1//! Types and traits associated with masking elements of vectors.
2//! Types representing
3#![allow(non_camel_case_types)]
4
5#[cfg_attr(
6 not(all(target_arch = "x86_64", target_feature = "avx512f")),
7 path = "masks/full_masks.rs"
8)]
9#[cfg_attr(
10 all(target_arch = "x86_64", target_feature = "avx512f"),
11 path = "masks/bitmask.rs"
12)]
13mod mask_impl;
14
15use crate::simd::{LaneCount, Simd, SimdCast, SimdElement, SupportedLaneCount};
16use core::cmp::Ordering;
17use core::{fmt, mem};
18
19mod sealed {
20 use super::*;
21
22 /// Not only does this seal the `MaskElement` trait, but these functions prevent other traits
23 /// from bleeding into the parent bounds.
24 ///
25 /// For example, `eq` could be provided by requiring `MaskElement: PartialEq`, but that would
26 /// prevent us from ever removing that bound, or from implementing `MaskElement` on
27 /// non-`PartialEq` types in the future.
28 pub trait Sealed {
29 fn valid<const N: usize>(values: Simd<Self, N>) -> bool
30 where
31 LaneCount<N>: SupportedLaneCount,
32 Self: SimdElement;
33
34 fn eq(self, other: Self) -> bool;
35
36 fn to_usize(self) -> usize;
37 fn max_unsigned() -> u64;
38
39 type Unsigned: SimdElement;
40
41 const TRUE: Self;
42
43 const FALSE: Self;
44 }
45}
46use sealed::Sealed;
47
48/// Marker trait for types that may be used as SIMD mask elements.
49///
50/// # Safety
51/// Type must be a signed integer.
52pub unsafe trait MaskElement: SimdElement<Mask = Self> + SimdCast + Sealed {}
53
54macro_rules! impl_element {
55 { $ty:ty, $unsigned:ty } => {
56 impl Sealed for $ty {
57 #[inline]
58 fn valid<const N: usize>(value: Simd<Self, N>) -> bool
59 where
60 LaneCount<N>: SupportedLaneCount,
61 {
62 // We can't use `Simd` directly, because `Simd`'s functions call this function and
63 // we will end up with an infinite loop.
64 // Safety: `value` is an integer vector
65 unsafe {
66 use core::intrinsics::simd;
67 let falses: Simd<Self, N> = simd::simd_eq(value, Simd::splat(0 as _));
68 let trues: Simd<Self, N> = simd::simd_eq(value, Simd::splat(-1 as _));
69 let valid: Simd<Self, N> = simd::simd_or(falses, trues);
70 simd::simd_reduce_all(valid)
71 }
72 }
73
74 #[inline]
75 fn eq(self, other: Self) -> bool { self == other }
76
77 #[inline]
78 fn to_usize(self) -> usize {
79 self as usize
80 }
81
82 #[inline]
83 fn max_unsigned() -> u64 {
84 <$unsigned>::MAX as u64
85 }
86
87 type Unsigned = $unsigned;
88
89 const TRUE: Self = -1;
90 const FALSE: Self = 0;
91 }
92
93 // Safety: this is a valid mask element type
94 unsafe impl MaskElement for $ty {}
95 }
96}
97
98impl_element! { i8, u8 }
99impl_element! { i16, u16 }
100impl_element! { i32, u32 }
101impl_element! { i64, u64 }
102impl_element! { isize, usize }
103
104/// A SIMD vector mask for `N` elements of width specified by `Element`.
105///
106/// Masks represent boolean inclusion/exclusion on a per-element basis.
107///
108/// The layout of this type is unspecified, and may change between platforms
109/// and/or Rust versions, and code should not assume that it is equivalent to
110/// `[T; N]`.
111#[repr(transparent)]
112pub struct Mask<T, const N: usize>(mask_impl::Mask<T, N>)
113where
114 T: MaskElement,
115 LaneCount<N>: SupportedLaneCount;
116
117impl<T, const N: usize> Copy for Mask<T, N>
118where
119 T: MaskElement,
120 LaneCount<N>: SupportedLaneCount,
121{
122}
123
124impl<T, const N: usize> Clone for Mask<T, N>
125where
126 T: MaskElement,
127 LaneCount<N>: SupportedLaneCount,
128{
129 #[inline]
130 fn clone(&self) -> Self {
131 *self
132 }
133}
134
135impl<T, const N: usize> Mask<T, N>
136where
137 T: MaskElement,
138 LaneCount<N>: SupportedLaneCount,
139{
140 /// Constructs a mask by setting all elements to the given value.
141 #[inline]
142 pub fn splat(value: bool) -> Self {
143 Self(mask_impl::Mask::splat(value))
144 }
145
146 /// Converts an array of bools to a SIMD mask.
147 #[inline]
148 pub fn from_array(array: [bool; N]) -> Self {
149 // SAFETY: Rust's bool has a layout of 1 byte (u8) with a value of
150 // true: 0b_0000_0001
151 // false: 0b_0000_0000
152 // Thus, an array of bools is also a valid array of bytes: [u8; N]
153 // This would be hypothetically valid as an "in-place" transmute,
154 // but these are "dependently-sized" types, so copy elision it is!
155 unsafe {
156 let bytes: [u8; N] = mem::transmute_copy(&array);
157 let bools: Simd<i8, N> =
158 core::intrinsics::simd::simd_ne(Simd::from_array(bytes), Simd::splat(0u8));
159 Mask::from_int_unchecked(core::intrinsics::simd::simd_cast(bools))
160 }
161 }
162
163 /// Converts a SIMD mask to an array of bools.
164 #[inline]
165 pub fn to_array(self) -> [bool; N] {
166 // This follows mostly the same logic as from_array.
167 // SAFETY: Rust's bool has a layout of 1 byte (u8) with a value of
168 // true: 0b_0000_0001
169 // false: 0b_0000_0000
170 // Thus, an array of bools is also a valid array of bytes: [u8; N]
171 // Since our masks are equal to integers where all bits are set,
172 // we can simply convert them to i8s, and then bitand them by the
173 // bitpattern for Rust's "true" bool.
174 // This would be hypothetically valid as an "in-place" transmute,
175 // but these are "dependently-sized" types, so copy elision it is!
176 unsafe {
177 let mut bytes: Simd<i8, N> = core::intrinsics::simd::simd_cast(self.to_int());
178 bytes &= Simd::splat(1i8);
179 mem::transmute_copy(&bytes)
180 }
181 }
182
183 /// Converts a vector of integers to a mask, where 0 represents `false` and -1
184 /// represents `true`.
185 ///
186 /// # Safety
187 /// All elements must be either 0 or -1.
188 #[inline]
189 #[must_use = "method returns a new mask and does not mutate the original value"]
190 pub unsafe fn from_int_unchecked(value: Simd<T, N>) -> Self {
191 // Safety: the caller must confirm this invariant
192 unsafe {
193 core::intrinsics::assume(<T as Sealed>::valid(value));
194 Self(mask_impl::Mask::from_int_unchecked(value))
195 }
196 }
197
198 /// Converts a vector of integers to a mask, where 0 represents `false` and -1
199 /// represents `true`.
200 ///
201 /// # Panics
202 /// Panics if any element is not 0 or -1.
203 #[inline]
204 #[must_use = "method returns a new mask and does not mutate the original value"]
205 #[track_caller]
206 pub fn from_int(value: Simd<T, N>) -> Self {
207 assert!(T::valid(value), "all values must be either 0 or -1",);
208 // Safety: the validity has been checked
209 unsafe { Self::from_int_unchecked(value) }
210 }
211
212 /// Converts the mask to a vector of integers, where 0 represents `false` and -1
213 /// represents `true`.
214 #[inline]
215 #[must_use = "method returns a new vector and does not mutate the original value"]
216 pub fn to_int(self) -> Simd<T, N> {
217 self.0.to_int()
218 }
219
220 /// Converts the mask to a mask of any other element size.
221 #[inline]
222 #[must_use = "method returns a new mask and does not mutate the original value"]
223 pub fn cast<U: MaskElement>(self) -> Mask<U, N> {
224 Mask(self.0.convert())
225 }
226
227 /// Tests the value of the specified element.
228 ///
229 /// # Safety
230 /// `index` must be less than `self.len()`.
231 #[inline]
232 #[must_use = "method returns a new bool and does not mutate the original value"]
233 pub unsafe fn test_unchecked(&self, index: usize) -> bool {
234 // Safety: the caller must confirm this invariant
235 unsafe { self.0.test_unchecked(index) }
236 }
237
238 /// Tests the value of the specified element.
239 ///
240 /// # Panics
241 /// Panics if `index` is greater than or equal to the number of elements in the vector.
242 #[inline]
243 #[must_use = "method returns a new bool and does not mutate the original value"]
244 #[track_caller]
245 pub fn test(&self, index: usize) -> bool {
246 assert!(index < N, "element index out of range");
247 // Safety: the element index has been checked
248 unsafe { self.test_unchecked(index) }
249 }
250
251 /// Sets the value of the specified element.
252 ///
253 /// # Safety
254 /// `index` must be less than `self.len()`.
255 #[inline]
256 pub unsafe fn set_unchecked(&mut self, index: usize, value: bool) {
257 // Safety: the caller must confirm this invariant
258 unsafe {
259 self.0.set_unchecked(index, value);
260 }
261 }
262
263 /// Sets the value of the specified element.
264 ///
265 /// # Panics
266 /// Panics if `index` is greater than or equal to the number of elements in the vector.
267 #[inline]
268 #[track_caller]
269 pub fn set(&mut self, index: usize, value: bool) {
270 assert!(index < N, "element index out of range");
271 // Safety: the element index has been checked
272 unsafe {
273 self.set_unchecked(index, value);
274 }
275 }
276
277 /// Returns true if any element is set, or false otherwise.
278 #[inline]
279 #[must_use = "method returns a new bool and does not mutate the original value"]
280 pub fn any(self) -> bool {
281 self.0.any()
282 }
283
284 /// Returns true if all elements are set, or false otherwise.
285 #[inline]
286 #[must_use = "method returns a new bool and does not mutate the original value"]
287 pub fn all(self) -> bool {
288 self.0.all()
289 }
290
291 /// Creates a bitmask from a mask.
292 ///
293 /// Each bit is set if the corresponding element in the mask is `true`.
294 /// If the mask contains more than 64 elements, the bitmask is truncated to the first 64.
295 #[inline]
296 #[must_use = "method returns a new integer and does not mutate the original value"]
297 pub fn to_bitmask(self) -> u64 {
298 self.0.to_bitmask_integer()
299 }
300
301 /// Creates a mask from a bitmask.
302 ///
303 /// For each bit, if it is set, the corresponding element in the mask is set to `true`.
304 /// If the mask contains more than 64 elements, the remainder are set to `false`.
305 #[inline]
306 #[must_use = "method returns a new mask and does not mutate the original value"]
307 pub fn from_bitmask(bitmask: u64) -> Self {
308 Self(mask_impl::Mask::from_bitmask_integer(bitmask))
309 }
310
311 /// Finds the index of the first set element.
312 ///
313 /// ```
314 /// # #![feature(portable_simd)]
315 /// # #[cfg(feature = "as_crate")] use core_simd::simd;
316 /// # #[cfg(not(feature = "as_crate"))] use core::simd;
317 /// # use simd::mask32x8;
318 /// assert_eq!(mask32x8::splat(false).first_set(), None);
319 /// assert_eq!(mask32x8::splat(true).first_set(), Some(0));
320 ///
321 /// let mask = mask32x8::from_array([false, true, false, false, true, false, false, true]);
322 /// assert_eq!(mask.first_set(), Some(1));
323 /// ```
324 #[inline]
325 #[must_use = "method returns the index and does not mutate the original value"]
326 pub fn first_set(self) -> Option<usize> {
327 // If bitmasks are efficient, using them is better
328 if cfg!(target_feature = "sse") && N <= 64 {
329 let tz = self.to_bitmask().trailing_zeros();
330 return if tz == 64 { None } else { Some(tz as usize) };
331 }
332
333 // To find the first set index:
334 // * create a vector 0..N
335 // * replace unset mask elements in that vector with -1
336 // * perform _unsigned_ reduce-min
337 // * check if the result is -1 or an index
338
339 let index = Simd::from_array(
340 const {
341 let mut index = [0; N];
342 let mut i = 0;
343 while i < N {
344 index[i] = i;
345 i += 1;
346 }
347 index
348 },
349 );
350
351 // Safety: the input and output are integer vectors
352 let index: Simd<T, N> = unsafe { core::intrinsics::simd::simd_cast(index) };
353
354 let masked_index = self.select(index, Self::splat(true).to_int());
355
356 // Safety: the input and output are integer vectors
357 let masked_index: Simd<T::Unsigned, N> =
358 unsafe { core::intrinsics::simd::simd_cast(masked_index) };
359
360 // Safety: the input is an integer vector
361 let min_index: T::Unsigned =
362 unsafe { core::intrinsics::simd::simd_reduce_min(masked_index) };
363
364 // Safety: the return value is the unsigned version of T
365 let min_index: T = unsafe { core::mem::transmute_copy(&min_index) };
366
367 if min_index.eq(T::TRUE) {
368 None
369 } else {
370 Some(min_index.to_usize())
371 }
372 }
373}
374
375// vector/array conversion
376impl<T, const N: usize> From<[bool; N]> for Mask<T, N>
377where
378 T: MaskElement,
379 LaneCount<N>: SupportedLaneCount,
380{
381 #[inline]
382 fn from(array: [bool; N]) -> Self {
383 Self::from_array(array)
384 }
385}
386
387impl<T, const N: usize> From<Mask<T, N>> for [bool; N]
388where
389 T: MaskElement,
390 LaneCount<N>: SupportedLaneCount,
391{
392 #[inline]
393 fn from(vector: Mask<T, N>) -> Self {
394 vector.to_array()
395 }
396}
397
398impl<T, const N: usize> Default for Mask<T, N>
399where
400 T: MaskElement,
401 LaneCount<N>: SupportedLaneCount,
402{
403 #[inline]
404 fn default() -> Self {
405 Self::splat(false)
406 }
407}
408
409impl<T, const N: usize> PartialEq for Mask<T, N>
410where
411 T: MaskElement + PartialEq,
412 LaneCount<N>: SupportedLaneCount,
413{
414 #[inline]
415 fn eq(&self, other: &Self) -> bool {
416 self.0 == other.0
417 }
418}
419
420impl<T, const N: usize> PartialOrd for Mask<T, N>
421where
422 T: MaskElement + PartialOrd,
423 LaneCount<N>: SupportedLaneCount,
424{
425 #[inline]
426 fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
427 self.0.partial_cmp(&other.0)
428 }
429}
430
431impl<T, const N: usize> fmt::Debug for Mask<T, N>
432where
433 T: MaskElement + fmt::Debug,
434 LaneCount<N>: SupportedLaneCount,
435{
436 #[inline]
437 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
438 f&mut DebugList<'_, '_>.debug_list()
439 .entries((0..N).map(|i: usize| self.test(index:i)))
440 .finish()
441 }
442}
443
444impl<T, const N: usize> core::ops::BitAnd for Mask<T, N>
445where
446 T: MaskElement,
447 LaneCount<N>: SupportedLaneCount,
448{
449 type Output = Self;
450 #[inline]
451 fn bitand(self, rhs: Self) -> Self {
452 Self(self.0 & rhs.0)
453 }
454}
455
456impl<T, const N: usize> core::ops::BitAnd<bool> for Mask<T, N>
457where
458 T: MaskElement,
459 LaneCount<N>: SupportedLaneCount,
460{
461 type Output = Self;
462 #[inline]
463 fn bitand(self, rhs: bool) -> Self {
464 self & Self::splat(rhs)
465 }
466}
467
468impl<T, const N: usize> core::ops::BitAnd<Mask<T, N>> for bool
469where
470 T: MaskElement,
471 LaneCount<N>: SupportedLaneCount,
472{
473 type Output = Mask<T, N>;
474 #[inline]
475 fn bitand(self, rhs: Mask<T, N>) -> Mask<T, N> {
476 Mask::splat(self) & rhs
477 }
478}
479
480impl<T, const N: usize> core::ops::BitOr for Mask<T, N>
481where
482 T: MaskElement,
483 LaneCount<N>: SupportedLaneCount,
484{
485 type Output = Self;
486 #[inline]
487 fn bitor(self, rhs: Self) -> Self {
488 Self(self.0 | rhs.0)
489 }
490}
491
492impl<T, const N: usize> core::ops::BitOr<bool> for Mask<T, N>
493where
494 T: MaskElement,
495 LaneCount<N>: SupportedLaneCount,
496{
497 type Output = Self;
498 #[inline]
499 fn bitor(self, rhs: bool) -> Self {
500 self | Self::splat(rhs)
501 }
502}
503
504impl<T, const N: usize> core::ops::BitOr<Mask<T, N>> for bool
505where
506 T: MaskElement,
507 LaneCount<N>: SupportedLaneCount,
508{
509 type Output = Mask<T, N>;
510 #[inline]
511 fn bitor(self, rhs: Mask<T, N>) -> Mask<T, N> {
512 Mask::splat(self) | rhs
513 }
514}
515
516impl<T, const N: usize> core::ops::BitXor for Mask<T, N>
517where
518 T: MaskElement,
519 LaneCount<N>: SupportedLaneCount,
520{
521 type Output = Self;
522 #[inline]
523 fn bitxor(self, rhs: Self) -> Self::Output {
524 Self(self.0 ^ rhs.0)
525 }
526}
527
528impl<T, const N: usize> core::ops::BitXor<bool> for Mask<T, N>
529where
530 T: MaskElement,
531 LaneCount<N>: SupportedLaneCount,
532{
533 type Output = Self;
534 #[inline]
535 fn bitxor(self, rhs: bool) -> Self::Output {
536 self ^ Self::splat(rhs)
537 }
538}
539
540impl<T, const N: usize> core::ops::BitXor<Mask<T, N>> for bool
541where
542 T: MaskElement,
543 LaneCount<N>: SupportedLaneCount,
544{
545 type Output = Mask<T, N>;
546 #[inline]
547 fn bitxor(self, rhs: Mask<T, N>) -> Self::Output {
548 Mask::splat(self) ^ rhs
549 }
550}
551
552impl<T, const N: usize> core::ops::Not for Mask<T, N>
553where
554 T: MaskElement,
555 LaneCount<N>: SupportedLaneCount,
556{
557 type Output = Mask<T, N>;
558 #[inline]
559 fn not(self) -> Self::Output {
560 Self(!self.0)
561 }
562}
563
564impl<T, const N: usize> core::ops::BitAndAssign for Mask<T, N>
565where
566 T: MaskElement,
567 LaneCount<N>: SupportedLaneCount,
568{
569 #[inline]
570 fn bitand_assign(&mut self, rhs: Self) {
571 self.0 = self.0 & rhs.0;
572 }
573}
574
575impl<T, const N: usize> core::ops::BitAndAssign<bool> for Mask<T, N>
576where
577 T: MaskElement,
578 LaneCount<N>: SupportedLaneCount,
579{
580 #[inline]
581 fn bitand_assign(&mut self, rhs: bool) {
582 *self &= Self::splat(rhs);
583 }
584}
585
586impl<T, const N: usize> core::ops::BitOrAssign for Mask<T, N>
587where
588 T: MaskElement,
589 LaneCount<N>: SupportedLaneCount,
590{
591 #[inline]
592 fn bitor_assign(&mut self, rhs: Self) {
593 self.0 = self.0 | rhs.0;
594 }
595}
596
597impl<T, const N: usize> core::ops::BitOrAssign<bool> for Mask<T, N>
598where
599 T: MaskElement,
600 LaneCount<N>: SupportedLaneCount,
601{
602 #[inline]
603 fn bitor_assign(&mut self, rhs: bool) {
604 *self |= Self::splat(rhs);
605 }
606}
607
608impl<T, const N: usize> core::ops::BitXorAssign for Mask<T, N>
609where
610 T: MaskElement,
611 LaneCount<N>: SupportedLaneCount,
612{
613 #[inline]
614 fn bitxor_assign(&mut self, rhs: Self) {
615 self.0 = self.0 ^ rhs.0;
616 }
617}
618
619impl<T, const N: usize> core::ops::BitXorAssign<bool> for Mask<T, N>
620where
621 T: MaskElement,
622 LaneCount<N>: SupportedLaneCount,
623{
624 #[inline]
625 fn bitxor_assign(&mut self, rhs: bool) {
626 *self ^= Self::splat(rhs);
627 }
628}
629
630macro_rules! impl_from {
631 { $from:ty => $($to:ty),* } => {
632 $(
633 impl<const N: usize> From<Mask<$from, N>> for Mask<$to, N>
634 where
635 LaneCount<N>: SupportedLaneCount,
636 {
637 #[inline]
638 fn from(value: Mask<$from, N>) -> Self {
639 value.cast()
640 }
641 }
642 )*
643 }
644}
645impl_from! { i8 => i16, i32, i64, isize }
646impl_from! { i16 => i32, i64, isize, i8 }
647impl_from! { i32 => i64, isize, i8, i16 }
648impl_from! { i64 => isize, i8, i16, i32 }
649impl_from! { isize => i8, i16, i32, i64 }
650