1 /*

     2  * Copyright (c) 2013, 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 SplittableRandom} supports methods for

    35  * producing pseudorandom numbers of type {@code int}, {@code long},

    36  * and {@code double} with similar usages as for class

    37  * {@link java.util.Random} but differs in the following ways:

    38  *

    39  * <ul>

    40  *

    41  * <li>Series of generated values pass the DieHarder suite testing

    42  * independence and uniformity properties of random number generators.

    43  * (Most recently validated with <a

    44  * href="http://www.phy.duke.edu/~rgb/General/dieharder.php"> version

    45  * 3.31.1</a>.) These tests validate only the methods for certain

    46  * types and ranges, but similar properties are expected to hold, at

    47  * least approximately, for others as well. The <em>period</em>

    48  * (length of any series of generated values before it repeats) is at

    49  * least 2<sup>64</sup>. </li>

    50  *

    51  * <li> Method {@link #split} constructs and returns a new

    52  * SplittableRandom instance that shares no mutable state with the

    53  * current instance. However, with very high probability, the

    54  * values collectively generated by the two objects have the same

    55  * statistical properties as if the same quantity of values were

    56  * generated by a single thread using a single {@code

    57  * SplittableRandom} object.  </li>

    58  *

    59  * <li>Instances of SplittableRandom are <em>not</em> thread-safe.

    60  * They are designed to be split, not shared, across threads. For

    61  * example, a {@link java.util.concurrent.ForkJoinTask

    62  * fork/join-style} computation using random numbers might include a

    63  * construction of the form {@code new

    64  * Subtask(aSplittableRandom.split()).fork()}.

    65  *

    66  * <li>This class provides additional methods for generating random

    67  * streams, that employ the above techniques when used in {@code

    68  * stream.parallel()} mode.</li>

    69  *

    70  * </ul>

    71  *

    72  * <p>Instances of {@code SplittableRandom} are not cryptographically

    73  * secure.  Consider instead using {@link java.security.SecureRandom}

    74  * in security-sensitive applications. Additionally,

    75  * default-constructed instances do not use a cryptographically random

    76  * seed unless the {@linkplain System#getProperty system property}

    77  * {@code java.util.secureRandomSeed} is set to {@code true}.

    78  *

    79  * @author  Guy Steele

    80  * @author  Doug Lea

    81  * @since   1.8

    82  */

    83 public final class SplittableRandom extends AbstractSplittableRng {

    84 

    85     /*

    86      * Implementation Overview.

    87      *

    88      * This algorithm was inspired by the "DotMix" algorithm by

    89      * Leiserson, Schardl, and Sukha "Deterministic Parallel

    90      * Random-Number Generation for Dynamic-Multithreading Platforms",

    91      * PPoPP 2012, as well as those in "Parallel random numbers: as

    92      * easy as 1, 2, 3" by Salmon, Morae, Dror, and Shaw, SC 2011.  It

    93      * differs mainly in simplifying and cheapening operations.

    94      *

    95      * The primary update step (method nextSeed()) is to add a

    96      * constant ("gamma") to the current (64 bit) seed, forming a

    97      * simple sequence.  The seed and the gamma values for any two

    98      * SplittableRandom instances are highly likely to be different.

    99      *

   100      * Methods nextLong, nextInt, and derivatives do not return the

   101      * sequence (seed) values, but instead a hash-like bit-mix of

   102      * their bits, producing more independently distributed sequences.

   103      * For nextLong, the mix64 function is based on David Stafford's

   104      * (http://zimbry.blogspot.com/2011/09/better-bit-mixing-improving-on.html)

   105      * "Mix13" variant of the "64-bit finalizer" function in Austin

   106      * Appleby's MurmurHash3 algorithm (see

   107      * http://code.google.com/p/smhasher/wiki/MurmurHash3). The mix32

   108      * function is based on Stafford's Mix04 mix function, but returns

   109      * the upper 32 bits cast as int.

   110      *

   111      * The split operation uses the current generator to form the seed

   112      * and gamma for another SplittableRandom.  To conservatively

   113      * avoid potential correlations between seed and value generation,

   114      * gamma selection (method mixGamma) uses different

   115      * (Murmurhash3's) mix constants.  To avoid potential weaknesses

   116      * in bit-mixing transformations, we restrict gammas to odd values

   117      * with at least 24 0-1 or 1-0 bit transitions.  Rather than

   118      * rejecting candidates with too few or too many bits set, method

   119      * mixGamma flips some bits (which has the effect of mapping at

   120      * most 4 to any given gamma value).  This reduces the effective

   121      * set of 64bit odd gamma values by about 2%, and serves as an

   122      * automated screening for sequence constant selection that is

   123      * left as an empirical decision in some other hashing and crypto

   124      * algorithms.

   125      *

   126      * The resulting generator thus transforms a sequence in which

   127      * (typically) many bits change on each step, with an inexpensive

   128      * mixer with good (but less than cryptographically secure)

   129      * avalanching.

   130      *

   131      * The default (no-argument) constructor, in essence, invokes

   132      * split() for a common "defaultGen" SplittableRandom.  Unlike

   133      * other cases, this split must be performed in a thread-safe

   134      * manner, so we use an AtomicLong to represent the seed rather

   135      * than use an explicit SplittableRandom. To bootstrap the

   136      * defaultGen, we start off using a seed based on current time

   137      * unless the java.util.secureRandomSeed property is set. This

   138      * serves as a slimmed-down (and insecure) variant of SecureRandom

   139      * that also avoids stalls that may occur when using /dev/random.

   140      *

   141      * It is a relatively simple matter to apply the basic design here

   142      * to use 128 bit seeds. However, emulating 128bit arithmetic and

   143      * carrying around twice the state add more overhead than appears

   144      * warranted for current usages.

   145      *

   146      * File organization: First the non-public methods that constitute

   147      * the main algorithm, then the main public methods, followed by

   148      * some custom spliterator classes needed for stream methods.

   149      */

   150 

   151     /**

   152      * The golden ratio scaled to 64bits, used as the initial gamma

   153      * value for (unsplit) SplittableRandoms.

   154      */

   155     private static final long GOLDEN_GAMMA = 0x9e3779b97f4a7c15L;

   156 

   157     /**

   158      * The seed. Updated only via method nextSeed.

   159      */

   160     private long seed;

   161 

   162     /**

   163      * The step value.

   164      */

   165     private final long gamma;

   166 

   167     /**

   168      * Internal constructor used by all others except default constructor.

   169      */

   170     private SplittableRandom(long seed, long gamma) {

   171         this.seed = seed;

   172         this.gamma = gamma;

   173     }

   174 

   175     /* The implementation of AbstractSplittableRng requires this. */

   176     // SplittableRandom getThis() { return this; }

   177 

   178     /**

   179      * Computes Stafford variant 13 of 64bit mix function.

   180      * http://zimbry.blogspot.com/2011/09/better-bit-mixing-improving-on.html

   181      */

   182     private static long mix64(long z) {

   183         z = (z ^ (z >>> 30)) * 0xbf58476d1ce4e5b9L;

   184         z = (z ^ (z >>> 27)) * 0x94d049bb133111ebL;

   185         return z ^ (z >>> 31);

   186     }

   187 

   188     /**

   189      * Returns the 32 high bits of Stafford variant 4 mix64 function as int.

   190      * http://zimbry.blogspot.com/2011/09/better-bit-mixing-improving-on.html

   191      */

   192     private static int mix32(long z) {

   193         z = (z ^ (z >>> 33)) * 0x62a9d9ed799705f5L;

   194         return (int)(((z ^ (z >>> 28)) * 0xcb24d0a5c88c35b3L) >>> 32);

   195     }

   196 

   197     /**

   198      * Returns the gamma value to use for a new split instance.

   199      * Uses the 64bit mix function from MurmurHash3.

   200      * https://github.com/aappleby/smhasher/wiki/MurmurHash3     

   201      */

   202     private static long mixGamma(long z) {

   203         z = (z ^ (z >>> 33)) * 0xff51afd7ed558ccdL; // MurmurHash3 mix constants

   204         z = (z ^ (z >>> 33)) * 0xc4ceb9fe1a85ec53L;

   205         z = (z ^ (z >>> 33)) | 1L;                  // force to be odd

   206         int n = Long.bitCount(z ^ (z >>> 1));       // ensure enough transitions

   207         return (n < 24) ? z ^ 0xaaaaaaaaaaaaaaaaL : z;

   208     }

   209 

   210     /**

   211      * Adds gamma to seed.

   212      */

   213     private long nextSeed() {

   214         return seed += gamma;

   215     }

   216 

   217     /**

   218      * The seed generator for default constructors.

   219      */

   220     private static final AtomicLong defaultGen = new AtomicLong(RngSupport.initialSeed());

   221     

   222     /* ---------------- public methods ---------------- */

   223 

   224     /**

   225      * Creates a new SplittableRandom instance using the specified

   226      * initial seed. SplittableRandom instances created with the same

   227      * seed in the same program generate identical sequences of values.

   228      *

   229      * @param seed the initial seed

   230      */

   231     public SplittableRandom(long seed) {

   232         this(seed, GOLDEN_GAMMA);

   233     }

   234 

   235     /**

   236      * Creates a new SplittableRandom instance that is likely to

   237      * generate sequences of values that are statistically independent

   238      * of those of any other instances in the current program; and

   239      * may, and typically does, vary across program invocations.

   240      */

   241     public SplittableRandom() { // emulate defaultGen.split()

   242         long s = defaultGen.getAndAdd(2 * GOLDEN_GAMMA);

   243         this.seed = mix64(s);

   244         this.gamma = mixGamma(s + GOLDEN_GAMMA);

   245     }

   246 

   247     //    public SplittableRandom copy() { return new SplittableRandom(seed, gamma); }

   248     

   249     /**

   250      * Constructs and returns a new SplittableRandom instance that

   251      * shares no mutable state with this instance. However, with very

   252      * high probability, the set of values collectively generated by

   253      * the two objects has the same statistical properties as if the

   254      * same quantity of values were generated by a single thread using

   255      * a single SplittableRandom object.  Either or both of the two

   256      * objects may be further split using the {@code split()} method,

   257      * and the same expected statistical properties apply to the

   258      * entire set of generators constructed by such recursive

   259      * splitting.

   260      *

   261      * @return the new SplittableRandom instance

   262      */

   263     public SplittableRandom split() {

   264         return new SplittableRandom(nextLong(), mixGamma(nextSeed()));

   265     }

   266 

   267     public SplittableRandom split(SplittableRng source) {

   268         return new SplittableRandom(source.nextLong(), mixGamma(source.nextLong()));

   269     }

   270 

   271     /**

   272      * Returns a pseudorandom {@code int} value.

   273      *

   274      * @return a pseudorandom {@code int} value

   275      */

   276     public int nextInt() {

   277         return mix32(nextSeed());

   278     }

   279 

   280     /**

   281      * Returns a pseudorandom {@code long} value.

   282      *

   283      * @return a pseudorandom {@code long} value

   284      */

   285     public long nextLong() {

   286         return mix64(nextSeed());

   287     }

   288 

   289     static final BigInteger thePeriod = BigInteger.ONE.shiftLeft(64);  // Period is 2**64

   290     public BigInteger period() { return thePeriod; }

   291 

   292 }