xorshift+pro.c

#include <immintrin.h>

// Global seed value
__m256i seed;

// Use SSE2 perform pesudo random for given seed
__m128i xor_shift_plus_pro()
{
  // t = 0
  __m128i t = _mm_setzero_si128();
  // s = 0
  __m128i s = _mm_setzero_si128();
  // t = seed >> 128
  // s = seed & 0xffffffffffffffff
  _mm256_storeu2_m128i(&t, &s, seed);
  // t ^= t << 23;
  t = _mm_xor_si128(t, _mm_bslli_si128(t, 23));
  // t ^= t >> 17;
  t = _mm_xor_si128(t, _mm_bsrli_si128(t, 17));
  // t ^= s ^ (t >> 26)
  t = _mm_xor_si128(t, _mm_xor_si128(s, _mm_bsrli_si128(t, 26)));
  // seed = (s << 128) | t
  seed = _mm256_set_m128i(s, t);
  return _mm_add_epi32(s, t);
}