001/* Written in 2016 by David Blackman and Sebastiano Vigna (vigna@acm.org) 002 003To the extent possible under law, the author has dedicated all copyright 004and related and neighboring rights to this software to the public domain 005worldwide. This software is distributed without any warranty. 006 007See <http://creativecommons.org/publicdomain/zero/1.0/>. */ 008package squidpony.squidmath; 009 010import squidpony.StringKit; 011 012import java.io.Serializable; 013 014/** 015 * A modification of Blackman and Vigna's xoroshiro128+ generator using two 32-bit ints of state instead of two 64-bit 016 * longs, as well as modifying the output with two additional operations on the existing state; this is both the fastest 017 * generator on GWT I have found without statistical failures, and a StatefulRandomness. This algorithm is sometimes 018 * called xoroshiro64++ and is mentioned in <a href="http://vigna.di.unimi.it/ftp/papers/ScrambledLinear.pdf">this paper 019 * by Blackman and Vigna</a> (in section 10.7; the 'r' constant is 10, which seems to usually do well on those same 020 * authors' HWD test). Lathe32RNG passes the full 32TB battery of PractRand's statistical tests, and does so with 3 021 * "unusual" anomalies, no more-serious anomalies, and no failures. It isn't especially likely that this can pass much 022 * more than 32TB of testing (judging by related attempts, 128TB would be a likely failure point), but because 023 * multi-threaded code is either impossible or impractical on GWT, actually using that many numbers would take a very 024 * long time (generating them would take about 3 nanoseconds per int, but it would take more than 2 to the 43 ints to 025 * start to approach detectable failures, and detecting the failures in anything but the worst case would take more than 026 * a day). In statistical testing, xoroshiro with the '+' scrambler always fails some binary matrix rank tests, but 027 * smaller-state versions fail other tests as well. The changes Lathe makes apply only to the output of xoroshiro64+ (in 028 * Vigna's and Blackman's terms, they are a scrambler), not its well-tested state transition, and these changes 029 * eliminate all statistical failures on 32TB of tested data, avoiding the failures the small-state variant of xoroshiro 030 * suffers on BinaryRank, BCFN, DC6, and FPF. It avoids multiplication (except in {@link #setSeed(int)}, which needs to 031 * use a different algorithm to spread a seed out across twice as much state), like xoroshiro and much of the xorshift 032 * family of generators, and any arithmetic it performs is safe for GWT. Lathe makes an extremely small set of changes 033 * to xoroshiro64+, running xoroshiro64+ as normal (holding on to the result as well as the initial stateA, called s[0] 034 * in the original xoroshiro code) and then bitwise-rotating the result and adding the (now previous) stateA. Although 035 * no bits of xoroshiro are truly free of artifacts, some are harder to find issues with 036 * (see <a href="http://www.pcg-random.org/posts/xoroshiro-fails-truncated.html">this article by PCG-Random's author</a> 037 * for more detail). It is unclear if the changes made here would improve the larger-state version, but they probably 038 * would help to some extent with at least the binary rank failures. The period is identical to xoroshiro with two 039 * 32-bit states, at 0xFFFFFFFFFFFFFFFF or 2 to the 64 minus 1. This generator is slightly slower than xoroshiro without 040 * the small extra steps applied to the output, but about as fast as {@link Oriole32RNG} (this has a smaller period and 041 * smaller state but implements StatefulRandomness). Some simple tests on bytes instead of ints showed that the 042 * technique used here produces all possible bytes with equal frequency when run on bytes as state, with the exception 043 * of producing 0 one less time (because both states cannot be 0 at the same time). This gives some confidence for the 044 * algorithm used here, but the algorithm is still only one-dimensionally equidistributed (the same as xoroshiro128+), 045 * meaning it produces some pairs of ints more frequently than others. You may want to prefer {@link Starfish32RNG}, 046 * which is the current default in {@link GWTRNG}, because it can produce all pairs of ints and all longs (except one), 047 * and has noticeably better quality even on some short generated sequences. 048 * <br> 049 * The name comes from a tool that rotates very quickly to remove undesirable parts of an object, akin to how this 050 * generator adds an extra bitwise rotation to xoroshiro64+ to remove several types of 051 * undesirable statistical failures from its test results. 052 * <br> 053 * <a href="http://xoroshiro.di.unimi.it/xoroshiro128plus.c">Original version here for xorshiro128+</a>; this version 054 * uses <a href="https://groups.google.com/d/msg/prng/Ll-KDIbpO8k/bfHK4FlUCwAJ">different constants</a> by the same 055 * author, Sebastiano Vigna. It does not use <a href="http://xoshiro.di.unimi.it/xoroshiro64star.c">the constants used 056 * in other xoroshiro64 scrambled generators</a>, instead using similar-quality ones from the earlier constants link. 057 * <br> 058 * Written in 2016 by David Blackman and Sebastiano Vigna (vigna@acm.org) 059 * Ported and modified in 2018 by Tommy Ettinger 060 * @author Sebastiano Vigna 061 * @author David Blackman 062 * @author Tommy Ettinger (if there's a flaw, use SquidLib's or Sarong's issues and don't bother Vigna or Blackman, it's probably a mistake in SquidLib's implementation) 063 */ 064public final class Lathe32RNG implements StatefulRandomness, Serializable { 065 066 private static final long serialVersionUID = 2L; 067 068 private int stateA, stateB; 069 070 /** 071 * Creates a new generator seeded using two calls to Math.random(). 072 */ 073 public Lathe32RNG() { 074 setState((int)((Math.random() * 2.0 - 1.0) * 0x80000000), (int)((Math.random() * 2.0 - 1.0) * 0x80000000)); 075 } 076 /** 077 * Constructs this Lathe32RNG by dispersing the bits of seed using {@link #setSeed(int)} across the two parts of state 078 * this has. 079 * @param seed an int that won't be used exactly, but will affect both components of state 080 */ 081 public Lathe32RNG(final int seed) { 082 setSeed(seed); 083 } 084 /** 085 * Constructs this Lathe32RNG by splitting the given seed across the two parts of state this has with 086 * {@link #setState(long)}. 087 * @param seed a long that will be split across both components of state 088 */ 089 public Lathe32RNG(final long seed) { 090 setState(seed); 091 } 092 /** 093 * Constructs this Lathe32RNG by calling {@link #setState(int, int)} on stateA and stateB as given; see that method 094 * for the specific details (stateA and stateB are kept as-is unless they are both 0). 095 * @param stateA the number to use as the first part of the state; this will be 1 instead if both seeds are 0 096 * @param stateB the number to use as the second part of the state 097 */ 098 public Lathe32RNG(final int stateA, final int stateB) { 099 setState(stateA, stateB); 100 } 101 102 @Override 103 public final int next(int bits) { 104 final int s0 = stateA; 105 int s1 = stateB; 106 final int result = s0 + s1; 107 s1 ^= s0; 108 stateA = (s0 << 13 | s0 >>> 19) ^ s1 ^ (s1 << 5); // a, b 109 stateB = (s1 << 28 | s1 >>> 4); // c 110 return (result << 10 | result >>> 22) + s0 >>> (32 - bits); 111 } 112 113 /** 114 * Can return any int, positive or negative, of any size permissible in a 32-bit signed integer. 115 * @return any int, all 32 bits are random 116 */ 117 public final int nextInt() { 118 final int s0 = stateA; 119 int s1 = stateB; 120 final int result = s0 + s1; 121 s1 ^= s0; 122 stateA = (s0 << 13 | s0 >>> 19) ^ s1 ^ (s1 << 5); // a, b 123 stateB = (s1 << 28 | s1 >>> 4); // c 124 return (result << 10 | result >>> 22) + s0; 125 } 126 127 @Override 128 public final long nextLong() { 129 final int s0 = stateA; 130 int s1 = stateB; 131 final int high = s0 + s1; 132 s1 ^= s0; 133 final int s00 = (s0 << 13 | s0 >>> 19) ^ s1 ^ (s1 << 5); // a, b 134 s1 = (s1 << 28 | s1 >>> 4); // c 135 final int low = s00 + s1; 136 s1 ^= s00; 137 stateA = (s00 << 13 | s00 >>> 19) ^ s1 ^ (s1 << 5); // a, b 138 stateB = (s1 << 28 | s1 >>> 4); // c 139 return ((high << 10 | high >>> 22) + s0 | 0L) << 32 | ((low << 10 | low >>> 22) + s00 & 0xFFFFFFFFL); 140 } 141 142 /** 143 * Produces a copy of this RandomnessSource that, if next() and/or nextLong() are called on this object and the 144 * copy, both will generate the same sequence of random numbers from the point copy() was called. This just needs to 145 * copy the state so it isn't shared, usually, and produce a new value with the same exact state. 146 * 147 * @return a copy of this RandomnessSource 148 */ 149 @Override 150 public Lathe32RNG copy() { 151 return new Lathe32RNG(stateA, stateB); 152 } 153 154 /** 155 * Sets the state of this generator using one int, running it through Zog32RNG's algorithm two times to get 156 * two ints. If the states would both be 0, state A is assigned 1 instead. 157 * @param seed the int to use to produce this generator's state 158 */ 159 public void setSeed(final int seed) { 160 int z = seed + 0xC74EAD55, a = seed ^ z; 161 a ^= a >>> 14; 162 z = (z ^ z >>> 10) * 0xA5CB3; 163 a ^= a >>> 15; 164 stateA = (z ^ z >>> 20) + (a ^= a << 13); 165 z = seed + 0x8E9D5AAA; 166 a ^= a >>> 14; 167 z = (z ^ z >>> 10) * 0xA5CB3; 168 a ^= a >>> 15; 169 stateB = (z ^ z >>> 20) + (a ^ a << 13); 170 if((stateA | stateB) == 0) 171 stateA = 1; 172 } 173 174 public int getStateA() 175 { 176 return stateA; 177 } 178 /** 179 * Sets the first part of the state to the given int. As a special case, if the parameter is 0 and stateB is 180 * already 0, this will set stateA to 1 instead, since both states cannot be 0 at the same time. Usually, you 181 * should use {@link #setState(int, int)} to set both states at once, but the result will be the same if you call 182 * setStateA() and then setStateB() or if you call setStateB() and then setStateA(). 183 * @param stateA any int 184 */ 185 186 public void setStateA(int stateA) 187 { 188 this.stateA = (stateA | stateB) == 0 ? 1 : stateA; 189 } 190 public int getStateB() 191 { 192 return stateB; 193 } 194 195 /** 196 * Sets the second part of the state to the given int. As a special case, if the parameter is 0 and stateA is 197 * already 0, this will set stateA to 1 and stateB to 0, since both cannot be 0 at the same time. Usually, you 198 * should use {@link #setState(int, int)} to set both states at once, but the result will be the same if you call 199 * setStateA() and then setStateB() or if you call setStateB() and then setStateA(). 200 * @param stateB any int 201 */ 202 public void setStateB(int stateB) 203 { 204 this.stateB = stateB; 205 if((stateB | stateA) == 0) stateA = 1; 206 } 207 208 /** 209 * Sets the current internal state of this Lathe32RNG with three ints, where stateA and stateB can each be any int 210 * unless they are both 0 (which will be treated as if stateA is 1 and stateB is 0). 211 * @param stateA any int (if stateA and stateB are both 0, this will be treated as 1) 212 * @param stateB any int 213 */ 214 public void setState(int stateA, int stateB) 215 { 216 this.stateA = (stateA | stateB) == 0 ? 1 : stateA; 217 this.stateB = stateB; 218 } 219 220 /** 221 * Get the current internal state of the StatefulRandomness as a long. 222 * 223 * @return the current internal state of this object. 224 */ 225 @Override 226 public long getState() { 227 return stateA & 0xFFFFFFFFL | ((long)stateB) << 32; 228 } 229 230 /** 231 * Set the current internal state of this StatefulRandomness with a long. 232 * 233 * @param state a 64-bit long. You should avoid passing 0; this implementation will treat it as 1. 234 */ 235 @Override 236 public void setState(long state) { 237 stateA = state == 0 ? 1 : (int)(state & 0xFFFFFFFFL); 238 stateB = (int)(state >>> 32); 239 } 240 241 @Override 242 public String toString() { 243 return "Lathe32RNG with stateA 0x" + StringKit.hex(stateA) + " and stateB 0x" + StringKit.hex(stateB); 244 } 245 246 @Override 247 public boolean equals(Object o) { 248 if (this == o) return true; 249 if (o == null || getClass() != o.getClass()) return false; 250 251 Lathe32RNG lathe32RNG = (Lathe32RNG) o; 252 253 return stateA == lathe32RNG.stateA && stateB == lathe32RNG.stateB; 254 } 255 256 @Override 257 public int hashCode() { 258 return 31 * stateA + stateB; 259 } 260}