| 1 | // Copyright 2018-2020 Developers of the Rand project. |
| 2 | // Copyright 2017 The Rust Project Developers. |
| 3 | // |
| 4 | // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or |
| 5 | // https://www.apache.org/licenses/LICENSE-2.0> or the MIT license |
| 6 | // <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your |
| 7 | // option. This file may not be copied, modified, or distributed |
| 8 | // except according to those terms. |
| 9 | |
| 10 | //! `UniformInt` implementation |
| 11 | |
| 12 | use super::{Error, SampleBorrow, SampleUniform, UniformSampler}; |
| 13 | use crate::distr::utils::WideningMultiply; |
| 14 | #[cfg (feature = "simd_support" )] |
| 15 | use crate::distr::{Distribution, StandardUniform}; |
| 16 | use crate::Rng; |
| 17 | |
| 18 | #[cfg (feature = "simd_support" )] |
| 19 | use core::simd::prelude::*; |
| 20 | #[cfg (feature = "simd_support" )] |
| 21 | use core::simd::{LaneCount, SupportedLaneCount}; |
| 22 | |
| 23 | #[cfg (feature = "serde" )] |
| 24 | use serde::{Deserialize, Serialize}; |
| 25 | |
| 26 | /// The back-end implementing [`UniformSampler`] for integer types. |
| 27 | /// |
| 28 | /// Unless you are implementing [`UniformSampler`] for your own type, this type |
| 29 | /// should not be used directly, use [`Uniform`] instead. |
| 30 | /// |
| 31 | /// # Implementation notes |
| 32 | /// |
| 33 | /// For simplicity, we use the same generic struct `UniformInt<X>` for all |
| 34 | /// integer types `X`. This gives us only one field type, `X`; to store unsigned |
| 35 | /// values of this size, we take use the fact that these conversions are no-ops. |
| 36 | /// |
| 37 | /// For a closed range, the number of possible numbers we should generate is |
| 38 | /// `range = (high - low + 1)`. To avoid bias, we must ensure that the size of |
| 39 | /// our sample space, `zone`, is a multiple of `range`; other values must be |
| 40 | /// rejected (by replacing with a new random sample). |
| 41 | /// |
| 42 | /// As a special case, we use `range = 0` to represent the full range of the |
| 43 | /// result type (i.e. for `new_inclusive($ty::MIN, $ty::MAX)`). |
| 44 | /// |
| 45 | /// The optimum `zone` is the largest product of `range` which fits in our |
| 46 | /// (unsigned) target type. We calculate this by calculating how many numbers we |
| 47 | /// must reject: `reject = (MAX + 1) % range = (MAX - range + 1) % range`. Any (large) |
| 48 | /// product of `range` will suffice, thus in `sample_single` we multiply by a |
| 49 | /// power of 2 via bit-shifting (faster but may cause more rejections). |
| 50 | /// |
| 51 | /// The smallest integer PRNGs generate is `u32`. For 8- and 16-bit outputs we |
| 52 | /// use `u32` for our `zone` and samples (because it's not slower and because |
| 53 | /// it reduces the chance of having to reject a sample). In this case we cannot |
| 54 | /// store `zone` in the target type since it is too large, however we know |
| 55 | /// `ints_to_reject < range <= $uty::MAX`. |
| 56 | /// |
| 57 | /// An alternative to using a modulus is widening multiply: After a widening |
| 58 | /// multiply by `range`, the result is in the high word. Then comparing the low |
| 59 | /// word against `zone` makes sure our distribution is uniform. |
| 60 | /// |
| 61 | /// # Bias |
| 62 | /// |
| 63 | /// Unless the `unbiased` feature flag is used, outputs may have a small bias. |
| 64 | /// In the worst case, bias affects 1 in `2^n` samples where n is |
| 65 | /// 56 (`i8` and `u8`), 48 (`i16` and `u16`), 96 (`i32` and `u32`), 64 (`i64` |
| 66 | /// and `u64`), 128 (`i128` and `u128`). |
| 67 | /// |
| 68 | /// [`Uniform`]: super::Uniform |
| 69 | #[derive(Clone, Copy, Debug, PartialEq, Eq)] |
| 70 | #[cfg_attr (feature = "serde" , derive(Serialize, Deserialize))] |
| 71 | pub struct UniformInt<X> { |
| 72 | pub(super) low: X, |
| 73 | pub(super) range: X, |
| 74 | thresh: X, // effectively 2.pow(max(64, uty_bits)) % range |
| 75 | } |
| 76 | |
| 77 | macro_rules! uniform_int_impl { |
| 78 | ($ty:ty, $uty:ty, $sample_ty:ident) => { |
| 79 | impl SampleUniform for $ty { |
| 80 | type Sampler = UniformInt<$ty>; |
| 81 | } |
| 82 | |
| 83 | impl UniformSampler for UniformInt<$ty> { |
| 84 | // We play free and fast with unsigned vs signed here |
| 85 | // (when $ty is signed), but that's fine, since the |
| 86 | // contract of this macro is for $ty and $uty to be |
| 87 | // "bit-equal", so casting between them is a no-op. |
| 88 | |
| 89 | type X = $ty; |
| 90 | |
| 91 | #[inline] // if the range is constant, this helps LLVM to do the |
| 92 | // calculations at compile-time. |
| 93 | fn new<B1, B2>(low_b: B1, high_b: B2) -> Result<Self, Error> |
| 94 | where |
| 95 | B1: SampleBorrow<Self::X> + Sized, |
| 96 | B2: SampleBorrow<Self::X> + Sized, |
| 97 | { |
| 98 | let low = *low_b.borrow(); |
| 99 | let high = *high_b.borrow(); |
| 100 | if !(low < high) { |
| 101 | return Err(Error::EmptyRange); |
| 102 | } |
| 103 | UniformSampler::new_inclusive(low, high - 1) |
| 104 | } |
| 105 | |
| 106 | #[inline] // if the range is constant, this helps LLVM to do the |
| 107 | // calculations at compile-time. |
| 108 | fn new_inclusive<B1, B2>(low_b: B1, high_b: B2) -> Result<Self, Error> |
| 109 | where |
| 110 | B1: SampleBorrow<Self::X> + Sized, |
| 111 | B2: SampleBorrow<Self::X> + Sized, |
| 112 | { |
| 113 | let low = *low_b.borrow(); |
| 114 | let high = *high_b.borrow(); |
| 115 | if !(low <= high) { |
| 116 | return Err(Error::EmptyRange); |
| 117 | } |
| 118 | |
| 119 | let range = high.wrapping_sub(low).wrapping_add(1) as $uty; |
| 120 | let thresh = if range > 0 { |
| 121 | let range = $sample_ty::from(range); |
| 122 | (range.wrapping_neg() % range) |
| 123 | } else { |
| 124 | 0 |
| 125 | }; |
| 126 | |
| 127 | Ok(UniformInt { |
| 128 | low, |
| 129 | range: range as $ty, // type: $uty |
| 130 | thresh: thresh as $uty as $ty, // type: $sample_ty |
| 131 | }) |
| 132 | } |
| 133 | |
| 134 | /// Sample from distribution, Lemire's method, unbiased |
| 135 | #[inline] |
| 136 | fn sample<R: Rng + ?Sized>(&self, rng: &mut R) -> Self::X { |
| 137 | let range = self.range as $uty as $sample_ty; |
| 138 | if range == 0 { |
| 139 | return rng.random(); |
| 140 | } |
| 141 | |
| 142 | let thresh = self.thresh as $uty as $sample_ty; |
| 143 | let hi = loop { |
| 144 | let (hi, lo) = rng.random::<$sample_ty>().wmul(range); |
| 145 | if lo >= thresh { |
| 146 | break hi; |
| 147 | } |
| 148 | }; |
| 149 | self.low.wrapping_add(hi as $ty) |
| 150 | } |
| 151 | |
| 152 | #[inline] |
| 153 | fn sample_single<R: Rng + ?Sized, B1, B2>( |
| 154 | low_b: B1, |
| 155 | high_b: B2, |
| 156 | rng: &mut R, |
| 157 | ) -> Result<Self::X, Error> |
| 158 | where |
| 159 | B1: SampleBorrow<Self::X> + Sized, |
| 160 | B2: SampleBorrow<Self::X> + Sized, |
| 161 | { |
| 162 | let low = *low_b.borrow(); |
| 163 | let high = *high_b.borrow(); |
| 164 | if !(low < high) { |
| 165 | return Err(Error::EmptyRange); |
| 166 | } |
| 167 | Self::sample_single_inclusive(low, high - 1, rng) |
| 168 | } |
| 169 | |
| 170 | /// Sample single value, Canon's method, biased |
| 171 | /// |
| 172 | /// In the worst case, bias affects 1 in `2^n` samples where n is |
| 173 | /// 56 (`i8`), 48 (`i16`), 96 (`i32`), 64 (`i64`), 128 (`i128`). |
| 174 | #[cfg(not(feature = "unbiased" ))] |
| 175 | #[inline] |
| 176 | fn sample_single_inclusive<R: Rng + ?Sized, B1, B2>( |
| 177 | low_b: B1, |
| 178 | high_b: B2, |
| 179 | rng: &mut R, |
| 180 | ) -> Result<Self::X, Error> |
| 181 | where |
| 182 | B1: SampleBorrow<Self::X> + Sized, |
| 183 | B2: SampleBorrow<Self::X> + Sized, |
| 184 | { |
| 185 | let low = *low_b.borrow(); |
| 186 | let high = *high_b.borrow(); |
| 187 | if !(low <= high) { |
| 188 | return Err(Error::EmptyRange); |
| 189 | } |
| 190 | let range = high.wrapping_sub(low).wrapping_add(1) as $uty as $sample_ty; |
| 191 | if range == 0 { |
| 192 | // Range is MAX+1 (unrepresentable), so we need a special case |
| 193 | return Ok(rng.random()); |
| 194 | } |
| 195 | |
| 196 | // generate a sample using a sensible integer type |
| 197 | let (mut result, lo_order) = rng.random::<$sample_ty>().wmul(range); |
| 198 | |
| 199 | // if the sample is biased... |
| 200 | if lo_order > range.wrapping_neg() { |
| 201 | // ...generate a new sample to reduce bias... |
| 202 | let (new_hi_order, _) = (rng.random::<$sample_ty>()).wmul(range as $sample_ty); |
| 203 | // ... incrementing result on overflow |
| 204 | let is_overflow = lo_order.checked_add(new_hi_order as $sample_ty).is_none(); |
| 205 | result += is_overflow as $sample_ty; |
| 206 | } |
| 207 | |
| 208 | Ok(low.wrapping_add(result as $ty)) |
| 209 | } |
| 210 | |
| 211 | /// Sample single value, Canon's method, unbiased |
| 212 | #[cfg(feature = "unbiased" )] |
| 213 | #[inline] |
| 214 | fn sample_single_inclusive<R: Rng + ?Sized, B1, B2>( |
| 215 | low_b: B1, |
| 216 | high_b: B2, |
| 217 | rng: &mut R, |
| 218 | ) -> Result<Self::X, Error> |
| 219 | where |
| 220 | B1: SampleBorrow<$ty> + Sized, |
| 221 | B2: SampleBorrow<$ty> + Sized, |
| 222 | { |
| 223 | let low = *low_b.borrow(); |
| 224 | let high = *high_b.borrow(); |
| 225 | if !(low <= high) { |
| 226 | return Err(Error::EmptyRange); |
| 227 | } |
| 228 | let range = high.wrapping_sub(low).wrapping_add(1) as $uty as $sample_ty; |
| 229 | if range == 0 { |
| 230 | // Range is MAX+1 (unrepresentable), so we need a special case |
| 231 | return Ok(rng.random()); |
| 232 | } |
| 233 | |
| 234 | let (mut result, mut lo) = rng.random::<$sample_ty>().wmul(range); |
| 235 | |
| 236 | // In contrast to the biased sampler, we use a loop: |
| 237 | while lo > range.wrapping_neg() { |
| 238 | let (new_hi, new_lo) = (rng.random::<$sample_ty>()).wmul(range); |
| 239 | match lo.checked_add(new_hi) { |
| 240 | Some(x) if x < $sample_ty::MAX => { |
| 241 | // Anything less than MAX: last term is 0 |
| 242 | break; |
| 243 | } |
| 244 | None => { |
| 245 | // Overflow: last term is 1 |
| 246 | result += 1; |
| 247 | break; |
| 248 | } |
| 249 | _ => { |
| 250 | // Unlikely case: must check next sample |
| 251 | lo = new_lo; |
| 252 | continue; |
| 253 | } |
| 254 | } |
| 255 | } |
| 256 | |
| 257 | Ok(low.wrapping_add(result as $ty)) |
| 258 | } |
| 259 | } |
| 260 | }; |
| 261 | } |
| 262 | |
| 263 | uniform_int_impl! { i8, u8, u32 } |
| 264 | uniform_int_impl! { i16, u16, u32 } |
| 265 | uniform_int_impl! { i32, u32, u32 } |
| 266 | uniform_int_impl! { i64, u64, u64 } |
| 267 | uniform_int_impl! { i128, u128, u128 } |
| 268 | uniform_int_impl! { u8, u8, u32 } |
| 269 | uniform_int_impl! { u16, u16, u32 } |
| 270 | uniform_int_impl! { u32, u32, u32 } |
| 271 | uniform_int_impl! { u64, u64, u64 } |
| 272 | uniform_int_impl! { u128, u128, u128 } |
| 273 | |
| 274 | #[cfg (feature = "simd_support" )] |
| 275 | macro_rules! uniform_simd_int_impl { |
| 276 | ($ty:ident, $unsigned:ident) => { |
| 277 | // The "pick the largest zone that can fit in an `u32`" optimization |
| 278 | // is less useful here. Multiple lanes complicate things, we don't |
| 279 | // know the PRNG's minimal output size, and casting to a larger vector |
| 280 | // is generally a bad idea for SIMD performance. The user can still |
| 281 | // implement it manually. |
| 282 | |
| 283 | #[cfg(feature = "simd_support" )] |
| 284 | impl<const LANES: usize> SampleUniform for Simd<$ty, LANES> |
| 285 | where |
| 286 | LaneCount<LANES>: SupportedLaneCount, |
| 287 | Simd<$unsigned, LANES>: |
| 288 | WideningMultiply<Output = (Simd<$unsigned, LANES>, Simd<$unsigned, LANES>)>, |
| 289 | StandardUniform: Distribution<Simd<$unsigned, LANES>>, |
| 290 | { |
| 291 | type Sampler = UniformInt<Simd<$ty, LANES>>; |
| 292 | } |
| 293 | |
| 294 | #[cfg(feature = "simd_support" )] |
| 295 | impl<const LANES: usize> UniformSampler for UniformInt<Simd<$ty, LANES>> |
| 296 | where |
| 297 | LaneCount<LANES>: SupportedLaneCount, |
| 298 | Simd<$unsigned, LANES>: |
| 299 | WideningMultiply<Output = (Simd<$unsigned, LANES>, Simd<$unsigned, LANES>)>, |
| 300 | StandardUniform: Distribution<Simd<$unsigned, LANES>>, |
| 301 | { |
| 302 | type X = Simd<$ty, LANES>; |
| 303 | |
| 304 | #[inline] // if the range is constant, this helps LLVM to do the |
| 305 | // calculations at compile-time. |
| 306 | fn new<B1, B2>(low_b: B1, high_b: B2) -> Result<Self, Error> |
| 307 | where B1: SampleBorrow<Self::X> + Sized, |
| 308 | B2: SampleBorrow<Self::X> + Sized |
| 309 | { |
| 310 | let low = *low_b.borrow(); |
| 311 | let high = *high_b.borrow(); |
| 312 | if !(low.simd_lt(high).all()) { |
| 313 | return Err(Error::EmptyRange); |
| 314 | } |
| 315 | UniformSampler::new_inclusive(low, high - Simd::splat(1)) |
| 316 | } |
| 317 | |
| 318 | #[inline] // if the range is constant, this helps LLVM to do the |
| 319 | // calculations at compile-time. |
| 320 | fn new_inclusive<B1, B2>(low_b: B1, high_b: B2) -> Result<Self, Error> |
| 321 | where B1: SampleBorrow<Self::X> + Sized, |
| 322 | B2: SampleBorrow<Self::X> + Sized |
| 323 | { |
| 324 | let low = *low_b.borrow(); |
| 325 | let high = *high_b.borrow(); |
| 326 | if !(low.simd_le(high).all()) { |
| 327 | return Err(Error::EmptyRange); |
| 328 | } |
| 329 | |
| 330 | // NOTE: all `Simd` operations are inherently wrapping, |
| 331 | // see https://doc.rust-lang.org/std/simd/struct.Simd.html |
| 332 | let range: Simd<$unsigned, LANES> = ((high - low) + Simd::splat(1)).cast(); |
| 333 | |
| 334 | // We must avoid divide-by-zero by using 0 % 1 == 0. |
| 335 | let not_full_range = range.simd_gt(Simd::splat(0)); |
| 336 | let modulo = not_full_range.select(range, Simd::splat(1)); |
| 337 | let ints_to_reject = range.wrapping_neg() % modulo; |
| 338 | |
| 339 | Ok(UniformInt { |
| 340 | low, |
| 341 | // These are really $unsigned values, but store as $ty: |
| 342 | range: range.cast(), |
| 343 | thresh: ints_to_reject.cast(), |
| 344 | }) |
| 345 | } |
| 346 | |
| 347 | fn sample<R: Rng + ?Sized>(&self, rng: &mut R) -> Self::X { |
| 348 | let range: Simd<$unsigned, LANES> = self.range.cast(); |
| 349 | let thresh: Simd<$unsigned, LANES> = self.thresh.cast(); |
| 350 | |
| 351 | // This might seem very slow, generating a whole new |
| 352 | // SIMD vector for every sample rejection. For most uses |
| 353 | // though, the chance of rejection is small and provides good |
| 354 | // general performance. With multiple lanes, that chance is |
| 355 | // multiplied. To mitigate this, we replace only the lanes of |
| 356 | // the vector which fail, iteratively reducing the chance of |
| 357 | // rejection. The replacement method does however add a little |
| 358 | // overhead. Benchmarking or calculating probabilities might |
| 359 | // reveal contexts where this replacement method is slower. |
| 360 | let mut v: Simd<$unsigned, LANES> = rng.random(); |
| 361 | loop { |
| 362 | let (hi, lo) = v.wmul(range); |
| 363 | let mask = lo.simd_ge(thresh); |
| 364 | if mask.all() { |
| 365 | let hi: Simd<$ty, LANES> = hi.cast(); |
| 366 | // wrapping addition |
| 367 | let result = self.low + hi; |
| 368 | // `select` here compiles to a blend operation |
| 369 | // When `range.eq(0).none()` the compare and blend |
| 370 | // operations are avoided. |
| 371 | let v: Simd<$ty, LANES> = v.cast(); |
| 372 | return range.simd_gt(Simd::splat(0)).select(result, v); |
| 373 | } |
| 374 | // Replace only the failing lanes |
| 375 | v = mask.select(v, rng.random()); |
| 376 | } |
| 377 | } |
| 378 | } |
| 379 | }; |
| 380 | |
| 381 | // bulk implementation |
| 382 | ($(($unsigned:ident, $signed:ident)),+) => { |
| 383 | $( |
| 384 | uniform_simd_int_impl!($unsigned, $unsigned); |
| 385 | uniform_simd_int_impl!($signed, $unsigned); |
| 386 | )+ |
| 387 | }; |
| 388 | } |
| 389 | |
| 390 | #[cfg (feature = "simd_support" )] |
| 391 | uniform_simd_int_impl! { (u8, i8), (u16, i16), (u32, i32), (u64, i64) } |
| 392 | |
| 393 | /// The back-end implementing [`UniformSampler`] for `usize`. |
| 394 | /// |
| 395 | /// # Implementation notes |
| 396 | /// |
| 397 | /// Sampling a `usize` value is usually used in relation to the length of an |
| 398 | /// array or other memory structure, thus it is reasonable to assume that the |
| 399 | /// vast majority of use-cases will have a maximum size under [`u32::MAX`]. |
| 400 | /// In part to optimise for this use-case, but mostly to ensure that results |
| 401 | /// are portable across 32-bit and 64-bit architectures (as far as is possible), |
| 402 | /// this implementation will use 32-bit sampling when possible. |
| 403 | #[cfg (any(target_pointer_width = "32" , target_pointer_width = "64" ))] |
| 404 | #[derive(Clone, Copy, Debug, PartialEq, Eq)] |
| 405 | #[cfg_attr (all(feature = "serde" ), derive(Serialize))] |
| 406 | // To be able to deserialize on 32-bit we need to replace this with a custom |
| 407 | // implementation of the Deserialize trait, to be able to: |
| 408 | // - panic when `mode64` is `true` on 32-bit, |
| 409 | // - assign the default value to `mode64` when it's missing on 64-bit, |
| 410 | // - panic when the `usize` fields are greater than `u32::MAX` on 32-bit. |
| 411 | #[cfg_attr ( |
| 412 | all(feature = "serde" , target_pointer_width = "64" ), |
| 413 | derive(Deserialize) |
| 414 | )] |
| 415 | pub struct UniformUsize { |
| 416 | /// The lowest possible value. |
| 417 | low: usize, |
| 418 | /// The number of possible values. `0` has a special meaning: all. |
| 419 | range: usize, |
| 420 | /// Threshold used when sampling to obtain a uniform distribution. |
| 421 | thresh: usize, |
| 422 | /// Whether the largest possible value is greater than `u32::MAX`. |
| 423 | #[cfg (target_pointer_width = "64" )] |
| 424 | // Handle missing field when deserializing on 64-bit an object serialized |
| 425 | // on 32-bit. Can be removed when switching to a custom deserializer. |
| 426 | #[cfg_attr (feature = "serde" , serde(default))] |
| 427 | mode64: bool, |
| 428 | } |
| 429 | |
| 430 | #[cfg (any(target_pointer_width = "32" , target_pointer_width = "64" ))] |
| 431 | impl SampleUniform for usize { |
| 432 | type Sampler = UniformUsize; |
| 433 | } |
| 434 | |
| 435 | #[cfg (any(target_pointer_width = "32" , target_pointer_width = "64" ))] |
| 436 | impl UniformSampler for UniformUsize { |
| 437 | type X = usize; |
| 438 | |
| 439 | #[inline ] // if the range is constant, this helps LLVM to do the |
| 440 | // calculations at compile-time. |
| 441 | fn new<B1, B2>(low_b: B1, high_b: B2) -> Result<Self, Error> |
| 442 | where |
| 443 | B1: SampleBorrow<Self::X> + Sized, |
| 444 | B2: SampleBorrow<Self::X> + Sized, |
| 445 | { |
| 446 | let low = *low_b.borrow(); |
| 447 | let high = *high_b.borrow(); |
| 448 | if !(low < high) { |
| 449 | return Err(Error::EmptyRange); |
| 450 | } |
| 451 | |
| 452 | UniformSampler::new_inclusive(low, high - 1) |
| 453 | } |
| 454 | |
| 455 | #[inline ] // if the range is constant, this helps LLVM to do the |
| 456 | // calculations at compile-time. |
| 457 | fn new_inclusive<B1, B2>(low_b: B1, high_b: B2) -> Result<Self, Error> |
| 458 | where |
| 459 | B1: SampleBorrow<Self::X> + Sized, |
| 460 | B2: SampleBorrow<Self::X> + Sized, |
| 461 | { |
| 462 | let low = *low_b.borrow(); |
| 463 | let high = *high_b.borrow(); |
| 464 | if !(low <= high) { |
| 465 | return Err(Error::EmptyRange); |
| 466 | } |
| 467 | |
| 468 | #[cfg (target_pointer_width = "64" )] |
| 469 | let mode64 = high > (u32::MAX as usize); |
| 470 | #[cfg (target_pointer_width = "32" )] |
| 471 | let mode64 = false; |
| 472 | |
| 473 | let (range, thresh); |
| 474 | if cfg!(target_pointer_width = "64" ) && !mode64 { |
| 475 | let range32 = (high as u32).wrapping_sub(low as u32).wrapping_add(1); |
| 476 | range = range32 as usize; |
| 477 | thresh = if range32 > 0 { |
| 478 | (range32.wrapping_neg() % range32) as usize |
| 479 | } else { |
| 480 | 0 |
| 481 | }; |
| 482 | } else { |
| 483 | range = high.wrapping_sub(low).wrapping_add(1); |
| 484 | thresh = if range > 0 { |
| 485 | range.wrapping_neg() % range |
| 486 | } else { |
| 487 | 0 |
| 488 | }; |
| 489 | } |
| 490 | |
| 491 | Ok(UniformUsize { |
| 492 | low, |
| 493 | range, |
| 494 | thresh, |
| 495 | #[cfg (target_pointer_width = "64" )] |
| 496 | mode64, |
| 497 | }) |
| 498 | } |
| 499 | |
| 500 | #[inline ] |
| 501 | fn sample<R: Rng + ?Sized>(&self, rng: &mut R) -> usize { |
| 502 | #[cfg (target_pointer_width = "32" )] |
| 503 | let mode32 = true; |
| 504 | #[cfg (target_pointer_width = "64" )] |
| 505 | let mode32 = !self.mode64; |
| 506 | |
| 507 | if mode32 { |
| 508 | let range = self.range as u32; |
| 509 | if range == 0 { |
| 510 | return rng.random::<u32>() as usize; |
| 511 | } |
| 512 | |
| 513 | let thresh = self.thresh as u32; |
| 514 | let hi = loop { |
| 515 | let (hi, lo) = rng.random::<u32>().wmul(range); |
| 516 | if lo >= thresh { |
| 517 | break hi; |
| 518 | } |
| 519 | }; |
| 520 | self.low.wrapping_add(hi as usize) |
| 521 | } else { |
| 522 | let range = self.range as u64; |
| 523 | if range == 0 { |
| 524 | return rng.random::<u64>() as usize; |
| 525 | } |
| 526 | |
| 527 | let thresh = self.thresh as u64; |
| 528 | let hi = loop { |
| 529 | let (hi, lo) = rng.random::<u64>().wmul(range); |
| 530 | if lo >= thresh { |
| 531 | break hi; |
| 532 | } |
| 533 | }; |
| 534 | self.low.wrapping_add(hi as usize) |
| 535 | } |
| 536 | } |
| 537 | |
| 538 | #[inline ] |
| 539 | fn sample_single<R: Rng + ?Sized, B1, B2>( |
| 540 | low_b: B1, |
| 541 | high_b: B2, |
| 542 | rng: &mut R, |
| 543 | ) -> Result<Self::X, Error> |
| 544 | where |
| 545 | B1: SampleBorrow<Self::X> + Sized, |
| 546 | B2: SampleBorrow<Self::X> + Sized, |
| 547 | { |
| 548 | let low = *low_b.borrow(); |
| 549 | let high = *high_b.borrow(); |
| 550 | if !(low < high) { |
| 551 | return Err(Error::EmptyRange); |
| 552 | } |
| 553 | |
| 554 | if cfg!(target_pointer_width = "64" ) && high > (u32::MAX as usize) { |
| 555 | return UniformInt::<u64>::sample_single(low as u64, high as u64, rng) |
| 556 | .map(|x| x as usize); |
| 557 | } |
| 558 | |
| 559 | UniformInt::<u32>::sample_single(low as u32, high as u32, rng).map(|x| x as usize) |
| 560 | } |
| 561 | |
| 562 | #[inline ] |
| 563 | fn sample_single_inclusive<R: Rng + ?Sized, B1, B2>( |
| 564 | low_b: B1, |
| 565 | high_b: B2, |
| 566 | rng: &mut R, |
| 567 | ) -> Result<Self::X, Error> |
| 568 | where |
| 569 | B1: SampleBorrow<Self::X> + Sized, |
| 570 | B2: SampleBorrow<Self::X> + Sized, |
| 571 | { |
| 572 | let low = *low_b.borrow(); |
| 573 | let high = *high_b.borrow(); |
| 574 | if !(low <= high) { |
| 575 | return Err(Error::EmptyRange); |
| 576 | } |
| 577 | |
| 578 | if cfg!(target_pointer_width = "64" ) && high > (u32::MAX as usize) { |
| 579 | return UniformInt::<u64>::sample_single_inclusive(low as u64, high as u64, rng) |
| 580 | .map(|x| x as usize); |
| 581 | } |
| 582 | |
| 583 | UniformInt::<u32>::sample_single_inclusive(low as u32, high as u32, rng).map(|x| x as usize) |
| 584 | } |
| 585 | } |
| 586 | |
| 587 | #[cfg (test)] |
| 588 | mod tests { |
| 589 | use super::*; |
| 590 | use crate::distr::{Distribution, Uniform}; |
| 591 | use core::fmt::Debug; |
| 592 | use core::ops::Add; |
| 593 | |
| 594 | #[test] |
| 595 | fn test_uniform_bad_limits_equal_int() { |
| 596 | assert_eq!(Uniform::new(10, 10), Err(Error::EmptyRange)); |
| 597 | } |
| 598 | |
| 599 | #[test] |
| 600 | fn test_uniform_good_limits_equal_int() { |
| 601 | let mut rng = crate::test::rng(804); |
| 602 | let dist = Uniform::new_inclusive(10, 10).unwrap(); |
| 603 | for _ in 0..20 { |
| 604 | assert_eq!(rng.sample(dist), 10); |
| 605 | } |
| 606 | } |
| 607 | |
| 608 | #[test] |
| 609 | fn test_uniform_bad_limits_flipped_int() { |
| 610 | assert_eq!(Uniform::new(10, 5), Err(Error::EmptyRange)); |
| 611 | } |
| 612 | |
| 613 | #[test] |
| 614 | #[cfg_attr (miri, ignore)] // Miri is too slow |
| 615 | fn test_integers() { |
| 616 | let mut rng = crate::test::rng(251); |
| 617 | macro_rules! t { |
| 618 | ($ty:ident, $v:expr, $le:expr, $lt:expr) => {{ |
| 619 | for &(low, high) in $v.iter() { |
| 620 | let my_uniform = Uniform::new(low, high).unwrap(); |
| 621 | for _ in 0..1000 { |
| 622 | let v: $ty = rng.sample(my_uniform); |
| 623 | assert!($le(low, v) && $lt(v, high)); |
| 624 | } |
| 625 | |
| 626 | let my_uniform = Uniform::new_inclusive(low, high).unwrap(); |
| 627 | for _ in 0..1000 { |
| 628 | let v: $ty = rng.sample(my_uniform); |
| 629 | assert!($le(low, v) && $le(v, high)); |
| 630 | } |
| 631 | |
| 632 | let my_uniform = Uniform::new(&low, high).unwrap(); |
| 633 | for _ in 0..1000 { |
| 634 | let v: $ty = rng.sample(my_uniform); |
| 635 | assert!($le(low, v) && $lt(v, high)); |
| 636 | } |
| 637 | |
| 638 | let my_uniform = Uniform::new_inclusive(&low, &high).unwrap(); |
| 639 | for _ in 0..1000 { |
| 640 | let v: $ty = rng.sample(my_uniform); |
| 641 | assert!($le(low, v) && $le(v, high)); |
| 642 | } |
| 643 | |
| 644 | for _ in 0..1000 { |
| 645 | let v = <$ty as SampleUniform>::Sampler::sample_single(low, high, &mut rng).unwrap(); |
| 646 | assert!($le(low, v) && $lt(v, high)); |
| 647 | } |
| 648 | |
| 649 | for _ in 0..1000 { |
| 650 | let v = <$ty as SampleUniform>::Sampler::sample_single_inclusive(low, high, &mut rng).unwrap(); |
| 651 | assert!($le(low, v) && $le(v, high)); |
| 652 | } |
| 653 | } |
| 654 | }}; |
| 655 | |
| 656 | // scalar bulk |
| 657 | ($($ty:ident),*) => {{ |
| 658 | $(t!( |
| 659 | $ty, |
| 660 | [(0, 10), (10, 127), ($ty::MIN, $ty::MAX)], |
| 661 | |x, y| x <= y, |
| 662 | |x, y| x < y |
| 663 | );)* |
| 664 | }}; |
| 665 | |
| 666 | // simd bulk |
| 667 | ($($ty:ident),* => $scalar:ident) => {{ |
| 668 | $(t!( |
| 669 | $ty, |
| 670 | [ |
| 671 | ($ty::splat(0), $ty::splat(10)), |
| 672 | ($ty::splat(10), $ty::splat(127)), |
| 673 | ($ty::splat($scalar::MIN), $ty::splat($scalar::MAX)), |
| 674 | ], |
| 675 | |x: $ty, y| x.simd_le(y).all(), |
| 676 | |x: $ty, y| x.simd_lt(y).all() |
| 677 | );)* |
| 678 | }}; |
| 679 | } |
| 680 | t!(i8, i16, i32, i64, i128, u8, u16, u32, u64, usize, u128); |
| 681 | |
| 682 | #[cfg (feature = "simd_support" )] |
| 683 | { |
| 684 | t!(u8x4, u8x8, u8x16, u8x32, u8x64 => u8); |
| 685 | t!(i8x4, i8x8, i8x16, i8x32, i8x64 => i8); |
| 686 | t!(u16x2, u16x4, u16x8, u16x16, u16x32 => u16); |
| 687 | t!(i16x2, i16x4, i16x8, i16x16, i16x32 => i16); |
| 688 | t!(u32x2, u32x4, u32x8, u32x16 => u32); |
| 689 | t!(i32x2, i32x4, i32x8, i32x16 => i32); |
| 690 | t!(u64x2, u64x4, u64x8 => u64); |
| 691 | t!(i64x2, i64x4, i64x8 => i64); |
| 692 | } |
| 693 | } |
| 694 | |
| 695 | #[test] |
| 696 | fn test_uniform_from_std_range() { |
| 697 | let r = Uniform::try_from(2u32..7).unwrap(); |
| 698 | assert_eq!(r.0.low, 2); |
| 699 | assert_eq!(r.0.range, 5); |
| 700 | } |
| 701 | |
| 702 | #[test] |
| 703 | fn test_uniform_from_std_range_bad_limits() { |
| 704 | #![allow (clippy::reversed_empty_ranges)] |
| 705 | assert!(Uniform::try_from(100..10).is_err()); |
| 706 | assert!(Uniform::try_from(100..100).is_err()); |
| 707 | } |
| 708 | |
| 709 | #[test] |
| 710 | fn test_uniform_from_std_range_inclusive() { |
| 711 | let r = Uniform::try_from(2u32..=6).unwrap(); |
| 712 | assert_eq!(r.0.low, 2); |
| 713 | assert_eq!(r.0.range, 5); |
| 714 | } |
| 715 | |
| 716 | #[test] |
| 717 | fn test_uniform_from_std_range_inclusive_bad_limits() { |
| 718 | #![allow (clippy::reversed_empty_ranges)] |
| 719 | assert!(Uniform::try_from(100..=10).is_err()); |
| 720 | assert!(Uniform::try_from(100..=99).is_err()); |
| 721 | } |
| 722 | |
| 723 | #[test] |
| 724 | fn value_stability() { |
| 725 | fn test_samples<T: SampleUniform + Copy + Debug + PartialEq + Add<T>>( |
| 726 | lb: T, |
| 727 | ub: T, |
| 728 | ub_excl: T, |
| 729 | expected: &[T], |
| 730 | ) where |
| 731 | Uniform<T>: Distribution<T>, |
| 732 | { |
| 733 | let mut rng = crate::test::rng(897); |
| 734 | let mut buf = [lb; 6]; |
| 735 | |
| 736 | for x in &mut buf[0..3] { |
| 737 | *x = T::Sampler::sample_single_inclusive(lb, ub, &mut rng).unwrap(); |
| 738 | } |
| 739 | |
| 740 | let distr = Uniform::new_inclusive(lb, ub).unwrap(); |
| 741 | for x in &mut buf[3..6] { |
| 742 | *x = rng.sample(&distr); |
| 743 | } |
| 744 | assert_eq!(&buf, expected); |
| 745 | |
| 746 | let mut rng = crate::test::rng(897); |
| 747 | |
| 748 | for x in &mut buf[0..3] { |
| 749 | *x = T::Sampler::sample_single(lb, ub_excl, &mut rng).unwrap(); |
| 750 | } |
| 751 | |
| 752 | let distr = Uniform::new(lb, ub_excl).unwrap(); |
| 753 | for x in &mut buf[3..6] { |
| 754 | *x = rng.sample(&distr); |
| 755 | } |
| 756 | assert_eq!(&buf, expected); |
| 757 | } |
| 758 | |
| 759 | test_samples(-105i8, 111, 112, &[-99, -48, 107, 72, -19, 56]); |
| 760 | test_samples(2i16, 1352, 1353, &[43, 361, 1325, 1109, 539, 1005]); |
| 761 | test_samples( |
| 762 | -313853i32, |
| 763 | 13513, |
| 764 | 13514, |
| 765 | &[-303803, -226673, 6912, -45605, -183505, -70668], |
| 766 | ); |
| 767 | test_samples( |
| 768 | 131521i64, |
| 769 | 6542165, |
| 770 | 6542166, |
| 771 | &[1838724, 5384489, 4893692, 3712948, 3951509, 4094926], |
| 772 | ); |
| 773 | test_samples( |
| 774 | -0x8000_0000_0000_0000_0000_0000_0000_0000i128, |
| 775 | -1, |
| 776 | 0, |
| 777 | &[ |
| 778 | -30725222750250982319765550926688025855, |
| 779 | -75088619368053423329503924805178012357, |
| 780 | -64950748766625548510467638647674468829, |
| 781 | -41794017901603587121582892414659436495, |
| 782 | -63623852319608406524605295913876414006, |
| 783 | -17404679390297612013597359206379189023, |
| 784 | ], |
| 785 | ); |
| 786 | test_samples(11u8, 218, 219, &[17, 66, 214, 181, 93, 165]); |
| 787 | test_samples(11u16, 218, 219, &[17, 66, 214, 181, 93, 165]); |
| 788 | test_samples(11u32, 218, 219, &[17, 66, 214, 181, 93, 165]); |
| 789 | test_samples(11u64, 218, 219, &[66, 181, 165, 127, 134, 139]); |
| 790 | test_samples(11u128, 218, 219, &[181, 127, 139, 167, 141, 197]); |
| 791 | test_samples(11usize, 218, 219, &[17, 66, 214, 181, 93, 165]); |
| 792 | |
| 793 | #[cfg (feature = "simd_support" )] |
| 794 | { |
| 795 | let lb = Simd::from([11u8, 0, 128, 127]); |
| 796 | let ub = Simd::from([218, 254, 254, 254]); |
| 797 | let ub_excl = ub + Simd::splat(1); |
| 798 | test_samples( |
| 799 | lb, |
| 800 | ub, |
| 801 | ub_excl, |
| 802 | &[ |
| 803 | Simd::from([13, 5, 237, 130]), |
| 804 | Simd::from([126, 186, 149, 161]), |
| 805 | Simd::from([103, 86, 234, 252]), |
| 806 | Simd::from([35, 18, 225, 231]), |
| 807 | Simd::from([106, 153, 246, 177]), |
| 808 | Simd::from([195, 168, 149, 222]), |
| 809 | ], |
| 810 | ); |
| 811 | } |
| 812 | } |
| 813 | |
| 814 | #[test] |
| 815 | fn test_uniform_usize_empty_range() { |
| 816 | assert_eq!(UniformUsize::new(10, 10), Err(Error::EmptyRange)); |
| 817 | assert!(UniformUsize::new(10, 11).is_ok()); |
| 818 | |
| 819 | assert_eq!(UniformUsize::new_inclusive(10, 9), Err(Error::EmptyRange)); |
| 820 | assert!(UniformUsize::new_inclusive(10, 10).is_ok()); |
| 821 | } |
| 822 | |
| 823 | #[test] |
| 824 | fn test_uniform_usize_constructors() { |
| 825 | assert_eq!( |
| 826 | UniformUsize::new_inclusive(u32::MAX as usize, u32::MAX as usize), |
| 827 | Ok(UniformUsize { |
| 828 | low: u32::MAX as usize, |
| 829 | range: 1, |
| 830 | thresh: 0, |
| 831 | #[cfg(target_pointer_width = "64" )] |
| 832 | mode64: false |
| 833 | }) |
| 834 | ); |
| 835 | assert_eq!( |
| 836 | UniformUsize::new_inclusive(0, u32::MAX as usize), |
| 837 | Ok(UniformUsize { |
| 838 | low: 0, |
| 839 | range: 0, |
| 840 | thresh: 0, |
| 841 | #[cfg(target_pointer_width = "64" )] |
| 842 | mode64: false |
| 843 | }) |
| 844 | ); |
| 845 | #[cfg (target_pointer_width = "64" )] |
| 846 | assert_eq!( |
| 847 | UniformUsize::new_inclusive(0, u32::MAX as usize + 1), |
| 848 | Ok(UniformUsize { |
| 849 | low: 0, |
| 850 | range: u32::MAX as usize + 2, |
| 851 | thresh: 1, |
| 852 | mode64: true |
| 853 | }) |
| 854 | ); |
| 855 | #[cfg (target_pointer_width = "64" )] |
| 856 | assert_eq!( |
| 857 | UniformUsize::new_inclusive(u32::MAX as usize, u64::MAX as usize), |
| 858 | Ok(UniformUsize { |
| 859 | low: u32::MAX as usize, |
| 860 | range: u64::MAX as usize - u32::MAX as usize + 1, |
| 861 | thresh: u32::MAX as usize, |
| 862 | mode64: true |
| 863 | }) |
| 864 | ); |
| 865 | } |
| 866 | |
| 867 | // This could be run also on 32-bit when deserialization is implemented. |
| 868 | #[cfg (all(feature = "serde" , target_pointer_width = "64" ))] |
| 869 | #[test] |
| 870 | fn test_uniform_usize_deserialization() { |
| 871 | use serde_json; |
| 872 | let original = UniformUsize::new_inclusive(10, 100).expect("creation" ); |
| 873 | let serialized = serde_json::to_string(&original).expect("serialization" ); |
| 874 | let deserialized: UniformUsize = |
| 875 | serde_json::from_str(&serialized).expect("deserialization" ); |
| 876 | assert_eq!(deserialized, original); |
| 877 | } |
| 878 | |
| 879 | #[cfg (all(feature = "serde" , target_pointer_width = "64" ))] |
| 880 | #[test] |
| 881 | fn test_uniform_usize_deserialization_from_32bit() { |
| 882 | use serde_json; |
| 883 | let serialized_on_32bit = r#"{"low":10,"range":91,"thresh":74}"# ; |
| 884 | let deserialized: UniformUsize = |
| 885 | serde_json::from_str(&serialized_on_32bit).expect("deserialization" ); |
| 886 | assert_eq!( |
| 887 | deserialized, |
| 888 | UniformUsize::new_inclusive(10, 100).expect("creation" ) |
| 889 | ); |
| 890 | } |
| 891 | |
| 892 | #[cfg (all(feature = "serde" , target_pointer_width = "64" ))] |
| 893 | #[test] |
| 894 | fn test_uniform_usize_deserialization_64bit() { |
| 895 | use serde_json; |
| 896 | let original = UniformUsize::new_inclusive(1, u64::MAX as usize - 1).expect("creation" ); |
| 897 | assert!(original.mode64); |
| 898 | let serialized = serde_json::to_string(&original).expect("serialization" ); |
| 899 | let deserialized: UniformUsize = |
| 900 | serde_json::from_str(&serialized).expect("deserialization" ); |
| 901 | assert_eq!(deserialized, original); |
| 902 | } |
| 903 | } |
| 904 | |