__global__ void kernel_medianR( short *output,
                                int i_dim, int j_dim, int r)
{   
  // absolute coordinates of the center pixel
  int j = __mul24(blockIdx.x,blockDim.x) + threadIdx.x ; 
  int i = __mul24(blockIdx.y,blockDim.y) + threadIdx.y ;
  
  short cpt, ic, jc ;
  short histogram[256] ; // 8 bit image    
  // zeroing histogram data
  for (ic =0; ic<256; ic++) histogram[ ic ]=0 ; 
  // histogram filling
  for(ic=i-r; ic<=i+r; ic++ )
	for(jc=j-r; jc<=j+r; jc++)
	  histogram[ tex2D(tex_img_ins, jc, ic) ]++ ;
  // histogram parsing
  cpt = 0 ;
  for(ic=0; ic<256; ic++)
	{
	  cpt += histogram[ ic ] ;
	  // selection of 50% percentile 
	  if ( cpt > ((2*r+1)*(2*r+1))>>1 ) break ; 
	} 
  output[ __mul24(i, j_dim) +j ] =  ic ;  
}