const int w_chunk  = w/16;         //w is width
    uint8_t*    i0       = (uint8_t*)( in );
    uint8_t*    i1       = (uint8_t*)( in ) + w_chunk*1;
    uint8_t*    i2       = (uint8_t*)( in ) + w_chunk*2;
    int16_t* result_h  = (int16_t*)( out_h ) + 2*w_chunk;
    int16_t* result_v  = (int16_t*)( out_v ) + 2*w_chunk;
    uint8_t* end_input = (uint8_t*)( in ) + w_chunk*h;
    for( ; i2 != end_input; i0+= 16, i1+= 16, i2+= 16, result_v+= 16, result_h+= 16 ) {
        some code }

for (int i=0; i<8;i++)
  {
   result_h[i] = 0;
   result_h[i + 8] = 0;
   result_v[i] = 0;
   result_v[i + 8] = 0;
       result_h[i]     = (int16_t) i0[i] + result_h[i] ;
       result_h[i + 8]     = (int16_t) i0[i + 8] + result_h[i + 8] ;
       result_v[i]     = (int16_t) i0[i] + result_v[i] ;
       result_v[i + 8]     = (int16_t) i0[i + 8] + result_v[i + 8] ;
  }