1 | // |
---|
2 | // Cforall Version 1.0.0 Copyright (C) 2022 University of Waterloo |
---|
3 | // |
---|
4 | // The contents of this file are covered under the licence agreement in the |
---|
5 | // file "LICENCE" distributed with Cforall. |
---|
6 | // |
---|
7 | // random.hfa -- |
---|
8 | // |
---|
9 | // Author : Peter A. Buhr |
---|
10 | // Created On : Fri Jan 14 07:18:11 2022 |
---|
11 | // Last Modified By : Peter A. Buhr |
---|
12 | // Last Modified On : Mon Mar 20 10:01:40 2023 |
---|
13 | // Update Count : 180 |
---|
14 | // |
---|
15 | |
---|
16 | #pragma once |
---|
17 | |
---|
18 | #include <stdint.h> // uintXX_t |
---|
19 | |
---|
20 | #define GLUE2( x, y ) x##y |
---|
21 | #define GLUE( x, y ) GLUE2( x, y ) |
---|
22 | |
---|
23 | // Set default PRNG for architecture size. |
---|
24 | #ifdef __x86_64__ // 64-bit architecture |
---|
25 | // 64-bit generators |
---|
26 | //#define LEHMER64 |
---|
27 | //#define XORSHIFT_12_25_27 |
---|
28 | #define XOSHIRO256PP |
---|
29 | //#define KISS_64 |
---|
30 | |
---|
31 | // 32-bit generators |
---|
32 | //#define XORSHIFT_6_21_7 |
---|
33 | #define XOSHIRO128PP |
---|
34 | // #define SPLITMIX_32 |
---|
35 | #else // 32-bit architecture |
---|
36 | // 64-bit generators |
---|
37 | //#define XORSHIFT_13_7_17 |
---|
38 | #define XOSHIRO256PP |
---|
39 | |
---|
40 | // 32-bit generators |
---|
41 | //#define XORSHIFT_6_21_7 |
---|
42 | #define XOSHIRO128PP |
---|
43 | // #define SPLITMIX_32 |
---|
44 | #endif // __x86_64__ |
---|
45 | |
---|
46 | // Define C/CFA PRNG name and random-state. |
---|
47 | |
---|
48 | #ifdef XOSHIRO256PP |
---|
49 | #define PRNG_NAME_64 xoshiro256pp |
---|
50 | #define PRNG_STATE_64_T GLUE(PRNG_NAME_64,_t) |
---|
51 | typedef struct { uint64_t s0, s1, s2, s3; } PRNG_STATE_64_T; |
---|
52 | #endif // XOSHIRO256PP |
---|
53 | |
---|
54 | #ifdef XOSHIRO128PP |
---|
55 | #define PRNG_NAME_32 xoshiro128pp |
---|
56 | #define PRNG_STATE_32_T GLUE(PRNG_NAME_32,_t) |
---|
57 | typedef struct { uint32_t s0, s1, s2, s3; } PRNG_STATE_32_T; |
---|
58 | #endif // XOSHIRO128PP |
---|
59 | |
---|
60 | #ifdef LEHMER64 |
---|
61 | #define PRNG_NAME_64 lehmer64 |
---|
62 | #define PRNG_STATE_64_T __uint128_t |
---|
63 | #endif // LEHMER64 |
---|
64 | |
---|
65 | #ifdef WYHASH64 |
---|
66 | #define PRNG_NAME_64 wyhash64 |
---|
67 | #define PRNG_STATE_64_T uint64_t |
---|
68 | #endif // LEHMER64 |
---|
69 | |
---|
70 | #ifdef XORSHIFT_13_7_17 |
---|
71 | #define PRNG_NAME_64 xorshift_13_7_17 |
---|
72 | #define PRNG_STATE_64_T uint64_t |
---|
73 | #endif // XORSHIFT_13_7_17 |
---|
74 | |
---|
75 | #ifdef XORSHIFT_6_21_7 |
---|
76 | #define PRNG_NAME_32 xorshift_6_21_7 |
---|
77 | #define PRNG_STATE_32_T uint32_t |
---|
78 | #endif // XORSHIFT_6_21_7 |
---|
79 | |
---|
80 | #ifdef SPLITMIX_32 |
---|
81 | #define PRNG_NAME_32 splitmix |
---|
82 | #define PRNG_STATE_32_T uint32_t |
---|
83 | #endif // SPLITMIX32 |
---|
84 | |
---|
85 | #ifdef XORSHIFT_12_25_27 |
---|
86 | #define PRNG_NAME_64 xorshift_12_25_27 |
---|
87 | #define PRNG_STATE_64_T uint64_t |
---|
88 | #endif // XORSHIFT_12_25_27 |
---|
89 | |
---|
90 | #ifdef KISS_64 |
---|
91 | #define PRNG_NAME_64 kiss_64 |
---|
92 | #define PRNG_STATE_64_T GLUE(PRNG_NAME_64,_t) |
---|
93 | typedef struct { uint64_t z, w, jsr, jcong; } PRNG_STATE_64_T; |
---|
94 | #endif // KISS_^64 |
---|
95 | |
---|
96 | #ifdef XORWOW |
---|
97 | #define PRNG_NAME_32 xorwow |
---|
98 | #define PRNG_STATE_32_T GLUE(PRNG_NAME_32,_t) |
---|
99 | typedef struct { uint32_t a, b, c, d, counter; } PRNG_STATE_32_T; |
---|
100 | #endif // XOSHIRO128PP |
---|
101 | |
---|
102 | #define PRNG_SET_SEED_64 GLUE(PRNG_NAME_64,_set_seed) |
---|
103 | #define PRNG_SET_SEED_32 GLUE(PRNG_NAME_32,_set_seed) |
---|
104 | |
---|
105 | |
---|
106 | // Default PRNG used by runtime. |
---|
107 | #ifdef __x86_64__ // 64-bit architecture |
---|
108 | #define PRNG_NAME PRNG_NAME_64 |
---|
109 | #define PRNG_STATE_T PRNG_STATE_64_T |
---|
110 | #else // 32-bit architecture |
---|
111 | #define PRNG_NAME PRNG_NAME_32 |
---|
112 | #define PRNG_STATE_T PRNG_STATE_32_T |
---|
113 | #endif // __x86_64__ |
---|
114 | |
---|
115 | #define PRNG_SET_SEED GLUE(PRNG_NAME,_set_seed) |
---|
116 | |
---|
117 | |
---|
118 | // ALL PRNG ALGORITHMS ARE OPTIMIZED SO THAT THE PRNG LOGIC CAN HAPPEN IN PARALLEL WITH THE USE OF THE RESULT. |
---|
119 | // Specifically, the current random state is copied for returning, before computing the next value. As a consequence, |
---|
120 | // the set_seed routine primes the PRNG by calling it with the state so the seed is not return as the first random |
---|
121 | // value. |
---|
122 | |
---|
123 | |
---|
124 | #ifdef __cforall // don't include in C code (invoke.h) |
---|
125 | |
---|
126 | // Splitmix32 |
---|
127 | // https://github.com/bryc/code/blob/master/jshash/PRNGs.md#splitmix32 |
---|
128 | // Splitmix32 is not recommended for demanding random number requirements, |
---|
129 | // but is often used to calculate initial states for other more complex |
---|
130 | // pseudo-random number generators. |
---|
131 | // SplitMix32 is a 32 bit variant of Splitmix64 |
---|
132 | |
---|
133 | static inline uint32_t splitmix32( uint32_t & state ) { |
---|
134 | state += 0x9e3779b9; |
---|
135 | uint64_t z = state; |
---|
136 | z = (z ^ (z >> 15)) * 0x85ebca6b; |
---|
137 | z = (z ^ (z >> 13)) * 0xc2b2ae35; |
---|
138 | return z ^ (z >> 16); |
---|
139 | } |
---|
140 | |
---|
141 | static inline void splitmix32_set_seed( uint32_t & state , uint64_t seed ) { |
---|
142 | state = seed; |
---|
143 | splitmix32( state ); // prime |
---|
144 | } // splitmix32_set_seed |
---|
145 | |
---|
146 | #ifdef __SIZEOF_INT128__ |
---|
147 | //-------------------------------------------------- |
---|
148 | static inline uint64_t lehmer64( __uint128_t & state ) { |
---|
149 | __uint128_t ret = state; |
---|
150 | state *= 0x_da94_2042_e4dd_58b5; |
---|
151 | return ret >> 64; |
---|
152 | } // lehmer64 |
---|
153 | |
---|
154 | static inline void lehmer64_set_seed( __uint128_t & state, uint64_t seed ) { |
---|
155 | // The seed needs to be coprime with the 2^64 modulus to get the largest period, so no factors of 2 in the seed. |
---|
156 | state = seed; |
---|
157 | lehmer64( state ); // prime |
---|
158 | } // lehmer64_set_seed |
---|
159 | |
---|
160 | //-------------------------------------------------- |
---|
161 | static inline uint64_t wyhash64( uint64_t & state ) { |
---|
162 | uint64_t ret = state; |
---|
163 | state += 0x_60be_e2be_e120_fc15; |
---|
164 | __uint128_t tmp; |
---|
165 | tmp = (__uint128_t) ret * 0x_a3b1_9535_4a39_b70d; |
---|
166 | uint64_t m1 = (tmp >> 64) ^ tmp; |
---|
167 | tmp = (__uint128_t)m1 * 0x_1b03_7387_12fa_d5c9; |
---|
168 | uint64_t m2 = (tmp >> 64) ^ tmp; |
---|
169 | return m2; |
---|
170 | } // wyhash64 |
---|
171 | |
---|
172 | static inline void wyhash64_set_seed( uint64_t & state, uint64_t seed ) { |
---|
173 | state = seed; |
---|
174 | wyhash64( state ); // prime |
---|
175 | } // wyhash64_set_seed |
---|
176 | #endif // __SIZEOF_INT128__ |
---|
177 | |
---|
178 | // https://prng.di.unimi.it/xoshiro256starstar.c |
---|
179 | // |
---|
180 | // This is xoshiro256++ 1.0, one of our all-purpose, rock-solid generators. It has excellent (sub-ns) speed, a state |
---|
181 | // (256 bits) that is large enough for any parallel application, and it passes all tests we are aware of. |
---|
182 | // |
---|
183 | // For generating just floating-point numbers, xoshiro256+ is even faster. |
---|
184 | // |
---|
185 | // The state must be seeded so that it is not everywhere zero. If you have a 64-bit seed, we suggest to seed a |
---|
186 | // splitmix64 generator and use its output to fill s. |
---|
187 | |
---|
188 | #ifndef XOSHIRO256PP |
---|
189 | typedef struct { uint64_t s0, s1, s2, s3; } xoshiro256pp_t; |
---|
190 | #endif // ! XOSHIRO256PP |
---|
191 | |
---|
192 | static inline uint64_t xoshiro256pp( xoshiro256pp_t & rs ) with(rs) { |
---|
193 | inline uint64_t rotl( const uint64_t x, int k ) { |
---|
194 | return (x << k) | (x >> (64 - k)); |
---|
195 | } // rotl |
---|
196 | |
---|
197 | const uint64_t result = rotl( s0 + s3, 23 ) + s0; |
---|
198 | const uint64_t t = s1 << 17; |
---|
199 | |
---|
200 | s2 ^= s0; |
---|
201 | s3 ^= s1; |
---|
202 | s1 ^= s2; |
---|
203 | s0 ^= s3; |
---|
204 | s2 ^= t; |
---|
205 | s3 = rotl( s3, 45 ); |
---|
206 | return result; |
---|
207 | } // xoshiro256pp |
---|
208 | |
---|
209 | static inline void xoshiro256pp_set_seed( xoshiro256pp_t & state, uint64_t seed ) { |
---|
210 | uint64_t state; |
---|
211 | wyhash64_set_seed( state, seed ); |
---|
212 | // these are done explicitly in this order to attain repeatable seeding. |
---|
213 | // do not call splitmix32 directly in the state init since order of argument evaluation |
---|
214 | // may not be consistent leading to irreproducible seeding |
---|
215 | uint64_t seed1 = wyhash64( state ); |
---|
216 | uint64_t seed2 = wyhash64( state ); |
---|
217 | uint64_t seed3 = wyhash64( state ); |
---|
218 | uint64_t seed4 = wyhash64( state ); |
---|
219 | state = (xoshiro256pp_t){ seed1, seed2, seed3, seed4 }; |
---|
220 | xoshiro256pp( state ); // prime |
---|
221 | } // xoshiro256pp_set_seed |
---|
222 | |
---|
223 | // https://prng.di.unimi.it/xoshiro128plusplus.c |
---|
224 | // |
---|
225 | // This is xoshiro128++ 1.0, one of our 32-bit all-purpose, rock-solid generators. It has excellent speed, a state size |
---|
226 | // (128 bits) that is large enough for mild parallelism, and it passes all tests we are aware of. |
---|
227 | // |
---|
228 | // For generating just single-precision (i.e., 32-bit) floating-point numbers, xoshiro128+ is even faster. |
---|
229 | // |
---|
230 | // The state must be seeded so that it is not everywhere zero. |
---|
231 | |
---|
232 | #ifndef XOSHIRO128PP |
---|
233 | typedef struct { uint32_t s0, s1, s2, s3; } xoshiro128pp_t; |
---|
234 | #endif // ! XOSHIRO128PP |
---|
235 | |
---|
236 | static inline uint32_t xoshiro128pp( xoshiro128pp_t & rs ) with(rs) { |
---|
237 | inline uint32_t rotl( const uint32_t x, int k ) { |
---|
238 | return (x << k) | (x >> (32 - k)); |
---|
239 | } // rotl |
---|
240 | |
---|
241 | const uint32_t result = rotl( s0 + s3, 7 ) + s0; |
---|
242 | const uint32_t t = s1 << 9; |
---|
243 | |
---|
244 | s2 ^= s0; |
---|
245 | s3 ^= s1; |
---|
246 | s1 ^= s2; |
---|
247 | s0 ^= s3; |
---|
248 | s2 ^= t; |
---|
249 | s3 = rotl( s3, 11 ); |
---|
250 | return result; |
---|
251 | } // xoshiro128pp |
---|
252 | |
---|
253 | static inline void xoshiro128pp_set_seed( xoshiro128pp_t & state, uint32_t seed ) { |
---|
254 | // these are done explicitly in this order to attain repeatable seeding. |
---|
255 | // do not call splitmix32 directly in the state init since order of argument evaluation |
---|
256 | // may not be consistent leading to irreproducible seeding |
---|
257 | uint32_t seed1 = splitmix32( seed ); |
---|
258 | uint32_t seed2 = splitmix32( seed ); |
---|
259 | uint32_t seed3 = splitmix32( seed ); |
---|
260 | uint32_t seed4 = splitmix32( seed ); |
---|
261 | state = (xoshiro128pp_t){ seed1, seed2, seed3, seed4 }; |
---|
262 | xoshiro128pp( state ); // prime |
---|
263 | } // xoshiro128pp_set_seed |
---|
264 | |
---|
265 | //-------------------------------------------------- |
---|
266 | static inline uint64_t xorshift_13_7_17( uint64_t & state ) { |
---|
267 | uint64_t ret = state; |
---|
268 | state ^= state << 13; |
---|
269 | state ^= state >> 7; |
---|
270 | state ^= state << 17; |
---|
271 | return ret; |
---|
272 | } // xorshift_13_7_17 |
---|
273 | |
---|
274 | static inline void xorshift_13_7_17_set_seed( uint64_t & state, uint64_t seed ) { |
---|
275 | state = seed; |
---|
276 | xorshift_13_7_17( state ); // prime |
---|
277 | } // xorshift_13_7_17_set_seed |
---|
278 | |
---|
279 | //-------------------------------------------------- |
---|
280 | // Marsaglia shift-XOR PRNG with thread-local state |
---|
281 | // Period is 4G-1 |
---|
282 | // 0 is absorbing and must be avoided |
---|
283 | // Low-order bits are not particularly random |
---|
284 | static inline uint32_t xorshift_6_21_7( uint32_t & state ) { |
---|
285 | uint32_t ret = state; |
---|
286 | state ^= state << 6; |
---|
287 | state ^= state >> 21; |
---|
288 | state ^= state << 7; |
---|
289 | return ret; |
---|
290 | } // xorshift_6_21_7 |
---|
291 | |
---|
292 | static inline void xorshift_6_21_7_set_seed( uint32_t & state, uint32_t seed ) { |
---|
293 | state = seed; |
---|
294 | xorshift_6_21_7( state ); // prime |
---|
295 | } // xorshift_6_21_7_set_seed |
---|
296 | |
---|
297 | //-------------------------------------------------- |
---|
298 | // The state must be seeded with a nonzero value. |
---|
299 | static inline uint64_t xorshift_12_25_27( uint64_t & state ) { |
---|
300 | uint64_t ret = state; |
---|
301 | state ^= state >> 12; |
---|
302 | state ^= state << 25; |
---|
303 | state ^= state >> 27; |
---|
304 | return ret * 0x_2545_F491_4F6C_DD1D; |
---|
305 | } // xorshift_12_25_27 |
---|
306 | |
---|
307 | static inline void xorshift_12_25_27_set_seed( uint64_t & state, uint64_t seed ) { |
---|
308 | state = seed; |
---|
309 | xorshift_12_25_27( state ); // prime |
---|
310 | } // xorshift_12_25_27_set_seed |
---|
311 | |
---|
312 | //-------------------------------------------------- |
---|
313 | // The state must be seeded with a nonzero value. |
---|
314 | #ifndef KISS_64 |
---|
315 | typedef struct { uint64_t z, w, jsr, jcong; } kiss_64_t; |
---|
316 | #endif // ! KISS_64 |
---|
317 | |
---|
318 | static inline uint64_t kiss_64( kiss_64_t & rs ) with(rs) { |
---|
319 | kiss_64_t ret = rs; |
---|
320 | z = 36969 * (z & 65535) + (z >> 16); |
---|
321 | w = 18000 * (w & 65535) + (w >> 16); |
---|
322 | jsr ^= (jsr << 13); |
---|
323 | jsr ^= (jsr >> 17); |
---|
324 | jsr ^= (jsr << 5); |
---|
325 | jcong = 69069 * jcong + 1234567; |
---|
326 | return (((ret.z << 16) + ret.w) ^ ret.jcong) + ret.jsr; |
---|
327 | } // kiss_64 |
---|
328 | |
---|
329 | static inline void kiss_64_set_seed( kiss_64_t & rs, uint64_t seed ) with(rs) { |
---|
330 | z = 1; w = 1; jsr = 4; jcong = seed; |
---|
331 | kiss_64( rs ); // prime |
---|
332 | } // kiss_64_set_seed |
---|
333 | |
---|
334 | //-------------------------------------------------- |
---|
335 | // The state array must be initialized to non-zero in the first four words. |
---|
336 | #ifndef XORWOW |
---|
337 | typedef struct { uint32_t a, b, c, d, counter; } xorwow_t; |
---|
338 | #endif // ! XORWOW |
---|
339 | |
---|
340 | static inline uint32_t xorwow( xorwow_t & rs ) with(rs) { |
---|
341 | // Algorithm "xorwow" from p. 5 of Marsaglia, "Xorshift RNGs". |
---|
342 | uint32_t ret = a + counter; |
---|
343 | uint32_t t = d; |
---|
344 | |
---|
345 | uint32_t const s = a; |
---|
346 | d = c; |
---|
347 | c = b; |
---|
348 | b = s; |
---|
349 | |
---|
350 | t ^= t >> 2; |
---|
351 | t ^= t << 1; |
---|
352 | t ^= s ^ (s << 4); |
---|
353 | a = t; |
---|
354 | counter += 362437; |
---|
355 | return ret; |
---|
356 | } // xorwow |
---|
357 | |
---|
358 | static inline void xorwow_set_seed( xorwow_t & rs, uint32_t seed ) { |
---|
359 | // these are done explicitly in this order to attain repeatable seeding. |
---|
360 | // do not call splitmix32 directly in the state init since order of argument evaluation |
---|
361 | // may not be consistent leading to irreproducible seeding |
---|
362 | uint32_t seed1 = splitmix32( seed ); |
---|
363 | uint32_t seed2 = splitmix32( seed ); |
---|
364 | uint32_t seed3 = splitmix32( seed ); |
---|
365 | uint32_t seed4 = splitmix32( seed ); |
---|
366 | rs = (xorwow_t){ seed1, seed2, seed3, seed4, 0 }; |
---|
367 | xorwow( rs ); // prime |
---|
368 | } // xorwow_set_seed |
---|
369 | |
---|
370 | //-------------------------------------------------- |
---|
371 | // Used in __tls_rand_fwd |
---|
372 | #define M (1_l64u << 48_l64u) |
---|
373 | #define A (25214903917_l64u) |
---|
374 | #define AI (18446708753438544741_l64u) |
---|
375 | #define C (11_l64u) |
---|
376 | #define D (16_l64u) |
---|
377 | |
---|
378 | // Bi-directional LCG random-number generator |
---|
379 | static inline uint32_t LCGBI_fwd( uint64_t & rs ) { |
---|
380 | rs = (A * rs + C) & (M - 1); |
---|
381 | return rs >> D; |
---|
382 | } // LCGBI_fwd |
---|
383 | |
---|
384 | static inline uint32_t LCGBI_bck( uint64_t & rs ) { |
---|
385 | unsigned int r = rs >> D; |
---|
386 | rs = AI * (rs - C) & (M - 1); |
---|
387 | return r; |
---|
388 | } // LCGBI_bck |
---|
389 | |
---|
390 | #undef M |
---|
391 | #undef A |
---|
392 | #undef AI |
---|
393 | #undef C |
---|
394 | #undef D |
---|
395 | |
---|
396 | #endif // __cforall |
---|