aboutsummaryrefslogtreecommitdiff
path: root/rand/src/distributions/utils.rs
blob: a2112fd49dc68ea2930e6c14e37bf1d7c9ad351d (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
// Copyright 2018 Developers of the Rand project.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// https://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.

//! Math helper functions

#[cfg(feature="simd_support")]
use packed_simd::*;
#[cfg(feature="std")]
use distributions::ziggurat_tables;
#[cfg(feature="std")]
use Rng;


pub trait WideningMultiply<RHS = Self> {
    type Output;

    fn wmul(self, x: RHS) -> Self::Output;
}

macro_rules! wmul_impl {
    ($ty:ty, $wide:ty, $shift:expr) => {
        impl WideningMultiply for $ty {
            type Output = ($ty, $ty);

            #[inline(always)]
            fn wmul(self, x: $ty) -> Self::Output {
                let tmp = (self as $wide) * (x as $wide);
                ((tmp >> $shift) as $ty, tmp as $ty)
            }
        }
    };

    // simd bulk implementation
    ($(($ty:ident, $wide:ident),)+, $shift:expr) => {
        $(
            impl WideningMultiply for $ty {
                type Output = ($ty, $ty);

                #[inline(always)]
                fn wmul(self, x: $ty) -> Self::Output {
                    // For supported vectors, this should compile to a couple
                    // supported multiply & swizzle instructions (no actual
                    // casting).
                    // TODO: optimize
                    let y: $wide = self.cast();
                    let x: $wide = x.cast();
                    let tmp = y * x;
                    let hi: $ty = (tmp >> $shift).cast();
                    let lo: $ty = tmp.cast();
                    (hi, lo)
                }
            }
        )+
    };
}
wmul_impl! { u8, u16, 8 }
wmul_impl! { u16, u32, 16 }
wmul_impl! { u32, u64, 32 }
#[cfg(rust_1_26)]
wmul_impl! { u64, u128, 64 }

// This code is a translation of the __mulddi3 function in LLVM's
// compiler-rt. It is an optimised variant of the common method
// `(a + b) * (c + d) = ac + ad + bc + bd`.
//
// For some reason LLVM can optimise the C version very well, but
// keeps shuffling registers in this Rust translation.
macro_rules! wmul_impl_large {
    ($ty:ty, $half:expr) => {
        impl WideningMultiply for $ty {
            type Output = ($ty, $ty);

            #[inline(always)]
            fn wmul(self, b: $ty) -> Self::Output {
                const LOWER_MASK: $ty = !0 >> $half;
                let mut low = (self & LOWER_MASK).wrapping_mul(b & LOWER_MASK);
                let mut t = low >> $half;
                low &= LOWER_MASK;
                t += (self >> $half).wrapping_mul(b & LOWER_MASK);
                low += (t & LOWER_MASK) << $half;
                let mut high = t >> $half;
                t = low >> $half;
                low &= LOWER_MASK;
                t += (b >> $half).wrapping_mul(self & LOWER_MASK);
                low += (t & LOWER_MASK) << $half;
                high += t >> $half;
                high += (self >> $half).wrapping_mul(b >> $half);

                (high, low)
            }
        }
    };

    // simd bulk implementation
    (($($ty:ty,)+) $scalar:ty, $half:expr) => {
        $(
            impl WideningMultiply for $ty {
                type Output = ($ty, $ty);

                #[inline(always)]
                fn wmul(self, b: $ty) -> Self::Output {
                    // needs wrapping multiplication
                    const LOWER_MASK: $scalar = !0 >> $half;
                    let mut low = (self & LOWER_MASK) * (b & LOWER_MASK);
                    let mut t = low >> $half;
                    low &= LOWER_MASK;
                    t += (self >> $half) * (b & LOWER_MASK);
                    low += (t & LOWER_MASK) << $half;
                    let mut high = t >> $half;
                    t = low >> $half;
                    low &= LOWER_MASK;
                    t += (b >> $half) * (self & LOWER_MASK);
                    low += (t & LOWER_MASK) << $half;
                    high += t >> $half;
                    high += (self >> $half) * (b >> $half);

                    (high, low)
                }
            }
        )+
    };
}
#[cfg(not(rust_1_26))]
wmul_impl_large! { u64, 32 }
#[cfg(rust_1_26)]
wmul_impl_large! { u128, 64 }

macro_rules! wmul_impl_usize {
    ($ty:ty) => {
        impl WideningMultiply for usize {
            type Output = (usize, usize);

            #[inline(always)]
            fn wmul(self, x: usize) -> Self::Output {
                let (high, low) = (self as $ty).wmul(x as $ty);
                (high as usize, low as usize)
            }
        }
    }
}
#[cfg(target_pointer_width = "32")]
wmul_impl_usize! { u32 }
#[cfg(target_pointer_width = "64")]
wmul_impl_usize! { u64 }

#[cfg(all(feature = "simd_support", feature = "nightly"))]
mod simd_wmul {
    #[cfg(target_arch = "x86")]
    use core::arch::x86::*;
    #[cfg(target_arch = "x86_64")]
    use core::arch::x86_64::*;
    use super::*;

    wmul_impl! {
        (u8x2, u16x2),
        (u8x4, u16x4),
        (u8x8, u16x8),
        (u8x16, u16x16),
        (u8x32, u16x32),,
        8
    }

    wmul_impl! { (u16x2, u32x2),, 16 }
    #[cfg(not(target_feature = "sse2"))]
    wmul_impl! { (u16x4, u32x4),, 16 }
    #[cfg(not(target_feature = "sse4.2"))]
    wmul_impl! { (u16x8, u32x8),, 16 }
    #[cfg(not(target_feature = "avx2"))]
    wmul_impl! { (u16x16, u32x16),, 16 }

    // 16-bit lane widths allow use of the x86 `mulhi` instructions, which
    // means `wmul` can be implemented with only two instructions.
    #[allow(unused_macros)]
    macro_rules! wmul_impl_16 {
        ($ty:ident, $intrinsic:ident, $mulhi:ident, $mullo:ident) => {
            impl WideningMultiply for $ty {
                type Output = ($ty, $ty);

                #[inline(always)]
                fn wmul(self, x: $ty) -> Self::Output {
                    let b = $intrinsic::from_bits(x);
                    let a = $intrinsic::from_bits(self);
                    let hi = $ty::from_bits(unsafe { $mulhi(a, b) });
                    let lo = $ty::from_bits(unsafe { $mullo(a, b) });
                    (hi, lo)
                }
            }
        };
    }

    #[cfg(target_feature = "sse2")]
    wmul_impl_16! { u16x4, __m64, _mm_mulhi_pu16, _mm_mullo_pi16 }
    #[cfg(target_feature = "sse4.2")]
    wmul_impl_16! { u16x8, __m128i, _mm_mulhi_epu16, _mm_mullo_epi16 }
    #[cfg(target_feature = "avx2")]
    wmul_impl_16! { u16x16, __m256i, _mm256_mulhi_epu16, _mm256_mullo_epi16 }
    // FIXME: there are no `__m512i` types in stdsimd yet, so `wmul::<u16x32>`
    // cannot use the same implementation.

    wmul_impl! {
        (u32x2, u64x2),
        (u32x4, u64x4),
        (u32x8, u64x8),,
        32
    }

    // TODO: optimize, this seems to seriously slow things down
    wmul_impl_large! { (u8x64,) u8, 4 }
    wmul_impl_large! { (u16x32,) u16, 8 }
    wmul_impl_large! { (u32x16,) u32, 16 }
    wmul_impl_large! { (u64x2, u64x4, u64x8,) u64, 32 }
}
#[cfg(all(feature = "simd_support", feature = "nightly"))]
pub use self::simd_wmul::*;


/// Helper trait when dealing with scalar and SIMD floating point types.
pub(crate) trait FloatSIMDUtils {
    // `PartialOrd` for vectors compares lexicographically. We want to compare all
    // the individual SIMD lanes instead, and get the combined result over all
    // lanes. This is possible using something like `a.lt(b).all()`, but we
    // implement it as a trait so we can write the same code for `f32` and `f64`.
    // Only the comparison functions we need are implemented.
    fn all_lt(self, other: Self) -> bool;
    fn all_le(self, other: Self) -> bool;
    fn all_finite(self) -> bool;

    type Mask;
    fn finite_mask(self) -> Self::Mask;
    fn gt_mask(self, other: Self) -> Self::Mask;
    fn ge_mask(self, other: Self) -> Self::Mask;

    // Decrease all lanes where the mask is `true` to the next lower value
    // representable by the floating-point type. At least one of the lanes
    // must be set.
    fn decrease_masked(self, mask: Self::Mask) -> Self;

    // Convert from int value. Conversion is done while retaining the numerical
    // value, not by retaining the binary representation.
    type UInt;
    fn cast_from_int(i: Self::UInt) -> Self;
}

/// Implement functions available in std builds but missing from core primitives
#[cfg(not(std))]
pub(crate) trait Float : Sized {
    type Bits;

    fn is_nan(self) -> bool;
    fn is_infinite(self) -> bool;
    fn is_finite(self) -> bool;
    fn to_bits(self) -> Self::Bits;
    fn from_bits(v: Self::Bits) -> Self;
}

/// Implement functions on f32/f64 to give them APIs similar to SIMD types
pub(crate) trait FloatAsSIMD : Sized {
    #[inline(always)]
    fn lanes() -> usize { 1 }
    #[inline(always)]
    fn splat(scalar: Self) -> Self { scalar }
    #[inline(always)]
    fn extract(self, index: usize) -> Self { debug_assert_eq!(index, 0); self }
    #[inline(always)]
    fn replace(self, index: usize, new_value: Self) -> Self { debug_assert_eq!(index, 0); new_value }
}

pub(crate) trait BoolAsSIMD : Sized {
    fn any(self) -> bool;
    fn all(self) -> bool;
    fn none(self) -> bool;
}

impl BoolAsSIMD for bool {
    #[inline(always)]
    fn any(self) -> bool { self }
    #[inline(always)]
    fn all(self) -> bool { self }
    #[inline(always)]
    fn none(self) -> bool { !self }
}

macro_rules! scalar_float_impl {
    ($ty:ident, $uty:ident) => {
        #[cfg(not(std))]
        impl Float for $ty {
            type Bits = $uty;

            #[inline]
            fn is_nan(self) -> bool {
                self != self
            }

            #[inline]
            fn is_infinite(self) -> bool {
                self == ::core::$ty::INFINITY || self == ::core::$ty::NEG_INFINITY
            }

            #[inline]
            fn is_finite(self) -> bool {
                !(self.is_nan() || self.is_infinite())
            }

            #[inline]
            fn to_bits(self) -> Self::Bits {
                unsafe { ::core::mem::transmute(self) }
            }

            #[inline]
            fn from_bits(v: Self::Bits) -> Self {
                // It turns out the safety issues with sNaN were overblown! Hooray!
                unsafe { ::core::mem::transmute(v) }
            }
        }

        impl FloatSIMDUtils for $ty {
            type Mask = bool;
            #[inline(always)]
            fn all_lt(self, other: Self) -> bool { self < other }
            #[inline(always)]
            fn all_le(self, other: Self) -> bool { self <= other }
            #[inline(always)]
            fn all_finite(self) -> bool { self.is_finite() }
            #[inline(always)]
            fn finite_mask(self) -> Self::Mask { self.is_finite() }
            #[inline(always)]
            fn gt_mask(self, other: Self) -> Self::Mask { self > other }
            #[inline(always)]
            fn ge_mask(self, other: Self) -> Self::Mask { self >= other }
            #[inline(always)]
            fn decrease_masked(self, mask: Self::Mask) -> Self {
                debug_assert!(mask, "At least one lane must be set");
                <$ty>::from_bits(self.to_bits() - 1)
            }
            type UInt = $uty;
            fn cast_from_int(i: Self::UInt) -> Self { i as $ty }
        }

        impl FloatAsSIMD for $ty {}
    }
}

scalar_float_impl!(f32, u32);
scalar_float_impl!(f64, u64);


#[cfg(feature="simd_support")]
macro_rules! simd_impl {
    ($ty:ident, $f_scalar:ident, $mty:ident, $uty:ident) => {
        impl FloatSIMDUtils for $ty {
            type Mask = $mty;
            #[inline(always)]
            fn all_lt(self, other: Self) -> bool { self.lt(other).all() }
            #[inline(always)]
            fn all_le(self, other: Self) -> bool { self.le(other).all() }
            #[inline(always)]
            fn all_finite(self) -> bool { self.finite_mask().all() }
            #[inline(always)]
            fn finite_mask(self) -> Self::Mask {
                // This can possibly be done faster by checking bit patterns
                let neg_inf = $ty::splat(::core::$f_scalar::NEG_INFINITY);
                let pos_inf = $ty::splat(::core::$f_scalar::INFINITY);
                self.gt(neg_inf) & self.lt(pos_inf)
            }
            #[inline(always)]
            fn gt_mask(self, other: Self) -> Self::Mask { self.gt(other) }
            #[inline(always)]
            fn ge_mask(self, other: Self) -> Self::Mask { self.ge(other) }
            #[inline(always)]
            fn decrease_masked(self, mask: Self::Mask) -> Self {
                // Casting a mask into ints will produce all bits set for
                // true, and 0 for false. Adding that to the binary
                // representation of a float means subtracting one from
                // the binary representation, resulting in the next lower
                // value representable by $ty. This works even when the
                // current value is infinity.
                debug_assert!(mask.any(), "At least one lane must be set");
                <$ty>::from_bits(<$uty>::from_bits(self) + <$uty>::from_bits(mask))
            }
            type UInt = $uty;
            fn cast_from_int(i: Self::UInt) -> Self { i.cast() }
        }
    }
}

#[cfg(feature="simd_support")] simd_impl! { f32x2, f32, m32x2, u32x2 }
#[cfg(feature="simd_support")] simd_impl! { f32x4, f32, m32x4, u32x4 }
#[cfg(feature="simd_support")] simd_impl! { f32x8, f32, m32x8, u32x8 }
#[cfg(feature="simd_support")] simd_impl! { f32x16, f32, m32x16, u32x16 }
#[cfg(feature="simd_support")] simd_impl! { f64x2, f64, m64x2, u64x2 }
#[cfg(feature="simd_support")] simd_impl! { f64x4, f64, m64x4, u64x4 }
#[cfg(feature="simd_support")] simd_impl! { f64x8, f64, m64x8, u64x8 }

/// Calculates ln(gamma(x)) (natural logarithm of the gamma
/// function) using the Lanczos approximation.
///
/// The approximation expresses the gamma function as:
/// `gamma(z+1) = sqrt(2*pi)*(z+g+0.5)^(z+0.5)*exp(-z-g-0.5)*Ag(z)`
/// `g` is an arbitrary constant; we use the approximation with `g=5`.
///
/// Noting that `gamma(z+1) = z*gamma(z)` and applying `ln` to both sides:
/// `ln(gamma(z)) = (z+0.5)*ln(z+g+0.5)-(z+g+0.5) + ln(sqrt(2*pi)*Ag(z)/z)`
///
/// `Ag(z)` is an infinite series with coefficients that can be calculated
/// ahead of time - we use just the first 6 terms, which is good enough
/// for most purposes.
#[cfg(feature="std")]
pub fn log_gamma(x: f64) -> f64 {
    // precalculated 6 coefficients for the first 6 terms of the series
    let coefficients: [f64; 6] = [
        76.18009172947146,
        -86.50532032941677,
        24.01409824083091,
        -1.231739572450155,
        0.1208650973866179e-2,
        -0.5395239384953e-5,
    ];

    // (x+0.5)*ln(x+g+0.5)-(x+g+0.5)
    let tmp = x + 5.5;
    let log = (x + 0.5) * tmp.ln() - tmp;

    // the first few terms of the series for Ag(x)
    let mut a = 1.000000000190015;
    let mut denom = x;
    for coeff in &coefficients {
        denom += 1.0;
        a += coeff / denom;
    }

    // get everything together
    // a is Ag(x)
    // 2.5066... is sqrt(2pi)
    log + (2.5066282746310005 * a / x).ln()
}

/// Sample a random number using the Ziggurat method (specifically the
/// ZIGNOR variant from Doornik 2005). Most of the arguments are
/// directly from the paper:
///
/// * `rng`: source of randomness
/// * `symmetric`: whether this is a symmetric distribution, or one-sided with P(x < 0) = 0.
/// * `X`: the $x_i$ abscissae.
/// * `F`: precomputed values of the PDF at the $x_i$, (i.e. $f(x_i)$)
/// * `F_DIFF`: precomputed values of $f(x_i) - f(x_{i+1})$
/// * `pdf`: the probability density function
/// * `zero_case`: manual sampling from the tail when we chose the
///    bottom box (i.e. i == 0)

// the perf improvement (25-50%) is definitely worth the extra code
// size from force-inlining.
#[cfg(feature="std")]
#[inline(always)]
pub fn ziggurat<R: Rng + ?Sized, P, Z>(
            rng: &mut R,
            symmetric: bool,
            x_tab: ziggurat_tables::ZigTable,
            f_tab: ziggurat_tables::ZigTable,
            mut pdf: P,
            mut zero_case: Z)
            -> f64 where P: FnMut(f64) -> f64, Z: FnMut(&mut R, f64) -> f64 {
    use distributions::float::IntoFloat;
    loop {
        // As an optimisation we re-implement the conversion to a f64.
        // From the remaining 12 most significant bits we use 8 to construct `i`.
        // This saves us generating a whole extra random number, while the added
        // precision of using 64 bits for f64 does not buy us much.
        let bits = rng.next_u64();
        let i = bits as usize & 0xff;

        let u = if symmetric {
            // Convert to a value in the range [2,4) and substract to get [-1,1)
            // We can't convert to an open range directly, that would require
            // substracting `3.0 - EPSILON`, which is not representable.
            // It is possible with an extra step, but an open range does not
            // seem neccesary for the ziggurat algorithm anyway.
            (bits >> 12).into_float_with_exponent(1) - 3.0
        } else {
            // Convert to a value in the range [1,2) and substract to get (0,1)
            (bits >> 12).into_float_with_exponent(0)
            - (1.0 - ::core::f64::EPSILON / 2.0)
        };
        let x = u * x_tab[i];

        let test_x = if symmetric { x.abs() } else {x};

        // algebraically equivalent to |u| < x_tab[i+1]/x_tab[i] (or u < x_tab[i+1]/x_tab[i])
        if test_x < x_tab[i + 1] {
            return x;
        }
        if i == 0 {
            return zero_case(rng, u);
        }
        // algebraically equivalent to f1 + DRanU()*(f0 - f1) < 1
        if f_tab[i + 1] + (f_tab[i] - f_tab[i + 1]) * rng.gen::<f64>() < pdf(x) {
            return x;
        }
    }
}