| 1 |
//! Contains utility functions and traits to convert between slices of [`u16`] bits and [`f16`] or |
| 2 |
//! [`bf16`] numbers. |
| 3 |
//! |
| 4 |
//! The utility [`HalfBitsSliceExt`] sealed extension trait is implemented for `[u16]` slices, |
| 5 |
//! while the utility [`HalfFloatSliceExt`] sealed extension trait is implemented for both `[f16]` |
| 6 |
//! and `[bf16]` slices. These traits provide efficient conversions and reinterpret casting of |
| 7 |
//! larger buffers of floating point values, and are automatically included in the |
| 8 |
//! [`prelude`][crate::prelude] module. |
| 9 |
|
| 10 |
use crate::{bf16, binary16::arch, f16}; |
| 11 |
#[cfg(feature = "alloc")] |
| 12 |
#[allow(unused_imports)] |
| 13 |
use alloc::{vec, vec::Vec}; |
| 14 |
use core::slice; |
| 15 |
|
| 16 |
/// Extensions to `[f16]` and `[bf16]` slices to support conversion and reinterpret operations. |
| 17 |
/// |
| 18 |
/// This trait is sealed and cannot be implemented outside of this crate. |
| 19 |
pub trait HalfFloatSliceExt: private::SealedHalfFloatSlice { |
| 20 |
/// Reinterprets a slice of [`f16`] or [`bf16`] numbers as a slice of [`u16`] bits. |
| 21 |
/// |
| 22 |
/// This is a zero-copy operation. The reinterpreted slice has the same lifetime and memory |
| 23 |
/// location as `self`. |
| 24 |
/// |
| 25 |
/// # Examples |
| 26 |
/// |
| 27 |
/// ```rust |
| 28 |
/// # use half::prelude::*; |
| 29 |
/// let float_buffer = [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.)]; |
| 30 |
/// let int_buffer = float_buffer.reinterpret_cast(); |
| 31 |
/// |
| 32 |
/// assert_eq!(int_buffer, [float_buffer[0].to_bits(), float_buffer[1].to_bits(), float_buffer[2].to_bits()]); |
| 33 |
/// ``` |
| 34 |
#[must_use] |
| 35 |
fn reinterpret_cast(&self) -> &[u16]; |
| 36 |
|
| 37 |
/// Reinterprets a mutable slice of [`f16`] or [`bf16`] numbers as a mutable slice of [`u16`]. |
| 38 |
/// bits |
| 39 |
/// |
| 40 |
/// This is a zero-copy operation. The transmuted slice has the same lifetime as the original, |
| 41 |
/// which prevents mutating `self` as long as the returned `&mut [u16]` is borrowed. |
| 42 |
/// |
| 43 |
/// # Examples |
| 44 |
/// |
| 45 |
/// ```rust |
| 46 |
/// # use half::prelude::*; |
| 47 |
/// let mut float_buffer = [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.)]; |
| 48 |
/// |
| 49 |
/// { |
| 50 |
/// let int_buffer = float_buffer.reinterpret_cast_mut(); |
| 51 |
/// |
| 52 |
/// assert_eq!(int_buffer, [f16::from_f32(1.).to_bits(), f16::from_f32(2.).to_bits(), f16::from_f32(3.).to_bits()]); |
| 53 |
/// |
| 54 |
/// // Mutating the u16 slice will mutating the original |
| 55 |
/// int_buffer[0] = 0; |
| 56 |
/// } |
| 57 |
/// |
| 58 |
/// // Note that we need to drop int_buffer before using float_buffer again or we will get a borrow error. |
| 59 |
/// assert_eq!(float_buffer, [f16::from_f32(0.), f16::from_f32(2.), f16::from_f32(3.)]); |
| 60 |
/// ``` |
| 61 |
#[must_use] |
| 62 |
fn reinterpret_cast_mut(&mut self) -> &mut [u16]; |
| 63 |
|
| 64 |
/// Converts all of the elements of a `[f32]` slice into [`f16`] or [`bf16`] values in `self`. |
| 65 |
/// |
| 66 |
/// The length of `src` must be the same as `self`. |
| 67 |
/// |
| 68 |
/// The conversion operation is vectorized over the slice, meaning the conversion may be more |
| 69 |
/// efficient than converting individual elements on some hardware that supports SIMD |
| 70 |
/// conversions. See [crate documentation](crate) for more information on hardware conversion |
| 71 |
/// support. |
| 72 |
/// |
| 73 |
/// # Panics |
| 74 |
/// |
| 75 |
/// This function will panic if the two slices have different lengths. |
| 76 |
/// |
| 77 |
/// # Examples |
| 78 |
/// ```rust |
| 79 |
/// # use half::prelude::*; |
| 80 |
/// // Initialize an empty buffer |
| 81 |
/// let mut buffer = [0u16; 4]; |
| 82 |
/// let buffer = buffer.reinterpret_cast_mut::<f16>(); |
| 83 |
/// |
| 84 |
/// let float_values = [1., 2., 3., 4.]; |
| 85 |
/// |
| 86 |
/// // Now convert |
| 87 |
/// buffer.convert_from_f32_slice(&float_values); |
| 88 |
/// |
| 89 |
/// assert_eq!(buffer, [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.), f16::from_f32(4.)]); |
| 90 |
/// ``` |
| 91 |
fn convert_from_f32_slice(&mut self, src: &[f32]); |
| 92 |
|
| 93 |
/// Converts all of the elements of a `[f64]` slice into [`f16`] or [`bf16`] values in `self`. |
| 94 |
/// |
| 95 |
/// The length of `src` must be the same as `self`. |
| 96 |
/// |
| 97 |
/// The conversion operation is vectorized over the slice, meaning the conversion may be more |
| 98 |
/// efficient than converting individual elements on some hardware that supports SIMD |
| 99 |
/// conversions. See [crate documentation](crate) for more information on hardware conversion |
| 100 |
/// support. |
| 101 |
/// |
| 102 |
/// # Panics |
| 103 |
/// |
| 104 |
/// This function will panic if the two slices have different lengths. |
| 105 |
/// |
| 106 |
/// # Examples |
| 107 |
/// ```rust |
| 108 |
/// # use half::prelude::*; |
| 109 |
/// // Initialize an empty buffer |
| 110 |
/// let mut buffer = [0u16; 4]; |
| 111 |
/// let buffer = buffer.reinterpret_cast_mut::<f16>(); |
| 112 |
/// |
| 113 |
/// let float_values = [1., 2., 3., 4.]; |
| 114 |
/// |
| 115 |
/// // Now convert |
| 116 |
/// buffer.convert_from_f64_slice(&float_values); |
| 117 |
/// |
| 118 |
/// assert_eq!(buffer, [f16::from_f64(1.), f16::from_f64(2.), f16::from_f64(3.), f16::from_f64(4.)]); |
| 119 |
/// ``` |
| 120 |
fn convert_from_f64_slice(&mut self, src: &[f64]); |
| 121 |
|
| 122 |
/// Converts all of the [`f16`] or [`bf16`] elements of `self` into [`f32`] values in `dst`. |
| 123 |
/// |
| 124 |
/// The length of `src` must be the same as `self`. |
| 125 |
/// |
| 126 |
/// The conversion operation is vectorized over the slice, meaning the conversion may be more |
| 127 |
/// efficient than converting individual elements on some hardware that supports SIMD |
| 128 |
/// conversions. See [crate documentation](crate) for more information on hardware conversion |
| 129 |
/// support. |
| 130 |
/// |
| 131 |
/// # Panics |
| 132 |
/// |
| 133 |
/// This function will panic if the two slices have different lengths. |
| 134 |
/// |
| 135 |
/// # Examples |
| 136 |
/// ```rust |
| 137 |
/// # use half::prelude::*; |
| 138 |
/// // Initialize an empty buffer |
| 139 |
/// let mut buffer = [0f32; 4]; |
| 140 |
/// |
| 141 |
/// let half_values = [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.), f16::from_f32(4.)]; |
| 142 |
/// |
| 143 |
/// // Now convert |
| 144 |
/// half_values.convert_to_f32_slice(&mut buffer); |
| 145 |
/// |
| 146 |
/// assert_eq!(buffer, [1., 2., 3., 4.]); |
| 147 |
/// ``` |
| 148 |
fn convert_to_f32_slice(&self, dst: &mut [f32]); |
| 149 |
|
| 150 |
/// Converts all of the [`f16`] or [`bf16`] elements of `self` into [`f64`] values in `dst`. |
| 151 |
/// |
| 152 |
/// The length of `src` must be the same as `self`. |
| 153 |
/// |
| 154 |
/// The conversion operation is vectorized over the slice, meaning the conversion may be more |
| 155 |
/// efficient than converting individual elements on some hardware that supports SIMD |
| 156 |
/// conversions. See [crate documentation](crate) for more information on hardware conversion |
| 157 |
/// support. |
| 158 |
/// |
| 159 |
/// # Panics |
| 160 |
/// |
| 161 |
/// This function will panic if the two slices have different lengths. |
| 162 |
/// |
| 163 |
/// # Examples |
| 164 |
/// ```rust |
| 165 |
/// # use half::prelude::*; |
| 166 |
/// // Initialize an empty buffer |
| 167 |
/// let mut buffer = [0f64; 4]; |
| 168 |
/// |
| 169 |
/// let half_values = [f16::from_f64(1.), f16::from_f64(2.), f16::from_f64(3.), f16::from_f64(4.)]; |
| 170 |
/// |
| 171 |
/// // Now convert |
| 172 |
/// half_values.convert_to_f64_slice(&mut buffer); |
| 173 |
/// |
| 174 |
/// assert_eq!(buffer, [1., 2., 3., 4.]); |
| 175 |
/// ``` |
| 176 |
fn convert_to_f64_slice(&self, dst: &mut [f64]); |
| 177 |
|
| 178 |
// Because trait is sealed, we can get away with different interfaces between features. |
| 179 |
|
| 180 |
/// Converts all of the [`f16`] or [`bf16`] elements of `self` into [`f32`] values in a new |
| 181 |
/// vector |
| 182 |
/// |
| 183 |
/// The conversion operation is vectorized over the slice, meaning the conversion may be more |
| 184 |
/// efficient than converting individual elements on some hardware that supports SIMD |
| 185 |
/// conversions. See [crate documentation](crate) for more information on hardware conversion |
| 186 |
/// support. |
| 187 |
/// |
| 188 |
/// This method is only available with the `std` or `alloc` feature. |
| 189 |
/// |
| 190 |
/// # Examples |
| 191 |
/// ```rust |
| 192 |
/// # use half::prelude::*; |
| 193 |
/// let half_values = [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.), f16::from_f32(4.)]; |
| 194 |
/// let vec = half_values.to_f32_vec(); |
| 195 |
/// |
| 196 |
/// assert_eq!(vec, vec![1., 2., 3., 4.]); |
| 197 |
/// ``` |
| 198 |
#[cfg(any(feature = "alloc", feature = "std"))] |
| 199 |
#[must_use] |
| 200 |
fn to_f32_vec(&self) -> Vec<f32>; |
| 201 |
|
| 202 |
/// Converts all of the [`f16`] or [`bf16`] elements of `self` into [`f64`] values in a new |
| 203 |
/// vector. |
| 204 |
/// |
| 205 |
/// The conversion operation is vectorized over the slice, meaning the conversion may be more |
| 206 |
/// efficient than converting individual elements on some hardware that supports SIMD |
| 207 |
/// conversions. See [crate documentation](crate) for more information on hardware conversion |
| 208 |
/// support. |
| 209 |
/// |
| 210 |
/// This method is only available with the `std` or `alloc` feature. |
| 211 |
/// |
| 212 |
/// # Examples |
| 213 |
/// ```rust |
| 214 |
/// # use half::prelude::*; |
| 215 |
/// let half_values = [f16::from_f64(1.), f16::from_f64(2.), f16::from_f64(3.), f16::from_f64(4.)]; |
| 216 |
/// let vec = half_values.to_f64_vec(); |
| 217 |
/// |
| 218 |
/// assert_eq!(vec, vec![1., 2., 3., 4.]); |
| 219 |
/// ``` |
| 220 |
#[cfg(feature = "alloc")] |
| 221 |
#[must_use] |
| 222 |
fn to_f64_vec(&self) -> Vec<f64>; |
| 223 |
} |
| 224 |
|
| 225 |
/// Extensions to `[u16]` slices to support reinterpret operations. |
| 226 |
/// |
| 227 |
/// This trait is sealed and cannot be implemented outside of this crate. |
| 228 |
pub trait HalfBitsSliceExt: private::SealedHalfBitsSlice { |
| 229 |
/// Reinterprets a slice of [`u16`] bits as a slice of [`f16`] or [`bf16`] numbers. |
| 230 |
/// |
| 231 |
/// `H` is the type to cast to, and must be either the [`f16`] or [`bf16`] type. |
| 232 |
/// |
| 233 |
/// This is a zero-copy operation. The reinterpreted slice has the same lifetime and memory |
| 234 |
/// location as `self`. |
| 235 |
/// |
| 236 |
/// # Examples |
| 237 |
/// |
| 238 |
/// ```rust |
| 239 |
/// # use half::prelude::*; |
| 240 |
/// let int_buffer = [f16::from_f32(1.).to_bits(), f16::from_f32(2.).to_bits(), f16::from_f32(3.).to_bits()]; |
| 241 |
/// let float_buffer: &[f16] = int_buffer.reinterpret_cast(); |
| 242 |
/// |
| 243 |
/// assert_eq!(float_buffer, [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.)]); |
| 244 |
/// |
| 245 |
/// // You may have to specify the cast type directly if the compiler can't infer the type. |
| 246 |
/// // The following is also valid in Rust. |
| 247 |
/// let typed_buffer = int_buffer.reinterpret_cast::<f16>(); |
| 248 |
/// ``` |
| 249 |
#[must_use] |
| 250 |
fn reinterpret_cast<H>(&self) -> &[H] |
| 251 |
where |
| 252 |
H: crate::private::SealedHalf; |
| 253 |
|
| 254 |
/// Reinterprets a mutable slice of [`u16`] bits as a mutable slice of [`f16`] or [`bf16`] |
| 255 |
/// numbers. |
| 256 |
/// |
| 257 |
/// `H` is the type to cast to, and must be either the [`f16`] or [`bf16`] type. |
| 258 |
/// |
| 259 |
/// This is a zero-copy operation. The transmuted slice has the same lifetime as the original, |
| 260 |
/// which prevents mutating `self` as long as the returned `&mut [f16]` is borrowed. |
| 261 |
/// |
| 262 |
/// # Examples |
| 263 |
/// |
| 264 |
/// ```rust |
| 265 |
/// # use half::prelude::*; |
| 266 |
/// let mut int_buffer = [f16::from_f32(1.).to_bits(), f16::from_f32(2.).to_bits(), f16::from_f32(3.).to_bits()]; |
| 267 |
/// |
| 268 |
/// { |
| 269 |
/// let float_buffer: &mut [f16] = int_buffer.reinterpret_cast_mut(); |
| 270 |
/// |
| 271 |
/// assert_eq!(float_buffer, [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.)]); |
| 272 |
/// |
| 273 |
/// // Mutating the f16 slice will mutating the original |
| 274 |
/// float_buffer[0] = f16::from_f32(0.); |
| 275 |
/// } |
| 276 |
/// |
| 277 |
/// // Note that we need to drop float_buffer before using int_buffer again or we will get a borrow error. |
| 278 |
/// assert_eq!(int_buffer, [f16::from_f32(0.).to_bits(), f16::from_f32(2.).to_bits(), f16::from_f32(3.).to_bits()]); |
| 279 |
/// |
| 280 |
/// // You may have to specify the cast type directly if the compiler can't infer the type. |
| 281 |
/// // The following is also valid in Rust. |
| 282 |
/// let typed_buffer = int_buffer.reinterpret_cast_mut::<f16>(); |
| 283 |
/// ``` |
| 284 |
#[must_use] |
| 285 |
fn reinterpret_cast_mut<H>(&mut self) -> &mut [H] |
| 286 |
where |
| 287 |
H: crate::private::SealedHalf; |
| 288 |
} |
| 289 |
|
| 290 |
mod private { |
| 291 |
use crate::{bf16, f16}; |
| 292 |
|
| 293 |
pub trait SealedHalfFloatSlice {} |
| 294 |
impl SealedHalfFloatSlice for [f16] {} |
| 295 |
impl SealedHalfFloatSlice for [bf16] {} |
| 296 |
|
| 297 |
pub trait SealedHalfBitsSlice {} |
| 298 |
impl SealedHalfBitsSlice for [u16] {} |
| 299 |
} |
| 300 |
|
| 301 |
impl HalfFloatSliceExt for [f16] { |
| 302 |
#[inline] |
| 303 |
fn reinterpret_cast(&self) -> &[u16] { |
| 304 |
let pointer = self.as_ptr() as *const u16; |
| 305 |
let length = self.len(); |
| 306 |
// SAFETY: We are reconstructing full length of original slice, using its same lifetime, |
| 307 |
// and the size of elements are identical |
| 308 |
unsafe { slice::from_raw_parts(pointer, length) } |
| 309 |
} |
| 310 |
|
| 311 |
#[inline] |
| 312 |
fn reinterpret_cast_mut(&mut self) -> &mut [u16] { |
| 313 |
let pointer = self.as_mut_ptr().cast::<u16>(); |
| 314 |
let length = self.len(); |
| 315 |
// SAFETY: We are reconstructing full length of original slice, using its same lifetime, |
| 316 |
// and the size of elements are identical |
| 317 |
unsafe { slice::from_raw_parts_mut(pointer, length) } |
| 318 |
} |
| 319 |
|
| 320 |
#[inline] |
| 321 |
fn convert_from_f32_slice(&mut self, src: &[f32]) { |
| 322 |
assert_eq!( |
| 323 |
self.len(), |
| 324 |
src.len(), |
| 325 |
"destination and source slices have different lengths" |
| 326 |
); |
| 327 |
|
| 328 |
arch::f32_to_f16_slice(src, self.reinterpret_cast_mut()) |
| 329 |
} |
| 330 |
|
| 331 |
#[inline] |
| 332 |
fn convert_from_f64_slice(&mut self, src: &[f64]) { |
| 333 |
assert_eq!( |
| 334 |
self.len(), |
| 335 |
src.len(), |
| 336 |
"destination and source slices have different lengths" |
| 337 |
); |
| 338 |
|
| 339 |
arch::f64_to_f16_slice(src, self.reinterpret_cast_mut()) |
| 340 |
} |
| 341 |
|
| 342 |
#[inline] |
| 343 |
fn convert_to_f32_slice(&self, dst: &mut [f32]) { |
| 344 |
assert_eq!( |
| 345 |
self.len(), |
| 346 |
dst.len(), |
| 347 |
"destination and source slices have different lengths" |
| 348 |
); |
| 349 |
|
| 350 |
arch::f16_to_f32_slice(self.reinterpret_cast(), dst) |
| 351 |
} |
| 352 |
|
| 353 |
#[inline] |
| 354 |
fn convert_to_f64_slice(&self, dst: &mut [f64]) { |
| 355 |
assert_eq!( |
| 356 |
self.len(), |
| 357 |
dst.len(), |
| 358 |
"destination and source slices have different lengths" |
| 359 |
); |
| 360 |
|
| 361 |
arch::f16_to_f64_slice(self.reinterpret_cast(), dst) |
| 362 |
} |
| 363 |
|
| 364 |
#[cfg(any(feature = "alloc", feature = "std"))] |
| 365 |
#[inline] |
| 366 |
#[allow(clippy::uninit_vec)] |
| 367 |
fn to_f32_vec(&self) -> Vec<f32> { |
| 368 |
let mut vec = vec![0f32; self.len()]; |
| 369 |
self.convert_to_f32_slice(&mut vec); |
| 370 |
vec |
| 371 |
} |
| 372 |
|
| 373 |
#[cfg(any(feature = "alloc", feature = "std"))] |
| 374 |
#[inline] |
| 375 |
#[allow(clippy::uninit_vec)] |
| 376 |
fn to_f64_vec(&self) -> Vec<f64> { |
| 377 |
let mut vec = vec![0f64; self.len()]; |
| 378 |
self.convert_to_f64_slice(&mut vec); |
| 379 |
vec |
| 380 |
} |
| 381 |
} |
| 382 |
|
| 383 |
impl HalfFloatSliceExt for [bf16] { |
| 384 |
#[inline] |
| 385 |
fn reinterpret_cast(&self) -> &[u16] { |
| 386 |
let pointer = self.as_ptr() as *const u16; |
| 387 |
let length = self.len(); |
| 388 |
// SAFETY: We are reconstructing full length of original slice, using its same lifetime, |
| 389 |
// and the size of elements are identical |
| 390 |
unsafe { slice::from_raw_parts(pointer, length) } |
| 391 |
} |
| 392 |
|
| 393 |
#[inline] |
| 394 |
fn reinterpret_cast_mut(&mut self) -> &mut [u16] { |
| 395 |
let pointer = self.as_mut_ptr().cast::<u16>(); |
| 396 |
let length = self.len(); |
| 397 |
// SAFETY: We are reconstructing full length of original slice, using its same lifetime, |
| 398 |
// and the size of elements are identical |
| 399 |
unsafe { slice::from_raw_parts_mut(pointer, length) } |
| 400 |
} |
| 401 |
|
| 402 |
#[inline] |
| 403 |
fn convert_from_f32_slice(&mut self, src: &[f32]) { |
| 404 |
assert_eq!( |
| 405 |
self.len(), |
| 406 |
src.len(), |
| 407 |
"destination and source slices have different lengths" |
| 408 |
); |
| 409 |
|
| 410 |
// Just use regular loop here until there's any bf16 SIMD support. |
| 411 |
for (i, f) in src.iter().enumerate() { |
| 412 |
self[i] = bf16::from_f32(*f); |
| 413 |
} |
| 414 |
} |
| 415 |
|
| 416 |
#[inline] |
| 417 |
fn convert_from_f64_slice(&mut self, src: &[f64]) { |
| 418 |
assert_eq!( |
| 419 |
self.len(), |
| 420 |
src.len(), |
| 421 |
"destination and source slices have different lengths" |
| 422 |
); |
| 423 |
|
| 424 |
// Just use regular loop here until there's any bf16 SIMD support. |
| 425 |
for (i, f) in src.iter().enumerate() { |
| 426 |
self[i] = bf16::from_f64(*f); |
| 427 |
} |
| 428 |
} |
| 429 |
|
| 430 |
#[inline] |
| 431 |
fn convert_to_f32_slice(&self, dst: &mut [f32]) { |
| 432 |
assert_eq!( |
| 433 |
self.len(), |
| 434 |
dst.len(), |
| 435 |
"destination and source slices have different lengths" |
| 436 |
); |
| 437 |
|
| 438 |
// Just use regular loop here until there's any bf16 SIMD support. |
| 439 |
for (i, f) in self.iter().enumerate() { |
| 440 |
dst[i] = f.to_f32(); |
| 441 |
} |
| 442 |
} |
| 443 |
|
| 444 |
#[inline] |
| 445 |
fn convert_to_f64_slice(&self, dst: &mut [f64]) { |
| 446 |
assert_eq!( |
| 447 |
self.len(), |
| 448 |
dst.len(), |
| 449 |
"destination and source slices have different lengths" |
| 450 |
); |
| 451 |
|
| 452 |
// Just use regular loop here until there's any bf16 SIMD support. |
| 453 |
for (i, f) in self.iter().enumerate() { |
| 454 |
dst[i] = f.to_f64(); |
| 455 |
} |
| 456 |
} |
| 457 |
|
| 458 |
#[cfg(any(feature = "alloc", feature = "std"))] |
| 459 |
#[inline] |
| 460 |
#[allow(clippy::uninit_vec)] |
| 461 |
fn to_f32_vec(&self) -> Vec<f32> { |
| 462 |
let mut vec = vec![0f32; self.len()]; |
| 463 |
self.convert_to_f32_slice(&mut vec); |
| 464 |
vec |
| 465 |
} |
| 466 |
|
| 467 |
#[cfg(any(feature = "alloc", feature = "std"))] |
| 468 |
#[inline] |
| 469 |
#[allow(clippy::uninit_vec)] |
| 470 |
fn to_f64_vec(&self) -> Vec<f64> { |
| 471 |
let mut vec = vec![0f64; self.len()]; |
| 472 |
self.convert_to_f64_slice(&mut vec); |
| 473 |
vec |
| 474 |
} |
| 475 |
} |
| 476 |
|
| 477 |
impl HalfBitsSliceExt for [u16] { |
| 478 |
// Since we sealed all the traits involved, these are safe. |
| 479 |
#[inline] |
| 480 |
fn reinterpret_cast<H>(&self) -> &[H] |
| 481 |
where |
| 482 |
H: crate::private::SealedHalf, |
| 483 |
{ |
| 484 |
let pointer = self.as_ptr() as *const H; |
| 485 |
let length = self.len(); |
| 486 |
// SAFETY: We are reconstructing full length of original slice, using its same lifetime, |
| 487 |
// and the size of elements are identical |
| 488 |
unsafe { slice::from_raw_parts(pointer, length) } |
| 489 |
} |
| 490 |
|
| 491 |
#[inline] |
| 492 |
fn reinterpret_cast_mut<H>(&mut self) -> &mut [H] |
| 493 |
where |
| 494 |
H: crate::private::SealedHalf, |
| 495 |
{ |
| 496 |
let pointer = self.as_mut_ptr() as *mut H; |
| 497 |
let length = self.len(); |
| 498 |
// SAFETY: We are reconstructing full length of original slice, using its same lifetime, |
| 499 |
// and the size of elements are identical |
| 500 |
unsafe { slice::from_raw_parts_mut(pointer, length) } |
| 501 |
} |
| 502 |
} |
| 503 |
|
| 504 |
#[allow(clippy::float_cmp)] |
| 505 |
#[cfg(test)] |
| 506 |
mod test { |
| 507 |
use super::{HalfBitsSliceExt, HalfFloatSliceExt}; |
| 508 |
use crate::{bf16, f16}; |
| 509 |
|
| 510 |
#[test] |
| 511 |
fn test_slice_conversions_f16() { |
| 512 |
let bits = &[ |
| 513 |
f16::E.to_bits(), |
| 514 |
f16::PI.to_bits(), |
| 515 |
f16::EPSILON.to_bits(), |
| 516 |
f16::FRAC_1_SQRT_2.to_bits(), |
| 517 |
]; |
| 518 |
let numbers = &[f16::E, f16::PI, f16::EPSILON, f16::FRAC_1_SQRT_2]; |
| 519 |
|
| 520 |
// Convert from bits to numbers |
| 521 |
let from_bits = bits.reinterpret_cast::<f16>(); |
| 522 |
assert_eq!(from_bits, numbers); |
| 523 |
|
| 524 |
// Convert from numbers back to bits |
| 525 |
let to_bits = from_bits.reinterpret_cast(); |
| 526 |
assert_eq!(to_bits, bits); |
| 527 |
} |
| 528 |
|
| 529 |
#[test] |
| 530 |
fn test_mutablility_f16() { |
| 531 |
let mut bits_array = [f16::PI.to_bits()]; |
| 532 |
let bits = &mut bits_array[..]; |
| 533 |
|
| 534 |
{ |
| 535 |
// would not compile without these braces |
| 536 |
let numbers = bits.reinterpret_cast_mut(); |
| 537 |
numbers[0] = f16::E; |
| 538 |
} |
| 539 |
|
| 540 |
assert_eq!(bits, &[f16::E.to_bits()]); |
| 541 |
|
| 542 |
bits[0] = f16::LN_2.to_bits(); |
| 543 |
assert_eq!(bits, &[f16::LN_2.to_bits()]); |
| 544 |
} |
| 545 |
|
| 546 |
#[test] |
| 547 |
fn test_slice_conversions_bf16() { |
| 548 |
let bits = &[ |
| 549 |
bf16::E.to_bits(), |
| 550 |
bf16::PI.to_bits(), |
| 551 |
bf16::EPSILON.to_bits(), |
| 552 |
bf16::FRAC_1_SQRT_2.to_bits(), |
| 553 |
]; |
| 554 |
let numbers = &[bf16::E, bf16::PI, bf16::EPSILON, bf16::FRAC_1_SQRT_2]; |
| 555 |
|
| 556 |
// Convert from bits to numbers |
| 557 |
let from_bits = bits.reinterpret_cast::<bf16>(); |
| 558 |
assert_eq!(from_bits, numbers); |
| 559 |
|
| 560 |
// Convert from numbers back to bits |
| 561 |
let to_bits = from_bits.reinterpret_cast(); |
| 562 |
assert_eq!(to_bits, bits); |
| 563 |
} |
| 564 |
|
| 565 |
#[test] |
| 566 |
fn test_mutablility_bf16() { |
| 567 |
let mut bits_array = [bf16::PI.to_bits()]; |
| 568 |
let bits = &mut bits_array[..]; |
| 569 |
|
| 570 |
{ |
| 571 |
// would not compile without these braces |
| 572 |
let numbers = bits.reinterpret_cast_mut(); |
| 573 |
numbers[0] = bf16::E; |
| 574 |
} |
| 575 |
|
| 576 |
assert_eq!(bits, &[bf16::E.to_bits()]); |
| 577 |
|
| 578 |
bits[0] = bf16::LN_2.to_bits(); |
| 579 |
assert_eq!(bits, &[bf16::LN_2.to_bits()]); |
| 580 |
} |
| 581 |
|
| 582 |
#[test] |
| 583 |
fn slice_convert_f16_f32() { |
| 584 |
// Exact chunks |
| 585 |
let vf32 = [1., 2., 3., 4., 5., 6., 7., 8.]; |
| 586 |
let vf16 = [ |
| 587 |
f16::from_f32(1.), |
| 588 |
f16::from_f32(2.), |
| 589 |
f16::from_f32(3.), |
| 590 |
f16::from_f32(4.), |
| 591 |
f16::from_f32(5.), |
| 592 |
f16::from_f32(6.), |
| 593 |
f16::from_f32(7.), |
| 594 |
f16::from_f32(8.), |
| 595 |
]; |
| 596 |
let mut buf32 = vf32; |
| 597 |
let mut buf16 = vf16; |
| 598 |
|
| 599 |
vf16.convert_to_f32_slice(&mut buf32); |
| 600 |
assert_eq!(&vf32, &buf32); |
| 601 |
|
| 602 |
buf16.convert_from_f32_slice(&vf32); |
| 603 |
assert_eq!(&vf16, &buf16); |
| 604 |
|
| 605 |
// Partial with chunks |
| 606 |
let vf32 = [1., 2., 3., 4., 5., 6., 7., 8., 9.]; |
| 607 |
let vf16 = [ |
| 608 |
f16::from_f32(1.), |
| 609 |
f16::from_f32(2.), |
| 610 |
f16::from_f32(3.), |
| 611 |
f16::from_f32(4.), |
| 612 |
f16::from_f32(5.), |
| 613 |
f16::from_f32(6.), |
| 614 |
f16::from_f32(7.), |
| 615 |
f16::from_f32(8.), |
| 616 |
f16::from_f32(9.), |
| 617 |
]; |
| 618 |
let mut buf32 = vf32; |
| 619 |
let mut buf16 = vf16; |
| 620 |
|
| 621 |
vf16.convert_to_f32_slice(&mut buf32); |
| 622 |
assert_eq!(&vf32, &buf32); |
| 623 |
|
| 624 |
buf16.convert_from_f32_slice(&vf32); |
| 625 |
assert_eq!(&vf16, &buf16); |
| 626 |
|
| 627 |
// Partial with chunks |
| 628 |
let vf32 = [1., 2.]; |
| 629 |
let vf16 = [f16::from_f32(1.), f16::from_f32(2.)]; |
| 630 |
let mut buf32 = vf32; |
| 631 |
let mut buf16 = vf16; |
| 632 |
|
| 633 |
vf16.convert_to_f32_slice(&mut buf32); |
| 634 |
assert_eq!(&vf32, &buf32); |
| 635 |
|
| 636 |
buf16.convert_from_f32_slice(&vf32); |
| 637 |
assert_eq!(&vf16, &buf16); |
| 638 |
} |
| 639 |
|
| 640 |
#[test] |
| 641 |
fn slice_convert_bf16_f32() { |
| 642 |
// Exact chunks |
| 643 |
let vf32 = [1., 2., 3., 4., 5., 6., 7., 8.]; |
| 644 |
let vf16 = [ |
| 645 |
bf16::from_f32(1.), |
| 646 |
bf16::from_f32(2.), |
| 647 |
bf16::from_f32(3.), |
| 648 |
bf16::from_f32(4.), |
| 649 |
bf16::from_f32(5.), |
| 650 |
bf16::from_f32(6.), |
| 651 |
bf16::from_f32(7.), |
| 652 |
bf16::from_f32(8.), |
| 653 |
]; |
| 654 |
let mut buf32 = vf32; |
| 655 |
let mut buf16 = vf16; |
| 656 |
|
| 657 |
vf16.convert_to_f32_slice(&mut buf32); |
| 658 |
assert_eq!(&vf32, &buf32); |
| 659 |
|
| 660 |
buf16.convert_from_f32_slice(&vf32); |
| 661 |
assert_eq!(&vf16, &buf16); |
| 662 |
|
| 663 |
// Partial with chunks |
| 664 |
let vf32 = [1., 2., 3., 4., 5., 6., 7., 8., 9.]; |
| 665 |
let vf16 = [ |
| 666 |
bf16::from_f32(1.), |
| 667 |
bf16::from_f32(2.), |
| 668 |
bf16::from_f32(3.), |
| 669 |
bf16::from_f32(4.), |
| 670 |
bf16::from_f32(5.), |
| 671 |
bf16::from_f32(6.), |
| 672 |
bf16::from_f32(7.), |
| 673 |
bf16::from_f32(8.), |
| 674 |
bf16::from_f32(9.), |
| 675 |
]; |
| 676 |
let mut buf32 = vf32; |
| 677 |
let mut buf16 = vf16; |
| 678 |
|
| 679 |
vf16.convert_to_f32_slice(&mut buf32); |
| 680 |
assert_eq!(&vf32, &buf32); |
| 681 |
|
| 682 |
buf16.convert_from_f32_slice(&vf32); |
| 683 |
assert_eq!(&vf16, &buf16); |
| 684 |
|
| 685 |
// Partial with chunks |
| 686 |
let vf32 = [1., 2.]; |
| 687 |
let vf16 = [bf16::from_f32(1.), bf16::from_f32(2.)]; |
| 688 |
let mut buf32 = vf32; |
| 689 |
let mut buf16 = vf16; |
| 690 |
|
| 691 |
vf16.convert_to_f32_slice(&mut buf32); |
| 692 |
assert_eq!(&vf32, &buf32); |
| 693 |
|
| 694 |
buf16.convert_from_f32_slice(&vf32); |
| 695 |
assert_eq!(&vf16, &buf16); |
| 696 |
} |
| 697 |
|
| 698 |
#[test] |
| 699 |
fn slice_convert_f16_f64() { |
| 700 |
// Exact chunks |
| 701 |
let vf64 = [1., 2., 3., 4., 5., 6., 7., 8.]; |
| 702 |
let vf16 = [ |
| 703 |
f16::from_f64(1.), |
| 704 |
f16::from_f64(2.), |
| 705 |
f16::from_f64(3.), |
| 706 |
f16::from_f64(4.), |
| 707 |
f16::from_f64(5.), |
| 708 |
f16::from_f64(6.), |
| 709 |
f16::from_f64(7.), |
| 710 |
f16::from_f64(8.), |
| 711 |
]; |
| 712 |
let mut buf64 = vf64; |
| 713 |
let mut buf16 = vf16; |
| 714 |
|
| 715 |
vf16.convert_to_f64_slice(&mut buf64); |
| 716 |
assert_eq!(&vf64, &buf64); |
| 717 |
|
| 718 |
buf16.convert_from_f64_slice(&vf64); |
| 719 |
assert_eq!(&vf16, &buf16); |
| 720 |
|
| 721 |
// Partial with chunks |
| 722 |
let vf64 = [1., 2., 3., 4., 5., 6., 7., 8., 9.]; |
| 723 |
let vf16 = [ |
| 724 |
f16::from_f64(1.), |
| 725 |
f16::from_f64(2.), |
| 726 |
f16::from_f64(3.), |
| 727 |
f16::from_f64(4.), |
| 728 |
f16::from_f64(5.), |
| 729 |
f16::from_f64(6.), |
| 730 |
f16::from_f64(7.), |
| 731 |
f16::from_f64(8.), |
| 732 |
f16::from_f64(9.), |
| 733 |
]; |
| 734 |
let mut buf64 = vf64; |
| 735 |
let mut buf16 = vf16; |
| 736 |
|
| 737 |
vf16.convert_to_f64_slice(&mut buf64); |
| 738 |
assert_eq!(&vf64, &buf64); |
| 739 |
|
| 740 |
buf16.convert_from_f64_slice(&vf64); |
| 741 |
assert_eq!(&vf16, &buf16); |
| 742 |
|
| 743 |
// Partial with chunks |
| 744 |
let vf64 = [1., 2.]; |
| 745 |
let vf16 = [f16::from_f64(1.), f16::from_f64(2.)]; |
| 746 |
let mut buf64 = vf64; |
| 747 |
let mut buf16 = vf16; |
| 748 |
|
| 749 |
vf16.convert_to_f64_slice(&mut buf64); |
| 750 |
assert_eq!(&vf64, &buf64); |
| 751 |
|
| 752 |
buf16.convert_from_f64_slice(&vf64); |
| 753 |
assert_eq!(&vf16, &buf16); |
| 754 |
} |
| 755 |
|
| 756 |
#[test] |
| 757 |
fn slice_convert_bf16_f64() { |
| 758 |
// Exact chunks |
| 759 |
let vf64 = [1., 2., 3., 4., 5., 6., 7., 8.]; |
| 760 |
let vf16 = [ |
| 761 |
bf16::from_f64(1.), |
| 762 |
bf16::from_f64(2.), |
| 763 |
bf16::from_f64(3.), |
| 764 |
bf16::from_f64(4.), |
| 765 |
bf16::from_f64(5.), |
| 766 |
bf16::from_f64(6.), |
| 767 |
bf16::from_f64(7.), |
| 768 |
bf16::from_f64(8.), |
| 769 |
]; |
| 770 |
let mut buf64 = vf64; |
| 771 |
let mut buf16 = vf16; |
| 772 |
|
| 773 |
vf16.convert_to_f64_slice(&mut buf64); |
| 774 |
assert_eq!(&vf64, &buf64); |
| 775 |
|
| 776 |
buf16.convert_from_f64_slice(&vf64); |
| 777 |
assert_eq!(&vf16, &buf16); |
| 778 |
|
| 779 |
// Partial with chunks |
| 780 |
let vf64 = [1., 2., 3., 4., 5., 6., 7., 8., 9.]; |
| 781 |
let vf16 = [ |
| 782 |
bf16::from_f64(1.), |
| 783 |
bf16::from_f64(2.), |
| 784 |
bf16::from_f64(3.), |
| 785 |
bf16::from_f64(4.), |
| 786 |
bf16::from_f64(5.), |
| 787 |
bf16::from_f64(6.), |
| 788 |
bf16::from_f64(7.), |
| 789 |
bf16::from_f64(8.), |
| 790 |
bf16::from_f64(9.), |
| 791 |
]; |
| 792 |
let mut buf64 = vf64; |
| 793 |
let mut buf16 = vf16; |
| 794 |
|
| 795 |
vf16.convert_to_f64_slice(&mut buf64); |
| 796 |
assert_eq!(&vf64, &buf64); |
| 797 |
|
| 798 |
buf16.convert_from_f64_slice(&vf64); |
| 799 |
assert_eq!(&vf16, &buf16); |
| 800 |
|
| 801 |
// Partial with chunks |
| 802 |
let vf64 = [1., 2.]; |
| 803 |
let vf16 = [bf16::from_f64(1.), bf16::from_f64(2.)]; |
| 804 |
let mut buf64 = vf64; |
| 805 |
let mut buf16 = vf16; |
| 806 |
|
| 807 |
vf16.convert_to_f64_slice(&mut buf64); |
| 808 |
assert_eq!(&vf64, &buf64); |
| 809 |
|
| 810 |
buf16.convert_from_f64_slice(&vf64); |
| 811 |
assert_eq!(&vf16, &buf16); |
| 812 |
} |
| 813 |
|
| 814 |
#[test] |
| 815 |
#[should_panic] |
| 816 |
fn convert_from_f32_slice_len_mismatch_panics() { |
| 817 |
let mut slice1 = [f16::ZERO; 3]; |
| 818 |
let slice2 = [0f32; 4]; |
| 819 |
slice1.convert_from_f32_slice(&slice2); |
| 820 |
} |
| 821 |
|
| 822 |
#[test] |
| 823 |
#[should_panic] |
| 824 |
fn convert_from_f64_slice_len_mismatch_panics() { |
| 825 |
let mut slice1 = [f16::ZERO; 3]; |
| 826 |
let slice2 = [0f64; 4]; |
| 827 |
slice1.convert_from_f64_slice(&slice2); |
| 828 |
} |
| 829 |
|
| 830 |
#[test] |
| 831 |
#[should_panic] |
| 832 |
fn convert_to_f32_slice_len_mismatch_panics() { |
| 833 |
let slice1 = [f16::ZERO; 3]; |
| 834 |
let mut slice2 = [0f32; 4]; |
| 835 |
slice1.convert_to_f32_slice(&mut slice2); |
| 836 |
} |
| 837 |
|
| 838 |
#[test] |
| 839 |
#[should_panic] |
| 840 |
fn convert_to_f64_slice_len_mismatch_panics() { |
| 841 |
let slice1 = [f16::ZERO; 3]; |
| 842 |
let mut slice2 = [0f64; 4]; |
| 843 |
slice1.convert_to_f64_slice(&mut slice2); |
| 844 |
} |
| 845 |
} |