rlm@1: /** 
rlm@1:  * @file SFMT-alti.h 
rlm@1:  *
rlm@1:  * @brief SIMD oriented Fast Mersenne Twister(SFMT)
rlm@1:  * pseudorandom number generator
rlm@1:  *
rlm@1:  * @author Mutsuo Saito (Hiroshima University)
rlm@1:  * @author Makoto Matsumoto (Hiroshima University)
rlm@1:  *
rlm@1:  * Copyright (C) 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.
rlm@1:  * see LICENSE.txt
rlm@1:  */
rlm@1: 
rlm@1: #ifndef SFMT_ALTI_H
rlm@1: #define SFMT_ALTI_H
rlm@1: 
rlm@1: inline static vector unsigned int vec_recursion(vector unsigned int a,
rlm@1: 						vector unsigned int b,
rlm@1: 						vector unsigned int c,
rlm@1: 						vector unsigned int d)
rlm@1:     ALWAYSINLINE;
rlm@1: 
rlm@1: /**
rlm@1:  * This function represents the recursion formula in AltiVec and BIG ENDIAN.
rlm@1:  * @param a a 128-bit part of the interal state array
rlm@1:  * @param b a 128-bit part of the interal state array
rlm@1:  * @param c a 128-bit part of the interal state array
rlm@1:  * @param d a 128-bit part of the interal state array
rlm@1:  * @return output
rlm@1:  */
rlm@1: inline static vector unsigned int vec_recursion(vector unsigned int a,
rlm@1: 						vector unsigned int b,
rlm@1: 						vector unsigned int c,
rlm@1: 						vector unsigned int d) {
rlm@1: 
rlm@1:     const vector unsigned int sl1 = ALTI_SL1;
rlm@1:     const vector unsigned int sr1 = ALTI_SR1;
rlm@1: #ifdef ONLY64
rlm@1:     const vector unsigned int mask = ALTI_MSK64;
rlm@1:     const vector unsigned char perm_sl = ALTI_SL2_PERM64;
rlm@1:     const vector unsigned char perm_sr = ALTI_SR2_PERM64;
rlm@1: #else
rlm@1:     const vector unsigned int mask = ALTI_MSK;
rlm@1:     const vector unsigned char perm_sl = ALTI_SL2_PERM;
rlm@1:     const vector unsigned char perm_sr = ALTI_SR2_PERM;
rlm@1: #endif
rlm@1:     vector unsigned int v, w, x, y, z;
rlm@1:     x = vec_perm(a, (vector unsigned int)perm_sl, perm_sl);
rlm@1:     v = a;
rlm@1:     y = vec_sr(b, sr1);
rlm@1:     z = vec_perm(c, (vector unsigned int)perm_sr, perm_sr);
rlm@1:     w = vec_sl(d, sl1);
rlm@1:     z = vec_xor(z, w);
rlm@1:     y = vec_and(y, mask);
rlm@1:     v = vec_xor(v, x);
rlm@1:     z = vec_xor(z, y);
rlm@1:     z = vec_xor(z, v);
rlm@1:     return z;
rlm@1: }
rlm@1: 
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:     vector unsigned int r, r1, r2;
rlm@1: 
rlm@1:     r1 = sfmt[N - 2].s;
rlm@1:     r2 = sfmt[N - 1].s;
rlm@1:     for (i = 0; i < N - POS1; i++) {
rlm@1: 	r = vec_recursion(sfmt[i].s, sfmt[i + POS1].s, r1, r2);
rlm@1: 	sfmt[i].s = r;
rlm@1: 	r1 = r2;
rlm@1: 	r2 = r;
rlm@1:     }
rlm@1:     for (; i < N; i++) {
rlm@1: 	r = vec_recursion(sfmt[i].s, sfmt[i + POS1 - N].s, r1, r2);
rlm@1: 	sfmt[i].s = r;
rlm@1: 	r1 = r2;
rlm@1: 	r2 = r;
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 pesudorandom 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:     vector unsigned int r, r1, r2;
rlm@1: 
rlm@1:     r1 = sfmt[N - 2].s;
rlm@1:     r2 = sfmt[N - 1].s;
rlm@1:     for (i = 0; i < N - POS1; i++) {
rlm@1: 	r = vec_recursion(sfmt[i].s, sfmt[i + POS1].s, r1, r2);
rlm@1: 	array[i].s = r;
rlm@1: 	r1 = r2;
rlm@1: 	r2 = r;
rlm@1:     }
rlm@1:     for (; i < N; i++) {
rlm@1: 	r = vec_recursion(sfmt[i].s, array[i + POS1 - N].s, r1, r2);
rlm@1: 	array[i].s = r;
rlm@1: 	r1 = r2;
rlm@1: 	r2 = r;
rlm@1:     }
rlm@1:     /* main loop */
rlm@1:     for (; i < size - N; i++) {
rlm@1: 	r = vec_recursion(array[i - N].s, array[i + POS1 - N].s, r1, r2);
rlm@1: 	array[i].s = r;
rlm@1: 	r1 = r2;
rlm@1: 	r2 = r;
rlm@1:     }
rlm@1:     for (j = 0; j < 2 * N - size; j++) {
rlm@1: 	sfmt[j].s = array[j + size - N].s;
rlm@1:     }
rlm@1:     for (; i < size; i++) {
rlm@1: 	r = vec_recursion(array[i - N].s, array[i + POS1 - N].s, r1, r2);
rlm@1: 	array[i].s = r;
rlm@1: 	sfmt[j++].s = r;
rlm@1: 	r1 = r2;
rlm@1: 	r2 = r;
rlm@1:     }
rlm@1: }
rlm@1: 
rlm@1: #ifndef ONLY64
rlm@1: #if defined(__APPLE__)
rlm@1: #define ALTI_SWAP (vector unsigned char) \
rlm@1: 	(4, 5, 6, 7, 0, 1, 2, 3, 12, 13, 14, 15, 8, 9, 10, 11)
rlm@1: #else
rlm@1: #define ALTI_SWAP {4, 5, 6, 7, 0, 1, 2, 3, 12, 13, 14, 15, 8, 9, 10, 11}
rlm@1: #endif
rlm@1: /**
rlm@1:  * This function swaps high and low 32-bit of 64-bit integers in user
rlm@1:  * specified array.
rlm@1:  *
rlm@1:  * @param array an 128-bit array to be swaped.
rlm@1:  * @param size size of 128-bit array.
rlm@1:  */
rlm@1: inline static void swap(w128_t *array, int size) {
rlm@1:     int i;
rlm@1:     const vector unsigned char perm = ALTI_SWAP;
rlm@1: 
rlm@1:     for (i = 0; i < size; i++) {
rlm@1: 	array[i].s = vec_perm(array[i].s, (vector unsigned int)perm, perm);
rlm@1:     }
rlm@1: }
rlm@1: #endif
rlm@1: 
rlm@1: #endif