Mercurial > vba-clojure

     1 /** 

     2  * @file  SFMT-sse2.h

     3  * @brief SIMD oriented Fast Mersenne Twister(SFMT) for Intel SSE2

     4  *

     5  * @author Mutsuo Saito (Hiroshima University)

     6  * @author Makoto Matsumoto (Hiroshima University)

     7  *

     8  * @note We assume LITTLE ENDIAN in this file

     9  *

    10  * Copyright (C) 2006, 2007 Mutsuo Saito, Makoto Matsumoto and Hiroshima

    11  * University. All rights reserved.

    12  *

    13  * The new BSD License is applied to this software, see LICENSE.txt

    14  */

    15 

    16 #ifndef SFMT_SSE2_H

    17 #define SFMT_SSE2_H

    18 

    19 PRE_ALWAYS static __m128i mm_recursion(__m128i *a, __m128i *b, __m128i c,

    20 				   __m128i d, __m128i mask) ALWAYSINLINE;

    21 

    22 /**

    23  * This function represents the recursion formula.

    24  * @param a a 128-bit part of the interal state array

    25  * @param b a 128-bit part of the interal state array

    26  * @param c a 128-bit part of the interal state array

    27  * @param d a 128-bit part of the interal state array

    28  * @param mask 128-bit mask

    29  * @return output

    30  */

    31 PRE_ALWAYS static __m128i mm_recursion(__m128i *a, __m128i *b, 

    32 				   __m128i c, __m128i d, __m128i mask) {

    33     __m128i v, x, y, z;

    34     

    35     x = _mm_load_si128(a);

    36     y = _mm_srli_epi32(*b, SR1);

    37     z = _mm_srli_si128(c, SR2);

    38     v = _mm_slli_epi32(d, SL1);

    39     z = _mm_xor_si128(z, x);

    40     z = _mm_xor_si128(z, v);

    41     x = _mm_slli_si128(x, SL2);

    42     y = _mm_and_si128(y, mask);

    43     z = _mm_xor_si128(z, x);

    44     z = _mm_xor_si128(z, y);

    45     return z;

    46 }

    47 

    48 /**

    49  * This function fills the internal state array with pseudorandom

    50  * integers.

    51  */

    52 inline static void gen_rand_all(void) {

    53     int i;

    54     __m128i r, r1, r2, mask;

    55     mask = _mm_set_epi32(MSK4, MSK3, MSK2, MSK1);

    56 

    57     r1 = _mm_load_si128(&sfmt[N - 2].si);

    58     r2 = _mm_load_si128(&sfmt[N - 1].si);

    59     for (i = 0; i < N - POS1; i++) {

    60 	r = mm_recursion(&sfmt[i].si, &sfmt[i + POS1].si, r1, r2, mask);

    61 	_mm_store_si128(&sfmt[i].si, r);

    62 	r1 = r2;

    63 	r2 = r;

    64     }

    65     for (; i < N; i++) {

    66 	r = mm_recursion(&sfmt[i].si, &sfmt[i + POS1 - N].si, r1, r2, mask);

    67 	_mm_store_si128(&sfmt[i].si, r);

    68 	r1 = r2;

    69 	r2 = r;

    70     }

    71 }

    72 

    73 /**

    74  * This function fills the user-specified array with pseudorandom

    75  * integers.

    76  *

    77  * @param array an 128-bit array to be filled by pseudorandom numbers.  

    78  * @param size number of 128-bit pesudorandom numbers to be generated.

    79  */

    80 inline static void gen_rand_array(w128_t *array, int size) {

    81     int i, j;

    82     __m128i r, r1, r2, mask;

    83     mask = _mm_set_epi32(MSK4, MSK3, MSK2, MSK1);

    84 

    85     r1 = _mm_load_si128(&sfmt[N - 2].si);

    86     r2 = _mm_load_si128(&sfmt[N - 1].si);

    87     for (i = 0; i < N - POS1; i++) {

    88 	r = mm_recursion(&sfmt[i].si, &sfmt[i + POS1].si, r1, r2, mask);

    89 	_mm_store_si128(&array[i].si, r);

    90 	r1 = r2;

    91 	r2 = r;

    92     }

    93     for (; i < N; i++) {

    94 	r = mm_recursion(&sfmt[i].si, &array[i + POS1 - N].si, r1, r2, mask);

    95 	_mm_store_si128(&array[i].si, r);

    96 	r1 = r2;

    97 	r2 = r;

    98     }

    99     /* main loop */

   100     for (; i < size - N; i++) {

   101 	r = mm_recursion(&array[i - N].si, &array[i + POS1 - N].si, r1, r2,

   102 			 mask);

   103 	_mm_store_si128(&array[i].si, r);

   104 	r1 = r2;

   105 	r2 = r;

   106     }

   107     for (j = 0; j < 2 * N - size; j++) {

   108 	r = _mm_load_si128(&array[j + size - N].si);

   109 	_mm_store_si128(&sfmt[j].si, r);

   110     }

   111     for (; i < size; i++) {

   112 	r = mm_recursion(&array[i - N].si, &array[i + POS1 - N].si, r1, r2,

   113 			 mask);

   114 	_mm_store_si128(&array[i].si, r);

   115 	_mm_store_si128(&sfmt[j++].si, r);

   116 	r1 = r2;

   117 	r2 = r;

   118     }

   119 }

   120 

   121 #endif
author	Dylan Holmes <ocsenave@gmail.com>
date	Mon, 19 Mar 2012 05:20:34 -0500
parents	f9f4f1b99eed
children