| 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 : Fri Jan 14 07:18:58 2022
 | 
|---|
| 13 | // Update Count     : 1
 | 
|---|
| 14 | // 
 | 
|---|
| 15 | 
 | 
|---|
| 16 | #pragma once
 | 
|---|
| 17 | 
 | 
|---|
| 18 | #include <stdint.h>
 | 
|---|
| 19 | 
 | 
|---|
| 20 | // Pipelined to allow out-of-order overlap with reduced dependencies. Critically, the current random state is returned
 | 
|---|
| 21 | // (copied), and then compute and store the next random value.
 | 
|---|
| 22 | 
 | 
|---|
| 23 | #if defined(__SIZEOF_INT128__)
 | 
|---|
| 24 | //--------------------------------------------------
 | 
|---|
| 25 |         static inline uint64_t lehmer64( __uint128_t & state ) {
 | 
|---|
| 26 |                 __uint128_t ret = state;
 | 
|---|
| 27 |                 state *= 0xda942042e4dd58b5;
 | 
|---|
| 28 |                 return ret >> 64;
 | 
|---|
| 29 |         }
 | 
|---|
| 30 | 
 | 
|---|
| 31 | //--------------------------------------------------
 | 
|---|
| 32 |         static inline uint64_t wyhash64( uint64_t & state ) {
 | 
|---|
| 33 |                 state += 0x60bee2bee120fc15;
 | 
|---|
| 34 |                 __uint128_t tmp;
 | 
|---|
| 35 |                 tmp = (__uint128_t) state * 0xa3b195354a39b70d;
 | 
|---|
| 36 |                 uint64_t m1 = (tmp >> 64) ^ tmp;
 | 
|---|
| 37 |                 tmp = (__uint128_t)m1 * 0x1b03738712fad5c9;
 | 
|---|
| 38 |                 uint64_t m2 = (tmp >> 64) ^ tmp;
 | 
|---|
| 39 |                 return m2;
 | 
|---|
| 40 |         }
 | 
|---|
| 41 | #endif
 | 
|---|
| 42 | 
 | 
|---|
| 43 | //--------------------------------------------------
 | 
|---|
| 44 | static inline uint64_t xorshift_13_7_17( uint64_t & state ) {
 | 
|---|
| 45 |         uint64_t ret = state;
 | 
|---|
| 46 |         state ^= state << 13;
 | 
|---|
| 47 |         state ^= state >> 7;
 | 
|---|
| 48 |         state ^= state << 17;
 | 
|---|
| 49 |         return ret;
 | 
|---|
| 50 | }
 | 
|---|
| 51 | 
 | 
|---|
| 52 | //--------------------------------------------------
 | 
|---|
| 53 | static inline uint32_t xorshift_6_21_7( uint32_t & state ) {
 | 
|---|
| 54 |         uint32_t ret = state;
 | 
|---|
| 55 |         state ^= state << 6;
 | 
|---|
| 56 |         state ^= state >> 21;
 | 
|---|
| 57 |         state ^= state << 7;
 | 
|---|
| 58 |         return ret;
 | 
|---|
| 59 | } // xorshift_6_21_7
 | 
|---|
| 60 | 
 | 
|---|
| 61 | //--------------------------------------------------
 | 
|---|
| 62 | typedef struct {
 | 
|---|
| 63 |   uint32_t a, b, c, d;
 | 
|---|
| 64 |   uint32_t counter;
 | 
|---|
| 65 | } xorwow__state_t;
 | 
|---|
| 66 | 
 | 
|---|
| 67 | // The state array must be initialized to not be all zero in the first four words.
 | 
|---|
| 68 | static inline uint32_t xorwow( xorwow__state_t & state ) {
 | 
|---|
| 69 |         // Algorithm "xorwow" from p. 5 of Marsaglia, "Xorshift RNGs".
 | 
|---|
| 70 |         uint32_t ret = state.a + state.counter;
 | 
|---|
| 71 |         uint32_t t = state.d;
 | 
|---|
| 72 | 
 | 
|---|
| 73 |         uint32_t const s = state.a;
 | 
|---|
| 74 |         state.d = state.c;
 | 
|---|
| 75 |         state.c = state.b;
 | 
|---|
| 76 |         state.b = s;
 | 
|---|
| 77 | 
 | 
|---|
| 78 |         t ^= t >> 2;
 | 
|---|
| 79 |         t ^= t << 1;
 | 
|---|
| 80 |         t ^= s ^ (s << 4);
 | 
|---|
| 81 |         state.a = t;
 | 
|---|
| 82 | 
 | 
|---|
| 83 |         state.counter += 362437;
 | 
|---|
| 84 |         return ret;
 | 
|---|
| 85 | }
 | 
|---|
| 86 | 
 | 
|---|
| 87 | //--------------------------------------------------
 | 
|---|
| 88 | static inline uint32_t LCG( uint32_t & state ) {                // linear congruential generator
 | 
|---|
| 89 |         uint32_t ret = state;
 | 
|---|
| 90 |         state = 36969 * (state & 65535) + (state >> 16);        // 36969 is NOT prime! No not change it!
 | 
|---|
| 91 |         return ret;
 | 
|---|
| 92 | } // LCG
 | 
|---|
| 93 | 
 | 
|---|
| 94 | //--------------------------------------------------
 | 
|---|
| 95 | #define M  (1_l64u << 48_l64u)
 | 
|---|
| 96 | #define A  (25214903917_l64u)
 | 
|---|
| 97 | #define AI (18446708753438544741_l64u)
 | 
|---|
| 98 | #define C  (11_l64u)
 | 
|---|
| 99 | #define D  (16_l64u)
 | 
|---|
| 100 | 
 | 
|---|
| 101 | // Bi-directional LCG random-number generator
 | 
|---|
| 102 | static inline uint32_t LCGBI_fwd( uint64_t & state ) {
 | 
|---|
| 103 |         state = (A * state + C) & (M - 1);
 | 
|---|
| 104 |         return state >> D;
 | 
|---|
| 105 | }
 | 
|---|
| 106 | 
 | 
|---|
| 107 | static inline uint32_t LCGBI_bck( uint64_t & state ) {
 | 
|---|
| 108 |         unsigned int r = state >> D;
 | 
|---|
| 109 |         state = AI * (state - C) & (M - 1);
 | 
|---|
| 110 |         return r;
 | 
|---|
| 111 | }
 | 
|---|
| 112 | 
 | 
|---|
| 113 | #undef M
 | 
|---|
| 114 | #undef A
 | 
|---|
| 115 | #undef AI
 | 
|---|
| 116 | #undef C
 | 
|---|
| 117 | #undef D
 | 
|---|