| 1 | //! Contains utility functions and traits to convert between vectors of [`u16`] bits and [`struct@f16`] or |
| 2 | //! [`bf16`] vectors. |
| 3 | //! |
| 4 | //! The utility [`HalfBitsVecExt`] sealed extension trait is implemented for [`Vec<u16>`] vectors, |
| 5 | //! while the utility [`HalfFloatVecExt`] sealed extension trait is implemented for both |
| 6 | //! [`Vec<f16>`] and [`Vec<bf16>`] vectors. These traits provide efficient conversions and |
| 7 | //! reinterpret casting of larger buffers of floating point values, and are automatically included |
| 8 | //! in the [`prelude`][crate::prelude] module. |
| 9 | //! |
| 10 | //! This module is only available with the `std` or `alloc` feature. |
| 11 | |
| 12 | use super::{bf16, f16, slice::HalfFloatSliceExt}; |
| 13 | #[cfg (feature = "alloc" )] |
| 14 | #[allow (unused_imports)] |
| 15 | use alloc::{vec, vec::Vec}; |
| 16 | use core::mem; |
| 17 | |
| 18 | /// Extensions to [`Vec<f16>`] and [`Vec<bf16>`] to support reinterpret operations. |
| 19 | /// |
| 20 | /// This trait is sealed and cannot be implemented outside of this crate. |
| 21 | pub trait HalfFloatVecExt: private::SealedHalfFloatVec { |
| 22 | /// Reinterprets a vector of [`struct@f16`]or [`bf16`] numbers as a vector of [`u16`] bits. |
| 23 | /// |
| 24 | /// This is a zero-copy operation. The reinterpreted vector has the same memory location as |
| 25 | /// `self`. |
| 26 | /// |
| 27 | /// # Examples |
| 28 | /// |
| 29 | /// ```rust |
| 30 | /// # use half::prelude::*; |
| 31 | /// let float_buffer = vec![f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.)]; |
| 32 | /// let int_buffer = float_buffer.reinterpret_into(); |
| 33 | /// |
| 34 | /// assert_eq!(int_buffer, [f16::from_f32(1.).to_bits(), f16::from_f32(2.).to_bits(), f16::from_f32(3.).to_bits()]); |
| 35 | /// ``` |
| 36 | #[must_use ] |
| 37 | fn reinterpret_into(self) -> Vec<u16>; |
| 38 | |
| 39 | /// Converts all of the elements of a `[f32]` slice into a new [`struct@f16`] or [`bf16`] vector. |
| 40 | /// |
| 41 | /// The conversion operation is vectorized over the slice, meaning the conversion may be more |
| 42 | /// efficient than converting individual elements on some hardware that supports SIMD |
| 43 | /// conversions. See [crate documentation][crate] for more information on hardware conversion |
| 44 | /// support. |
| 45 | /// |
| 46 | /// # Examples |
| 47 | /// ```rust |
| 48 | /// # use half::prelude::*; |
| 49 | /// let float_values = [1., 2., 3., 4.]; |
| 50 | /// let vec: Vec<f16> = Vec::from_f32_slice(&float_values); |
| 51 | /// |
| 52 | /// assert_eq!(vec, vec![f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.), f16::from_f32(4.)]); |
| 53 | /// ``` |
| 54 | #[must_use ] |
| 55 | fn from_f32_slice(slice: &[f32]) -> Self; |
| 56 | |
| 57 | /// Converts all of the elements of a `[f64]` slice into a new [`struct@f16`] or [`bf16`] vector. |
| 58 | /// |
| 59 | /// The conversion operation is vectorized over the slice, meaning the conversion may be more |
| 60 | /// efficient than converting individual elements on some hardware that supports SIMD |
| 61 | /// conversions. See [crate documentation][crate] for more information on hardware conversion |
| 62 | /// support. |
| 63 | /// |
| 64 | /// # Examples |
| 65 | /// ```rust |
| 66 | /// # use half::prelude::*; |
| 67 | /// let float_values = [1., 2., 3., 4.]; |
| 68 | /// let vec: Vec<f16> = Vec::from_f64_slice(&float_values); |
| 69 | /// |
| 70 | /// assert_eq!(vec, vec![f16::from_f64(1.), f16::from_f64(2.), f16::from_f64(3.), f16::from_f64(4.)]); |
| 71 | /// ``` |
| 72 | #[must_use ] |
| 73 | fn from_f64_slice(slice: &[f64]) -> Self; |
| 74 | } |
| 75 | |
| 76 | /// Extensions to [`Vec<u16>`] to support reinterpret operations. |
| 77 | /// |
| 78 | /// This trait is sealed and cannot be implemented outside of this crate. |
| 79 | pub trait HalfBitsVecExt: private::SealedHalfBitsVec { |
| 80 | /// Reinterprets a vector of [`u16`] bits as a vector of [`struct@f16`] or [`bf16`] numbers. |
| 81 | /// |
| 82 | /// `H` is the type to cast to, and must be either the [`struct@f16`] or [`bf16`] type. |
| 83 | /// |
| 84 | /// This is a zero-copy operation. The reinterpreted vector has the same memory location as |
| 85 | /// `self`. |
| 86 | /// |
| 87 | /// # Examples |
| 88 | /// |
| 89 | /// ```rust |
| 90 | /// # use half::prelude::*; |
| 91 | /// let int_buffer = vec![f16::from_f32(1.).to_bits(), f16::from_f32(2.).to_bits(), f16::from_f32(3.).to_bits()]; |
| 92 | /// let float_buffer = int_buffer.reinterpret_into::<f16>(); |
| 93 | /// |
| 94 | /// assert_eq!(float_buffer, [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.)]); |
| 95 | /// ``` |
| 96 | #[must_use ] |
| 97 | fn reinterpret_into<H>(self) -> Vec<H> |
| 98 | where |
| 99 | H: crate::private::SealedHalf; |
| 100 | } |
| 101 | |
| 102 | mod private { |
| 103 | use crate::{bf16, f16}; |
| 104 | #[cfg (feature = "alloc" )] |
| 105 | #[allow (unused_imports)] |
| 106 | use alloc::vec::Vec; |
| 107 | |
| 108 | pub trait SealedHalfFloatVec {} |
| 109 | impl SealedHalfFloatVec for Vec<f16> {} |
| 110 | impl SealedHalfFloatVec for Vec<bf16> {} |
| 111 | |
| 112 | pub trait SealedHalfBitsVec {} |
| 113 | impl SealedHalfBitsVec for Vec<u16> {} |
| 114 | } |
| 115 | |
| 116 | impl HalfFloatVecExt for Vec<f16> { |
| 117 | #[inline ] |
| 118 | fn reinterpret_into(mut self) -> Vec<u16> { |
| 119 | // An f16 array has same length and capacity as u16 array |
| 120 | let length = self.len(); |
| 121 | let capacity = self.capacity(); |
| 122 | |
| 123 | // Actually reinterpret the contents of the Vec<f16> as u16, |
| 124 | // knowing that structs are represented as only their members in memory, |
| 125 | // which is the u16 part of `f16(u16)` |
| 126 | let pointer = self.as_mut_ptr() as *mut u16; |
| 127 | |
| 128 | // Prevent running a destructor on the old Vec<u16>, so the pointer won't be deleted |
| 129 | mem::forget(self); |
| 130 | |
| 131 | // Finally construct a new Vec<f16> from the raw pointer |
| 132 | // SAFETY: We are reconstructing full length and capacity of original vector, |
| 133 | // using its original pointer, and the size of elements are identical. |
| 134 | unsafe { Vec::from_raw_parts(pointer, length, capacity) } |
| 135 | } |
| 136 | |
| 137 | #[allow (clippy::uninit_vec)] |
| 138 | fn from_f32_slice(slice: &[f32]) -> Self { |
| 139 | let mut vec = vec![f16::from_bits(0); slice.len()]; |
| 140 | vec.convert_from_f32_slice(slice); |
| 141 | vec |
| 142 | } |
| 143 | |
| 144 | #[allow (clippy::uninit_vec)] |
| 145 | fn from_f64_slice(slice: &[f64]) -> Self { |
| 146 | let mut vec = vec![f16::from_bits(0); slice.len()]; |
| 147 | vec.convert_from_f64_slice(slice); |
| 148 | vec |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | impl HalfFloatVecExt for Vec<bf16> { |
| 153 | #[inline ] |
| 154 | fn reinterpret_into(mut self) -> Vec<u16> { |
| 155 | // An f16 array has same length and capacity as u16 array |
| 156 | let length = self.len(); |
| 157 | let capacity = self.capacity(); |
| 158 | |
| 159 | // Actually reinterpret the contents of the Vec<f16> as u16, |
| 160 | // knowing that structs are represented as only their members in memory, |
| 161 | // which is the u16 part of `f16(u16)` |
| 162 | let pointer = self.as_mut_ptr() as *mut u16; |
| 163 | |
| 164 | // Prevent running a destructor on the old Vec<u16>, so the pointer won't be deleted |
| 165 | mem::forget(self); |
| 166 | |
| 167 | // Finally construct a new Vec<f16> from the raw pointer |
| 168 | // SAFETY: We are reconstructing full length and capacity of original vector, |
| 169 | // using its original pointer, and the size of elements are identical. |
| 170 | unsafe { Vec::from_raw_parts(pointer, length, capacity) } |
| 171 | } |
| 172 | |
| 173 | #[allow (clippy::uninit_vec)] |
| 174 | fn from_f32_slice(slice: &[f32]) -> Self { |
| 175 | let mut vec = vec![bf16::from_bits(0); slice.len()]; |
| 176 | vec.convert_from_f32_slice(slice); |
| 177 | vec |
| 178 | } |
| 179 | |
| 180 | #[allow (clippy::uninit_vec)] |
| 181 | fn from_f64_slice(slice: &[f64]) -> Self { |
| 182 | let mut vec = vec![bf16::from_bits(0); slice.len()]; |
| 183 | vec.convert_from_f64_slice(slice); |
| 184 | vec |
| 185 | } |
| 186 | } |
| 187 | |
| 188 | impl HalfBitsVecExt for Vec<u16> { |
| 189 | // This is safe because all traits are sealed |
| 190 | #[inline ] |
| 191 | fn reinterpret_into<H>(mut self) -> Vec<H> |
| 192 | where |
| 193 | H: crate::private::SealedHalf, |
| 194 | { |
| 195 | // An f16 array has same length and capacity as u16 array |
| 196 | let length = self.len(); |
| 197 | let capacity = self.capacity(); |
| 198 | |
| 199 | // Actually reinterpret the contents of the Vec<u16> as f16, |
| 200 | // knowing that structs are represented as only their members in memory, |
| 201 | // which is the u16 part of `f16(u16)` |
| 202 | let pointer = self.as_mut_ptr() as *mut H; |
| 203 | |
| 204 | // Prevent running a destructor on the old Vec<u16>, so the pointer won't be deleted |
| 205 | mem::forget(self); |
| 206 | |
| 207 | // Finally construct a new Vec<f16> from the raw pointer |
| 208 | // SAFETY: We are reconstructing full length and capacity of original vector, |
| 209 | // using its original pointer, and the size of elements are identical. |
| 210 | unsafe { Vec::from_raw_parts(pointer, length, capacity) } |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | #[cfg (test)] |
| 215 | mod test { |
| 216 | use super::{HalfBitsVecExt, HalfFloatVecExt}; |
| 217 | use crate::{bf16, f16}; |
| 218 | #[cfg (all(feature = "alloc" , not(feature = "std" )))] |
| 219 | use alloc::vec; |
| 220 | |
| 221 | #[test ] |
| 222 | fn test_vec_conversions_f16() { |
| 223 | let numbers = vec![f16::E, f16::PI, f16::EPSILON, f16::FRAC_1_SQRT_2]; |
| 224 | let bits = vec![ |
| 225 | f16::E.to_bits(), |
| 226 | f16::PI.to_bits(), |
| 227 | f16::EPSILON.to_bits(), |
| 228 | f16::FRAC_1_SQRT_2.to_bits(), |
| 229 | ]; |
| 230 | let bits_cloned = bits.clone(); |
| 231 | |
| 232 | // Convert from bits to numbers |
| 233 | let from_bits = bits.reinterpret_into::<f16>(); |
| 234 | assert_eq!(&from_bits[..], &numbers[..]); |
| 235 | |
| 236 | // Convert from numbers back to bits |
| 237 | let to_bits = from_bits.reinterpret_into(); |
| 238 | assert_eq!(&to_bits[..], &bits_cloned[..]); |
| 239 | } |
| 240 | |
| 241 | #[test ] |
| 242 | fn test_vec_conversions_bf16() { |
| 243 | let numbers = vec![bf16::E, bf16::PI, bf16::EPSILON, bf16::FRAC_1_SQRT_2]; |
| 244 | let bits = vec![ |
| 245 | bf16::E.to_bits(), |
| 246 | bf16::PI.to_bits(), |
| 247 | bf16::EPSILON.to_bits(), |
| 248 | bf16::FRAC_1_SQRT_2.to_bits(), |
| 249 | ]; |
| 250 | let bits_cloned = bits.clone(); |
| 251 | |
| 252 | // Convert from bits to numbers |
| 253 | let from_bits = bits.reinterpret_into::<bf16>(); |
| 254 | assert_eq!(&from_bits[..], &numbers[..]); |
| 255 | |
| 256 | // Convert from numbers back to bits |
| 257 | let to_bits = from_bits.reinterpret_into(); |
| 258 | assert_eq!(&to_bits[..], &bits_cloned[..]); |
| 259 | } |
| 260 | } |
| 261 | |