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::convert, f16}; |
11 | #[cfg (feature = "alloc" )] |
12 | use alloc::vec::Vec; |
13 | use core::slice; |
14 | |
15 | /// Extensions to `[f16]` and `[bf16]` slices to support conversion and reinterpret operations. |
16 | /// |
17 | /// This trait is sealed and cannot be implemented outside of this crate. |
18 | pub trait HalfFloatSliceExt: private::SealedHalfFloatSlice { |
19 | /// Reinterprets a slice of [`f16`] or [`bf16`] numbers as a slice of [`u16`] bits. |
20 | /// |
21 | /// This is a zero-copy operation. The reinterpreted slice has the same lifetime and memory |
22 | /// location as `self`. |
23 | /// |
24 | /// # Examples |
25 | /// |
26 | /// ```rust |
27 | /// # use half::prelude::*; |
28 | /// let float_buffer = [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.)]; |
29 | /// let int_buffer = float_buffer.reinterpret_cast(); |
30 | /// |
31 | /// assert_eq!(int_buffer, [float_buffer[0].to_bits(), float_buffer[1].to_bits(), float_buffer[2].to_bits()]); |
32 | /// ``` |
33 | fn reinterpret_cast(&self) -> &[u16]; |
34 | |
35 | /// Reinterprets a mutable slice of [`f16`] or [`bf16`] numbers as a mutable slice of [`u16`]. |
36 | /// bits |
37 | /// |
38 | /// This is a zero-copy operation. The transmuted slice has the same lifetime as the original, |
39 | /// which prevents mutating `self` as long as the returned `&mut [u16]` is borrowed. |
40 | /// |
41 | /// # Examples |
42 | /// |
43 | /// ```rust |
44 | /// # use half::prelude::*; |
45 | /// let mut float_buffer = [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.)]; |
46 | /// |
47 | /// { |
48 | /// let int_buffer = float_buffer.reinterpret_cast_mut(); |
49 | /// |
50 | /// assert_eq!(int_buffer, [f16::from_f32(1.).to_bits(), f16::from_f32(2.).to_bits(), f16::from_f32(3.).to_bits()]); |
51 | /// |
52 | /// // Mutating the u16 slice will mutating the original |
53 | /// int_buffer[0] = 0; |
54 | /// } |
55 | /// |
56 | /// // Note that we need to drop int_buffer before using float_buffer again or we will get a borrow error. |
57 | /// assert_eq!(float_buffer, [f16::from_f32(0.), f16::from_f32(2.), f16::from_f32(3.)]); |
58 | /// ``` |
59 | fn reinterpret_cast_mut(&mut self) -> &mut [u16]; |
60 | |
61 | /// Converts all of the elements of a `[f32]` slice into [`f16`] or [`bf16`] values in `self`. |
62 | /// |
63 | /// The length of `src` must be the same as `self`. |
64 | /// |
65 | /// The conversion operation is vectorized over the slice, meaning the conversion may be more |
66 | /// efficient than converting individual elements on some hardware that supports SIMD |
67 | /// conversions. See [crate documentation](crate) for more information on hardware conversion |
68 | /// support. |
69 | /// |
70 | /// # Panics |
71 | /// |
72 | /// This function will panic if the two slices have different lengths. |
73 | /// |
74 | /// # Examples |
75 | /// ```rust |
76 | /// # use half::prelude::*; |
77 | /// // Initialize an empty buffer |
78 | /// let mut buffer = [0u16; 4]; |
79 | /// let buffer = buffer.reinterpret_cast_mut::<f16>(); |
80 | /// |
81 | /// let float_values = [1., 2., 3., 4.]; |
82 | /// |
83 | /// // Now convert |
84 | /// buffer.convert_from_f32_slice(&float_values); |
85 | /// |
86 | /// assert_eq!(buffer, [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.), f16::from_f32(4.)]); |
87 | /// ``` |
88 | fn convert_from_f32_slice(&mut self, src: &[f32]); |
89 | |
90 | /// Converts all of the elements of a `[f64]` slice into [`f16`] or [`bf16`] values in `self`. |
91 | /// |
92 | /// The length of `src` must be the same as `self`. |
93 | /// |
94 | /// The conversion operation is vectorized over the slice, meaning the conversion may be more |
95 | /// efficient than converting individual elements on some hardware that supports SIMD |
96 | /// conversions. See [crate documentation](crate) for more information on hardware conversion |
97 | /// support. |
98 | /// |
99 | /// # Panics |
100 | /// |
101 | /// This function will panic if the two slices have different lengths. |
102 | /// |
103 | /// # Examples |
104 | /// ```rust |
105 | /// # use half::prelude::*; |
106 | /// // Initialize an empty buffer |
107 | /// let mut buffer = [0u16; 4]; |
108 | /// let buffer = buffer.reinterpret_cast_mut::<f16>(); |
109 | /// |
110 | /// let float_values = [1., 2., 3., 4.]; |
111 | /// |
112 | /// // Now convert |
113 | /// buffer.convert_from_f64_slice(&float_values); |
114 | /// |
115 | /// assert_eq!(buffer, [f16::from_f64(1.), f16::from_f64(2.), f16::from_f64(3.), f16::from_f64(4.)]); |
116 | /// ``` |
117 | fn convert_from_f64_slice(&mut self, src: &[f64]); |
118 | |
119 | /// Converts all of the [`f16`] or [`bf16`] elements of `self` into [`f32`] values in `dst`. |
120 | /// |
121 | /// The length of `src` must be the same as `self`. |
122 | /// |
123 | /// The conversion operation is vectorized over the slice, meaning the conversion may be more |
124 | /// efficient than converting individual elements on some hardware that supports SIMD |
125 | /// conversions. See [crate documentation](crate) for more information on hardware conversion |
126 | /// support. |
127 | /// |
128 | /// # Panics |
129 | /// |
130 | /// This function will panic if the two slices have different lengths. |
131 | /// |
132 | /// # Examples |
133 | /// ```rust |
134 | /// # use half::prelude::*; |
135 | /// // Initialize an empty buffer |
136 | /// let mut buffer = [0f32; 4]; |
137 | /// |
138 | /// let half_values = [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.), f16::from_f32(4.)]; |
139 | /// |
140 | /// // Now convert |
141 | /// half_values.convert_to_f32_slice(&mut buffer); |
142 | /// |
143 | /// assert_eq!(buffer, [1., 2., 3., 4.]); |
144 | /// ``` |
145 | fn convert_to_f32_slice(&self, dst: &mut [f32]); |
146 | |
147 | /// Converts all of the [`f16`] or [`bf16`] elements of `self` into [`f64`] values in `dst`. |
148 | /// |
149 | /// The length of `src` must be the same as `self`. |
150 | /// |
151 | /// The conversion operation is vectorized over the slice, meaning the conversion may be more |
152 | /// efficient than converting individual elements on some hardware that supports SIMD |
153 | /// conversions. See [crate documentation](crate) for more information on hardware conversion |
154 | /// support. |
155 | /// |
156 | /// # Panics |
157 | /// |
158 | /// This function will panic if the two slices have different lengths. |
159 | /// |
160 | /// # Examples |
161 | /// ```rust |
162 | /// # use half::prelude::*; |
163 | /// // Initialize an empty buffer |
164 | /// let mut buffer = [0f64; 4]; |
165 | /// |
166 | /// let half_values = [f16::from_f64(1.), f16::from_f64(2.), f16::from_f64(3.), f16::from_f64(4.)]; |
167 | /// |
168 | /// // Now convert |
169 | /// half_values.convert_to_f64_slice(&mut buffer); |
170 | /// |
171 | /// assert_eq!(buffer, [1., 2., 3., 4.]); |
172 | /// ``` |
173 | fn convert_to_f64_slice(&self, dst: &mut [f64]); |
174 | |
175 | // Because trait is sealed, we can get away with different interfaces between features. |
176 | |
177 | /// Converts all of the [`f16`] or [`bf16`] elements of `self` into [`f32`] values in a new |
178 | /// vector |
179 | /// |
180 | /// The conversion operation is vectorized over the slice, meaning the conversion may be more |
181 | /// efficient than converting individual elements on some hardware that supports SIMD |
182 | /// conversions. See [crate documentation](crate) for more information on hardware conversion |
183 | /// support. |
184 | /// |
185 | /// This method is only available with the `std` or `alloc` feature. |
186 | /// |
187 | /// # Examples |
188 | /// ```rust |
189 | /// # use half::prelude::*; |
190 | /// let half_values = [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.), f16::from_f32(4.)]; |
191 | /// let vec = half_values.to_f32_vec(); |
192 | /// |
193 | /// assert_eq!(vec, vec![1., 2., 3., 4.]); |
194 | /// ``` |
195 | #[cfg (any(feature = "alloc" , feature = "std" ))] |
196 | #[cfg_attr (docsrs, doc(cfg(feature = "alloc" )))] |
197 | fn to_f32_vec(&self) -> Vec<f32>; |
198 | |
199 | /// Converts all of the [`f16`] or [`bf16`] elements of `self` into [`f64`] values in a new |
200 | /// vector. |
201 | /// |
202 | /// The conversion operation is vectorized over the slice, meaning the conversion may be more |
203 | /// efficient than converting individual elements on some hardware that supports SIMD |
204 | /// conversions. See [crate documentation](crate) for more information on hardware conversion |
205 | /// support. |
206 | /// |
207 | /// This method is only available with the `std` or `alloc` feature. |
208 | /// |
209 | /// # Examples |
210 | /// ```rust |
211 | /// # use half::prelude::*; |
212 | /// let half_values = [f16::from_f64(1.), f16::from_f64(2.), f16::from_f64(3.), f16::from_f64(4.)]; |
213 | /// let vec = half_values.to_f64_vec(); |
214 | /// |
215 | /// assert_eq!(vec, vec![1., 2., 3., 4.]); |
216 | /// ``` |
217 | #[cfg (feature = "alloc" )] |
218 | #[cfg_attr (docsrs, doc(cfg(feature = "alloc" )))] |
219 | fn to_f64_vec(&self) -> Vec<f64>; |
220 | } |
221 | |
222 | /// Extensions to `[u16]` slices to support reinterpret operations. |
223 | /// |
224 | /// This trait is sealed and cannot be implemented outside of this crate. |
225 | pub trait HalfBitsSliceExt: private::SealedHalfBitsSlice { |
226 | /// Reinterprets a slice of [`u16`] bits as a slice of [`f16`] or [`bf16`] numbers. |
227 | /// |
228 | /// `H` is the type to cast to, and must be either the [`f16`] or [`bf16`] type. |
229 | /// |
230 | /// This is a zero-copy operation. The reinterpreted slice has the same lifetime and memory |
231 | /// location as `self`. |
232 | /// |
233 | /// # Examples |
234 | /// |
235 | /// ```rust |
236 | /// # use half::prelude::*; |
237 | /// let int_buffer = [f16::from_f32(1.).to_bits(), f16::from_f32(2.).to_bits(), f16::from_f32(3.).to_bits()]; |
238 | /// let float_buffer: &[f16] = int_buffer.reinterpret_cast(); |
239 | /// |
240 | /// assert_eq!(float_buffer, [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.)]); |
241 | /// |
242 | /// // You may have to specify the cast type directly if the compiler can't infer the type. |
243 | /// // The following is also valid in Rust. |
244 | /// let typed_buffer = int_buffer.reinterpret_cast::<f16>(); |
245 | /// ``` |
246 | fn reinterpret_cast<H>(&self) -> &[H] |
247 | where |
248 | H: crate::private::SealedHalf; |
249 | |
250 | /// Reinterprets a mutable slice of [`u16`] bits as a mutable slice of [`f16`] or [`bf16`] |
251 | /// numbers. |
252 | /// |
253 | /// `H` is the type to cast to, and must be either the [`f16`] or [`bf16`] type. |
254 | /// |
255 | /// This is a zero-copy operation. The transmuted slice has the same lifetime as the original, |
256 | /// which prevents mutating `self` as long as the returned `&mut [f16]` is borrowed. |
257 | /// |
258 | /// # Examples |
259 | /// |
260 | /// ```rust |
261 | /// # use half::prelude::*; |
262 | /// let mut int_buffer = [f16::from_f32(1.).to_bits(), f16::from_f32(2.).to_bits(), f16::from_f32(3.).to_bits()]; |
263 | /// |
264 | /// { |
265 | /// let float_buffer: &mut [f16] = int_buffer.reinterpret_cast_mut(); |
266 | /// |
267 | /// assert_eq!(float_buffer, [f16::from_f32(1.), f16::from_f32(2.), f16::from_f32(3.)]); |
268 | /// |
269 | /// // Mutating the f16 slice will mutating the original |
270 | /// float_buffer[0] = f16::from_f32(0.); |
271 | /// } |
272 | /// |
273 | /// // Note that we need to drop float_buffer before using int_buffer again or we will get a borrow error. |
274 | /// assert_eq!(int_buffer, [f16::from_f32(0.).to_bits(), f16::from_f32(2.).to_bits(), f16::from_f32(3.).to_bits()]); |
275 | /// |
276 | /// // You may have to specify the cast type directly if the compiler can't infer the type. |
277 | /// // The following is also valid in Rust. |
278 | /// let typed_buffer = int_buffer.reinterpret_cast_mut::<f16>(); |
279 | /// ``` |
280 | fn reinterpret_cast_mut<H>(&mut self) -> &mut [H] |
281 | where |
282 | H: crate::private::SealedHalf; |
283 | } |
284 | |
285 | mod private { |
286 | use crate::{bf16, f16}; |
287 | |
288 | pub trait SealedHalfFloatSlice {} |
289 | impl SealedHalfFloatSlice for [f16] {} |
290 | impl SealedHalfFloatSlice for [bf16] {} |
291 | |
292 | pub trait SealedHalfBitsSlice {} |
293 | impl SealedHalfBitsSlice for [u16] {} |
294 | } |
295 | |
296 | impl HalfFloatSliceExt for [f16] { |
297 | #[inline ] |
298 | fn reinterpret_cast(&self) -> &[u16] { |
299 | let pointer = self.as_ptr() as *const u16; |
300 | let length = self.len(); |
301 | // SAFETY: We are reconstructing full length of original slice, using its same lifetime, |
302 | // and the size of elements are identical |
303 | unsafe { slice::from_raw_parts(pointer, length) } |
304 | } |
305 | |
306 | #[inline ] |
307 | fn reinterpret_cast_mut(&mut self) -> &mut [u16] { |
308 | let pointer = self.as_ptr() as *mut u16; |
309 | let length = self.len(); |
310 | // SAFETY: We are reconstructing full length of original slice, using its same lifetime, |
311 | // and the size of elements are identical |
312 | unsafe { slice::from_raw_parts_mut(pointer, length) } |
313 | } |
314 | |
315 | fn convert_from_f32_slice(&mut self, src: &[f32]) { |
316 | assert_eq!( |
317 | self.len(), |
318 | src.len(), |
319 | "destination and source slices have different lengths" |
320 | ); |
321 | |
322 | let mut chunks = src.chunks_exact(4); |
323 | let mut chunk_count = 0usize; // Not using .enumerate() because we need this value for remainder |
324 | for chunk in &mut chunks { |
325 | let vec = convert::f32x4_to_f16x4(chunk); |
326 | let dst_idx = chunk_count * 4; |
327 | self[dst_idx..dst_idx + 4].copy_from_slice(vec.reinterpret_cast()); |
328 | chunk_count += 1; |
329 | } |
330 | |
331 | // Process remainder |
332 | if !chunks.remainder().is_empty() { |
333 | let mut buf = [0f32; 4]; |
334 | buf[..chunks.remainder().len()].copy_from_slice(chunks.remainder()); |
335 | let vec = convert::f32x4_to_f16x4(&buf); |
336 | let dst_idx = chunk_count * 4; |
337 | self[dst_idx..dst_idx + chunks.remainder().len()] |
338 | .copy_from_slice(vec[..chunks.remainder().len()].reinterpret_cast()); |
339 | } |
340 | } |
341 | |
342 | fn convert_from_f64_slice(&mut self, src: &[f64]) { |
343 | assert_eq!( |
344 | self.len(), |
345 | src.len(), |
346 | "destination and source slices have different lengths" |
347 | ); |
348 | |
349 | let mut chunks = src.chunks_exact(4); |
350 | let mut chunk_count = 0usize; // Not using .enumerate() because we need this value for remainder |
351 | for chunk in &mut chunks { |
352 | let vec = convert::f64x4_to_f16x4(chunk); |
353 | let dst_idx = chunk_count * 4; |
354 | self[dst_idx..dst_idx + 4].copy_from_slice(vec.reinterpret_cast()); |
355 | chunk_count += 1; |
356 | } |
357 | |
358 | // Process remainder |
359 | if !chunks.remainder().is_empty() { |
360 | let mut buf = [0f64; 4]; |
361 | buf[..chunks.remainder().len()].copy_from_slice(chunks.remainder()); |
362 | let vec = convert::f64x4_to_f16x4(&buf); |
363 | let dst_idx = chunk_count * 4; |
364 | self[dst_idx..dst_idx + chunks.remainder().len()] |
365 | .copy_from_slice(vec[..chunks.remainder().len()].reinterpret_cast()); |
366 | } |
367 | } |
368 | |
369 | fn convert_to_f32_slice(&self, dst: &mut [f32]) { |
370 | assert_eq!( |
371 | self.len(), |
372 | dst.len(), |
373 | "destination and source slices have different lengths" |
374 | ); |
375 | |
376 | let mut chunks = self.chunks_exact(4); |
377 | let mut chunk_count = 0usize; // Not using .enumerate() because we need this value for remainder |
378 | for chunk in &mut chunks { |
379 | let vec = convert::f16x4_to_f32x4(chunk.reinterpret_cast()); |
380 | let dst_idx = chunk_count * 4; |
381 | dst[dst_idx..dst_idx + 4].copy_from_slice(&vec); |
382 | chunk_count += 1; |
383 | } |
384 | |
385 | // Process remainder |
386 | if !chunks.remainder().is_empty() { |
387 | let mut buf = [0u16; 4]; |
388 | buf[..chunks.remainder().len()].copy_from_slice(chunks.remainder().reinterpret_cast()); |
389 | let vec = convert::f16x4_to_f32x4(&buf); |
390 | let dst_idx = chunk_count * 4; |
391 | dst[dst_idx..dst_idx + chunks.remainder().len()] |
392 | .copy_from_slice(&vec[..chunks.remainder().len()]); |
393 | } |
394 | } |
395 | |
396 | fn convert_to_f64_slice(&self, dst: &mut [f64]) { |
397 | assert_eq!( |
398 | self.len(), |
399 | dst.len(), |
400 | "destination and source slices have different lengths" |
401 | ); |
402 | |
403 | let mut chunks = self.chunks_exact(4); |
404 | let mut chunk_count = 0usize; // Not using .enumerate() because we need this value for remainder |
405 | for chunk in &mut chunks { |
406 | let vec = convert::f16x4_to_f64x4(chunk.reinterpret_cast()); |
407 | let dst_idx = chunk_count * 4; |
408 | dst[dst_idx..dst_idx + 4].copy_from_slice(&vec); |
409 | chunk_count += 1; |
410 | } |
411 | |
412 | // Process remainder |
413 | if !chunks.remainder().is_empty() { |
414 | let mut buf = [0u16; 4]; |
415 | buf[..chunks.remainder().len()].copy_from_slice(chunks.remainder().reinterpret_cast()); |
416 | let vec = convert::f16x4_to_f64x4(&buf); |
417 | let dst_idx = chunk_count * 4; |
418 | dst[dst_idx..dst_idx + chunks.remainder().len()] |
419 | .copy_from_slice(&vec[..chunks.remainder().len()]); |
420 | } |
421 | } |
422 | |
423 | #[cfg (any(feature = "alloc" , feature = "std" ))] |
424 | #[inline ] |
425 | fn to_f32_vec(&self) -> Vec<f32> { |
426 | let mut vec = Vec::with_capacity(self.len()); |
427 | // SAFETY: convert will initialize every value in the vector without reading them, |
428 | // so this is safe to do instead of double initialize from resize, and we're setting it to |
429 | // same value as capacity. |
430 | unsafe { vec.set_len(self.len()) }; |
431 | self.convert_to_f32_slice(&mut vec); |
432 | vec |
433 | } |
434 | |
435 | #[cfg (any(feature = "alloc" , feature = "std" ))] |
436 | #[inline ] |
437 | fn to_f64_vec(&self) -> Vec<f64> { |
438 | let mut vec = Vec::with_capacity(self.len()); |
439 | // SAFETY: convert will initialize every value in the vector without reading them, |
440 | // so this is safe to do instead of double initialize from resize, and we're setting it to |
441 | // same value as capacity. |
442 | unsafe { vec.set_len(self.len()) }; |
443 | self.convert_to_f64_slice(&mut vec); |
444 | vec |
445 | } |
446 | } |
447 | |
448 | impl HalfFloatSliceExt for [bf16] { |
449 | #[inline ] |
450 | fn reinterpret_cast(&self) -> &[u16] { |
451 | let pointer = self.as_ptr() as *const u16; |
452 | let length = self.len(); |
453 | // SAFETY: We are reconstructing full length of original slice, using its same lifetime, |
454 | // and the size of elements are identical |
455 | unsafe { slice::from_raw_parts(pointer, length) } |
456 | } |
457 | |
458 | #[inline ] |
459 | fn reinterpret_cast_mut(&mut self) -> &mut [u16] { |
460 | let pointer = self.as_ptr() as *mut u16; |
461 | let length = self.len(); |
462 | // SAFETY: We are reconstructing full length of original slice, using its same lifetime, |
463 | // and the size of elements are identical |
464 | unsafe { slice::from_raw_parts_mut(pointer, length) } |
465 | } |
466 | |
467 | fn convert_from_f32_slice(&mut self, src: &[f32]) { |
468 | assert_eq!( |
469 | self.len(), |
470 | src.len(), |
471 | "destination and source slices have different lengths" |
472 | ); |
473 | |
474 | // Just use regular loop here until there's any bf16 SIMD support. |
475 | for (i, f) in src.iter().enumerate() { |
476 | self[i] = bf16::from_f32(*f); |
477 | } |
478 | } |
479 | |
480 | fn convert_from_f64_slice(&mut self, src: &[f64]) { |
481 | assert_eq!( |
482 | self.len(), |
483 | src.len(), |
484 | "destination and source slices have different lengths" |
485 | ); |
486 | |
487 | // Just use regular loop here until there's any bf16 SIMD support. |
488 | for (i, f) in src.iter().enumerate() { |
489 | self[i] = bf16::from_f64(*f); |
490 | } |
491 | } |
492 | |
493 | fn convert_to_f32_slice(&self, dst: &mut [f32]) { |
494 | assert_eq!( |
495 | self.len(), |
496 | dst.len(), |
497 | "destination and source slices have different lengths" |
498 | ); |
499 | |
500 | // Just use regular loop here until there's any bf16 SIMD support. |
501 | for (i, f) in self.iter().enumerate() { |
502 | dst[i] = f.to_f32(); |
503 | } |
504 | } |
505 | |
506 | fn convert_to_f64_slice(&self, dst: &mut [f64]) { |
507 | assert_eq!( |
508 | self.len(), |
509 | dst.len(), |
510 | "destination and source slices have different lengths" |
511 | ); |
512 | |
513 | // Just use regular loop here until there's any bf16 SIMD support. |
514 | for (i, f) in self.iter().enumerate() { |
515 | dst[i] = f.to_f64(); |
516 | } |
517 | } |
518 | |
519 | #[cfg (any(feature = "alloc" , feature = "std" ))] |
520 | #[inline ] |
521 | fn to_f32_vec(&self) -> Vec<f32> { |
522 | let mut vec = Vec::with_capacity(self.len()); |
523 | // SAFETY: convert will initialize every value in the vector without reading them, |
524 | // so this is safe to do instead of double initialize from resize, and we're setting it to |
525 | // same value as capacity. |
526 | unsafe { vec.set_len(self.len()) }; |
527 | self.convert_to_f32_slice(&mut vec); |
528 | vec |
529 | } |
530 | |
531 | #[cfg (any(feature = "alloc" , feature = "std" ))] |
532 | #[inline ] |
533 | fn to_f64_vec(&self) -> Vec<f64> { |
534 | let mut vec = Vec::with_capacity(self.len()); |
535 | // SAFETY: convert will initialize every value in the vector without reading them, |
536 | // so this is safe to do instead of double initialize from resize, and we're setting it to |
537 | // same value as capacity. |
538 | unsafe { vec.set_len(self.len()) }; |
539 | self.convert_to_f64_slice(&mut vec); |
540 | vec |
541 | } |
542 | } |
543 | |
544 | impl HalfBitsSliceExt for [u16] { |
545 | // Since we sealed all the traits involved, these are safe. |
546 | #[inline ] |
547 | fn reinterpret_cast<H>(&self) -> &[H] |
548 | where |
549 | H: crate::private::SealedHalf, |
550 | { |
551 | let pointer = self.as_ptr() as *const H; |
552 | let length = self.len(); |
553 | // SAFETY: We are reconstructing full length of original slice, using its same lifetime, |
554 | // and the size of elements are identical |
555 | unsafe { slice::from_raw_parts(pointer, length) } |
556 | } |
557 | |
558 | #[inline ] |
559 | fn reinterpret_cast_mut<H>(&mut self) -> &mut [H] |
560 | where |
561 | H: crate::private::SealedHalf, |
562 | { |
563 | let pointer = self.as_mut_ptr() as *mut H; |
564 | let length = self.len(); |
565 | // SAFETY: We are reconstructing full length of original slice, using its same lifetime, |
566 | // and the size of elements are identical |
567 | unsafe { slice::from_raw_parts_mut(pointer, length) } |
568 | } |
569 | } |
570 | |
571 | #[doc (hidden)] |
572 | #[deprecated ( |
573 | since = "1.4.0" , |
574 | note = "use `HalfBitsSliceExt::reinterpret_cast_mut` instead" |
575 | )] |
576 | #[inline ] |
577 | pub fn from_bits_mut(bits: &mut [u16]) -> &mut [f16] { |
578 | bits.reinterpret_cast_mut() |
579 | } |
580 | |
581 | #[doc (hidden)] |
582 | #[deprecated ( |
583 | since = "1.4.0" , |
584 | note = "use `HalfFloatSliceExt::reinterpret_cast_mut` instead" |
585 | )] |
586 | #[inline ] |
587 | pub fn to_bits_mut(bits: &mut [f16]) -> &mut [u16] { |
588 | bits.reinterpret_cast_mut() |
589 | } |
590 | |
591 | #[doc (hidden)] |
592 | #[deprecated ( |
593 | since = "1.4.0" , |
594 | note = "use `HalfBitsSliceExt::reinterpret_cast` instead" |
595 | )] |
596 | #[inline ] |
597 | pub fn from_bits(bits: &[u16]) -> &[f16] { |
598 | bits.reinterpret_cast() |
599 | } |
600 | |
601 | #[doc (hidden)] |
602 | #[deprecated ( |
603 | since = "1.4.0" , |
604 | note = "use `HalfFloatSliceExt::reinterpret_cast` instead" |
605 | )] |
606 | #[inline ] |
607 | pub fn to_bits(bits: &[f16]) -> &[u16] { |
608 | bits.reinterpret_cast() |
609 | } |
610 | |
611 | #[allow (clippy::float_cmp)] |
612 | #[cfg (test)] |
613 | mod test { |
614 | use super::{HalfBitsSliceExt, HalfFloatSliceExt}; |
615 | use crate::{bf16, f16}; |
616 | |
617 | #[test] |
618 | fn test_slice_conversions_f16() { |
619 | let bits = &[ |
620 | f16::E.to_bits(), |
621 | f16::PI.to_bits(), |
622 | f16::EPSILON.to_bits(), |
623 | f16::FRAC_1_SQRT_2.to_bits(), |
624 | ]; |
625 | let numbers = &[f16::E, f16::PI, f16::EPSILON, f16::FRAC_1_SQRT_2]; |
626 | |
627 | // Convert from bits to numbers |
628 | let from_bits = bits.reinterpret_cast::<f16>(); |
629 | assert_eq!(from_bits, numbers); |
630 | |
631 | // Convert from numbers back to bits |
632 | let to_bits = from_bits.reinterpret_cast(); |
633 | assert_eq!(to_bits, bits); |
634 | } |
635 | |
636 | #[test] |
637 | fn test_mutablility_f16() { |
638 | let mut bits_array = [f16::PI.to_bits()]; |
639 | let bits = &mut bits_array[..]; |
640 | |
641 | { |
642 | // would not compile without these braces |
643 | let numbers = bits.reinterpret_cast_mut(); |
644 | numbers[0] = f16::E; |
645 | } |
646 | |
647 | assert_eq!(bits, &[f16::E.to_bits()]); |
648 | |
649 | bits[0] = f16::LN_2.to_bits(); |
650 | assert_eq!(bits, &[f16::LN_2.to_bits()]); |
651 | } |
652 | |
653 | #[test] |
654 | fn test_slice_conversions_bf16() { |
655 | let bits = &[ |
656 | bf16::E.to_bits(), |
657 | bf16::PI.to_bits(), |
658 | bf16::EPSILON.to_bits(), |
659 | bf16::FRAC_1_SQRT_2.to_bits(), |
660 | ]; |
661 | let numbers = &[bf16::E, bf16::PI, bf16::EPSILON, bf16::FRAC_1_SQRT_2]; |
662 | |
663 | // Convert from bits to numbers |
664 | let from_bits = bits.reinterpret_cast::<bf16>(); |
665 | assert_eq!(from_bits, numbers); |
666 | |
667 | // Convert from numbers back to bits |
668 | let to_bits = from_bits.reinterpret_cast(); |
669 | assert_eq!(to_bits, bits); |
670 | } |
671 | |
672 | #[test] |
673 | fn test_mutablility_bf16() { |
674 | let mut bits_array = [bf16::PI.to_bits()]; |
675 | let bits = &mut bits_array[..]; |
676 | |
677 | { |
678 | // would not compile without these braces |
679 | let numbers = bits.reinterpret_cast_mut(); |
680 | numbers[0] = bf16::E; |
681 | } |
682 | |
683 | assert_eq!(bits, &[bf16::E.to_bits()]); |
684 | |
685 | bits[0] = bf16::LN_2.to_bits(); |
686 | assert_eq!(bits, &[bf16::LN_2.to_bits()]); |
687 | } |
688 | |
689 | #[test] |
690 | fn slice_convert_f16_f32() { |
691 | // Exact chunks |
692 | let vf32 = [1., 2., 3., 4., 5., 6., 7., 8.]; |
693 | let vf16 = [ |
694 | f16::from_f32(1.), |
695 | f16::from_f32(2.), |
696 | f16::from_f32(3.), |
697 | f16::from_f32(4.), |
698 | f16::from_f32(5.), |
699 | f16::from_f32(6.), |
700 | f16::from_f32(7.), |
701 | f16::from_f32(8.), |
702 | ]; |
703 | let mut buf32 = vf32; |
704 | let mut buf16 = vf16; |
705 | |
706 | vf16.convert_to_f32_slice(&mut buf32); |
707 | assert_eq!(&vf32, &buf32); |
708 | |
709 | buf16.convert_from_f32_slice(&vf32); |
710 | assert_eq!(&vf16, &buf16); |
711 | |
712 | // Partial with chunks |
713 | let vf32 = [1., 2., 3., 4., 5., 6., 7., 8., 9.]; |
714 | let vf16 = [ |
715 | f16::from_f32(1.), |
716 | f16::from_f32(2.), |
717 | f16::from_f32(3.), |
718 | f16::from_f32(4.), |
719 | f16::from_f32(5.), |
720 | f16::from_f32(6.), |
721 | f16::from_f32(7.), |
722 | f16::from_f32(8.), |
723 | f16::from_f32(9.), |
724 | ]; |
725 | let mut buf32 = vf32; |
726 | let mut buf16 = vf16; |
727 | |
728 | vf16.convert_to_f32_slice(&mut buf32); |
729 | assert_eq!(&vf32, &buf32); |
730 | |
731 | buf16.convert_from_f32_slice(&vf32); |
732 | assert_eq!(&vf16, &buf16); |
733 | |
734 | // Partial with chunks |
735 | let vf32 = [1., 2.]; |
736 | let vf16 = [f16::from_f32(1.), f16::from_f32(2.)]; |
737 | let mut buf32 = vf32; |
738 | let mut buf16 = vf16; |
739 | |
740 | vf16.convert_to_f32_slice(&mut buf32); |
741 | assert_eq!(&vf32, &buf32); |
742 | |
743 | buf16.convert_from_f32_slice(&vf32); |
744 | assert_eq!(&vf16, &buf16); |
745 | } |
746 | |
747 | #[test] |
748 | fn slice_convert_bf16_f32() { |
749 | // Exact chunks |
750 | let vf32 = [1., 2., 3., 4., 5., 6., 7., 8.]; |
751 | let vf16 = [ |
752 | bf16::from_f32(1.), |
753 | bf16::from_f32(2.), |
754 | bf16::from_f32(3.), |
755 | bf16::from_f32(4.), |
756 | bf16::from_f32(5.), |
757 | bf16::from_f32(6.), |
758 | bf16::from_f32(7.), |
759 | bf16::from_f32(8.), |
760 | ]; |
761 | let mut buf32 = vf32; |
762 | let mut buf16 = vf16; |
763 | |
764 | vf16.convert_to_f32_slice(&mut buf32); |
765 | assert_eq!(&vf32, &buf32); |
766 | |
767 | buf16.convert_from_f32_slice(&vf32); |
768 | assert_eq!(&vf16, &buf16); |
769 | |
770 | // Partial with chunks |
771 | let vf32 = [1., 2., 3., 4., 5., 6., 7., 8., 9.]; |
772 | let vf16 = [ |
773 | bf16::from_f32(1.), |
774 | bf16::from_f32(2.), |
775 | bf16::from_f32(3.), |
776 | bf16::from_f32(4.), |
777 | bf16::from_f32(5.), |
778 | bf16::from_f32(6.), |
779 | bf16::from_f32(7.), |
780 | bf16::from_f32(8.), |
781 | bf16::from_f32(9.), |
782 | ]; |
783 | let mut buf32 = vf32; |
784 | let mut buf16 = vf16; |
785 | |
786 | vf16.convert_to_f32_slice(&mut buf32); |
787 | assert_eq!(&vf32, &buf32); |
788 | |
789 | buf16.convert_from_f32_slice(&vf32); |
790 | assert_eq!(&vf16, &buf16); |
791 | |
792 | // Partial with chunks |
793 | let vf32 = [1., 2.]; |
794 | let vf16 = [bf16::from_f32(1.), bf16::from_f32(2.)]; |
795 | let mut buf32 = vf32; |
796 | let mut buf16 = vf16; |
797 | |
798 | vf16.convert_to_f32_slice(&mut buf32); |
799 | assert_eq!(&vf32, &buf32); |
800 | |
801 | buf16.convert_from_f32_slice(&vf32); |
802 | assert_eq!(&vf16, &buf16); |
803 | } |
804 | |
805 | #[test] |
806 | fn slice_convert_f16_f64() { |
807 | // Exact chunks |
808 | let vf64 = [1., 2., 3., 4., 5., 6., 7., 8.]; |
809 | let vf16 = [ |
810 | f16::from_f64(1.), |
811 | f16::from_f64(2.), |
812 | f16::from_f64(3.), |
813 | f16::from_f64(4.), |
814 | f16::from_f64(5.), |
815 | f16::from_f64(6.), |
816 | f16::from_f64(7.), |
817 | f16::from_f64(8.), |
818 | ]; |
819 | let mut buf64 = vf64; |
820 | let mut buf16 = vf16; |
821 | |
822 | vf16.convert_to_f64_slice(&mut buf64); |
823 | assert_eq!(&vf64, &buf64); |
824 | |
825 | buf16.convert_from_f64_slice(&vf64); |
826 | assert_eq!(&vf16, &buf16); |
827 | |
828 | // Partial with chunks |
829 | let vf64 = [1., 2., 3., 4., 5., 6., 7., 8., 9.]; |
830 | let vf16 = [ |
831 | f16::from_f64(1.), |
832 | f16::from_f64(2.), |
833 | f16::from_f64(3.), |
834 | f16::from_f64(4.), |
835 | f16::from_f64(5.), |
836 | f16::from_f64(6.), |
837 | f16::from_f64(7.), |
838 | f16::from_f64(8.), |
839 | f16::from_f64(9.), |
840 | ]; |
841 | let mut buf64 = vf64; |
842 | let mut buf16 = vf16; |
843 | |
844 | vf16.convert_to_f64_slice(&mut buf64); |
845 | assert_eq!(&vf64, &buf64); |
846 | |
847 | buf16.convert_from_f64_slice(&vf64); |
848 | assert_eq!(&vf16, &buf16); |
849 | |
850 | // Partial with chunks |
851 | let vf64 = [1., 2.]; |
852 | let vf16 = [f16::from_f64(1.), f16::from_f64(2.)]; |
853 | let mut buf64 = vf64; |
854 | let mut buf16 = vf16; |
855 | |
856 | vf16.convert_to_f64_slice(&mut buf64); |
857 | assert_eq!(&vf64, &buf64); |
858 | |
859 | buf16.convert_from_f64_slice(&vf64); |
860 | assert_eq!(&vf16, &buf16); |
861 | } |
862 | |
863 | #[test] |
864 | fn slice_convert_bf16_f64() { |
865 | // Exact chunks |
866 | let vf64 = [1., 2., 3., 4., 5., 6., 7., 8.]; |
867 | let vf16 = [ |
868 | bf16::from_f64(1.), |
869 | bf16::from_f64(2.), |
870 | bf16::from_f64(3.), |
871 | bf16::from_f64(4.), |
872 | bf16::from_f64(5.), |
873 | bf16::from_f64(6.), |
874 | bf16::from_f64(7.), |
875 | bf16::from_f64(8.), |
876 | ]; |
877 | let mut buf64 = vf64; |
878 | let mut buf16 = vf16; |
879 | |
880 | vf16.convert_to_f64_slice(&mut buf64); |
881 | assert_eq!(&vf64, &buf64); |
882 | |
883 | buf16.convert_from_f64_slice(&vf64); |
884 | assert_eq!(&vf16, &buf16); |
885 | |
886 | // Partial with chunks |
887 | let vf64 = [1., 2., 3., 4., 5., 6., 7., 8., 9.]; |
888 | let vf16 = [ |
889 | bf16::from_f64(1.), |
890 | bf16::from_f64(2.), |
891 | bf16::from_f64(3.), |
892 | bf16::from_f64(4.), |
893 | bf16::from_f64(5.), |
894 | bf16::from_f64(6.), |
895 | bf16::from_f64(7.), |
896 | bf16::from_f64(8.), |
897 | bf16::from_f64(9.), |
898 | ]; |
899 | let mut buf64 = vf64; |
900 | let mut buf16 = vf16; |
901 | |
902 | vf16.convert_to_f64_slice(&mut buf64); |
903 | assert_eq!(&vf64, &buf64); |
904 | |
905 | buf16.convert_from_f64_slice(&vf64); |
906 | assert_eq!(&vf16, &buf16); |
907 | |
908 | // Partial with chunks |
909 | let vf64 = [1., 2.]; |
910 | let vf16 = [bf16::from_f64(1.), bf16::from_f64(2.)]; |
911 | let mut buf64 = vf64; |
912 | let mut buf16 = vf16; |
913 | |
914 | vf16.convert_to_f64_slice(&mut buf64); |
915 | assert_eq!(&vf64, &buf64); |
916 | |
917 | buf16.convert_from_f64_slice(&vf64); |
918 | assert_eq!(&vf16, &buf16); |
919 | } |
920 | |
921 | #[test] |
922 | #[should_panic ] |
923 | fn convert_from_f32_slice_len_mismatch_panics() { |
924 | let mut slice1 = [f16::ZERO; 3]; |
925 | let slice2 = [0f32; 4]; |
926 | slice1.convert_from_f32_slice(&slice2); |
927 | } |
928 | |
929 | #[test] |
930 | #[should_panic ] |
931 | fn convert_from_f64_slice_len_mismatch_panics() { |
932 | let mut slice1 = [f16::ZERO; 3]; |
933 | let slice2 = [0f64; 4]; |
934 | slice1.convert_from_f64_slice(&slice2); |
935 | } |
936 | |
937 | #[test] |
938 | #[should_panic ] |
939 | fn convert_to_f32_slice_len_mismatch_panics() { |
940 | let slice1 = [f16::ZERO; 3]; |
941 | let mut slice2 = [0f32; 4]; |
942 | slice1.convert_to_f32_slice(&mut slice2); |
943 | } |
944 | |
945 | #[test] |
946 | #[should_panic ] |
947 | fn convert_to_f64_slice_len_mismatch_panics() { |
948 | let slice1 = [f16::ZERO; 3]; |
949 | let mut slice2 = [0f64; 4]; |
950 | slice1.convert_to_f64_slice(&mut slice2); |
951 | } |
952 | } |
953 | |