SSE2.cpp

#include <iostream>
#include "emmintrin.h"
 
using namespace std;
 
int main(int argc, char* argv[]) {
  const LENGTH = 1 << 10;
  const SSE_LENGTH = LENGTH >> 2;
  
  // Set constant vectors
  __m128i A = _mm_set1_epi32(1);
  __m128i B = _mm_set1_epi32(2);
  
  // Allocate result array and cast as vector
  unsigned int* C = (unsigned int*) _aligned_malloc(LENGTH * sizeof(unsigned int), 32);
  __m128i* result = (__m128i*) C;
  
  // Process vectorized addition: four items will be set per cycle
  for (size_t i = 0; i < SSE_LENGTH; i += 4)
    result[i] = _mm_add_epi32(A, B);
    
  // Print the result array
  for (size_t i = 0; i < LENGTH; ++i)
      cout << C[i] << endl;
 
  return 0;
}