rlm@1: /** 
rlm@1:  * @file  SFMT.c
rlm@1:  * @brief SIMD oriented Fast Mersenne Twister(SFMT)
rlm@1:  *
rlm@1:  * @author Mutsuo Saito (Hiroshima University)
rlm@1:  * @author Makoto Matsumoto (Hiroshima University)
rlm@1:  *
rlm@1:  * Copyright (C) 2006,2007 Mutsuo Saito, Makoto Matsumoto and Hiroshima
rlm@1:  * University. All rights reserved.
rlm@1:  *
rlm@1:  * The new BSD License is applied to this software, see LICENSE.txt
rlm@1:  */
rlm@1: #include <string.h>
rlm@1: #include <assert.h>
rlm@1: #include "SFMT.h"
rlm@1: #include "SFMT-params.h"
rlm@1: 
rlm@1: #if defined(__BIG_ENDIAN__) && !defined(__amd64) && !defined(BIG_ENDIAN64)
rlm@1: #define BIG_ENDIAN64 1
rlm@1: #endif
rlm@1: #if defined(HAVE_ALTIVEC) && !defined(BIG_ENDIAN64)
rlm@1: #define BIG_ENDIAN64 1
rlm@1: #endif
rlm@1: #if defined(ONLY64) && !defined(BIG_ENDIAN64)
rlm@1:   #if defined(__GNUC__)
rlm@1:     #error "-DONLY64 must be specified with -DBIG_ENDIAN64"
rlm@1:   #endif
rlm@1: #undef ONLY64
rlm@1: #endif
rlm@1: /*------------------------------------------------------
rlm@1:   128-bit SIMD data type for Altivec, SSE2 or standard C
rlm@1:   ------------------------------------------------------*/
rlm@1: #if defined(HAVE_ALTIVEC)
rlm@1:   #if !defined(__APPLE__)
rlm@1:     #include <altivec.h>
rlm@1:   #endif
rlm@1: /** 128-bit data structure */
rlm@1: union W128_T {
rlm@1:     vector unsigned int s;
rlm@1:     uint32_t u[4];
rlm@1: };
rlm@1: /** 128-bit data type */
rlm@1: typedef union W128_T w128_t;
rlm@1: 
rlm@1: #elif defined(HAVE_SSE2)
rlm@1:   #include <emmintrin.h>
rlm@1: 
rlm@1: /** 128-bit data structure */
rlm@1: union W128_T {
rlm@1:     __m128i si;
rlm@1:     uint32_t u[4];
rlm@1: };
rlm@1: /** 128-bit data type */
rlm@1: typedef union W128_T w128_t;
rlm@1: 
rlm@1: #else
rlm@1: 
rlm@1: /** 128-bit data structure */
rlm@1: struct W128_T {
rlm@1:     uint32_t u[4];
rlm@1: };
rlm@1: /** 128-bit data type */
rlm@1: typedef struct W128_T w128_t;
rlm@1: 
rlm@1: #endif
rlm@1: 
rlm@1: /*--------------------------------------
rlm@1:   FILE GLOBAL VARIABLES
rlm@1:   internal state, index counter and flag 
rlm@1:   --------------------------------------*/
rlm@1: /** the 128-bit internal state array */
rlm@1: static w128_t sfmt[N];
rlm@1: /** the 32bit integer pointer to the 128-bit internal state array */
rlm@1: static uint32_t *psfmt32 = &sfmt[0].u[0];
rlm@1: #if !defined(BIG_ENDIAN64) || defined(ONLY64)
rlm@1: /** the 64bit integer pointer to the 128-bit internal state array */
rlm@1: static uint64_t *psfmt64 = (uint64_t *)&sfmt[0].u[0];
rlm@1: #endif
rlm@1: /** index counter to the 32-bit internal state array */
rlm@1: static int idx;
rlm@1: /** a flag: it is 0 if and only if the internal state is not yet
rlm@1:  * initialized. */
rlm@1: static int initialized = 0;
rlm@1: /** a parity check vector which certificate the period of 2^{MEXP} */
rlm@1: static uint32_t parity[4] = {PARITY1, PARITY2, PARITY3, PARITY4};
rlm@1: 
rlm@1: /*----------------
rlm@1:   STATIC FUNCTIONS
rlm@1:   ----------------*/
rlm@1: inline static int idxof(int i);
rlm@1: inline static void rshift128(w128_t *out,  w128_t const *in, int shift);
rlm@1: inline static void lshift128(w128_t *out,  w128_t const *in, int shift);
rlm@1: inline static void gen_rand_all(void);
rlm@1: inline static void gen_rand_array(w128_t *array, int size);
rlm@1: inline static uint32_t func1(uint32_t x);
rlm@1: inline static uint32_t func2(uint32_t x);
rlm@1: static void period_certification(void);
rlm@1: #if defined(BIG_ENDIAN64) && !defined(ONLY64)
rlm@1: inline static void swap(w128_t *array, int size);
rlm@1: #endif
rlm@1: 
rlm@1: #if defined(HAVE_ALTIVEC)
rlm@1:   #include "SFMT-alti.h"
rlm@1: #elif defined(HAVE_SSE2)
rlm@1:   #include "SFMT-sse2.h"
rlm@1: #endif
rlm@1: 
rlm@1: /**
rlm@1:  * This function simulate a 64-bit index of LITTLE ENDIAN 
rlm@1:  * in BIG ENDIAN machine.
rlm@1:  */
rlm@1: #ifdef ONLY64
rlm@1: inline static int idxof(int i) {
rlm@1:     return i ^ 1;
rlm@1: }
rlm@1: #else
rlm@1: inline static int idxof(int i) {
rlm@1:     return i;
rlm@1: }
rlm@1: #endif
rlm@1: /**
rlm@1:  * This function simulates SIMD 128-bit right shift by the standard C.
rlm@1:  * The 128-bit integer given in in is shifted by (shift * 8) bits.
rlm@1:  * This function simulates the LITTLE ENDIAN SIMD.
rlm@1:  * @param out the output of this function
rlm@1:  * @param in the 128-bit data to be shifted
rlm@1:  * @param shift the shift value
rlm@1:  */
rlm@1: #ifdef ONLY64
rlm@1: inline static void rshift128(w128_t *out, w128_t const *in, int shift) {
rlm@1:     uint64_t th, tl, oh, ol;
rlm@1: 
rlm@1:     th = ((uint64_t)in->u[2] << 32) | ((uint64_t)in->u[3]);
rlm@1:     tl = ((uint64_t)in->u[0] << 32) | ((uint64_t)in->u[1]);
rlm@1: 
rlm@1:     oh = th >> (shift * 8);
rlm@1:     ol = tl >> (shift * 8);
rlm@1:     ol |= th << (64 - shift * 8);
rlm@1:     out->u[0] = (uint32_t)(ol >> 32);
rlm@1:     out->u[1] = (uint32_t)(ol & 0xffffffff);
rlm@1:     out->u[2] = (uint32_t)(oh >> 32);
rlm@1:     out->u[3] = (uint32_t)(oh & 0xffffffff);
rlm@1: }
rlm@1: #else
rlm@1: inline static void rshift128(w128_t *out, w128_t const *in, int shift) {
rlm@1:     uint64_t th, tl, oh, ol;
rlm@1: 
rlm@1:     th = ((uint64_t)in->u[3] << 32) | ((uint64_t)in->u[2]);
rlm@1:     tl = ((uint64_t)in->u[1] << 32) | ((uint64_t)in->u[0]);
rlm@1: 
rlm@1:     oh = th >> (shift * 8);
rlm@1:     ol = tl >> (shift * 8);
rlm@1:     ol |= th << (64 - shift * 8);
rlm@1:     out->u[1] = (uint32_t)(ol >> 32);
rlm@1:     out->u[0] = (uint32_t)(ol & 0xffffffff);
rlm@1:     out->u[3] = (uint32_t)(oh >> 32);
rlm@1:     out->u[2] = (uint32_t)(oh & 0xffffffff);
rlm@1: }
rlm@1: #endif
rlm@1: /**
rlm@1:  * This function simulates SIMD 128-bit left shift by the standard C.
rlm@1:  * The 128-bit integer given in in is shifted by (shift * 8) bits.
rlm@1:  * This function simulates the LITTLE ENDIAN SIMD.
rlm@1:  * @param out the output of this function
rlm@1:  * @param in the 128-bit data to be shifted
rlm@1:  * @param shift the shift value
rlm@1:  */
rlm@1: #ifdef ONLY64
rlm@1: inline static void lshift128(w128_t *out, w128_t const *in, int shift) {
rlm@1:     uint64_t th, tl, oh, ol;
rlm@1: 
rlm@1:     th = ((uint64_t)in->u[2] << 32) | ((uint64_t)in->u[3]);
rlm@1:     tl = ((uint64_t)in->u[0] << 32) | ((uint64_t)in->u[1]);
rlm@1: 
rlm@1:     oh = th << (shift * 8);
rlm@1:     ol = tl << (shift * 8);
rlm@1:     oh |= tl >> (64 - shift * 8);
rlm@1:     out->u[0] = (uint32_t)(ol >> 32);
rlm@1:     out->u[1] = (uint32_t)(ol & 0xffffffff);
rlm@1:     out->u[2] = (uint32_t)(oh >> 32);
rlm@1:     out->u[3] = (uint32_t)(oh & 0xffffffff);
rlm@1: }
rlm@1: #else
rlm@1: inline static void lshift128(w128_t *out, w128_t const *in, int shift) {
rlm@1:     uint64_t th, tl, oh, ol;
rlm@1: 
rlm@1:     th = ((uint64_t)in->u[3] << 32) | ((uint64_t)in->u[2]);
rlm@1:     tl = ((uint64_t)in->u[1] << 32) | ((uint64_t)in->u[0]);
rlm@1: 
rlm@1:     oh = th << (shift * 8);
rlm@1:     ol = tl << (shift * 8);
rlm@1:     oh |= tl >> (64 - shift * 8);
rlm@1:     out->u[1] = (uint32_t)(ol >> 32);
rlm@1:     out->u[0] = (uint32_t)(ol & 0xffffffff);
rlm@1:     out->u[3] = (uint32_t)(oh >> 32);
rlm@1:     out->u[2] = (uint32_t)(oh & 0xffffffff);
rlm@1: }
rlm@1: #endif
rlm@1: 
rlm@1: /**
rlm@1:  * This function represents the recursion formula.
rlm@1:  * @param r output
rlm@1:  * @param a a 128-bit part of the internal state array
rlm@1:  * @param b a 128-bit part of the internal state array
rlm@1:  * @param c a 128-bit part of the internal state array
rlm@1:  * @param d a 128-bit part of the internal state array
rlm@1:  */
rlm@1: #if (!defined(HAVE_ALTIVEC)) && (!defined(HAVE_SSE2))
rlm@1: #ifdef ONLY64
rlm@1: inline static void do_recursion(w128_t *r, w128_t *a, w128_t *b, w128_t *c,
rlm@1: 				w128_t *d) {
rlm@1:     w128_t x;
rlm@1:     w128_t y;
rlm@1: 
rlm@1:     lshift128(&x, a, SL2);
rlm@1:     rshift128(&y, c, SR2);
rlm@1:     r->u[0] = a->u[0] ^ x.u[0] ^ ((b->u[0] >> SR1) & MSK2) ^ y.u[0] 
rlm@1: 	^ (d->u[0] << SL1);
rlm@1:     r->u[1] = a->u[1] ^ x.u[1] ^ ((b->u[1] >> SR1) & MSK1) ^ y.u[1] 
rlm@1: 	^ (d->u[1] << SL1);
rlm@1:     r->u[2] = a->u[2] ^ x.u[2] ^ ((b->u[2] >> SR1) & MSK4) ^ y.u[2] 
rlm@1: 	^ (d->u[2] << SL1);
rlm@1:     r->u[3] = a->u[3] ^ x.u[3] ^ ((b->u[3] >> SR1) & MSK3) ^ y.u[3] 
rlm@1: 	^ (d->u[3] << SL1);
rlm@1: }
rlm@1: #else
rlm@1: inline static void do_recursion(w128_t *r, w128_t *a, w128_t *b, w128_t *c,
rlm@1: 				w128_t *d) {
rlm@1:     w128_t x;
rlm@1:     w128_t y;
rlm@1: 
rlm@1:     lshift128(&x, a, SL2);
rlm@1:     rshift128(&y, c, SR2);
rlm@1:     r->u[0] = a->u[0] ^ x.u[0] ^ ((b->u[0] >> SR1) & MSK1) ^ y.u[0] 
rlm@1: 	^ (d->u[0] << SL1);
rlm@1:     r->u[1] = a->u[1] ^ x.u[1] ^ ((b->u[1] >> SR1) & MSK2) ^ y.u[1] 
rlm@1: 	^ (d->u[1] << SL1);
rlm@1:     r->u[2] = a->u[2] ^ x.u[2] ^ ((b->u[2] >> SR1) & MSK3) ^ y.u[2] 
rlm@1: 	^ (d->u[2] << SL1);
rlm@1:     r->u[3] = a->u[3] ^ x.u[3] ^ ((b->u[3] >> SR1) & MSK4) ^ y.u[3] 
rlm@1: 	^ (d->u[3] << SL1);
rlm@1: }
rlm@1: #endif
rlm@1: #endif
rlm@1: 
rlm@1: #if (!defined(HAVE_ALTIVEC)) && (!defined(HAVE_SSE2))
rlm@1: /**
rlm@1:  * This function fills the internal state array with pseudorandom
rlm@1:  * integers.
rlm@1:  */
rlm@1: inline static void gen_rand_all(void) {
rlm@1:     int i;
rlm@1:     w128_t *r1, *r2;
rlm@1: 
rlm@1:     r1 = &sfmt[N - 2];
rlm@1:     r2 = &sfmt[N - 1];
rlm@1:     for (i = 0; i < N - POS1; i++) {
rlm@1: 	do_recursion(&sfmt[i], &sfmt[i], &sfmt[i + POS1], r1, r2);
rlm@1: 	r1 = r2;
rlm@1: 	r2 = &sfmt[i];
rlm@1:     }
rlm@1:     for (; i < N; i++) {
rlm@1: 	do_recursion(&sfmt[i], &sfmt[i], &sfmt[i + POS1 - N], r1, r2);
rlm@1: 	r1 = r2;
rlm@1: 	r2 = &sfmt[i];
rlm@1:     }
rlm@1: }
rlm@1: 
rlm@1: /**
rlm@1:  * This function fills the user-specified array with pseudorandom
rlm@1:  * integers.
rlm@1:  *
rlm@1:  * @param array an 128-bit array to be filled by pseudorandom numbers.  
rlm@1:  * @param size number of 128-bit pseudorandom numbers to be generated.
rlm@1:  */
rlm@1: inline static void gen_rand_array(w128_t *array, int size) {
rlm@1:     int i, j;
rlm@1:     w128_t *r1, *r2;
rlm@1: 
rlm@1:     r1 = &sfmt[N - 2];
rlm@1:     r2 = &sfmt[N - 1];
rlm@1:     for (i = 0; i < N - POS1; i++) {
rlm@1: 	do_recursion(&array[i], &sfmt[i], &sfmt[i + POS1], r1, r2);
rlm@1: 	r1 = r2;
rlm@1: 	r2 = &array[i];
rlm@1:     }
rlm@1:     for (; i < N; i++) {
rlm@1: 	do_recursion(&array[i], &sfmt[i], &array[i + POS1 - N], r1, r2);
rlm@1: 	r1 = r2;
rlm@1: 	r2 = &array[i];
rlm@1:     }
rlm@1:     for (; i < size - N; i++) {
rlm@1: 	do_recursion(&array[i], &array[i - N], &array[i + POS1 - N], r1, r2);
rlm@1: 	r1 = r2;
rlm@1: 	r2 = &array[i];
rlm@1:     }
rlm@1:     for (j = 0; j < 2 * N - size; j++) {
rlm@1: 	sfmt[j] = array[j + size - N];
rlm@1:     }
rlm@1:     for (; i < size; i++, j++) {
rlm@1: 	do_recursion(&array[i], &array[i - N], &array[i + POS1 - N], r1, r2);
rlm@1: 	r1 = r2;
rlm@1: 	r2 = &array[i];
rlm@1: 	sfmt[j] = array[i];
rlm@1:     }
rlm@1: }
rlm@1: #endif
rlm@1: 
rlm@1: #if defined(BIG_ENDIAN64) && !defined(ONLY64) && !defined(HAVE_ALTIVEC)
rlm@1: inline static void swap(w128_t *array, int size) {
rlm@1:     int i;
rlm@1:     uint32_t x, y;
rlm@1: 
rlm@1:     for (i = 0; i < size; i++) {
rlm@1: 	x = array[i].u[0];
rlm@1: 	y = array[i].u[2];
rlm@1: 	array[i].u[0] = array[i].u[1];
rlm@1: 	array[i].u[2] = array[i].u[3];
rlm@1: 	array[i].u[1] = x;
rlm@1: 	array[i].u[3] = y;
rlm@1:     }
rlm@1: }
rlm@1: #endif
rlm@1: /**
rlm@1:  * This function represents a function used in the initialization
rlm@1:  * by init_by_array
rlm@1:  * @param x 32-bit integer
rlm@1:  * @return 32-bit integer
rlm@1:  */
rlm@1: static uint32_t func1(uint32_t x) {
rlm@1:     return (x ^ (x >> 27)) * (uint32_t)1664525UL;
rlm@1: }
rlm@1: 
rlm@1: /**
rlm@1:  * This function represents a function used in the initialization
rlm@1:  * by init_by_array
rlm@1:  * @param x 32-bit integer
rlm@1:  * @return 32-bit integer
rlm@1:  */
rlm@1: static uint32_t func2(uint32_t x) {
rlm@1:     return (x ^ (x >> 27)) * (uint32_t)1566083941UL;
rlm@1: }
rlm@1: 
rlm@1: /**
rlm@1:  * This function certificate the period of 2^{MEXP}
rlm@1:  */
rlm@1: static void period_certification(void) {
rlm@1:     int inner = 0;
rlm@1:     int i, j;
rlm@1:     uint32_t work;
rlm@1: 
rlm@1:     for (i = 0; i < 4; i++)
rlm@1: 	inner ^= psfmt32[idxof(i)] & parity[i];
rlm@1:     for (i = 16; i > 0; i >>= 1)
rlm@1: 	inner ^= inner >> i;
rlm@1:     inner &= 1;
rlm@1:     /* check OK */
rlm@1:     if (inner == 1) {
rlm@1: 	return;
rlm@1:     }
rlm@1:     /* check NG, and modification */
rlm@1:     for (i = 0; i < 4; i++) {
rlm@1: 	work = 1;
rlm@1: 	for (j = 0; j < 32; j++) {
rlm@1: 	    if ((work & parity[i]) != 0) {
rlm@1: 		psfmt32[idxof(i)] ^= work;
rlm@1: 		return;
rlm@1: 	    }
rlm@1: 	    work = work << 1;
rlm@1: 	}
rlm@1:     }
rlm@1: }
rlm@1: 
rlm@1: /*----------------
rlm@1:   PUBLIC FUNCTIONS
rlm@1:   ----------------*/
rlm@1: /**
rlm@1:  * This function returns the identification string.
rlm@1:  * The string shows the word size, the Mersenne exponent,
rlm@1:  * and all parameters of this generator.
rlm@1:  */
rlm@1: const char *get_idstring(void) {
rlm@1:     return IDSTR;
rlm@1: }
rlm@1: 
rlm@1: /**
rlm@1:  * This function returns the minimum size of array used for \b
rlm@1:  * fill_array32() function.
rlm@1:  * @return minimum size of array used for fill_array32() function.
rlm@1:  */
rlm@1: int get_min_array_size32(void) {
rlm@1:     return N32;
rlm@1: }
rlm@1: 
rlm@1: /**
rlm@1:  * This function returns the minimum size of array used for \b
rlm@1:  * fill_array64() function.
rlm@1:  * @return minimum size of array used for fill_array64() function.
rlm@1:  */
rlm@1: int get_min_array_size64(void) {
rlm@1:     return N64;
rlm@1: }
rlm@1: 
rlm@1: #ifndef ONLY64
rlm@1: /**
rlm@1:  * This function generates and returns 32-bit pseudorandom number.
rlm@1:  * init_gen_rand or init_by_array must be called before this function.
rlm@1:  * @return 32-bit pseudorandom number
rlm@1:  */
rlm@1: uint32_t gen_rand32(void) {
rlm@1:     uint32_t r;
rlm@1: 
rlm@1:     assert(initialized);
rlm@1:     if (idx >= N32) {
rlm@1: 	gen_rand_all();
rlm@1: 	idx = 0;
rlm@1:     }
rlm@1:     r = psfmt32[idx++];
rlm@1:     return r;
rlm@1: }
rlm@1: #endif
rlm@1: /**
rlm@1:  * This function generates and returns 64-bit pseudorandom number.
rlm@1:  * init_gen_rand or init_by_array must be called before this function.
rlm@1:  * The function gen_rand64 should not be called after gen_rand32,
rlm@1:  * unless an initialization is again executed. 
rlm@1:  * @return 64-bit pseudorandom number
rlm@1:  */
rlm@1: uint64_t gen_rand64(void) {
rlm@1: #if defined(BIG_ENDIAN64) && !defined(ONLY64)
rlm@1:     uint32_t r1, r2;
rlm@1: #else
rlm@1:     uint64_t r;
rlm@1: #endif
rlm@1: 
rlm@1:     assert(initialized);
rlm@1:     assert(idx % 2 == 0);
rlm@1: 
rlm@1:     if (idx >= N32) {
rlm@1: 	gen_rand_all();
rlm@1: 	idx = 0;
rlm@1:     }
rlm@1: #if defined(BIG_ENDIAN64) && !defined(ONLY64)
rlm@1:     r1 = psfmt32[idx];
rlm@1:     r2 = psfmt32[idx + 1];
rlm@1:     idx += 2;
rlm@1:     return ((uint64_t)r2 << 32) | r1;
rlm@1: #else
rlm@1:     r = psfmt64[idx / 2];
rlm@1:     idx += 2;
rlm@1:     return r;
rlm@1: #endif
rlm@1: }
rlm@1: 
rlm@1: #ifndef ONLY64
rlm@1: /**
rlm@1:  * This function generates pseudorandom 32-bit integers in the
rlm@1:  * specified array[] by one call. The number of pseudorandom integers
rlm@1:  * is specified by the argument size, which must be at least 624 and a
rlm@1:  * multiple of four.  The generation by this function is much faster
rlm@1:  * than the following gen_rand function.
rlm@1:  *
rlm@1:  * For initialization, init_gen_rand or init_by_array must be called
rlm@1:  * before the first call of this function. This function can not be
rlm@1:  * used after calling gen_rand function, without initialization.
rlm@1:  *
rlm@1:  * @param array an array where pseudorandom 32-bit integers are filled
rlm@1:  * by this function.  The pointer to the array must be \b "aligned"
rlm@1:  * (namely, must be a multiple of 16) in the SIMD version, since it
rlm@1:  * refers to the address of a 128-bit integer.  In the standard C
rlm@1:  * version, the pointer is arbitrary.
rlm@1:  *
rlm@1:  * @param size the number of 32-bit pseudorandom integers to be
rlm@1:  * generated.  size must be a multiple of 4, and greater than or equal
rlm@1:  * to (MEXP / 128 + 1) * 4.
rlm@1:  *
rlm@1:  * @note \b memalign or \b posix_memalign is available to get aligned
rlm@1:  * memory. Mac OSX doesn't have these functions, but \b malloc of OSX
rlm@1:  * returns the pointer to the aligned memory block.
rlm@1:  */
rlm@1: void fill_array32(uint32_t *array, int size) {
rlm@1:     assert(initialized);
rlm@1:     assert(idx == N32);
rlm@1:     assert(size % 4 == 0);
rlm@1:     assert(size >= N32);
rlm@1: 
rlm@1:     gen_rand_array((w128_t *)array, size / 4);
rlm@1:     idx = N32;
rlm@1: }
rlm@1: #endif
rlm@1: 
rlm@1: /**
rlm@1:  * This function generates pseudorandom 64-bit integers in the
rlm@1:  * specified array[] by one call. The number of pseudorandom integers
rlm@1:  * is specified by the argument size, which must be at least 312 and a
rlm@1:  * multiple of two.  The generation by this function is much faster
rlm@1:  * than the following gen_rand function.
rlm@1:  *
rlm@1:  * For initialization, init_gen_rand or init_by_array must be called
rlm@1:  * before the first call of this function. This function can not be
rlm@1:  * used after calling gen_rand function, without initialization.
rlm@1:  *
rlm@1:  * @param array an array where pseudorandom 64-bit integers are filled
rlm@1:  * by this function.  The pointer to the array must be "aligned"
rlm@1:  * (namely, must be a multiple of 16) in the SIMD version, since it
rlm@1:  * refers to the address of a 128-bit integer.  In the standard C
rlm@1:  * version, the pointer is arbitrary.
rlm@1:  *
rlm@1:  * @param size the number of 64-bit pseudorandom integers to be
rlm@1:  * generated.  size must be a multiple of 2, and greater than or equal
rlm@1:  * to (MEXP / 128 + 1) * 2
rlm@1:  *
rlm@1:  * @note \b memalign or \b posix_memalign is available to get aligned
rlm@1:  * memory. Mac OSX doesn't have these functions, but \b malloc of OSX
rlm@1:  * returns the pointer to the aligned memory block.
rlm@1:  */
rlm@1: void fill_array64(uint64_t *array, int size) {
rlm@1:     assert(initialized);
rlm@1:     assert(idx == N32);
rlm@1:     assert(size % 2 == 0);
rlm@1:     assert(size >= N64);
rlm@1: 
rlm@1:     gen_rand_array((w128_t *)array, size / 2);
rlm@1:     idx = N32;
rlm@1: 
rlm@1: #if defined(BIG_ENDIAN64) && !defined(ONLY64)
rlm@1:     swap((w128_t *)array, size /2);
rlm@1: #endif
rlm@1: }
rlm@1: 
rlm@1: /**
rlm@1:  * This function initializes the internal state array with a 32-bit
rlm@1:  * integer seed.
rlm@1:  *
rlm@1:  * @param seed a 32-bit integer used as the seed.
rlm@1:  */
rlm@1: void init_gen_rand(uint32_t seed) {
rlm@1:     int i;
rlm@1: 
rlm@1:     psfmt32[idxof(0)] = seed;
rlm@1:     for (i = 1; i < N32; i++) {
rlm@1: 	psfmt32[idxof(i)] = 1812433253UL * (psfmt32[idxof(i - 1)] 
rlm@1: 					    ^ (psfmt32[idxof(i - 1)] >> 30))
rlm@1: 	    + i;
rlm@1:     }
rlm@1:     idx = N32;
rlm@1:     period_certification();
rlm@1:     initialized = 1;
rlm@1: }
rlm@1: 
rlm@1: /**
rlm@1:  * This function initializes the internal state array,
rlm@1:  * with an array of 32-bit integers used as the seeds
rlm@1:  * @param init_key the array of 32-bit integers, used as a seed.
rlm@1:  * @param key_length the length of init_key.
rlm@1:  */
rlm@1: void init_by_array(uint32_t *init_key, int key_length) {
rlm@1:     int i, j, count;
rlm@1:     uint32_t r;
rlm@1:     int lag;
rlm@1:     int mid;
rlm@1:     int size = N * 4;
rlm@1: 
rlm@1:     if (size >= 623) {
rlm@1: 	lag = 11;
rlm@1:     } else if (size >= 68) {
rlm@1: 	lag = 7;
rlm@1:     } else if (size >= 39) {
rlm@1: 	lag = 5;
rlm@1:     } else {
rlm@1: 	lag = 3;
rlm@1:     }
rlm@1:     mid = (size - lag) / 2;
rlm@1: 
rlm@1:     memset(sfmt, 0x8b, sizeof(sfmt));
rlm@1:     if (key_length + 1 > N32) {
rlm@1: 	count = key_length + 1;
rlm@1:     } else {
rlm@1: 	count = N32;
rlm@1:     }
rlm@1:     r = func1(psfmt32[idxof(0)] ^ psfmt32[idxof(mid)] 
rlm@1: 	      ^ psfmt32[idxof(N32 - 1)]);
rlm@1:     psfmt32[idxof(mid)] += r;
rlm@1:     r += key_length;
rlm@1:     psfmt32[idxof(mid + lag)] += r;
rlm@1:     psfmt32[idxof(0)] = r;
rlm@1: 
rlm@1:     count--;
rlm@1:     for (i = 1, j = 0; (j < count) && (j < key_length); j++) {
rlm@1: 	r = func1(psfmt32[idxof(i)] ^ psfmt32[idxof((i + mid) % N32)] 
rlm@1: 		  ^ psfmt32[idxof((i + N32 - 1) % N32)]);
rlm@1: 	psfmt32[idxof((i + mid) % N32)] += r;
rlm@1: 	r += init_key[j] + i;
rlm@1: 	psfmt32[idxof((i + mid + lag) % N32)] += r;
rlm@1: 	psfmt32[idxof(i)] = r;
rlm@1: 	i = (i + 1) % N32;
rlm@1:     }
rlm@1:     for (; j < count; j++) {
rlm@1: 	r = func1(psfmt32[idxof(i)] ^ psfmt32[idxof((i + mid) % N32)] 
rlm@1: 		  ^ psfmt32[idxof((i + N32 - 1) % N32)]);
rlm@1: 	psfmt32[idxof((i + mid) % N32)] += r;
rlm@1: 	r += i;
rlm@1: 	psfmt32[idxof((i + mid + lag) % N32)] += r;
rlm@1: 	psfmt32[idxof(i)] = r;
rlm@1: 	i = (i + 1) % N32;
rlm@1:     }
rlm@1:     for (j = 0; j < N32; j++) {
rlm@1: 	r = func2(psfmt32[idxof(i)] + psfmt32[idxof((i + mid) % N32)] 
rlm@1: 		  + psfmt32[idxof((i + N32 - 1) % N32)]);
rlm@1: 	psfmt32[idxof((i + mid) % N32)] ^= r;
rlm@1: 	r -= i;
rlm@1: 	psfmt32[idxof((i + mid + lag) % N32)] ^= r;
rlm@1: 	psfmt32[idxof(i)] = r;
rlm@1: 	i = (i + 1) % N32;
rlm@1:     }
rlm@1: 
rlm@1:     idx = N32;
rlm@1:     period_certification();
rlm@1:     initialized = 1;
rlm@1: }