psubw instruction bug

#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]; // output
   int n = 0;
   for (int i = 0; i < 4; i++)
     for (int j = 0; j < 4; j++, n++)
        B2[i][j] = n;

   __asm{
      movq mm0, B2
      movq mm1, B2+8
      movq mm2, B2+16
      movq mm3, B2+24

      // add-sub register element
      paddw mm0, mm3 // mm0+mm3, result correct
      paddw mm1, mm2 // mm1+mm2, result correct
      psubw mm2, mm1 // mm1-mm2, result not correct
      psubw mm3, mm0 // mm0-mm3, result not correct
  
      // move result in B3
      movq B3, mm0
      movq B3+8, mm1
      movq B3+16, mm2
      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;
}