|
1 /* |
|
2 * Copyright (c) 2016, 2019, Oracle and/or its affiliates. All rights reserved. |
|
3 * ORACLE PROPRIETARY/CONFIDENTIAL. Use is subject to license terms. |
|
4 * |
|
5 * |
|
6 * |
|
7 * |
|
8 * |
|
9 * |
|
10 * |
|
11 * |
|
12 * |
|
13 * |
|
14 * |
|
15 * |
|
16 * |
|
17 * |
|
18 * |
|
19 * |
|
20 * |
|
21 * |
|
22 * |
|
23 * |
|
24 */ |
|
25 |
|
26 // package java.util; |
|
27 |
|
28 import java.math.BigInteger; |
|
29 import java.util.concurrent.atomic.AtomicLong; |
|
30 |
|
31 /** |
|
32 * A generator of uniform pseudorandom values applicable for use in |
|
33 * (among other contexts) isolated parallel computations that may |
|
34 * generate subtasks. Class {@code Xoroshiro128StarStar} implements |
|
35 * interfaces {@link java.util.Rng} and {@link java.util.LeapableRng}, |
|
36 * and therefore supports methods for producing pseudorandomly chosen |
|
37 * numbers of type {@code int}, {@code long}, {@code float}, and {@code double} |
|
38 * as well as creating new {@code Xoroshiro128StarStar} objects |
|
39 * by "jumping" or "leaping". |
|
40 * |
|
41 * <p>Series of generated values pass the TestU01 BigCrush and PractRand test suites |
|
42 * that measure independence and uniformity properties of random number generators. |
|
43 * |
|
44 * <p>The class {@code Xoroshiro128StarStar} uses the {@code xoroshiro128} algorithm, |
|
45 * version 1.0 (parameters 24, 16, 37), with the "**" scrambler (a mixing function). |
|
46 * Its state consists of two {@code long} fields {@code x0} and {@code x1}, |
|
47 * which can take on any values provided that they are not both zero. |
|
48 * The period of this generator is 2<sup>128</sup>-1. |
|
49 * |
|
50 * <p>The 64-bit values produced by the {@code nextLong()} method are equidistributed. |
|
51 * To be precise, over the course of the cycle of length 2<sup>128</sup>-1, |
|
52 * each nonzero {@code long} value is generated 2<sup>64</sup> times, |
|
53 * but the value 0 is generated only 2<sup>64</sup>-1 times. |
|
54 * The values produced by the {@code nextInt()}, {@code nextFloat()}, and {@code nextDouble()} |
|
55 * methods are likewise equidistributed. |
|
56 * |
|
57 * <p>In fact, the 64-bit values produced by the {@code nextLong()} method are 2-equidistributed. |
|
58 * To be precise: consider the (overlapping) length-2 subsequences of the cycle of 64-bit |
|
59 * values produced by {@code nextLong()} (assuming no other methods are called that would |
|
60 * affect the state). There are 2<sup>128</sup>-1 such subsequences, and each subsequence, |
|
61 * which consists of 2 64-bit values, can have one of 2<sup>128</sup> values. Of those |
|
62 * 2<sup>128</sup> subsequence values, each one is generated exactly once over the course |
|
63 * of the entire cycle, except that the subsequence (0, 0) never appears. |
|
64 * The values produced by the {@code nextInt()}, {@code nextFloat()}, and {@code nextDouble()} |
|
65 * methods are likewise 2-equidistributed, but note that that the subsequence (0, 0) |
|
66 * can also appear (but occurring somewhat less frequently than all other subsequences), |
|
67 * because the values produced by those methods have fewer than 64 randomly chosen bits. |
|
68 * |
|
69 * <p>Instances {@code Xoroshiro128StarStar} are <em>not</em> thread-safe. |
|
70 * They are designed to be used so that each thread as its own instance. |
|
71 * The methods {@link #jump} and {@link #leap} and {@link #jumps} and {@link #leaps} |
|
72 * can be used to construct new instances of {@code Xoroshiro128StarStar} that traverse |
|
73 * other parts of the state cycle. |
|
74 * |
|
75 * <p>Instances of {@code Xoroshiro128StarStar} are not cryptographically |
|
76 * secure. Consider instead using {@link java.security.SecureRandom} |
|
77 * in security-sensitive applications. Additionally, |
|
78 * default-constructed instances do not use a cryptographically random |
|
79 * seed unless the {@linkplain System#getProperty system property} |
|
80 * {@code java.util.secureRandomSeed} is set to {@code true}. |
|
81 * |
|
82 * @author Guy Steele |
|
83 * @author Doug Lea |
|
84 * @since 1.8 |
|
85 */ |
|
86 public final class Xoroshiro128StarStar implements LeapableRng { |
|
87 |
|
88 /* |
|
89 * Implementation Overview. |
|
90 * |
|
91 * This is an implementation of the xoroshiro128** algorithm written |
|
92 * in 2016 by David Blackman and Sebastiano Vigna (vigna@acm.org), |
|
93 * and updated with improved parameters in 2018. |
|
94 * See http://xoshiro.di.unimi.it and these two papers: |
|
95 * |
|
96 * Sebastiano Vigna. 2016. An Experimental Exploration of Marsaglia's |
|
97 * xorshift Generators, Scrambled. ACM Transactions on Mathematical |
|
98 * Software 42, 4, Article 30 (June 2016), 23 pages. |
|
99 * https://doi.org/10.1145/2845077 |
|
100 * |
|
101 * David Blackman and Sebastiano Vigna. 2018. Scrambled Linear |
|
102 * Pseudorandom Number Generators. Computing Research Repository (CoRR). |
|
103 * http://arxiv.org/abs/1805.01407 |
|
104 * |
|
105 * The jump operation moves the current generator forward by 2*64 |
|
106 * steps; this has the same effect as calling nextLong() 2**64 |
|
107 * times, but is much faster. Similarly, the leap operation moves |
|
108 * the current generator forward by 2*96 steps; this has the same |
|
109 * effect as calling nextLong() 2**96 times, but is much faster. |
|
110 * The copy method may be used to make a copy of the current |
|
111 * generator. Thus one may repeatedly and cumulatively copy and |
|
112 * jump to produce a sequence of generators whose states are well |
|
113 * spaced apart along the overall state cycle (indeed, the jumps() |
|
114 * and leaps() methods each produce a stream of such generators). |
|
115 * The generators can then be parceled out to other threads. |
|
116 * |
|
117 * File organization: First the non-public methods that constitute the |
|
118 * main algorithm, then the public methods. Note that many methods are |
|
119 * defined by classes {@code AbstractJumpableRng} and {@code AbstractRng}. |
|
120 */ |
|
121 |
|
122 /* ---------------- static fields ---------------- */ |
|
123 |
|
124 /** |
|
125 * The seed generator for default constructors. |
|
126 */ |
|
127 private static final AtomicLong defaultGen = new AtomicLong(RngSupport.initialSeed()); |
|
128 |
|
129 /* |
|
130 * The period of this generator, which is 2**128 - 1. |
|
131 */ |
|
132 private static final BigInteger thePeriod = |
|
133 BigInteger.ONE.shiftLeft(128).subtract(BigInteger.ONE); |
|
134 |
|
135 /* ---------------- instance fields ---------------- */ |
|
136 |
|
137 /** |
|
138 * The per-instance state. |
|
139 * At least one of the two fields x0 and x1 must be nonzero. |
|
140 */ |
|
141 private long x0, x1; |
|
142 |
|
143 /* ---------------- constructors ---------------- */ |
|
144 |
|
145 /** |
|
146 * Basic constructor that initializes all fields from parameters. |
|
147 * It then adjusts the field values if necessary to ensure that |
|
148 * all constraints on the values of fields are met. |
|
149 */ |
|
150 public Xoroshiro128StarStar(long x0, long x1) { |
|
151 this.x0 = x0; |
|
152 this.x1 = x1; |
|
153 // If x0 and x1 are both zero, we must choose nonzero values. |
|
154 if ((x0 | x1) == 0) { |
|
155 // At least one of the two values generated here will be nonzero. |
|
156 this.x0 = RngSupport.mixStafford13(x0 += RngSupport.GOLDEN_RATIO_64); |
|
157 this.x1 = (x0 += RngSupport.GOLDEN_RATIO_64); |
|
158 } |
|
159 } |
|
160 |
|
161 /** |
|
162 * Creates a new instance of {@code Xoroshiro128StarStar} using the |
|
163 * specified {@code long} value as the initial seed. Instances of |
|
164 * {@code Xoroshiro128StarStar} created with the same seed in the same |
|
165 * program generate identical sequences of values. |
|
166 * |
|
167 * @param seed the initial seed |
|
168 */ |
|
169 public Xoroshiro128StarStar(long seed) { |
|
170 // Using a value with irregularly spaced 1-bits to xor the seed |
|
171 // argument tends to improve "pedestrian" seeds such as 0 or |
|
172 // other small integers. We may as well use SILVER_RATIO_64. |
|
173 // |
|
174 // The x values are then filled in as if by a SplitMix PRNG with |
|
175 // GOLDEN_RATIO_64 as the gamma value and Stafford13 as the mixer. |
|
176 this(RngSupport.mixStafford13(seed ^= RngSupport.SILVER_RATIO_64), |
|
177 RngSupport.mixStafford13(seed + RngSupport.GOLDEN_RATIO_64)); |
|
178 } |
|
179 |
|
180 /** |
|
181 * Creates a new instance of {@code Xoroshiro128StarStar} that is likely to |
|
182 * generate sequences of values that are statistically independent |
|
183 * of those of any other instances in the current program execution, |
|
184 * but may, and typically does, vary across program invocations. |
|
185 */ |
|
186 public Xoroshiro128StarStar() { |
|
187 // Using GOLDEN_RATIO_64 here gives us a good Weyl sequence of values. |
|
188 this(defaultGen.getAndAdd(RngSupport.GOLDEN_RATIO_64)); |
|
189 } |
|
190 |
|
191 /** |
|
192 * Creates a new instance of {@code Xoroshiro128StarStar} using the specified array of |
|
193 * initial seed bytes. Instances of {@code Xoroshiro128StarStar} created with the same |
|
194 * seed array in the same program execution generate identical sequences of values. |
|
195 * |
|
196 * @param seed the initial seed |
|
197 */ |
|
198 public Xoroshiro128StarStar(byte[] seed) { |
|
199 // Convert the seed to 2 long values, which are not both zero. |
|
200 long[] data = RngSupport.convertSeedBytesToLongs(seed, 2, 2); |
|
201 long x0 = data[0], x1 = data[1]; |
|
202 this.x0 = x0; |
|
203 this.x1 = x1; |
|
204 } |
|
205 |
|
206 /* ---------------- public methods ---------------- */ |
|
207 |
|
208 public Xoroshiro128StarStar copy() { return new Xoroshiro128StarStar(x0, x1); } |
|
209 |
|
210 /* |
|
211 |
|
212 To the extent possible under law, the author has dedicated all copyright |
|
213 and related and neighboring rights to this software to the public domain |
|
214 worldwide. This software is distributed without any warranty. |
|
215 |
|
216 See <http://creativecommons.org/publicdomain/zero/1.0/>. */ |
|
217 |
|
218 /* This is the successor to xorshift128+. It is the fastest full-period |
|
219 generator passing BigCrush without systematic failures, but due to the |
|
220 relatively short period it is acceptable only for applications with a |
|
221 mild amount of parallelism; otherwise, use a xorshift1024* generator. |
|
222 |
|
223 Beside passing BigCrush, this generator passes the PractRand test suite |
|
224 up to (and included) 16TB, with the exception of binary rank tests, |
|
225 which fail due to the lowest bit being an LFSR; all other bits pass all |
|
226 tests. We suggest to use a sign test to extract a random Boolean value. |
|
227 |
|
228 Note that the generator uses a simulated rotate operation, which most C |
|
229 compilers will turn into a single instruction. In Java, you can use |
|
230 Long.rotateLeft(). In languages that do not make low-level rotation |
|
231 instructions accessible xorshift128+ could be faster. |
|
232 |
|
233 The state must be seeded so that it is not everywhere zero. If you have |
|
234 a 64-bit seed, we suggest to seed a splitmix64 generator and use its |
|
235 output to fill s. */ |
|
236 |
|
237 |
|
238 /** |
|
239 * Returns a pseudorandom {@code long} value. |
|
240 * |
|
241 * @return a pseudorandom {@code long} value |
|
242 */ |
|
243 public long nextLong() { |
|
244 final long s0 = x0; |
|
245 long s1 = x1; |
|
246 final long z = s0; |
|
247 |
|
248 s1 ^= s0; |
|
249 x0 = Long.rotateLeft(s0, 24) ^ s1 ^ (s1 << 16); // a, b |
|
250 x1 = Long.rotateLeft(s1, 37); // c |
|
251 |
|
252 return Long.rotateLeft(z * 5, 7) * 9; // "starstar" mixing function |
|
253 } |
|
254 |
|
255 public BigInteger period() { return thePeriod; } |
|
256 |
|
257 public double defaultJumpDistance() { return 0x1.0p64; } |
|
258 |
|
259 public double defaultLeapDistance() { return 0x1.0p96; } |
|
260 |
|
261 private static final long[] JUMP_TABLE = { 0xdf900294d8f554a5L, 0x170865df4b3201fcL }; |
|
262 |
|
263 private static final long[] LEAP_TABLE = { 0xd2a98b26625eee7bL, 0xdddf9b1090aa7ac1L }; |
|
264 |
|
265 /* This is the jump function for the generator. It is equivalent |
|
266 to 2**64 calls to nextLong(); it can be used to generate 2**64 |
|
267 non-overlapping subsequences for parallel computations. */ |
|
268 |
|
269 public void jump() { jumpAlgorithm(JUMP_TABLE); } |
|
270 |
|
271 /* This is the long-jump function for the generator. It is equivalent to |
|
272 2**96 calls to next(); it can be used to generate 2**32 starting points, |
|
273 from each of which jump() will generate 2**32 non-overlapping |
|
274 subsequences for parallel distributed computations. */ |
|
275 |
|
276 public void leap() { jumpAlgorithm(LEAP_TABLE); } |
|
277 |
|
278 private void jumpAlgorithm(long[] table) { |
|
279 long s0 = 0, s1 = 0; |
|
280 for (int i = 0; i < table.length; i++) { |
|
281 for (int b = 0; b < 64; b++) { |
|
282 if ((table[i] & (1L << b)) != 0) { |
|
283 s0 ^= x0; |
|
284 s1 ^= x1; |
|
285 } |
|
286 nextLong(); |
|
287 } |
|
288 x0 = s0; |
|
289 x1 = s1; |
|
290 } |
|
291 } |
|
292 } |