/rust/registry/src/index.crates.io-1949cf8c6b5b557f/half-2.7.0/src/slice.rs
Line | Count | Source |
1 | | //! Contains utility functions and traits to convert between slices of [`u16`] bits and [`struct@f16`] or |
2 | | //! [`struct@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 zerocopy::{transmute_mut, transmute_ref}; |
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 [`struct@f16`] or [`struct@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 [`struct@f16`] or [`struct@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 [`struct@f16`] or [`struct@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 [`struct@f16`] or [`struct@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 [`struct@f16`] or [`struct@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 [`struct@f16`] or [`struct@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 [`struct@f16`] or [`struct@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 [`struct@f16`] or [`struct@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 [`struct@f16`] or [`struct@bf16`] numbers. |
230 | | /// |
231 | | /// `H` is the type to cast to, and must be either the [`struct@f16`] or [`struct@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 [`struct@f16`] or [`struct@bf16`] |
255 | | /// numbers. |
256 | | /// |
257 | | /// `H` is the type to cast to, and must be either the [`struct@f16`] or [`struct@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 | 0 | fn reinterpret_cast(&self) -> &[u16] { |
304 | 0 | transmute_ref!(self) |
305 | 0 | } Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast |
306 | | |
307 | | #[inline] |
308 | 0 | fn reinterpret_cast_mut(&mut self) -> &mut [u16] { |
309 | 0 | transmute_mut!(self) |
310 | 0 | } Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::reinterpret_cast_mut |
311 | | |
312 | | #[inline] |
313 | 0 | fn convert_from_f32_slice(&mut self, src: &[f32]) { |
314 | 0 | assert_eq!( |
315 | 0 | self.len(), |
316 | 0 | src.len(), |
317 | 0 | "destination and source slices have different lengths" |
318 | | ); |
319 | | |
320 | 0 | arch::f32_to_f16_slice(src, self.reinterpret_cast_mut()) |
321 | 0 | } |
322 | | |
323 | | #[inline] |
324 | 0 | fn convert_from_f64_slice(&mut self, src: &[f64]) { |
325 | 0 | assert_eq!( |
326 | 0 | self.len(), |
327 | 0 | src.len(), |
328 | 0 | "destination and source slices have different lengths" |
329 | | ); |
330 | | |
331 | 0 | arch::f64_to_f16_slice(src, self.reinterpret_cast_mut()) |
332 | 0 | } |
333 | | |
334 | | #[inline] |
335 | 0 | fn convert_to_f32_slice(&self, dst: &mut [f32]) { |
336 | 0 | assert_eq!( |
337 | 0 | self.len(), |
338 | 0 | dst.len(), |
339 | 0 | "destination and source slices have different lengths" |
340 | | ); |
341 | | |
342 | 0 | arch::f16_to_f32_slice(self.reinterpret_cast(), dst) |
343 | 0 | } Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice Unexecuted instantiation: <[half::binary16::f16] as half::slice::HalfFloatSliceExt>::convert_to_f32_slice |
344 | | |
345 | | #[inline] |
346 | 0 | fn convert_to_f64_slice(&self, dst: &mut [f64]) { |
347 | 0 | assert_eq!( |
348 | 0 | self.len(), |
349 | 0 | dst.len(), |
350 | 0 | "destination and source slices have different lengths" |
351 | | ); |
352 | | |
353 | 0 | arch::f16_to_f64_slice(self.reinterpret_cast(), dst) |
354 | 0 | } |
355 | | |
356 | | #[cfg(any(feature = "alloc", feature = "std"))] |
357 | | #[inline] |
358 | | #[allow(clippy::uninit_vec)] |
359 | 0 | fn to_f32_vec(&self) -> Vec<f32> { |
360 | 0 | let mut vec = vec![0f32; self.len()]; |
361 | 0 | self.convert_to_f32_slice(&mut vec); |
362 | 0 | vec |
363 | 0 | } |
364 | | |
365 | | #[cfg(any(feature = "alloc", feature = "std"))] |
366 | | #[inline] |
367 | | #[allow(clippy::uninit_vec)] |
368 | 0 | fn to_f64_vec(&self) -> Vec<f64> { |
369 | 0 | let mut vec = vec![0f64; self.len()]; |
370 | 0 | self.convert_to_f64_slice(&mut vec); |
371 | 0 | vec |
372 | 0 | } |
373 | | } |
374 | | |
375 | | impl HalfFloatSliceExt for [bf16] { |
376 | | #[inline] |
377 | 0 | fn reinterpret_cast(&self) -> &[u16] { |
378 | 0 | transmute_ref!(self) |
379 | 0 | } |
380 | | |
381 | | #[inline] |
382 | 0 | fn reinterpret_cast_mut(&mut self) -> &mut [u16] { |
383 | 0 | transmute_mut!(self) |
384 | 0 | } |
385 | | |
386 | | #[inline] |
387 | 0 | fn convert_from_f32_slice(&mut self, src: &[f32]) { |
388 | 0 | assert_eq!( |
389 | 0 | self.len(), |
390 | 0 | src.len(), |
391 | 0 | "destination and source slices have different lengths" |
392 | | ); |
393 | | |
394 | | // Just use regular loop here until there's any bf16 SIMD support. |
395 | 0 | for (i, f) in src.iter().enumerate() { |
396 | 0 | self[i] = bf16::from_f32(*f); |
397 | 0 | } |
398 | 0 | } |
399 | | |
400 | | #[inline] |
401 | 0 | fn convert_from_f64_slice(&mut self, src: &[f64]) { |
402 | 0 | assert_eq!( |
403 | 0 | self.len(), |
404 | 0 | src.len(), |
405 | 0 | "destination and source slices have different lengths" |
406 | | ); |
407 | | |
408 | | // Just use regular loop here until there's any bf16 SIMD support. |
409 | 0 | for (i, f) in src.iter().enumerate() { |
410 | 0 | self[i] = bf16::from_f64(*f); |
411 | 0 | } |
412 | 0 | } |
413 | | |
414 | | #[inline] |
415 | 0 | fn convert_to_f32_slice(&self, dst: &mut [f32]) { |
416 | 0 | assert_eq!( |
417 | 0 | self.len(), |
418 | 0 | dst.len(), |
419 | 0 | "destination and source slices have different lengths" |
420 | | ); |
421 | | |
422 | | // Just use regular loop here until there's any bf16 SIMD support. |
423 | 0 | for (i, f) in self.iter().enumerate() { |
424 | 0 | dst[i] = f.to_f32(); |
425 | 0 | } |
426 | 0 | } |
427 | | |
428 | | #[inline] |
429 | 0 | fn convert_to_f64_slice(&self, dst: &mut [f64]) { |
430 | 0 | assert_eq!( |
431 | 0 | self.len(), |
432 | 0 | dst.len(), |
433 | 0 | "destination and source slices have different lengths" |
434 | | ); |
435 | | |
436 | | // Just use regular loop here until there's any bf16 SIMD support. |
437 | 0 | for (i, f) in self.iter().enumerate() { |
438 | 0 | dst[i] = f.to_f64(); |
439 | 0 | } |
440 | 0 | } |
441 | | |
442 | | #[cfg(any(feature = "alloc", feature = "std"))] |
443 | | #[inline] |
444 | | #[allow(clippy::uninit_vec)] |
445 | 0 | fn to_f32_vec(&self) -> Vec<f32> { |
446 | 0 | let mut vec = vec![0f32; self.len()]; |
447 | 0 | self.convert_to_f32_slice(&mut vec); |
448 | 0 | vec |
449 | 0 | } |
450 | | |
451 | | #[cfg(any(feature = "alloc", feature = "std"))] |
452 | | #[inline] |
453 | | #[allow(clippy::uninit_vec)] |
454 | 0 | fn to_f64_vec(&self) -> Vec<f64> { |
455 | 0 | let mut vec = vec![0f64; self.len()]; |
456 | 0 | self.convert_to_f64_slice(&mut vec); |
457 | 0 | vec |
458 | 0 | } |
459 | | } |
460 | | |
461 | | impl HalfBitsSliceExt for [u16] { |
462 | | // Since we sealed all the traits involved, these are safe. |
463 | | #[inline] |
464 | 0 | fn reinterpret_cast<H>(&self) -> &[H] |
465 | 0 | where |
466 | 0 | H: crate::private::SealedHalf, |
467 | | { |
468 | 0 | transmute_ref!(self) |
469 | 0 | } |
470 | | |
471 | | #[inline] |
472 | 0 | fn reinterpret_cast_mut<H>(&mut self) -> &mut [H] |
473 | 0 | where |
474 | 0 | H: crate::private::SealedHalf, |
475 | | { |
476 | 0 | transmute_mut!(self) |
477 | 0 | } |
478 | | } |
479 | | |
480 | | #[allow(clippy::float_cmp)] |
481 | | #[cfg(test)] |
482 | | mod test { |
483 | | use super::{HalfBitsSliceExt, HalfFloatSliceExt}; |
484 | | use crate::{bf16, f16}; |
485 | | |
486 | | #[test] |
487 | | fn test_slice_conversions_f16() { |
488 | | let bits = &[ |
489 | | f16::E.to_bits(), |
490 | | f16::PI.to_bits(), |
491 | | f16::EPSILON.to_bits(), |
492 | | f16::FRAC_1_SQRT_2.to_bits(), |
493 | | ]; |
494 | | let numbers = &[f16::E, f16::PI, f16::EPSILON, f16::FRAC_1_SQRT_2]; |
495 | | |
496 | | // Convert from bits to numbers |
497 | | let from_bits = bits.reinterpret_cast::<f16>(); |
498 | | assert_eq!(from_bits, numbers); |
499 | | |
500 | | // Convert from numbers back to bits |
501 | | let to_bits = from_bits.reinterpret_cast(); |
502 | | assert_eq!(to_bits, bits); |
503 | | } |
504 | | |
505 | | #[test] |
506 | | fn test_mutablility_f16() { |
507 | | let mut bits_array = [f16::PI.to_bits()]; |
508 | | let bits = &mut bits_array[..]; |
509 | | |
510 | | { |
511 | | // would not compile without these braces |
512 | | let numbers = bits.reinterpret_cast_mut(); |
513 | | numbers[0] = f16::E; |
514 | | } |
515 | | |
516 | | assert_eq!(bits, &[f16::E.to_bits()]); |
517 | | |
518 | | bits[0] = f16::LN_2.to_bits(); |
519 | | assert_eq!(bits, &[f16::LN_2.to_bits()]); |
520 | | } |
521 | | |
522 | | #[test] |
523 | | fn test_slice_conversions_bf16() { |
524 | | let bits = &[ |
525 | | bf16::E.to_bits(), |
526 | | bf16::PI.to_bits(), |
527 | | bf16::EPSILON.to_bits(), |
528 | | bf16::FRAC_1_SQRT_2.to_bits(), |
529 | | ]; |
530 | | let numbers = &[bf16::E, bf16::PI, bf16::EPSILON, bf16::FRAC_1_SQRT_2]; |
531 | | |
532 | | // Convert from bits to numbers |
533 | | let from_bits = bits.reinterpret_cast::<bf16>(); |
534 | | assert_eq!(from_bits, numbers); |
535 | | |
536 | | // Convert from numbers back to bits |
537 | | let to_bits = from_bits.reinterpret_cast(); |
538 | | assert_eq!(to_bits, bits); |
539 | | } |
540 | | |
541 | | #[test] |
542 | | fn test_mutablility_bf16() { |
543 | | let mut bits_array = [bf16::PI.to_bits()]; |
544 | | let bits = &mut bits_array[..]; |
545 | | |
546 | | { |
547 | | // would not compile without these braces |
548 | | let numbers = bits.reinterpret_cast_mut(); |
549 | | numbers[0] = bf16::E; |
550 | | } |
551 | | |
552 | | assert_eq!(bits, &[bf16::E.to_bits()]); |
553 | | |
554 | | bits[0] = bf16::LN_2.to_bits(); |
555 | | assert_eq!(bits, &[bf16::LN_2.to_bits()]); |
556 | | } |
557 | | |
558 | | #[test] |
559 | | fn slice_convert_f16_f32() { |
560 | | // Exact chunks |
561 | | let vf32 = [1., 2., 3., 4., 5., 6., 7., 8.]; |
562 | | let vf16 = [ |
563 | | f16::from_f32(1.), |
564 | | f16::from_f32(2.), |
565 | | f16::from_f32(3.), |
566 | | f16::from_f32(4.), |
567 | | f16::from_f32(5.), |
568 | | f16::from_f32(6.), |
569 | | f16::from_f32(7.), |
570 | | f16::from_f32(8.), |
571 | | ]; |
572 | | let mut buf32 = vf32; |
573 | | let mut buf16 = vf16; |
574 | | |
575 | | vf16.convert_to_f32_slice(&mut buf32); |
576 | | assert_eq!(&vf32, &buf32); |
577 | | |
578 | | buf16.convert_from_f32_slice(&vf32); |
579 | | assert_eq!(&vf16, &buf16); |
580 | | |
581 | | // Partial with chunks |
582 | | let vf32 = [1., 2., 3., 4., 5., 6., 7., 8., 9.]; |
583 | | let vf16 = [ |
584 | | f16::from_f32(1.), |
585 | | f16::from_f32(2.), |
586 | | f16::from_f32(3.), |
587 | | f16::from_f32(4.), |
588 | | f16::from_f32(5.), |
589 | | f16::from_f32(6.), |
590 | | f16::from_f32(7.), |
591 | | f16::from_f32(8.), |
592 | | f16::from_f32(9.), |
593 | | ]; |
594 | | let mut buf32 = vf32; |
595 | | let mut buf16 = vf16; |
596 | | |
597 | | vf16.convert_to_f32_slice(&mut buf32); |
598 | | assert_eq!(&vf32, &buf32); |
599 | | |
600 | | buf16.convert_from_f32_slice(&vf32); |
601 | | assert_eq!(&vf16, &buf16); |
602 | | |
603 | | // Partial with chunks |
604 | | let vf32 = [1., 2.]; |
605 | | let vf16 = [f16::from_f32(1.), f16::from_f32(2.)]; |
606 | | let mut buf32 = vf32; |
607 | | let mut buf16 = vf16; |
608 | | |
609 | | vf16.convert_to_f32_slice(&mut buf32); |
610 | | assert_eq!(&vf32, &buf32); |
611 | | |
612 | | buf16.convert_from_f32_slice(&vf32); |
613 | | assert_eq!(&vf16, &buf16); |
614 | | } |
615 | | |
616 | | #[test] |
617 | | fn slice_convert_bf16_f32() { |
618 | | // Exact chunks |
619 | | let vf32 = [1., 2., 3., 4., 5., 6., 7., 8.]; |
620 | | let vf16 = [ |
621 | | bf16::from_f32(1.), |
622 | | bf16::from_f32(2.), |
623 | | bf16::from_f32(3.), |
624 | | bf16::from_f32(4.), |
625 | | bf16::from_f32(5.), |
626 | | bf16::from_f32(6.), |
627 | | bf16::from_f32(7.), |
628 | | bf16::from_f32(8.), |
629 | | ]; |
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 | | // Partial with chunks |
640 | | let vf32 = [1., 2., 3., 4., 5., 6., 7., 8., 9.]; |
641 | | let vf16 = [ |
642 | | bf16::from_f32(1.), |
643 | | bf16::from_f32(2.), |
644 | | bf16::from_f32(3.), |
645 | | bf16::from_f32(4.), |
646 | | bf16::from_f32(5.), |
647 | | bf16::from_f32(6.), |
648 | | bf16::from_f32(7.), |
649 | | bf16::from_f32(8.), |
650 | | bf16::from_f32(9.), |
651 | | ]; |
652 | | let mut buf32 = vf32; |
653 | | let mut buf16 = vf16; |
654 | | |
655 | | vf16.convert_to_f32_slice(&mut buf32); |
656 | | assert_eq!(&vf32, &buf32); |
657 | | |
658 | | buf16.convert_from_f32_slice(&vf32); |
659 | | assert_eq!(&vf16, &buf16); |
660 | | |
661 | | // Partial with chunks |
662 | | let vf32 = [1., 2.]; |
663 | | let vf16 = [bf16::from_f32(1.), bf16::from_f32(2.)]; |
664 | | let mut buf32 = vf32; |
665 | | let mut buf16 = vf16; |
666 | | |
667 | | vf16.convert_to_f32_slice(&mut buf32); |
668 | | assert_eq!(&vf32, &buf32); |
669 | | |
670 | | buf16.convert_from_f32_slice(&vf32); |
671 | | assert_eq!(&vf16, &buf16); |
672 | | } |
673 | | |
674 | | #[test] |
675 | | fn slice_convert_f16_f64() { |
676 | | // Exact chunks |
677 | | let vf64 = [1., 2., 3., 4., 5., 6., 7., 8.]; |
678 | | let vf16 = [ |
679 | | f16::from_f64(1.), |
680 | | f16::from_f64(2.), |
681 | | f16::from_f64(3.), |
682 | | f16::from_f64(4.), |
683 | | f16::from_f64(5.), |
684 | | f16::from_f64(6.), |
685 | | f16::from_f64(7.), |
686 | | f16::from_f64(8.), |
687 | | ]; |
688 | | let mut buf64 = vf64; |
689 | | let mut buf16 = vf16; |
690 | | |
691 | | vf16.convert_to_f64_slice(&mut buf64); |
692 | | assert_eq!(&vf64, &buf64); |
693 | | |
694 | | buf16.convert_from_f64_slice(&vf64); |
695 | | assert_eq!(&vf16, &buf16); |
696 | | |
697 | | // Partial with chunks |
698 | | let vf64 = [1., 2., 3., 4., 5., 6., 7., 8., 9.]; |
699 | | let vf16 = [ |
700 | | f16::from_f64(1.), |
701 | | f16::from_f64(2.), |
702 | | f16::from_f64(3.), |
703 | | f16::from_f64(4.), |
704 | | f16::from_f64(5.), |
705 | | f16::from_f64(6.), |
706 | | f16::from_f64(7.), |
707 | | f16::from_f64(8.), |
708 | | f16::from_f64(9.), |
709 | | ]; |
710 | | let mut buf64 = vf64; |
711 | | let mut buf16 = vf16; |
712 | | |
713 | | vf16.convert_to_f64_slice(&mut buf64); |
714 | | assert_eq!(&vf64, &buf64); |
715 | | |
716 | | buf16.convert_from_f64_slice(&vf64); |
717 | | assert_eq!(&vf16, &buf16); |
718 | | |
719 | | // Partial with chunks |
720 | | let vf64 = [1., 2.]; |
721 | | let vf16 = [f16::from_f64(1.), f16::from_f64(2.)]; |
722 | | let mut buf64 = vf64; |
723 | | let mut buf16 = vf16; |
724 | | |
725 | | vf16.convert_to_f64_slice(&mut buf64); |
726 | | assert_eq!(&vf64, &buf64); |
727 | | |
728 | | buf16.convert_from_f64_slice(&vf64); |
729 | | assert_eq!(&vf16, &buf16); |
730 | | } |
731 | | |
732 | | #[test] |
733 | | fn slice_convert_bf16_f64() { |
734 | | // Exact chunks |
735 | | let vf64 = [1., 2., 3., 4., 5., 6., 7., 8.]; |
736 | | let vf16 = [ |
737 | | bf16::from_f64(1.), |
738 | | bf16::from_f64(2.), |
739 | | bf16::from_f64(3.), |
740 | | bf16::from_f64(4.), |
741 | | bf16::from_f64(5.), |
742 | | bf16::from_f64(6.), |
743 | | bf16::from_f64(7.), |
744 | | bf16::from_f64(8.), |
745 | | ]; |
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 | | // Partial with chunks |
756 | | let vf64 = [1., 2., 3., 4., 5., 6., 7., 8., 9.]; |
757 | | let vf16 = [ |
758 | | bf16::from_f64(1.), |
759 | | bf16::from_f64(2.), |
760 | | bf16::from_f64(3.), |
761 | | bf16::from_f64(4.), |
762 | | bf16::from_f64(5.), |
763 | | bf16::from_f64(6.), |
764 | | bf16::from_f64(7.), |
765 | | bf16::from_f64(8.), |
766 | | bf16::from_f64(9.), |
767 | | ]; |
768 | | let mut buf64 = vf64; |
769 | | let mut buf16 = vf16; |
770 | | |
771 | | vf16.convert_to_f64_slice(&mut buf64); |
772 | | assert_eq!(&vf64, &buf64); |
773 | | |
774 | | buf16.convert_from_f64_slice(&vf64); |
775 | | assert_eq!(&vf16, &buf16); |
776 | | |
777 | | // Partial with chunks |
778 | | let vf64 = [1., 2.]; |
779 | | let vf16 = [bf16::from_f64(1.), bf16::from_f64(2.)]; |
780 | | let mut buf64 = vf64; |
781 | | let mut buf16 = vf16; |
782 | | |
783 | | vf16.convert_to_f64_slice(&mut buf64); |
784 | | assert_eq!(&vf64, &buf64); |
785 | | |
786 | | buf16.convert_from_f64_slice(&vf64); |
787 | | assert_eq!(&vf16, &buf16); |
788 | | } |
789 | | |
790 | | #[test] |
791 | | #[should_panic] |
792 | | fn convert_from_f32_slice_len_mismatch_panics() { |
793 | | let mut slice1 = [f16::ZERO; 3]; |
794 | | let slice2 = [0f32; 4]; |
795 | | slice1.convert_from_f32_slice(&slice2); |
796 | | } |
797 | | |
798 | | #[test] |
799 | | #[should_panic] |
800 | | fn convert_from_f64_slice_len_mismatch_panics() { |
801 | | let mut slice1 = [f16::ZERO; 3]; |
802 | | let slice2 = [0f64; 4]; |
803 | | slice1.convert_from_f64_slice(&slice2); |
804 | | } |
805 | | |
806 | | #[test] |
807 | | #[should_panic] |
808 | | fn convert_to_f32_slice_len_mismatch_panics() { |
809 | | let slice1 = [f16::ZERO; 3]; |
810 | | let mut slice2 = [0f32; 4]; |
811 | | slice1.convert_to_f32_slice(&mut slice2); |
812 | | } |
813 | | |
814 | | #[test] |
815 | | #[should_panic] |
816 | | fn convert_to_f64_slice_len_mismatch_panics() { |
817 | | let slice1 = [f16::ZERO; 3]; |
818 | | let mut slice2 = [0f64; 4]; |
819 | | slice1.convert_to_f64_slice(&mut slice2); |
820 | | } |
821 | | } |