhs: Self) -> Self::Output {
Self(dispatch::add_uf8(self.0, rhs.0))
}
}
impl Sub for Uf8E4M4 {
type Output = Self;
fn sub(self, rhs: Self) -> Self::Output {
Self(dispatch::sub_uf8(self.0, rhs.0))
}
}
impl Mul for Uf8E4M4 {
type Output = Self;
fn mul(self, rhs: Self) -> Self::Output {
Self(dispatch::mul_uf8(self.0, rhs.0))
}
}
impl Div for Uf8E4M4 {
type Output = Self;
fn div(self, rhs: Self) -> Self::Output {
Self(dispatch::div_uf8(self.0, rhs.0))
}
}
impl fmt::Debug for Uf8E4M4 {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
f.debug_tuple("Uf8E4M4").field(&self.to_f32()).finish()
}
}
impl_float_format!(Uf8E4M4, to_f32);
#[derive(Clone, Copy, Default, Hash, PartialEq, Eq)]
#[repr(transparent)]
pub struct Uf8E5M3(u8);
impl Uf8E5M3 {
pub const EXPONENT_BITS: u32 = 5;
pub const MANTISSA_BITS: u32 = 3;
pub const EXPONENT_BIAS: i32 = 15;
pub const EXPONENT_MASK: u8 = 0xf8;
pub const MANTISSA_MASK: u8 = 0x07;
pub const ZERO: Self = Self(0);
pub const ONE: Self = Self(0x78);
pub const INFINITY: Self = Self(0xf8);
pub const NAN: Self = Self(0xfc);
pub const MAX: Self = Self(0xf7);
pub const MIN_POSITIVE: Self = Self(0x01);
pub const MIN_NORMAL: Self = Self(0x08);
pub const fn from_bits(bits: u8) -> Self {
Self(bits)
}
pub const fn to_bits(self) -> u8 {
self.0
}
pub fn from_f32(value: f32) -> Self {
Self(dispatch::f32_to_uf8_e5m3(value))
}
pub fn to_f32(self) -> f32 {
dispatch::uf8_e5m3_to_f32(self.0)
}
pub fn from_f64(value: f64) -> Self {
Self::from_f32(value as f32)
}
pub fn to_f64(self) -> f64 {
self.to_f32() as f64
}
pub fn try_from_f64(value: f64) -> Result<Self, ConversionError> {
crate::convert::check_finite_non_negative(value)?;
let encoded = Self::from_f64(value);
crate::convert::check_encoded(value, encoded.is_zero(), encoded.is_infinite())?;
Ok(encoded)
}
#[cfg(feature = "f16")]
pub fn from_f16(value: f16) -> Self {
Self::from_f32(value as f32)
}
#[cfg(feature = "f16")]
pub fn to_f16(self) -> f16 {
self.to_f32() as f16
}
pub const fn exponent(self) -> u8 {
(self.0 & Self::EXPONENT_MASK) >> Self::MANTISSA_BITS
}
pub const fn mantissa(self) -> u8 {
self.0 & Self::MANTISSA_MASK
}
pub const fn is_zero(self) -> bool {
self.0 == 0
}
pub const fn is_nan(self) -> bool {
self.exponent() == 0x1f && self.mantissa() != 0
}
pub const fn is_infinite(self) -> bool {
self.0 == Self::INFINITY.0
}
pub const fn is_finite(self) -> bool {
self.exponent() != 0x1f
}
pub const fn is_subnormal(self) -> bool {
self.exponent() == 0 && self.mantissa() != 0
}
}
impl From<f32> for Uf8E5M3 {
fn from(value: f32) -> Self {
Self::from_f32(value)
}
}
#[cfg(feature = "f16")]
impl From<f16> for Uf8E5M3 {
fn from(value: f16) -> Self {
Self::from_f16(value)
}
}
impl From<Uf8E5M3> for f32 {
fn from(value: Uf8E5M3) -> Self {
value.to_f32()
}
}
#[cfg(feature = "f16")]
impl From<Uf8E5M3> for f16 {
fn from(value: Uf8E5M3) -> Self {
value.to_f16()
}
}
impl From<Uf8E5M3> for f64 {
fn from(value: Uf8E5M3) -> Self {
value.to_f64()
}
}
impl Ord for Uf8E5M3 {
fn cmp(&self, other: &Self) -> Ordering {
self.0.cmp(&other.0)
}
}
impl PartialOrd for Uf8E5M3 {
fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
Some(self.cmp(other))
}
}
impl Add for Uf8E5M3 {
type Output = Self;
fn add(self, rhs: Self) -> Self::Output {
Self(dispatch::add_uf8_e5m3(self.0, rhs.0))
}
}
impl Sub for Uf8E5M3 {
type Output = Self;
fn sub(self, rhs: Self) -> Self::Output {
Self(dispatch::sub_uf8_e5m3(self.0, rhs.0))
}
}
impl Mul for Uf8E5M3 {
type Output = Self;
fn mul(self, rhs: Self) -> Self::Output {
Self(dispatch::mul_uf8_e5m3(self.0, rhs.0))
}
}
impl Div for Uf8E5M3 {
type Output = Self;
fn div(self, rhs: Self) -> Self::Output {
Self(dispatch::div_uf8_e5m3(self.0, rhs.0))
}
}
impl fmt::Debug for Uf8E5M3 {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
f.debug_tuple("Uf8E5M3").field(&self.to_f32()).finish()
}
}
impl_float_format!(Uf8E5M3, to_f32);