__global__ void kernel_median3_2pix( short *output,
									 int i_dim, int j_dim)
{
                // j base coordinate = 2*(thread index)
  int j = __mul24(__mul24(blockIdx.x,blockDim.x) + threadIdx.x,2) ; 
  int i = __mul24(blockIdx.y,blockDim.y) + threadIdx.y ;
  int a0, a1, a2, a3, a4, a5 ;       // for left window
  int b0, b1, b2, b3, b4, b5 ;      // for right window
  
  a0 = tex2D(tex_img_ins, j  , i-1); // 6 common pixels
  a1 = tex2D(tex_img_ins, j+1, i-1);
  a2 = tex2D(tex_img_ins, j  , i  );
  a3 = tex2D(tex_img_ins, j+1, i  );
  a4 = tex2D(tex_img_ins, j  , i+1);
  a5 = tex2D(tex_img_ins, j+1, i+1);

  minmax6(&a0, &a1, &a2, &a3, &a4, &a5);// common minmax
  b0=a0; b1=a1; b2=a2; b3=a3; b4=a4; b5=a5;// separation
 
  a5 = tex2D(tex_img_ins, j-1, i);    //separate process
  b5 = tex2D(tex_img_ins, j+2, i);
  minmax5(&a1, &a2, &a3, &a4, &a5);
  minmax5(&b1, &b2, &b3, &b4, &b5);
  a5 = tex2D(tex_img_ins, j-1, i-1);
  b5 = tex2D(tex_img_ins, j+2, i-1);
  minmax4(&a2, &a3, &a4, &a5);
  minmax4(&b2, &b3, &b4, &b5);
  a5 = tex2D(tex_img_ins, j-1, i+1);
  b5 = tex2D(tex_img_ins, j+2, i+1);  
  minmax3(&a3, &a4, &a5);
  minmax3(&b3, &b4, &b5);
  
  output[ __mul24(i, j_dim) +j ] = a4 ;      //2 outputs
  output[ __mul24(i, j_dim) +j+1 ] = b4 ;
 
}