|
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 L64X128Random} implements |
|
35 * interfaces {@link java.util.Rng} and {@link java.util.SplittableRng}, |
|
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 split-off {@code L64X128Random} objects, |
|
39 * with similar usages as for class {@link java.util.SplittableRandom}. |
|
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 * (Most recently validated with |
|
44 * <a href="http://simul.iro.umontreal.ca/testu01/tu01.html">version 1.2.3 of TestU01</a> |
|
45 * and <a href="http://pracrand.sourceforge.net">version 0.90 of PractRand</a>. |
|
46 * Note that TestU01 BigCrush was used to test not only values produced by the {@code nextLong()} |
|
47 * method but also the result of bit-reversing each value produced by {@code nextLong()}.) |
|
48 * These tests validate only the methods for certain |
|
49 * types and ranges, but similar properties are expected to hold, at |
|
50 * least approximately, for others as well. |
|
51 * |
|
52 * <p>{@code L64X128Random} is a specific member of the LXM family of algorithms |
|
53 * for pseudorandom number generators. Every LXM generator consists of two |
|
54 * subgenerators; one is an LCG (Linear Congruential Generator) and the other is |
|
55 * an Xorshift generator. Each output of an LXM generator is the sum of one |
|
56 * output from each subgenerator, possibly processed by a final mixing function |
|
57 * (but {@code L64X128Random} does not use a mixing function). |
|
58 * |
|
59 * <p>The LCG subgenerator for {@code L64X128Random} has an update step of the |
|
60 * form {@code s = m * s + a}, where {@code s}, {@code m}, and {@code a} are all |
|
61 * of type {@code long}; {@code s} is the mutable state, the multiplier {@code m} |
|
62 * is fixed (the same for all instances of {@code L64X128Random}}) and the addend |
|
63 * {@code a} is a parameter (a final field of the instance). The parameter |
|
64 * {@code a} is required to be odd (this allows the LCG to have the maximal |
|
65 * period, namely 2<sup>64</sup>); therefore there are 2<sup>63</sup> distinct choices |
|
66 * of parameter. |
|
67 * |
|
68 * <p>The Xorshift subgenerator for {@code L64X128Random} is the {@code xoroshiro128} algorithm, |
|
69 * version 1.0 (parameters 24, 16, 37), without any final scrambler such as "+" or "**". |
|
70 * Its state consists of two {@code long} fields {@code x0} and {@code x1}, |
|
71 * which can take on any values provided that they are not both zero. |
|
72 * The period of this subgenerator is 2<sup>128</sup>-1. |
|
73 * |
|
74 * <p> Because the periods 2<sup>64</sup> and 2<sup>128</sup>-1 of the two subgenerators |
|
75 * are relatively prime, the <em>period</em> of any single {@code L64X128Random} object |
|
76 * (the length of the series of generated 64-bit values before it repeats) is the product |
|
77 * of the periods of the subgenerators, that is, 2<sup>64</sup>(2<sup>128</sup>-1), |
|
78 * which is just slightly smaller than 2<sup>192</sup>. Moreover, if two distinct |
|
79 * {@code L64X128Random} objects have different {@code a} parameters, then their |
|
80 * cycles of produced values will be different. |
|
81 * |
|
82 * <p>The 64-bit values produced by the {@code nextLong()} method are exactly equidistributed. |
|
83 * For any specific instance of {@code L64X128Random}, over the course of its cycle each |
|
84 * of the 2<sup>64</sup> possible {@code long} values will be produced 2<sup>128</sup>-1 times. |
|
85 * The values produced by the {@code nextInt()}, {@code nextFloat()}, and {@code nextDouble()} |
|
86 * methods are likewise exactly equidistributed. |
|
87 * |
|
88 * <p>In fact, the 64-bit values produced by the {@code nextLong()} method are 2-equidistributed. |
|
89 * To be precise: for any specific instance of {@code L64X128Random}, consider |
|
90 * the (overlapping) length-2 subsequences of the cycle of 64-bit values produced by |
|
91 * {@code nextLong()} (assuming no other methods are called that would affect the state). |
|
92 * There are 2<sup>64</sup>(2<sup>128</sup>-1) such subsequences, and each subsequence, |
|
93 * which consists of 2 64-bit values, can have one of 2<sup>128</sup> values. Of those |
|
94 * 2<sup>128</sup> subsequence values, nearly all of them (2<sup>128</sup>-2<sup>64</sup>) |
|
95 * occur 2<sup>64</sup> times over the course of the entire cycle, and the other |
|
96 * 2<sup>64</sup> subsequence values occur only 2<sup>64</sup>-1 times. So the ratio |
|
97 * of the probability of getting one of the less common subsequence values and the |
|
98 * probability of getting one of the more common subsequence values is 1-2<sup>-64</sup>. |
|
99 * (Note that the set of 2<sup>64</sup> less-common subsequence values will differ from |
|
100 * one instance of {@code L64X128Random} to another, as a function of the additive |
|
101 * parameter of the LCG.) The values produced by the {@code nextInt()}, {@code nextFloat()}, |
|
102 * and {@code nextDouble()} methods are likewise 2-equidistributed. |
|
103 * |
|
104 * <p>Method {@link #split} constructs and returns a new {@code L64X128Random} |
|
105 * instance that shares no mutable state with the current instance. However, with |
|
106 * very high probability, the values collectively generated by the two objects |
|
107 * have the same statistical properties as if the same quantity of values were |
|
108 * generated by a single thread using a single {@code L64X128Random} object. |
|
109 * This is because, with high probability, distinct {@code L64X128Random} objects |
|
110 * have distinct {@code a} parameters and therefore use distinct members of the |
|
111 * algorithmic family; and even if their {@code a} parameters are the same, with |
|
112 * very high probability they will traverse different parts of their common state |
|
113 * cycle. |
|
114 * |
|
115 * <p>As with {@link java.util.SplittableRandom}, instances of |
|
116 * {@code L64X128Random} are <em>not</em> thread-safe. |
|
117 * They are designed to be split, not shared, across threads. For |
|
118 * example, a {@link java.util.concurrent.ForkJoinTask} fork/join-style |
|
119 * computation using random numbers might include a construction |
|
120 * of the form {@code new Subtask(someL64X128Random.split()).fork()}. |
|
121 * |
|
122 * <p>This class provides additional methods for generating random |
|
123 * streams, that employ the above techniques when used in |
|
124 * {@code stream.parallel()} mode. |
|
125 * |
|
126 * <p>Instances of {@code L64X128Random} are not cryptographically |
|
127 * secure. Consider instead using {@link java.security.SecureRandom} |
|
128 * in security-sensitive applications. Additionally, |
|
129 * default-constructed instances do not use a cryptographically random |
|
130 * seed unless the {@linkplain System#getProperty system property} |
|
131 * {@code java.util.secureRandomSeed} is set to {@code true}. |
|
132 * |
|
133 * @author Guy Steele |
|
134 * @since 1.9 |
|
135 */ |
|
136 public final class L64X128Random extends AbstractSplittableRng { |
|
137 |
|
138 /* |
|
139 * Implementation Overview. |
|
140 * |
|
141 * The split operation uses the current generator to choose four new 64-bit |
|
142 * long values that are then used to initialize the parameter `a` and the |
|
143 * state variables `s`, `x0`, and `x1` for a newly constructed generator. |
|
144 * |
|
145 * With extremely high probability, no two generators so chosen |
|
146 * will have the same `a` parameter, and testing has indicated |
|
147 * that the values generated by two instances of {@code L64X128Random} |
|
148 * will be (approximately) independent if have different values for `a`. |
|
149 * |
|
150 * The default (no-argument) constructor, in essence, uses |
|
151 * "defaultGen" to generate four new 64-bit values for the same |
|
152 * purpose. Multiple generators created in this way will certainly |
|
153 * differ in their `a` parameters. The defaultGen state must be accessed |
|
154 * in a thread-safe manner, so we use an AtomicLong to represent |
|
155 * this state. To bootstrap the defaultGen, we start off using a |
|
156 * seed based on current time unless the |
|
157 * java.util.secureRandomSeed property is set. This serves as a |
|
158 * slimmed-down (and insecure) variant of SecureRandom that also |
|
159 * avoids stalls that may occur when using /dev/random. |
|
160 * |
|
161 * File organization: First static fields, then instance |
|
162 * fields, then constructors, then instance methods. |
|
163 */ |
|
164 |
|
165 /* ---------------- static fields ---------------- */ |
|
166 |
|
167 /** |
|
168 * The seed generator for default constructors. |
|
169 */ |
|
170 private static final AtomicLong defaultGen = new AtomicLong(RngSupport.initialSeed()); |
|
171 |
|
172 /* |
|
173 * The period of this generator, which is (2**128 - 1) * 2**64. |
|
174 */ |
|
175 private static final BigInteger thePeriod = |
|
176 BigInteger.ONE.shiftLeft(128).subtract(BigInteger.ONE).shiftLeft(64); |
|
177 |
|
178 /* |
|
179 * Multiplier used in the LCG portion of the algorithm, taken from |
|
180 * Pierre L'Ecuyer, Tables of linear congruential generators of |
|
181 * different sizes and good lattice structure, <em>Mathematics of |
|
182 * Computation</em> 68, 225 (January 1999), pages 249–260, |
|
183 * Table 4 (first multiplier for size 2<sup>64</sup>). |
|
184 */ |
|
185 |
|
186 private static final long m = 2862933555777941757L; |
|
187 |
|
188 /* ---------------- instance fields ---------------- */ |
|
189 |
|
190 /** |
|
191 * The parameter that is used as an additive constant for the LCG. |
|
192 * Must be odd. |
|
193 */ |
|
194 private final long a; |
|
195 |
|
196 /** |
|
197 * The per-instance state: s for the LCG; x0 and x1 for the xorshift. |
|
198 * At least one of x0 and x1 must be nonzero. |
|
199 */ |
|
200 private long s, x0, x1; |
|
201 |
|
202 /* ---------------- constructors ---------------- */ |
|
203 |
|
204 /** |
|
205 * Basic constructor that initializes all fields from parameters. |
|
206 * It then adjusts the field values if necessary to ensure that |
|
207 * all constraints on the values of fields are met. |
|
208 */ |
|
209 public L64X128Random(long a, long s, long x0, long x1) { |
|
210 // Force a to be odd. |
|
211 this.a = a | 1; |
|
212 this.s = s; |
|
213 // If x0 and x1 are both zero, we must choose nonzero values. |
|
214 if ((x0 | x1) == 0) { |
|
215 // At least one of the two values generated here will be nonzero. |
|
216 this.x0 = RngSupport.mixStafford13(s += RngSupport.GOLDEN_RATIO_64); |
|
217 this.x1 = RngSupport.mixStafford13(s + RngSupport.GOLDEN_RATIO_64); |
|
218 } |
|
219 } |
|
220 |
|
221 /** |
|
222 * Creates a new instance of {@code L64X128Random} using the |
|
223 * specified {@code long} value as the initial seed. Instances of |
|
224 * {@code L64X128Random} created with the same seed in the same |
|
225 * program generate identical sequences of values. |
|
226 * |
|
227 * @param seed the initial seed |
|
228 */ |
|
229 public L64X128Random(long seed) { |
|
230 // Using a value with irregularly spaced 1-bits to xor the seed |
|
231 // argument tends to improve "pedestrian" seeds such as 0 or |
|
232 // other small integers. We may as well use SILVER_RATIO_64. |
|
233 // |
|
234 // The seed is hashed by mixMurmur64 to produce the `a` parameter. |
|
235 // The seed is hashed by mixStafford13 to produce the initial `x0`, |
|
236 // which will then be used to produce the first generated value. |
|
237 // Then x1 is filled in as if by a SplitMix PRNG with |
|
238 // GOLDEN_RATIO_64 as the gamma value and Stafford13 as the mixer. |
|
239 this(RngSupport.mixMurmur64(seed ^= RngSupport.SILVER_RATIO_64), |
|
240 1, |
|
241 RngSupport.mixStafford13(seed), |
|
242 RngSupport.mixStafford13(seed + RngSupport.GOLDEN_RATIO_64)); |
|
243 } |
|
244 |
|
245 /** |
|
246 * Creates a new instance of {@code L64X128Random} that is likely to |
|
247 * generate sequences of values that are statistically independent |
|
248 * of those of any other instances in the current program execution, |
|
249 * but may, and typically does, vary across program invocations. |
|
250 */ |
|
251 public L64X128Random() { |
|
252 // Using GOLDEN_RATIO_64 here gives us a good Weyl sequence of values. |
|
253 this(defaultGen.getAndAdd(RngSupport.GOLDEN_RATIO_64)); |
|
254 } |
|
255 |
|
256 /** |
|
257 * Creates a new instance of {@code L64X128MixRandom} using the specified array of |
|
258 * initial seed bytes. Instances of {@code L64X128MixRandom} created with the same |
|
259 * seed array in the same program execution generate identical sequences of values. |
|
260 * |
|
261 * @param seed the initial seed |
|
262 */ |
|
263 public L64X128Random(byte[] seed) { |
|
264 // Convert the seed to 4 long values, of which the last 2 are not all zero. |
|
265 long[] data = RngSupport.convertSeedBytesToLongs(seed, 4, 2); |
|
266 long a = data[0], s = data[1], x0 = data[2], x1 = data[3]; |
|
267 // Force a to be odd. |
|
268 this.a = a | 1; |
|
269 this.s = s; |
|
270 this.x0 = x0; |
|
271 this.x1 = x1; |
|
272 } |
|
273 |
|
274 /* ---------------- public methods ---------------- */ |
|
275 |
|
276 /** |
|
277 * Constructs and returns a new instance of {@code L64X128Random} |
|
278 * that shares no mutable state with this instance. |
|
279 * However, with very high probability, the set of values collectively |
|
280 * generated by the two objects has the same statistical properties as if |
|
281 * same the quantity of values were generated by a single thread using |
|
282 * a single {@code L64X128Random} object. Either or both of the two |
|
283 * objects may be further split using the {@code split} method, |
|
284 * and the same expected statistical properties apply to the |
|
285 * entire set of generators constructed by such recursive splitting. |
|
286 * |
|
287 * @param source a {@code SplittableRng} instance to be used instead |
|
288 * of this one as a source of pseudorandom bits used to |
|
289 * initialize the state of the new ones. |
|
290 * @return a new instance of {@code L64X128Random} |
|
291 */ |
|
292 public L64X128Random split(SplittableRng source) { |
|
293 // Literally pick a new instance "at random". |
|
294 return new L64X128Random(source.nextLong(), source.nextLong(), |
|
295 source.nextLong(), source.nextLong()); |
|
296 } |
|
297 |
|
298 /** |
|
299 * Returns a pseudorandom {@code long} value. |
|
300 * |
|
301 * @return a pseudorandom {@code long} value |
|
302 */ |
|
303 |
|
304 public long nextLong() { |
|
305 final long z = s + x0; |
|
306 s = m * s + a; // LCG |
|
307 long q0 = x0, q1 = x1; |
|
308 { q1 ^= q0; q0 = Long.rotateLeft(q0, 24); q0 = q0 ^ q1 ^ (q1 << 16); q1 = Long.rotateLeft(q1, 37); } // xoroshiro128v1_0 |
|
309 x0 = q0; x1 = q1; |
|
310 return z; |
|
311 } |
|
312 |
|
313 public BigInteger period() { return thePeriod; } |
|
314 } |