| 1 |
// Copyright 2018 Developers of the Rand project. |
| 2 |
// |
| 3 |
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or |
| 4 |
// https://www.apache.org/licenses/LICENSE-2.0> or the MIT license |
| 5 |
// <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your |
| 6 |
// option. This file may not be copied, modified, or distributed |
| 7 |
// except according to those terms. |
| 8 |
|
| 9 |
//! Math helper functions |
| 10 |
|
| 11 |
#[cfg(feature = "simd_support")] |
| 12 |
use core::simd::prelude::*; |
| 13 |
#[cfg(feature = "simd_support")] |
| 14 |
use core::simd::{LaneCount, SimdElement, SupportedLaneCount}; |
| 15 |
|
| 16 |
pub(crate) trait WideningMultiply<RHS = Self> { |
| 17 |
type Output; |
| 18 |
|
| 19 |
fn wmul(self, x: RHS) -> Self::Output; |
| 20 |
} |
| 21 |
|
| 22 |
macro_rules! wmul_impl { |
| 23 |
($ty:ty, $wide:ty, $shift:expr) => { |
| 24 |
impl WideningMultiply for $ty { |
| 25 |
type Output = ($ty, $ty); |
| 26 |
|
| 27 |
#[inline(always)] |
| 28 |
fn wmul(self, x: $ty) -> Self::Output { |
| 29 |
let tmp = (self as $wide) * (x as $wide); |
| 30 |
((tmp >> $shift) as $ty, tmp as $ty) |
| 31 |
} |
| 32 |
} |
| 33 |
}; |
| 34 |
|
| 35 |
// simd bulk implementation |
| 36 |
($(($ty:ident, $wide:ty),)+, $shift:expr) => { |
| 37 |
$( |
| 38 |
impl WideningMultiply for $ty { |
| 39 |
type Output = ($ty, $ty); |
| 40 |
|
| 41 |
#[inline(always)] |
| 42 |
fn wmul(self, x: $ty) -> Self::Output { |
| 43 |
// For supported vectors, this should compile to a couple |
| 44 |
// supported multiply & swizzle instructions (no actual |
| 45 |
// casting). |
| 46 |
// TODO: optimize |
| 47 |
let y: $wide = self.cast(); |
| 48 |
let x: $wide = x.cast(); |
| 49 |
let tmp = y * x; |
| 50 |
let hi: $ty = (tmp >> Simd::splat($shift)).cast(); |
| 51 |
let lo: $ty = tmp.cast(); |
| 52 |
(hi, lo) |
| 53 |
} |
| 54 |
} |
| 55 |
)+ |
| 56 |
}; |
| 57 |
} |
| 58 |
wmul_impl! { u8, u16, 8 } |
| 59 |
wmul_impl! { u16, u32, 16 } |
| 60 |
wmul_impl! { u32, u64, 32 } |
| 61 |
wmul_impl! { u64, u128, 64 } |
| 62 |
|
| 63 |
// This code is a translation of the __mulddi3 function in LLVM's |
| 64 |
// compiler-rt. It is an optimised variant of the common method |
| 65 |
// `(a + b) * (c + d) = ac + ad + bc + bd`. |
| 66 |
// |
| 67 |
// For some reason LLVM can optimise the C version very well, but |
| 68 |
// keeps shuffling registers in this Rust translation. |
| 69 |
macro_rules! wmul_impl_large { |
| 70 |
($ty:ty, $half:expr) => { |
| 71 |
impl WideningMultiply for $ty { |
| 72 |
type Output = ($ty, $ty); |
| 73 |
|
| 74 |
#[inline(always)] |
| 75 |
fn wmul(self, b: $ty) -> Self::Output { |
| 76 |
const LOWER_MASK: $ty = !0 >> $half; |
| 77 |
let mut low = (self & LOWER_MASK).wrapping_mul(b & LOWER_MASK); |
| 78 |
let mut t = low >> $half; |
| 79 |
low &= LOWER_MASK; |
| 80 |
t += (self >> $half).wrapping_mul(b & LOWER_MASK); |
| 81 |
low += (t & LOWER_MASK) << $half; |
| 82 |
let mut high = t >> $half; |
| 83 |
t = low >> $half; |
| 84 |
low &= LOWER_MASK; |
| 85 |
t += (b >> $half).wrapping_mul(self & LOWER_MASK); |
| 86 |
low += (t & LOWER_MASK) << $half; |
| 87 |
high += t >> $half; |
| 88 |
high += (self >> $half).wrapping_mul(b >> $half); |
| 89 |
|
| 90 |
(high, low) |
| 91 |
} |
| 92 |
} |
| 93 |
}; |
| 94 |
|
| 95 |
// simd bulk implementation |
| 96 |
(($($ty:ty,)+) $scalar:ty, $half:expr) => { |
| 97 |
$( |
| 98 |
impl WideningMultiply for $ty { |
| 99 |
type Output = ($ty, $ty); |
| 100 |
|
| 101 |
#[inline(always)] |
| 102 |
fn wmul(self, b: $ty) -> Self::Output { |
| 103 |
// needs wrapping multiplication |
| 104 |
let lower_mask = <$ty>::splat(!0 >> $half); |
| 105 |
let half = <$ty>::splat($half); |
| 106 |
let mut low = (self & lower_mask) * (b & lower_mask); |
| 107 |
let mut t = low >> half; |
| 108 |
low &= lower_mask; |
| 109 |
t += (self >> half) * (b & lower_mask); |
| 110 |
low += (t & lower_mask) << half; |
| 111 |
let mut high = t >> half; |
| 112 |
t = low >> half; |
| 113 |
low &= lower_mask; |
| 114 |
t += (b >> half) * (self & lower_mask); |
| 115 |
low += (t & lower_mask) << half; |
| 116 |
high += t >> half; |
| 117 |
high += (self >> half) * (b >> half); |
| 118 |
|
| 119 |
(high, low) |
| 120 |
} |
| 121 |
} |
| 122 |
)+ |
| 123 |
}; |
| 124 |
} |
| 125 |
wmul_impl_large! { u128, 64 } |
| 126 |
|
| 127 |
macro_rules! wmul_impl_usize { |
| 128 |
($ty:ty) => { |
| 129 |
impl WideningMultiply for usize { |
| 130 |
type Output = (usize, usize); |
| 131 |
|
| 132 |
#[inline(always)] |
| 133 |
fn wmul(self, x: usize) -> Self::Output { |
| 134 |
let (high, low) = (self as $ty).wmul(x as $ty); |
| 135 |
(high as usize, low as usize) |
| 136 |
} |
| 137 |
} |
| 138 |
}; |
| 139 |
} |
| 140 |
#[cfg(target_pointer_width = "16")] |
| 141 |
wmul_impl_usize! { u16 } |
| 142 |
#[cfg(target_pointer_width = "32")] |
| 143 |
wmul_impl_usize! { u32 } |
| 144 |
#[cfg(target_pointer_width = "64")] |
| 145 |
wmul_impl_usize! { u64 } |
| 146 |
|
| 147 |
#[cfg(feature = "simd_support")] |
| 148 |
mod simd_wmul { |
| 149 |
use super::*; |
| 150 |
#[cfg(target_arch = "x86")] |
| 151 |
use core::arch::x86::*; |
| 152 |
#[cfg(target_arch = "x86_64")] |
| 153 |
use core::arch::x86_64::*; |
| 154 |
|
| 155 |
wmul_impl! { |
| 156 |
(u8x4, u16x4), |
| 157 |
(u8x8, u16x8), |
| 158 |
(u8x16, u16x16), |
| 159 |
(u8x32, u16x32), |
| 160 |
(u8x64, Simd<u16, 64>),, |
| 161 |
8 |
| 162 |
} |
| 163 |
|
| 164 |
wmul_impl! { (u16x2, u32x2),, 16 } |
| 165 |
wmul_impl! { (u16x4, u32x4),, 16 } |
| 166 |
#[cfg(not(target_feature = "sse2"))] |
| 167 |
wmul_impl! { (u16x8, u32x8),, 16 } |
| 168 |
#[cfg(not(target_feature = "avx2"))] |
| 169 |
wmul_impl! { (u16x16, u32x16),, 16 } |
| 170 |
#[cfg(not(target_feature = "avx512bw"))] |
| 171 |
wmul_impl! { (u16x32, Simd<u32, 32>),, 16 } |
| 172 |
|
| 173 |
// 16-bit lane widths allow use of the x86 `mulhi` instructions, which |
| 174 |
// means `wmul` can be implemented with only two instructions. |
| 175 |
#[allow(unused_macros)] |
| 176 |
macro_rules! wmul_impl_16 { |
| 177 |
($ty:ident, $mulhi:ident, $mullo:ident) => { |
| 178 |
impl WideningMultiply for $ty { |
| 179 |
type Output = ($ty, $ty); |
| 180 |
|
| 181 |
#[inline(always)] |
| 182 |
fn wmul(self, x: $ty) -> Self::Output { |
| 183 |
let hi = unsafe { $mulhi(self.into(), x.into()) }.into(); |
| 184 |
let lo = unsafe { $mullo(self.into(), x.into()) }.into(); |
| 185 |
(hi, lo) |
| 186 |
} |
| 187 |
} |
| 188 |
}; |
| 189 |
} |
| 190 |
|
| 191 |
#[cfg(target_feature = "sse2")] |
| 192 |
wmul_impl_16! { u16x8, _mm_mulhi_epu16, _mm_mullo_epi16 } |
| 193 |
#[cfg(target_feature = "avx2")] |
| 194 |
wmul_impl_16! { u16x16, _mm256_mulhi_epu16, _mm256_mullo_epi16 } |
| 195 |
#[cfg(target_feature = "avx512bw")] |
| 196 |
wmul_impl_16! { u16x32, _mm512_mulhi_epu16, _mm512_mullo_epi16 } |
| 197 |
|
| 198 |
wmul_impl! { |
| 199 |
(u32x2, u64x2), |
| 200 |
(u32x4, u64x4), |
| 201 |
(u32x8, u64x8), |
| 202 |
(u32x16, Simd<u64, 16>),, |
| 203 |
32 |
| 204 |
} |
| 205 |
|
| 206 |
wmul_impl_large! { (u64x2, u64x4, u64x8,) u64, 32 } |
| 207 |
} |
| 208 |
|
| 209 |
/// Helper trait when dealing with scalar and SIMD floating point types. |
| 210 |
pub(crate) trait FloatSIMDUtils { |
| 211 |
// `PartialOrd` for vectors compares lexicographically. We want to compare all |
| 212 |
// the individual SIMD lanes instead, and get the combined result over all |
| 213 |
// lanes. This is possible using something like `a.lt(b).all()`, but we |
| 214 |
// implement it as a trait so we can write the same code for `f32` and `f64`. |
| 215 |
// Only the comparison functions we need are implemented. |
| 216 |
fn all_lt(self, other: Self) -> bool; |
| 217 |
fn all_le(self, other: Self) -> bool; |
| 218 |
fn all_finite(self) -> bool; |
| 219 |
|
| 220 |
type Mask; |
| 221 |
fn gt_mask(self, other: Self) -> Self::Mask; |
| 222 |
|
| 223 |
// Decrease all lanes where the mask is `true` to the next lower value |
| 224 |
// representable by the floating-point type. At least one of the lanes |
| 225 |
// must be set. |
| 226 |
fn decrease_masked(self, mask: Self::Mask) -> Self; |
| 227 |
|
| 228 |
// Convert from int value. Conversion is done while retaining the numerical |
| 229 |
// value, not by retaining the binary representation. |
| 230 |
type UInt; |
| 231 |
fn cast_from_int(i: Self::UInt) -> Self; |
| 232 |
} |
| 233 |
|
| 234 |
#[cfg(test)] |
| 235 |
pub(crate) trait FloatSIMDScalarUtils: FloatSIMDUtils { |
| 236 |
type Scalar; |
| 237 |
|
| 238 |
fn replace(self, index: usize, new_value: Self::Scalar) -> Self; |
| 239 |
fn extract_lane(self, index: usize) -> Self::Scalar; |
| 240 |
} |
| 241 |
|
| 242 |
/// Implement functions on f32/f64 to give them APIs similar to SIMD types |
| 243 |
pub(crate) trait FloatAsSIMD: Sized { |
| 244 |
#[cfg(test)] |
| 245 |
const LEN: usize = 1; |
| 246 |
|
| 247 |
#[inline(always)] |
| 248 |
fn splat(scalar: Self) -> Self { |
| 249 |
scalar |
| 250 |
} |
| 251 |
} |
| 252 |
|
| 253 |
pub(crate) trait IntAsSIMD: Sized { |
| 254 |
#[inline(always)] |
| 255 |
fn splat(scalar: Self) -> Self { |
| 256 |
scalar |
| 257 |
} |
| 258 |
} |
| 259 |
|
| 260 |
impl IntAsSIMD for u32 {} |
| 261 |
impl IntAsSIMD for u64 {} |
| 262 |
|
| 263 |
pub(crate) trait BoolAsSIMD: Sized { |
| 264 |
fn any(self) -> bool; |
| 265 |
} |
| 266 |
|
| 267 |
impl BoolAsSIMD for bool { |
| 268 |
#[inline(always)] |
| 269 |
fn any(self) -> bool { |
| 270 |
self |
| 271 |
} |
| 272 |
} |
| 273 |
|
| 274 |
macro_rules! scalar_float_impl { |
| 275 |
($ty:ident, $uty:ident) => { |
| 276 |
impl FloatSIMDUtils for $ty { |
| 277 |
type Mask = bool; |
| 278 |
type UInt = $uty; |
| 279 |
|
| 280 |
#[inline(always)] |
| 281 |
fn all_lt(self, other: Self) -> bool { |
| 282 |
self < other |
| 283 |
} |
| 284 |
|
| 285 |
#[inline(always)] |
| 286 |
fn all_le(self, other: Self) -> bool { |
| 287 |
self <= other |
| 288 |
} |
| 289 |
|
| 290 |
#[inline(always)] |
| 291 |
fn all_finite(self) -> bool { |
| 292 |
self.is_finite() |
| 293 |
} |
| 294 |
|
| 295 |
#[inline(always)] |
| 296 |
fn gt_mask(self, other: Self) -> Self::Mask { |
| 297 |
self > other |
| 298 |
} |
| 299 |
|
| 300 |
#[inline(always)] |
| 301 |
fn decrease_masked(self, mask: Self::Mask) -> Self { |
| 302 |
debug_assert!(mask, "At least one lane must be set"); |
| 303 |
<$ty>::from_bits(self.to_bits() - 1) |
| 304 |
} |
| 305 |
|
| 306 |
#[inline] |
| 307 |
fn cast_from_int(i: Self::UInt) -> Self { |
| 308 |
i as $ty |
| 309 |
} |
| 310 |
} |
| 311 |
|
| 312 |
#[cfg(test)] |
| 313 |
impl FloatSIMDScalarUtils for $ty { |
| 314 |
type Scalar = $ty; |
| 315 |
|
| 316 |
#[inline] |
| 317 |
fn replace(self, index: usize, new_value: Self::Scalar) -> Self { |
| 318 |
debug_assert_eq!(index, 0); |
| 319 |
new_value |
| 320 |
} |
| 321 |
|
| 322 |
#[inline] |
| 323 |
fn extract_lane(self, index: usize) -> Self::Scalar { |
| 324 |
debug_assert_eq!(index, 0); |
| 325 |
self |
| 326 |
} |
| 327 |
} |
| 328 |
|
| 329 |
impl FloatAsSIMD for $ty {} |
| 330 |
}; |
| 331 |
} |
| 332 |
|
| 333 |
scalar_float_impl!(f32, u32); |
| 334 |
scalar_float_impl!(f64, u64); |
| 335 |
|
| 336 |
#[cfg(feature = "simd_support")] |
| 337 |
macro_rules! simd_impl { |
| 338 |
($fty:ident, $uty:ident) => { |
| 339 |
impl<const LANES: usize> FloatSIMDUtils for Simd<$fty, LANES> |
| 340 |
where |
| 341 |
LaneCount<LANES>: SupportedLaneCount, |
| 342 |
{ |
| 343 |
type Mask = Mask<<$fty as SimdElement>::Mask, LANES>; |
| 344 |
type UInt = Simd<$uty, LANES>; |
| 345 |
|
| 346 |
#[inline(always)] |
| 347 |
fn all_lt(self, other: Self) -> bool { |
| 348 |
self.simd_lt(other).all() |
| 349 |
} |
| 350 |
|
| 351 |
#[inline(always)] |
| 352 |
fn all_le(self, other: Self) -> bool { |
| 353 |
self.simd_le(other).all() |
| 354 |
} |
| 355 |
|
| 356 |
#[inline(always)] |
| 357 |
fn all_finite(self) -> bool { |
| 358 |
self.is_finite().all() |
| 359 |
} |
| 360 |
|
| 361 |
#[inline(always)] |
| 362 |
fn gt_mask(self, other: Self) -> Self::Mask { |
| 363 |
self.simd_gt(other) |
| 364 |
} |
| 365 |
|
| 366 |
#[inline(always)] |
| 367 |
fn decrease_masked(self, mask: Self::Mask) -> Self { |
| 368 |
// Casting a mask into ints will produce all bits set for |
| 369 |
// true, and 0 for false. Adding that to the binary |
| 370 |
// representation of a float means subtracting one from |
| 371 |
// the binary representation, resulting in the next lower |
| 372 |
// value representable by $fty. This works even when the |
| 373 |
// current value is infinity. |
| 374 |
debug_assert!(mask.any(), "At least one lane must be set"); |
| 375 |
Self::from_bits(self.to_bits() + mask.to_int().cast()) |
| 376 |
} |
| 377 |
|
| 378 |
#[inline] |
| 379 |
fn cast_from_int(i: Self::UInt) -> Self { |
| 380 |
i.cast() |
| 381 |
} |
| 382 |
} |
| 383 |
|
| 384 |
#[cfg(test)] |
| 385 |
impl<const LANES: usize> FloatSIMDScalarUtils for Simd<$fty, LANES> |
| 386 |
where |
| 387 |
LaneCount<LANES>: SupportedLaneCount, |
| 388 |
{ |
| 389 |
type Scalar = $fty; |
| 390 |
|
| 391 |
#[inline] |
| 392 |
fn replace(mut self, index: usize, new_value: Self::Scalar) -> Self { |
| 393 |
self.as_mut_array()[index] = new_value; |
| 394 |
self |
| 395 |
} |
| 396 |
|
| 397 |
#[inline] |
| 398 |
fn extract_lane(self, index: usize) -> Self::Scalar { |
| 399 |
self.as_array()[index] |
| 400 |
} |
| 401 |
} |
| 402 |
}; |
| 403 |
} |
| 404 |
|
| 405 |
#[cfg(feature = "simd_support")] |
| 406 |
simd_impl!(f32, u32); |
| 407 |
#[cfg(feature = "simd_support")] |
| 408 |
simd_impl!(f64, u64); |