wide/
u32x16_.rs

1use super::*;
2
3pick! {
4  if #[cfg(target_feature="avx512f")] {
5    #[derive(Default, Clone, Copy, PartialEq, Eq)]
6    #[repr(C, align(64))]
7    pub struct u32x16 { pub(crate) avx512: m512i }
8  } else {
9    #[derive(Default, Clone, Copy, PartialEq, Eq)]
10    #[repr(C, align(64))]
11    pub struct u32x16 { pub(crate) a : u32x8, pub(crate) b : u32x8 }
12  }
13}
14
15int_uint_consts!(u32, 16, u32x16, 512);
16
17unsafe impl Zeroable for u32x16 {}
18unsafe impl Pod for u32x16 {}
19
20impl AlignTo for u32x16 {
21  type Elem = u32;
22}
23
24impl Add for u32x16 {
25  type Output = Self;
26  #[inline]
27  fn add(self, rhs: Self) -> Self::Output {
28    pick! {
29      if #[cfg(target_feature="avx512f")] {
30        Self { avx512: add_i32_m512i(self.avx512, rhs.avx512) }
31      } else {
32        Self {
33          a : self.a.add(rhs.a),
34          b : self.b.add(rhs.b),
35        }
36      }
37    }
38  }
39}
40
41impl Sub for u32x16 {
42  type Output = Self;
43  #[inline]
44  fn sub(self, rhs: Self) -> Self::Output {
45    pick! {
46      if #[cfg(target_feature="avx512f")] {
47        Self { avx512: sub_i32_m512i(self.avx512, rhs.avx512) }
48      } else {
49        Self {
50          a : self.a.sub(rhs.a),
51          b : self.b.sub(rhs.b),
52        }
53      }
54    }
55  }
56}
57
58impl Add<u32> for u32x16 {
59  type Output = Self;
60  #[inline]
61  fn add(self, rhs: u32) -> Self::Output {
62    self.add(Self::splat(rhs))
63  }
64}
65
66impl Sub<u32> for u32x16 {
67  type Output = Self;
68  #[inline]
69  fn sub(self, rhs: u32) -> Self::Output {
70    self.sub(Self::splat(rhs))
71  }
72}
73
74impl Mul<u32> for u32x16 {
75  type Output = Self;
76  #[inline]
77  fn mul(self, rhs: u32) -> Self::Output {
78    self * Self::splat(rhs)
79  }
80}
81
82impl Add<u32x16> for u32 {
83  type Output = u32x16;
84  #[inline]
85  fn add(self, rhs: u32x16) -> Self::Output {
86    u32x16::splat(self).add(rhs)
87  }
88}
89
90impl Sub<u32x16> for u32 {
91  type Output = u32x16;
92  #[inline]
93  fn sub(self, rhs: u32x16) -> Self::Output {
94    u32x16::splat(self).sub(rhs)
95  }
96}
97
98impl Mul<u32x16> for u32 {
99  type Output = u32x16;
100
101  #[inline]
102  fn mul(self, rhs: u32x16) -> Self::Output {
103    u32x16::splat(self) * rhs
104  }
105}
106
107impl Mul for u32x16 {
108  type Output = Self;
109  #[inline]
110  fn mul(self, rhs: Self) -> Self::Output {
111    pick! {
112      if #[cfg(target_feature="avx512f")] {
113        Self { avx512: mul_i32_keep_low_m512i(self.avx512, rhs.avx512) }
114      } else {
115        Self {
116          a : self.a.mul(rhs.a),
117          b : self.b.mul(rhs.b),
118        }
119      }
120    }
121  }
122}
123
124integer_impl_div_rem!(
125  u32,
126  u32x16,
127  [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15],
128);
129
130impl BitAnd for u32x16 {
131  type Output = Self;
132  #[inline]
133  fn bitand(self, rhs: Self) -> Self::Output {
134    pick! {
135      if #[cfg(target_feature="avx512f")] {
136        Self { avx512: bitand_m512i(self.avx512, rhs.avx512) }
137      } else {
138        Self {
139          a : self.a.bitand(rhs.a),
140          b : self.b.bitand(rhs.b),
141        }
142      }
143    }
144  }
145}
146
147impl BitOr for u32x16 {
148  type Output = Self;
149  #[inline]
150  fn bitor(self, rhs: Self) -> Self::Output {
151    pick! {
152    if #[cfg(target_feature="avx512f")] {
153        Self { avx512: bitor_m512i(self.avx512, rhs.avx512) }
154      } else {
155        Self {
156          a : self.a.bitor(rhs.a),
157          b : self.b.bitor(rhs.b),
158        }
159      }
160    }
161  }
162}
163
164impl BitXor for u32x16 {
165  type Output = Self;
166  #[inline]
167  fn bitxor(self, rhs: Self) -> Self::Output {
168    pick! {
169      if #[cfg(target_feature="avx512f")] {
170        Self { avx512: bitxor_m512i(self.avx512, rhs.avx512) }
171      } else {
172        Self {
173          a : self.a.bitxor(rhs.a),
174          b : self.b.bitxor(rhs.b),
175        }
176      }
177    }
178  }
179}
180
181impl From<u16x16> for u32x16 {
182  /// Widens and zero-extends each u16 lane to u32
183  #[inline]
184  fn from(v: u16x16) -> Self {
185    pick! {
186      if #[cfg(target_feature = "avx512f")] {
187        Self {
188          avx512: convert_to_u32_m512i_from_u16_m256i(v.avx2)
189        }
190      } else if #[cfg(target_feature = "avx2")] {
191        let lo: m128i = extract_m128i_from_m256i::<0>(v.avx2);
192        let hi: m128i = extract_m128i_from_m256i::<1>(v.avx2);
193        Self {
194          a: u32x8 { avx2: convert_to_i32_m256i_from_u16_m128i(lo) },
195          b: u32x8 { avx2: convert_to_i32_m256i_from_u16_m128i(hi) },
196        }
197      } else if #[cfg(target_feature = "sse2")] {
198        Self {
199          a: u32x8 {
200            a: u32x4 {
201              sse: shr_imm_u32_m128i::<16>(unpack_low_i16_m128i(v.a.sse, v.a.sse))
202            },
203            b: u32x4 {
204              sse: shr_imm_u32_m128i::<16>(unpack_high_i16_m128i(v.a.sse, v.a.sse))
205            },
206          },
207          b: u32x8 {
208            a: u32x4 {
209              sse: shr_imm_u32_m128i::<16>(unpack_low_i16_m128i(v.b.sse, v.b.sse))
210            },
211            b: u32x4 {
212              sse: shr_imm_u32_m128i::<16>(unpack_high_i16_m128i(v.b.sse, v.b.sse))
213            },
214          },
215        }
216      } else {
217        // Portable fallback
218        let arr = v.as_array();
219        Self::new([
220          arr[0] as u32,  arr[1] as u32,  arr[2] as u32,  arr[3] as u32,
221          arr[4] as u32,  arr[5] as u32,  arr[6] as u32,  arr[7] as u32,
222          arr[8] as u32,  arr[9] as u32,  arr[10] as u32, arr[11] as u32,
223          arr[12] as u32, arr[13] as u32, arr[14] as u32, arr[15] as u32,
224        ])
225      }
226    }
227  }
228}
229
230macro_rules! impl_shl_t_for_u32x16 {
231  ($($shift_type:ty),+ $(,)?) => {
232    $(impl Shl<$shift_type> for u32x16 {
233      type Output = Self;
234      /// Shifts all lanes by the value given.
235      #[inline]
236      fn shl(self, rhs: $shift_type) -> Self::Output {
237        pick! {
238          if #[cfg(target_feature="avx512f")] {
239            let shift = cast(rhs as u32);
240            Self { avx512: shl_all_u32_m512i(self.avx512, shift) }
241          } else {
242            Self {
243              a : self.a.shl(rhs),
244              b : self.b.shl(rhs),
245            }
246          }
247        }
248      }
249    })+
250  };
251}
252impl_shl_t_for_u32x16!(i8, u8, i16, u16, i32, u32, i64, u64, i128, u128);
253
254macro_rules! impl_shr_t_for_u32x16 {
255  ($($shift_type:ty),+ $(,)?) => {
256    $(impl Shr<$shift_type> for u32x16 {
257      type Output = Self;
258      /// Shifts all lanes by the value given.
259      #[inline]
260      fn shr(self, rhs: $shift_type) -> Self::Output {
261        pick! {
262          if #[cfg(target_feature="avx512f")] {
263            let shift = cast(rhs as u32);
264            Self { avx512: shr_all_u32_m512i(self.avx512, shift) }
265          } else {
266            Self {
267              a : self.a.shr(rhs),
268              b : self.b.shr(rhs),
269            }
270          }
271        }
272      }
273    })+
274  };
275}
276impl_shr_t_for_u32x16!(i8, u8, i16, u16, i32, u32, i64, u64, i128, u128);
277
278/// Shifts lanes by the corresponding lane.
279///
280/// Bitwise shift-right; yields `self >> mask(rhs)`, where mask removes any
281/// high-order bits of `rhs` that would cause the shift to exceed the bitwidth
282/// of the type. (same as `wrapping_shr`)
283impl Shr<u32x16> for u32x16 {
284  type Output = Self;
285
286  #[inline]
287  fn shr(self, rhs: u32x16) -> Self::Output {
288    pick! {
289      if #[cfg(target_feature="avx512f")] {
290        let shift_by = bitand_m512i(rhs.avx512, set_splat_i32_m512i(31));
291        Self { avx512: shr_each_u32_m512i(self.avx512, shift_by ) }
292      } else {
293        Self {
294          a : self.a.shr(rhs.a),
295          b : self.b.shr(rhs.b),
296        }
297      }
298    }
299  }
300}
301
302/// Shifts lanes by the corresponding lane.
303///
304/// Bitwise shift-left; yields `self << mask(rhs)`, where mask removes any
305/// high-order bits of `rhs` that would cause the shift to exceed the bitwidth
306/// of the type. (same as `wrapping_shl`)
307impl Shl<u32x16> for u32x16 {
308  type Output = Self;
309
310  #[inline]
311  fn shl(self, rhs: u32x16) -> Self::Output {
312    pick! {
313      if #[cfg(target_feature="avx512f")] {
314        let shift_by = bitand_m512i(rhs.avx512, set_splat_i32_m512i(31));
315        Self { avx512: shl_each_u32_m512i(self.avx512, shift_by) }
316      } else {
317        Self {
318          a : self.a.shl(rhs.a),
319          b : self.b.shl(rhs.b),
320        }
321      }
322    }
323  }
324}
325
326#[expect(deprecated)]
327impl CmpEq for u32x16 {
328  type Output = Self;
329  #[inline]
330  fn simd_eq(self, rhs: Self) -> Self::Output {
331    pick! {
332      if #[cfg(target_feature="avx512f")] {
333        Self { avx512: cmp_op_mask_u32_m512i::<{cmp_int_op!(Eq)}>(self.avx512, rhs.avx512) }
334      } else {
335        Self {
336          a : self.a.simd_eq(rhs.a),
337          b : self.b.simd_eq(rhs.b),
338        }
339      }
340    }
341  }
342}
343
344#[expect(deprecated)]
345impl CmpGt for u32x16 {
346  type Output = Self;
347  #[inline]
348  fn simd_gt(self, rhs: Self) -> Self::Output {
349    pick! {
350      if #[cfg(target_feature="avx512f")] {
351        Self { avx512: cmp_op_mask_u32_m512i::<{cmp_int_op!(Nle)}>(self.avx512, rhs.avx512) }
352      } else {
353        Self {
354          a : self.a.simd_gt(rhs.a),
355          b : self.b.simd_gt(rhs.b),
356        }
357      }
358    }
359  }
360}
361
362#[expect(deprecated)]
363impl CmpLt for u32x16 {
364  type Output = Self;
365  #[inline]
366  fn simd_lt(self, rhs: Self) -> Self::Output {
367    pick! {
368      if #[cfg(target_feature="avx512f")] {
369        Self { avx512: cmp_op_mask_u32_m512i::<{cmp_int_op!(Lt)}>(self.avx512, rhs.avx512) }
370      } else {
371        Self {
372          a : rhs.a.simd_gt(self.a),
373          b : rhs.b.simd_gt(self.b),
374        }
375      }
376    }
377  }
378}
379
380#[expect(deprecated)]
381impl CmpNe for u32x16 {
382  type Output = Self;
383  #[inline]
384  fn simd_ne(self, rhs: Self) -> Self::Output {
385    pick! {
386      if #[cfg(target_feature="avx512f")] {
387        Self { avx512: cmp_op_mask_u32_m512i::<{cmp_int_op!(Ne)}>(self.avx512, rhs.avx512) }
388      } else {
389        Self {
390          a : self.a.simd_ne(rhs.a),
391          b : self.b.simd_ne(rhs.b),
392        }
393      }
394    }
395  }
396}
397
398#[expect(deprecated)]
399impl CmpLe for u32x16 {
400  type Output = Self;
401  #[inline]
402  fn simd_le(self, rhs: Self) -> Self::Output {
403    pick! {
404      if #[cfg(target_feature="avx512f")] {
405        Self { avx512: cmp_op_mask_u32_m512i::<{cmp_int_op!(Le)}>(self.avx512, rhs.avx512) }
406      } else {
407        Self {
408          a : self.a.simd_le(rhs.a),
409          b : self.b.simd_le(rhs.b),
410        }
411      }
412    }
413  }
414}
415
416#[expect(deprecated)]
417impl CmpGe for u32x16 {
418  type Output = Self;
419  #[inline]
420  fn simd_ge(self, rhs: Self) -> Self::Output {
421    pick! {
422      if #[cfg(target_feature="avx512f")] {
423        Self { avx512: cmp_op_mask_u32_m512i::<{cmp_int_op!(Nlt)}>(self.avx512, rhs.avx512) }
424      } else {
425        Self {
426          a : self.a.simd_ge(rhs.a),
427          b : self.b.simd_ge(rhs.b),
428        }
429      }
430    }
431  }
432}
433
434impl u32x16 {
435  #[inline]
436  #[must_use]
437  pub const fn new(array: [u32; 16]) -> Self {
438    unsafe { core::mem::transmute(array) }
439  }
440
441  simd_comparison_fns!();
442
443  #[inline]
444  #[must_use]
445  pub fn blend(self, t: Self, f: Self) -> Self {
446    pick! {
447      if #[cfg(target_feature="avx512f")] {
448        Self { avx512: blend_varying_i8_m512i(f.avx512,t.avx512,movepi8_mask_m512i(self.avx512)) }
449      } else {
450        Self {
451          a : self.a.blend(t.a, f.a),
452          b : self.b.blend(t.b, f.b),
453        }
454      }
455    }
456  }
457
458  #[inline]
459  #[must_use]
460  pub fn reduce_add(self) -> u32 {
461    cast(i32x16::reduce_add(cast(self)))
462  }
463
464  #[inline]
465  #[must_use]
466  pub fn reduce_max(self) -> u32 {
467    let array: [u32x8; 2] = cast(self);
468    array[0].max(array[1]).reduce_max()
469  }
470
471  #[inline]
472  #[must_use]
473  pub fn reduce_min(self) -> u32 {
474    let array: [u32x8; 2] = cast(self);
475    array[0].min(array[1]).reduce_min()
476  }
477
478  #[inline]
479  #[must_use]
480  pub fn min(self, rhs: Self) -> Self {
481    pick! {
482      if #[cfg(target_feature="avx512f")] {
483        Self { avx512: min_u32_m512i(self.avx512, rhs.avx512) }
484      } else {
485        Self {
486          a: self.a.min(rhs.a),
487          b: self.b.min(rhs.b),
488        }
489      }
490    }
491  }
492
493  #[inline]
494  #[must_use]
495  pub fn max(self, rhs: Self) -> Self {
496    pick! {
497      if #[cfg(target_feature="avx512f")] {
498        Self { avx512: max_u32_m512i(self.avx512, rhs.avx512) }
499      } else {
500        Self {
501          a: self.a.max(rhs.a),
502          b: self.b.max(rhs.b),
503        }
504      }
505    }
506  }
507
508  integer_fn_clamp!();
509
510  #[inline]
511  #[must_use]
512  pub fn saturating_add(self, rhs: Self) -> Self {
513    pick! {
514      if #[cfg(target_feature="avx512f")] {
515        let result = self + rhs;
516        result.simd_lt(self).blend(Self::MAX, result)
517      } else {
518        Self {
519          a: self.a.saturating_add(rhs.a),
520          b: self.b.saturating_add(rhs.b),
521        }
522      }
523    }
524  }
525
526  #[inline]
527  #[must_use]
528  pub fn saturating_sub(self, rhs: Self) -> Self {
529    pick! {
530      if #[cfg(target_feature="avx512f")] {
531        let result = self - rhs;
532        result.simd_gt(self).blend(Self::MIN, result)
533      } else {
534        Self {
535          a: self.a.saturating_sub(rhs.a),
536          b: self.b.saturating_sub(rhs.b),
537        }
538      }
539    }
540  }
541
542  /// Lanewise saturating multiply.
543  #[inline]
544  #[must_use]
545  pub fn saturating_mul(self, rhs: Self) -> Self {
546    pick! {
547      if #[cfg(all(target_feature="avx512f", target_feature="avx512dq"))] {
548        #[cfg(target_arch = "x86")]
549        use core::arch::x86::{_mm512_unpackhi_epi64, _mm512_unpacklo_epi64};
550        #[cfg(target_arch = "x86_64")]
551        use core::arch::x86_64::{_mm512_unpackhi_epi64, _mm512_unpacklo_epi64};
552
553        let even_wide_mul = mul_u32_wide_m512i(self.avx512, rhs.avx512);
554        let odd_wide_mul = mul_u32_wide_m512i(
555          shuffle_i32_m512i::<0b_00_11_00_01>(self.avx512),
556          shuffle_i32_m512i::<0b_00_11_00_01>(rhs.avx512),
557        );
558
559        let ll_hh_1 = unpack_low_i32_m512i(even_wide_mul, odd_wide_mul);
560        let ll_hh_2 = unpack_high_i32_m512i(even_wide_mul, odd_wide_mul);
561        // TODO(safe_arch): Add `_mm512_unpacklo_epi64` and `_mm512_unpackhi_epi64`.
562        let low = Self {
563          avx512: m512i(unsafe { _mm512_unpacklo_epi64(ll_hh_1.0, ll_hh_2.0) }),
564        };
565        let high = Self {
566          avx512: m512i(unsafe { _mm512_unpackhi_epi64(ll_hh_1.0, ll_hh_2.0) }),
567        };
568
569        let no_overflow = high.simd_eq(Self::ZERO);
570        no_overflow.blend(low, Self::MAX)
571      } else {
572        let [self_a, self_b]: [u32x8; 2] = cast(self);
573        let [rhs_a, rhs_b]: [u32x8; 2] = cast(rhs);
574
575        cast([self_a.saturating_mul(rhs_a), self_b.saturating_mul(rhs_b)])
576      }
577    }
578  }
579
580  integer_fn_saturating_div!([
581    0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
582  ]);
583
584  #[inline]
585  #[must_use]
586  pub fn mul_keep_high(self, rhs: Self) -> Self {
587    pick! {
588      if #[cfg(target_feature="avx512f")] {
589        let alo = extract_m256i32_from_m512i::<0>(self.avx512);
590        let ahi = extract_m256i32_from_m512i::<1>(self.avx512);
591        let blo = extract_m256i32_from_m512i::<0>(rhs.avx512);
592        let bhi = extract_m256i32_from_m512i::<1>(rhs.avx512);
593
594        let lo_res: m256i = {
595          let a8 = u32x8 { avx2: alo };
596          let b8 = u32x8 { avx2: blo };
597          a8.mul_keep_high(b8).avx2
598        };
599        let hi_res: m256i = {
600          let a8 = u32x8 { avx2: ahi };
601          let b8 = u32x8 { avx2: bhi };
602          a8.mul_keep_high(b8).avx2
603        };
604
605        let zero = zeroed_m512i();
606        let with_lo = insert_m256i32_to_m512i::<0>(zero, lo_res);
607        let combined = insert_m256i32_to_m512i::<1>(with_lo, hi_res);
608
609        Self { avx512: combined }
610      } else {
611        Self {
612          a: self.a.mul_keep_high(rhs.a),
613          b: self.b.mul_keep_high(rhs.b),
614        }
615      }
616    }
617  }
618
619  #[inline]
620  #[must_use]
621  #[doc(alias("movemask", "move_mask"))]
622  pub fn to_bitmask(self) -> u32 {
623    i32x16::to_bitmask(cast(self))
624  }
625
626  #[inline]
627  #[must_use]
628  pub fn any(self) -> bool {
629    i32x16::any(cast(self))
630  }
631
632  #[inline]
633  #[must_use]
634  pub fn all(self) -> bool {
635    i32x16::all(cast(self))
636  }
637
638  #[inline]
639  #[must_use]
640  pub fn none(self) -> bool {
641    i32x16::none(cast(self))
642  }
643
644  /// Transpose matrix of 16x16 `u32` matrix. Currently not accelerated.
645  #[must_use]
646  #[inline]
647  pub fn transpose(data: [u32x16; 16]) -> [u32x16; 16] {
648    cast(i32x16::transpose(cast(data)))
649  }
650
651  #[inline]
652  pub fn to_array(self) -> [u32; 16] {
653    cast(self)
654  }
655
656  #[inline]
657  pub fn as_array(&self) -> &[u32; 16] {
658    cast_ref(self)
659  }
660
661  #[inline]
662  pub fn as_mut_array(&mut self) -> &mut [u32; 16] {
663    cast_mut(self)
664  }
665}
666
667impl Not for u32x16 {
668  type Output = Self;
669  #[inline]
670  fn not(self) -> Self::Output {
671    pick! {
672      if #[cfg(target_feature="avx512f")] {
673        Self { avx512: bitxor_m512i(self.avx512, set_splat_i32_m512i(-1)) }
674      } else {
675        Self {
676          a : self.a.not(),
677          b : self.b.not(),
678        }
679      }
680    }
681  }
682}
wide/u32x16_.rs

wide/
u32x16_.rs
