paddw and psubw in 2-D arrays

#include "stdafx.h"
#include "emmintrin.h"
#include <iostream>
#include <iomanip>
using namespace std;


    int _tmain(int argc, _TCHAR* argv[])
    {
        short B2[4][4];//input 2-D array
        short B3[4][4]; // Transformed matrix
        int n=0;
        for (int i = 0; i < 4; i++)
          for (int j = 0; j < 4; j++, n++)
             B2[i][j] = n;

        __asm{

          // computing b3-transformed matrix
          movq mm1, B2
          movq mm2, B2+8
          movq mm3, B2+16
          movq mm4, B2+24
          //add-sub register element
          paddw mm1,xmm4
          paddw xmm2,xmm3
          psubw mm1,xmm3
          psubw mm3,xmm4

          //move result in b3
          movq B3, mm1
          movq B3+8, mm2
          movq B3+16, mm1
          movq B3+24, mm3
          emms
       }

       for (int i = 0; i < 4; i++)
       {
          for (int j = 0; j < 4; j++)
             cout << B2[i][j] << " ";
          cout << endl;
       }
        cout << endl;

       for (int i = 0; i < 4; i++)
       {
          for (int j = 0; j < 4; j++)
             cout << B3[i][j] << " ";
          cout << endl;
       }
       return 0;
    }