source: libcfa/src/bits/random.hfa@ 0e16a2d

ADT ast-experimental
Last change on this file since 0e16a2d was 12b006c, checked in by caparsons <caparson@…>, 3 years ago

small comment cleanup
  • Property mode set to 100644
File size: 12.5 KB
Line 
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 // #define SPLITMIX_64
31
32 // 32-bit generators
33 //#define XORSHIFT_6_21_7
34 #define XOSHIRO128PP
35 // #define SPLITMIX_32
36#else // 32-bit architecture
37 // 64-bit generators
38 //#define XORSHIFT_13_7_17
39 #define XOSHIRO256PP
40 // #define SPLITMIX_64
41
42 // 32-bit generators
43 //#define XORSHIFT_6_21_7
44 #define XOSHIRO128PP
45 // #define SPLITMIX_32
46#endif // __x86_64__
47
48// Define C/CFA PRNG name and random-state.
49
50#ifdef XOSHIRO256PP
51#define PRNG_NAME_64 xoshiro256pp
52#define PRNG_STATE_64_T GLUE(PRNG_NAME_64,_t)
53typedef struct { uint64_t s0, s1, s2, s3; } PRNG_STATE_64_T;
54#endif // XOSHIRO256PP
55
56#ifdef XOSHIRO128PP
57#define PRNG_NAME_32 xoshiro128pp
58#define PRNG_STATE_32_T GLUE(PRNG_NAME_32,_t)
59typedef struct { uint32_t s0, s1, s2, s3; } PRNG_STATE_32_T;
60#endif // XOSHIRO128PP
61
62#ifdef LEHMER64
63#define PRNG_NAME_64 lehmer64
64#define PRNG_STATE_64_T __uint128_t
65#endif // LEHMER64
66
67#ifdef WYHASH64
68#define PRNG_NAME_64 wyhash64
69#define PRNG_STATE_64_T uint64_t
70#endif // LEHMER64
71
72#ifdef XORSHIFT_13_7_17
73#define PRNG_NAME_64 xorshift_13_7_17
74#define PRNG_STATE_64_T uint64_t
75#endif // XORSHIFT_13_7_17
76
77#ifdef XORSHIFT_6_21_7
78#define PRNG_NAME_32 xorshift_6_21_7
79#define PRNG_STATE_32_T uint32_t
80#endif // XORSHIFT_6_21_7
81
82#ifdef XORSHIFT_12_25_27
83#define PRNG_NAME_64 xorshift_12_25_27
84#define PRNG_STATE_64_T uint64_t
85#endif // XORSHIFT_12_25_27
86
87#ifdef SPLITMIX_64
88#define PRNG_NAME_64 splitmix64
89#define PRNG_STATE_64_T uint64_t
90#endif // SPLITMIX32
91
92#ifdef SPLITMIX_32
93#define PRNG_NAME_32 splitmix32
94#define PRNG_STATE_32_T uint32_t
95#endif // SPLITMIX32
96
97#ifdef KISS_64
98#define PRNG_NAME_64 kiss_64
99#define PRNG_STATE_64_T GLUE(PRNG_NAME_64,_t)
100typedef struct { uint64_t z, w, jsr, jcong; } PRNG_STATE_64_T;
101#endif // KISS_^64
102
103#ifdef XORWOW
104#define PRNG_NAME_32 xorwow
105#define PRNG_STATE_32_T GLUE(PRNG_NAME_32,_t)
106typedef struct { uint32_t a, b, c, d, counter; } PRNG_STATE_32_T;
107#endif // XOSHIRO128PP
108
109#define PRNG_SET_SEED_64 GLUE(PRNG_NAME_64,_set_seed)
110#define PRNG_SET_SEED_32 GLUE(PRNG_NAME_32,_set_seed)
111
112
113// Default PRNG used by runtime.
114#ifdef __x86_64__ // 64-bit architecture
115#define PRNG_NAME PRNG_NAME_64
116#define PRNG_STATE_T PRNG_STATE_64_T
117#else // 32-bit architecture
118#define PRNG_NAME PRNG_NAME_32
119#define PRNG_STATE_T PRNG_STATE_32_T
120#endif // __x86_64__
121
122#define PRNG_SET_SEED GLUE(PRNG_NAME,_set_seed)
123
124
125// ALL PRNG ALGORITHMS ARE OPTIMIZED SO THAT THE PRNG LOGIC CAN HAPPEN IN PARALLEL WITH THE USE OF THE RESULT.
126// Specifically, the current random state is copied for returning, before computing the next value. As a consequence,
127// the set_seed routine primes the PRNG by calling it with the state so the seed is not return as the first random
128// value.
129
130
131#ifdef __cforall // don't include in C code (invoke.h)
132
133// Splitmix64
134// https://rosettacode.org/wiki/Pseudo-random_numbers/Splitmix64
135// Splitmix64 is not recommended for demanding random number requirements,
136// but is often used to calculate initial states for other more complex
137// pseudo-random number generators.
138static inline uint64_t splitmix64( uint64_t & state ) {
139 state += 0x9e3779b97f4a7c15;
140 uint64_t z = state;
141 z = (z ^ (z >> 30)) * 0xbf58476d1ce4e5b9;
142 z = (z ^ (z >> 27)) * 0x94d049bb133111eb;
143 return z ^ (z >> 31);
144} // splitmix64
145
146static inline void splitmix64_set_seed( uint64_t & state , uint64_t seed ) {
147 state = seed;
148 splitmix64( state ); // prime
149} // splitmix64_set_seed
150
151// Splitmix32
152// https://github.com/bryc/code/blob/master/jshash/PRNGs.md#splitmix32
153// Splitmix32 is not recommended for demanding random number requirements,
154// but is often used to calculate initial states for other more complex
155// pseudo-random number generators.
156// SplitMix32 is a 32 bit variant of Splitmix64
157static inline uint32_t splitmix32( uint32_t & state ) {
158 state += 0x9e3779b9;
159 uint64_t z = state;
160 z = (z ^ (z >> 15)) * 0x85ebca6b;
161 z = (z ^ (z >> 13)) * 0xc2b2ae35;
162 return z ^ (z >> 16);
163} // splitmix32
164
165static inline void splitmix32_set_seed( uint32_t & state, uint64_t seed ) {
166 state = seed;
167 splitmix32( state ); // prime
168} // splitmix32_set_seed
169
170#ifdef __SIZEOF_INT128__
171 //--------------------------------------------------
172 static inline uint64_t lehmer64( __uint128_t & state ) {
173 __uint128_t ret = state;
174 state *= 0x_da94_2042_e4dd_58b5;
175 return ret >> 64;
176 } // lehmer64
177
178 static inline void lehmer64_set_seed( __uint128_t & state, uint64_t seed ) {
179 // The seed needs to be coprime with the 2^64 modulus to get the largest period, so no factors of 2 in the seed.
180 state = seed;
181 lehmer64( state ); // prime
182 } // lehmer64_set_seed
183
184 //--------------------------------------------------
185 static inline uint64_t wyhash64( uint64_t & state ) {
186 uint64_t ret = state;
187 state += 0x_60be_e2be_e120_fc15;
188 __uint128_t tmp;
189 tmp = (__uint128_t) ret * 0x_a3b1_9535_4a39_b70d;
190 uint64_t m1 = (tmp >> 64) ^ tmp;
191 tmp = (__uint128_t)m1 * 0x_1b03_7387_12fa_d5c9;
192 uint64_t m2 = (tmp >> 64) ^ tmp;
193 return m2;
194 } // wyhash64
195#endif // __SIZEOF_INT128__
196
197// https://prng.di.unimi.it/xoshiro256starstar.c
198//
199// This is xoshiro256++ 1.0, one of our all-purpose, rock-solid generators. It has excellent (sub-ns) speed, a state
200// (256 bits) that is large enough for any parallel application, and it passes all tests we are aware of.
201//
202// For generating just floating-point numbers, xoshiro256+ is even faster.
203//
204// The state must be seeded so that it is not everywhere zero. If you have a 64-bit seed, we suggest to seed a
205// splitmix64 generator and use its output to fill s.
206
207#ifndef XOSHIRO256PP
208typedef struct { uint64_t s0, s1, s2, s3; } xoshiro256pp_t;
209#endif // ! XOSHIRO256PP
210
211static inline uint64_t xoshiro256pp( xoshiro256pp_t & rs ) with(rs) {
212 inline uint64_t rotl( const uint64_t x, int k ) {
213 return (x << k) | (x >> (64 - k));
214 } // rotl
215
216 const uint64_t result = rotl( s0 + s3, 23 ) + s0;
217 const uint64_t t = s1 << 17;
218
219 s2 ^= s0;
220 s3 ^= s1;
221 s1 ^= s2;
222 s0 ^= s3;
223 s2 ^= t;
224 s3 = rotl( s3, 45 );
225 return result;
226} // xoshiro256pp
227
228static inline void xoshiro256pp_set_seed( xoshiro256pp_t & state, uint64_t seed ) {
229 // these are done explicitly in this order to attain repeatable seeding.
230 // do not call splitmix64 directly in the state init since order of argument evaluation
231 // may not be consistent leading to irreproducible seeding
232 uint64_t seed1 = splitmix64( seed );
233 uint64_t seed2 = splitmix64( seed );
234 uint64_t seed3 = splitmix64( seed );
235 uint64_t seed4 = splitmix64( seed );
236 state = (xoshiro256pp_t){ seed1, seed2, seed3, seed4 };
237 xoshiro256pp( state ); // prime
238} // xoshiro256pp_set_seed
239
240// https://prng.di.unimi.it/xoshiro128plusplus.c
241//
242// This is xoshiro128++ 1.0, one of our 32-bit all-purpose, rock-solid generators. It has excellent speed, a state size
243// (128 bits) that is large enough for mild parallelism, and it passes all tests we are aware of.
244//
245// For generating just single-precision (i.e., 32-bit) floating-point numbers, xoshiro128+ is even faster.
246//
247// The state must be seeded so that it is not everywhere zero.
248
249#ifndef XOSHIRO128PP
250typedef struct { uint32_t s0, s1, s2, s3; } xoshiro128pp_t;
251#endif // ! XOSHIRO128PP
252
253static inline uint32_t xoshiro128pp( xoshiro128pp_t & rs ) with(rs) {
254 inline uint32_t rotl( const uint32_t x, int k ) {
255 return (x << k) | (x >> (32 - k));
256 } // rotl
257
258 const uint32_t result = rotl( s0 + s3, 7 ) + s0;
259 const uint32_t t = s1 << 9;
260
261 s2 ^= s0;
262 s3 ^= s1;
263 s1 ^= s2;
264 s0 ^= s3;
265 s2 ^= t;
266 s3 = rotl( s3, 11 );
267 return result;
268} // xoshiro128pp
269
270static inline void xoshiro128pp_set_seed( xoshiro128pp_t & state, uint32_t seed ) {
271 // these are done explicitly in this order to attain repeatable seeding.
272 // do not call splitmix32 directly in the state init since order of argument evaluation
273 // may not be consistent leading to irreproducible seeding
274 uint32_t seed1 = splitmix32( seed );
275 uint32_t seed2 = splitmix32( seed );
276 uint32_t seed3 = splitmix32( seed );
277 uint32_t seed4 = splitmix32( seed );
278 state = (xoshiro128pp_t){ seed1, seed2, seed3, seed4 };
279 xoshiro128pp( state ); // prime
280} // xoshiro128pp_set_seed
281
282//--------------------------------------------------
283static inline uint64_t xorshift_13_7_17( uint64_t & state ) {
284 uint64_t ret = state;
285 state ^= state << 13;
286 state ^= state >> 7;
287 state ^= state << 17;
288 return ret;
289} // xorshift_13_7_17
290
291static inline void xorshift_13_7_17_set_seed( uint64_t & state, uint64_t seed ) {
292 state = seed;
293 xorshift_13_7_17( state ); // prime
294} // xorshift_13_7_17_set_seed
295
296//--------------------------------------------------
297// Marsaglia shift-XOR PRNG with thread-local state
298// Period is 4G-1
299// 0 is absorbing and must be avoided
300// Low-order bits are not particularly random
301static inline uint32_t xorshift_6_21_7( uint32_t & state ) {
302 uint32_t ret = state;
303 state ^= state << 6;
304 state ^= state >> 21;
305 state ^= state << 7;
306 return ret;
307} // xorshift_6_21_7
308
309static inline void xorshift_6_21_7_set_seed( uint32_t & state, uint32_t seed ) {
310 state = seed;
311 xorshift_6_21_7( state ); // prime
312} // xorshift_6_21_7_set_seed
313
314//--------------------------------------------------
315// The state must be seeded with a nonzero value.
316static inline uint64_t xorshift_12_25_27( uint64_t & state ) {
317 uint64_t ret = state;
318 state ^= state >> 12;
319 state ^= state << 25;
320 state ^= state >> 27;
321 return ret * 0x_2545_F491_4F6C_DD1D;
322} // xorshift_12_25_27
323
324static inline void xorshift_12_25_27_set_seed( uint64_t & state, uint64_t seed ) {
325 state = seed;
326 xorshift_12_25_27( state ); // prime
327} // xorshift_12_25_27_set_seed
328
329//--------------------------------------------------
330// The state must be seeded with a nonzero value.
331#ifndef KISS_64
332typedef struct { uint64_t z, w, jsr, jcong; } kiss_64_t;
333#endif // ! KISS_64
334
335static inline uint64_t kiss_64( kiss_64_t & rs ) with(rs) {
336 kiss_64_t ret = rs;
337 z = 36969 * (z & 65535) + (z >> 16);
338 w = 18000 * (w & 65535) + (w >> 16);
339 jsr ^= (jsr << 13);
340 jsr ^= (jsr >> 17);
341 jsr ^= (jsr << 5);
342 jcong = 69069 * jcong + 1234567;
343 return (((ret.z << 16) + ret.w) ^ ret.jcong) + ret.jsr;
344} // kiss_64
345
346static inline void kiss_64_set_seed( kiss_64_t & rs, uint64_t seed ) with(rs) {
347 z = 1; w = 1; jsr = 4; jcong = seed;
348 kiss_64( rs ); // prime
349} // kiss_64_set_seed
350
351//--------------------------------------------------
352// The state array must be initialized to non-zero in the first four words.
353#ifndef XORWOW
354typedef struct { uint32_t a, b, c, d, counter; } xorwow_t;
355#endif // ! XORWOW
356
357static inline uint32_t xorwow( xorwow_t & rs ) with(rs) {
358 // Algorithm "xorwow" from p. 5 of Marsaglia, "Xorshift RNGs".
359 uint32_t ret = a + counter;
360 uint32_t t = d;
361
362 uint32_t const s = a;
363 d = c;
364 c = b;
365 b = s;
366
367 t ^= t >> 2;
368 t ^= t << 1;
369 t ^= s ^ (s << 4);
370 a = t;
371 counter += 362437;
372 return ret;
373} // xorwow
374
375static inline void xorwow_set_seed( xorwow_t & rs, uint32_t seed ) {
376 // these are done explicitly in this order to attain repeatable seeding.
377 // do not call splitmix32 directly in the state init since order of argument evaluation
378 // may not be consistent leading to irreproducible seeding
379 uint32_t seed1 = splitmix32( seed );
380 uint32_t seed2 = splitmix32( seed );
381 uint32_t seed3 = splitmix32( seed );
382 uint32_t seed4 = splitmix32( seed );
383 rs = (xorwow_t){ seed1, seed2, seed3, seed4, 0 };
384 xorwow( rs ); // prime
385} // xorwow_set_seed
386
387//--------------------------------------------------
388// Used in __tls_rand_fwd
389#define M (1_l64u << 48_l64u)
390#define A (25214903917_l64u)
391#define AI (18446708753438544741_l64u)
392#define C (11_l64u)
393#define D (16_l64u)
394
395// Bi-directional LCG random-number generator
396static inline uint32_t LCGBI_fwd( uint64_t & rs ) {
397 rs = (A * rs + C) & (M - 1);
398 return rs >> D;
399} // LCGBI_fwd
400
401static inline uint32_t LCGBI_bck( uint64_t & rs ) {
402 unsigned int r = rs >> D;
403 rs = AI * (rs - C) & (M - 1);
404 return r;
405} // LCGBI_bck
406
407#undef M
408#undef A
409#undef AI
410#undef C
411#undef D
412
413#endif // __cforall
Note: See TracBrowser for help on using the repository browser.