rlm@1: /** 
rlm@1:  * @file  SFMT-sse2.h
rlm@1:  * @brief SIMD oriented Fast Mersenne Twister(SFMT) for Intel SSE2
rlm@1:  *
rlm@1:  * @author Mutsuo Saito (Hiroshima University)
rlm@1:  * @author Makoto Matsumoto (Hiroshima University)
rlm@1:  *
rlm@1:  * @note We assume LITTLE ENDIAN in this file
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: 
rlm@1: #ifndef SFMT_SSE2_H
rlm@1: #define SFMT_SSE2_H
rlm@1: 
rlm@1: PRE_ALWAYS static __m128i mm_recursion(__m128i *a, __m128i *b, __m128i c,
rlm@1: 				   __m128i d, __m128i mask) ALWAYSINLINE;
rlm@1: 
rlm@1: /**
rlm@1:  * This function represents the recursion formula.
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:  * @param mask 128-bit mask
rlm@1:  * @return output
rlm@1:  */
rlm@1: PRE_ALWAYS static __m128i mm_recursion(__m128i *a, __m128i *b, 
rlm@1: 				   __m128i c, __m128i d, __m128i mask) {
rlm@1:     __m128i v, x, y, z;
rlm@1:     
rlm@1:     x = _mm_load_si128(a);
rlm@1:     y = _mm_srli_epi32(*b, SR1);
rlm@1:     z = _mm_srli_si128(c, SR2);
rlm@1:     v = _mm_slli_epi32(d, SL1);
rlm@1:     z = _mm_xor_si128(z, x);
rlm@1:     z = _mm_xor_si128(z, v);
rlm@1:     x = _mm_slli_si128(x, SL2);
rlm@1:     y = _mm_and_si128(y, mask);
rlm@1:     z = _mm_xor_si128(z, x);
rlm@1:     z = _mm_xor_si128(z, y);
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:     __m128i r, r1, r2, mask;
rlm@1:     mask = _mm_set_epi32(MSK4, MSK3, MSK2, MSK1);
rlm@1: 
rlm@1:     r1 = _mm_load_si128(&sfmt[N - 2].si);
rlm@1:     r2 = _mm_load_si128(&sfmt[N - 1].si);
rlm@1:     for (i = 0; i < N - POS1; i++) {
rlm@1: 	r = mm_recursion(&sfmt[i].si, &sfmt[i + POS1].si, r1, r2, mask);
rlm@1: 	_mm_store_si128(&sfmt[i].si, r);
rlm@1: 	r1 = r2;
rlm@1: 	r2 = r;
rlm@1:     }
rlm@1:     for (; i < N; i++) {
rlm@1: 	r = mm_recursion(&sfmt[i].si, &sfmt[i + POS1 - N].si, r1, r2, mask);
rlm@1: 	_mm_store_si128(&sfmt[i].si, 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:     __m128i r, r1, r2, mask;
rlm@1:     mask = _mm_set_epi32(MSK4, MSK3, MSK2, MSK1);
rlm@1: 
rlm@1:     r1 = _mm_load_si128(&sfmt[N - 2].si);
rlm@1:     r2 = _mm_load_si128(&sfmt[N - 1].si);
rlm@1:     for (i = 0; i < N - POS1; i++) {
rlm@1: 	r = mm_recursion(&sfmt[i].si, &sfmt[i + POS1].si, r1, r2, mask);
rlm@1: 	_mm_store_si128(&array[i].si, r);
rlm@1: 	r1 = r2;
rlm@1: 	r2 = r;
rlm@1:     }
rlm@1:     for (; i < N; i++) {
rlm@1: 	r = mm_recursion(&sfmt[i].si, &array[i + POS1 - N].si, r1, r2, mask);
rlm@1: 	_mm_store_si128(&array[i].si, 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 = mm_recursion(&array[i - N].si, &array[i + POS1 - N].si, r1, r2,
rlm@1: 			 mask);
rlm@1: 	_mm_store_si128(&array[i].si, r);
rlm@1: 	r1 = r2;
rlm@1: 	r2 = r;
rlm@1:     }
rlm@1:     for (j = 0; j < 2 * N - size; j++) {
rlm@1: 	r = _mm_load_si128(&array[j + size - N].si);
rlm@1: 	_mm_store_si128(&sfmt[j].si, r);
rlm@1:     }
rlm@1:     for (; i < size; i++) {
rlm@1: 	r = mm_recursion(&array[i - N].si, &array[i + POS1 - N].si, r1, r2,
rlm@1: 			 mask);
rlm@1: 	_mm_store_si128(&array[i].si, r);
rlm@1: 	_mm_store_si128(&sfmt[j++].si, r);
rlm@1: 	r1 = r2;
rlm@1: 	r2 = r;
rlm@1:     }
rlm@1: }
rlm@1: 
rlm@1: #endif