|
| 1 | +use crate::simd::{ |
| 2 | + intrinsics, LaneCount, Mask, Simd, SimdElement, SimdPartialOrd, SupportedLaneCount, |
| 3 | +}; |
| 4 | + |
| 5 | +/// Operations on SIMD vectors of signed integers. |
| 6 | +pub trait SimdInt: Sized { |
| 7 | + /// Mask type used for manipulating this SIMD vector type. |
| 8 | + type Mask; |
| 9 | + |
| 10 | + /// Scalar type contained by this SIMD vector type. |
| 11 | + type Scalar; |
| 12 | + |
| 13 | + /// Lanewise saturating add. |
| 14 | + /// |
| 15 | + /// # Examples |
| 16 | + /// ``` |
| 17 | + /// # #![feature(portable_simd)] |
| 18 | + /// # use core::simd::Simd; |
| 19 | + /// use core::i32::{MIN, MAX}; |
| 20 | + /// let x = Simd::from_array([MIN, 0, 1, MAX]); |
| 21 | + /// let max = Simd::splat(MAX); |
| 22 | + /// let unsat = x + max; |
| 23 | + /// let sat = x.saturating_add(max); |
| 24 | + /// assert_eq!(unsat, Simd::from_array([-1, MAX, MIN, -2])); |
| 25 | + /// assert_eq!(sat, Simd::from_array([-1, MAX, MAX, MAX])); |
| 26 | + /// ``` |
| 27 | + fn saturating_add(self, second: Self) -> Self; |
| 28 | + |
| 29 | + /// Lanewise saturating subtract. |
| 30 | + /// |
| 31 | + /// # Examples |
| 32 | + /// ``` |
| 33 | + /// # #![feature(portable_simd)] |
| 34 | + /// # use core::simd::Simd; |
| 35 | + /// use core::i32::{MIN, MAX}; |
| 36 | + /// let x = Simd::from_array([MIN, -2, -1, MAX]); |
| 37 | + /// let max = Simd::splat(MAX); |
| 38 | + /// let unsat = x - max; |
| 39 | + /// let sat = x.saturating_sub(max); |
| 40 | + /// assert_eq!(unsat, Simd::from_array([1, MAX, MIN, 0])); |
| 41 | + /// assert_eq!(sat, Simd::from_array([MIN, MIN, MIN, 0])); |
| 42 | + fn saturating_sub(self, second: Self) -> Self; |
| 43 | + |
| 44 | + /// Lanewise absolute value, implemented in Rust. |
| 45 | + /// Every lane becomes its absolute value. |
| 46 | + /// |
| 47 | + /// # Examples |
| 48 | + /// ``` |
| 49 | + /// # #![feature(portable_simd)] |
| 50 | + /// # use core::simd::Simd; |
| 51 | + /// use core::i32::{MIN, MAX}; |
| 52 | + /// let xs = Simd::from_array([MIN, MIN +1, -5, 0]); |
| 53 | + /// assert_eq!(xs.abs(), Simd::from_array([MIN, MAX, 5, 0])); |
| 54 | + /// ``` |
| 55 | + fn abs(self) -> Self; |
| 56 | + |
| 57 | + /// Lanewise saturating absolute value, implemented in Rust. |
| 58 | + /// As abs(), except the MIN value becomes MAX instead of itself. |
| 59 | + /// |
| 60 | + /// # Examples |
| 61 | + /// ``` |
| 62 | + /// # #![feature(portable_simd)] |
| 63 | + /// # use core::simd::Simd; |
| 64 | + /// use core::i32::{MIN, MAX}; |
| 65 | + /// let xs = Simd::from_array([MIN, -2, 0, 3]); |
| 66 | + /// let unsat = xs.abs(); |
| 67 | + /// let sat = xs.saturating_abs(); |
| 68 | + /// assert_eq!(unsat, Simd::from_array([MIN, 2, 0, 3])); |
| 69 | + /// assert_eq!(sat, Simd::from_array([MAX, 2, 0, 3])); |
| 70 | + /// ``` |
| 71 | + fn saturating_abs(self) -> Self; |
| 72 | + |
| 73 | + /// Lanewise saturating negation, implemented in Rust. |
| 74 | + /// As neg(), except the MIN value becomes MAX instead of itself. |
| 75 | + /// |
| 76 | + /// # Examples |
| 77 | + /// ``` |
| 78 | + /// # #![feature(portable_simd)] |
| 79 | + /// # use core::simd::Simd; |
| 80 | + /// use core::i32::{MIN, MAX}; |
| 81 | + /// let x = Simd::from_array([MIN, -2, 3, MAX]); |
| 82 | + /// let unsat = -x; |
| 83 | + /// let sat = x.saturating_neg(); |
| 84 | + /// assert_eq!(unsat, Simd::from_array([MIN, 2, -3, MIN + 1])); |
| 85 | + /// assert_eq!(sat, Simd::from_array([MAX, 2, -3, MIN + 1])); |
| 86 | + /// ``` |
| 87 | + fn saturating_neg(self) -> Self; |
| 88 | + |
| 89 | + /// Returns true for each positive lane and false if it is zero or negative. |
| 90 | + fn is_positive(self) -> Self::Mask; |
| 91 | + |
| 92 | + /// Returns true for each negative lane and false if it is zero or positive. |
| 93 | + fn is_negative(self) -> Self::Mask; |
| 94 | + |
| 95 | + /// Returns numbers representing the sign of each lane. |
| 96 | + /// * `0` if the number is zero |
| 97 | + /// * `1` if the number is positive |
| 98 | + /// * `-1` if the number is negative |
| 99 | + fn signum(self) -> Self; |
| 100 | + |
| 101 | + /// Returns the sum of the lanes of the vector, with wrapping addition. |
| 102 | + /// |
| 103 | + /// # Examples |
| 104 | + /// |
| 105 | + /// ``` |
| 106 | + /// # #![feature(portable_simd)] |
| 107 | + /// # use core::simd::i32x4; |
| 108 | + /// let v = i32x4::from_array([1, 2, 3, 4]); |
| 109 | + /// assert_eq!(v.reduce_sum(), 10); |
| 110 | + /// |
| 111 | + /// // SIMD integer addition is always wrapping |
| 112 | + /// let v = i32x4::from_array([i32::MAX, 1, 0, 0]); |
| 113 | + /// assert_eq!(v.reduce_sum(), i32::MIN); |
| 114 | + /// ``` |
| 115 | + fn reduce_sum(self) -> Self::Scalar; |
| 116 | + |
| 117 | + /// Returns the product of the lanes of the vector, with wrapping multiplication. |
| 118 | + /// |
| 119 | + /// # Examples |
| 120 | + /// |
| 121 | + /// ``` |
| 122 | + /// # #![feature(portable_simd)] |
| 123 | + /// # use core::simd::i32x4; |
| 124 | + /// let v = i32x4::from_array([1, 2, 3, 4]); |
| 125 | + /// assert_eq!(v.reduce_product(), 24); |
| 126 | + /// |
| 127 | + /// // SIMD integer multiplication is always wrapping |
| 128 | + /// let v = i32x4::from_array([i32::MAX, 2, 1, 1]); |
| 129 | + /// assert!(v.reduce_product() < i32::MAX); |
| 130 | + /// ``` |
| 131 | + fn reduce_product(self) -> Self::Scalar; |
| 132 | + |
| 133 | + /// Returns the maximum lane in the vector. |
| 134 | + /// |
| 135 | + /// # Examples |
| 136 | + /// |
| 137 | + /// ``` |
| 138 | + /// # #![feature(portable_simd)] |
| 139 | + /// # use core::simd::i32x4; |
| 140 | + /// let v = i32x4::from_array([1, 2, 3, 4]); |
| 141 | + /// assert_eq!(v.reduce_max(), 4); |
| 142 | + /// ``` |
| 143 | + fn reduce_max(self) -> Self::Scalar; |
| 144 | + |
| 145 | + /// Returns the minimum lane in the vector. |
| 146 | + /// |
| 147 | + /// # Examples |
| 148 | + /// |
| 149 | + /// ``` |
| 150 | + /// # #![feature(portable_simd)] |
| 151 | + /// # use core::simd::i32x4; |
| 152 | + /// let v = i32x4::from_array([1, 2, 3, 4]); |
| 153 | + /// assert_eq!(v.reduce_min(), 1); |
| 154 | + /// ``` |
| 155 | + fn reduce_min(self) -> Self::Scalar; |
| 156 | + |
| 157 | + /// Returns the cumulative bitwise "and" across the lanes of the vector. |
| 158 | + fn reduce_and(self) -> Self::Scalar; |
| 159 | + |
| 160 | + /// Returns the cumulative bitwise "or" across the lanes of the vector. |
| 161 | + fn reduce_or(self) -> Self::Scalar; |
| 162 | + |
| 163 | + /// Returns the cumulative bitwise "xor" across the lanes of the vector. |
| 164 | + fn reduce_xor(self) -> Self::Scalar; |
| 165 | +} |
| 166 | + |
| 167 | +macro_rules! impl_trait { |
| 168 | + { $($ty:ty),* } => { |
| 169 | + $( |
| 170 | + impl<const LANES: usize> SimdInt for Simd<$ty, LANES> |
| 171 | + where |
| 172 | + LaneCount<LANES>: SupportedLaneCount, |
| 173 | + { |
| 174 | + type Mask = Mask<<$ty as SimdElement>::Mask, LANES>; |
| 175 | + type Scalar = $ty; |
| 176 | + |
| 177 | + #[inline] |
| 178 | + fn saturating_add(self, second: Self) -> Self { |
| 179 | + // Safety: `self` is a vector |
| 180 | + unsafe { intrinsics::simd_saturating_add(self, second) } |
| 181 | + } |
| 182 | + |
| 183 | + #[inline] |
| 184 | + fn saturating_sub(self, second: Self) -> Self { |
| 185 | + // Safety: `self` is a vector |
| 186 | + unsafe { intrinsics::simd_saturating_sub(self, second) } |
| 187 | + } |
| 188 | + |
| 189 | + #[inline] |
| 190 | + fn abs(self) -> Self { |
| 191 | + const SHR: $ty = <$ty>::BITS as $ty - 1; |
| 192 | + let m = self >> Simd::splat(SHR); |
| 193 | + (self^m) - m |
| 194 | + } |
| 195 | + |
| 196 | + #[inline] |
| 197 | + fn saturating_abs(self) -> Self { |
| 198 | + // arith shift for -1 or 0 mask based on sign bit, giving 2s complement |
| 199 | + const SHR: $ty = <$ty>::BITS as $ty - 1; |
| 200 | + let m = self >> Simd::splat(SHR); |
| 201 | + (self^m).saturating_sub(m) |
| 202 | + } |
| 203 | + |
| 204 | + #[inline] |
| 205 | + fn saturating_neg(self) -> Self { |
| 206 | + Self::splat(0).saturating_sub(self) |
| 207 | + } |
| 208 | + |
| 209 | + #[inline] |
| 210 | + fn is_positive(self) -> Self::Mask { |
| 211 | + self.simd_gt(Self::splat(0)) |
| 212 | + } |
| 213 | + |
| 214 | + #[inline] |
| 215 | + fn is_negative(self) -> Self::Mask { |
| 216 | + self.simd_lt(Self::splat(0)) |
| 217 | + } |
| 218 | + |
| 219 | + #[inline] |
| 220 | + fn signum(self) -> Self { |
| 221 | + self.is_positive().select( |
| 222 | + Self::splat(1), |
| 223 | + self.is_negative().select(Self::splat(-1), Self::splat(0)) |
| 224 | + ) |
| 225 | + } |
| 226 | + |
| 227 | + #[inline] |
| 228 | + fn reduce_sum(self) -> Self::Scalar { |
| 229 | + // Safety: `self` is an integer vector |
| 230 | + unsafe { intrinsics::simd_reduce_add_ordered(self, 0) } |
| 231 | + } |
| 232 | + |
| 233 | + #[inline] |
| 234 | + fn reduce_product(self) -> Self::Scalar { |
| 235 | + // Safety: `self` is an integer vector |
| 236 | + unsafe { intrinsics::simd_reduce_mul_ordered(self, 1) } |
| 237 | + } |
| 238 | + |
| 239 | + #[inline] |
| 240 | + fn reduce_max(self) -> Self::Scalar { |
| 241 | + // Safety: `self` is an integer vector |
| 242 | + unsafe { intrinsics::simd_reduce_max(self) } |
| 243 | + } |
| 244 | + |
| 245 | + #[inline] |
| 246 | + fn reduce_min(self) -> Self::Scalar { |
| 247 | + // Safety: `self` is an integer vector |
| 248 | + unsafe { intrinsics::simd_reduce_min(self) } |
| 249 | + } |
| 250 | + |
| 251 | + #[inline] |
| 252 | + fn reduce_and(self) -> Self::Scalar { |
| 253 | + // Safety: `self` is an integer vector |
| 254 | + unsafe { intrinsics::simd_reduce_and(self) } |
| 255 | + } |
| 256 | + |
| 257 | + #[inline] |
| 258 | + fn reduce_or(self) -> Self::Scalar { |
| 259 | + // Safety: `self` is an integer vector |
| 260 | + unsafe { intrinsics::simd_reduce_or(self) } |
| 261 | + } |
| 262 | + |
| 263 | + #[inline] |
| 264 | + fn reduce_xor(self) -> Self::Scalar { |
| 265 | + // Safety: `self` is an integer vector |
| 266 | + unsafe { intrinsics::simd_reduce_xor(self) } |
| 267 | + } |
| 268 | + } |
| 269 | + )* |
| 270 | + } |
| 271 | +} |
| 272 | + |
| 273 | +impl_trait! { i8, i16, i32, i64, isize } |
0 commit comments