// #define TEST
#define X	95
#define Y	22

//#define USE_DBL
#ifdef USE_DBL
#define TYPE_FLT  double
#define TYPE_INT  long
#define MASK      0xFFFF
#define SHIFT     16
#define EPSILSON  4.94065645841247E-324
#else
#define TYPE_FLT  float
#define TYPE_INT  int
#define MASK      0x00FF
#define SHIFT     8
#define EPSILSON  1.401298E-45
#endif

constant sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE |
                                CLK_ADDRESS_CLAMP_TO_EDGE |
                                CLK_FILTER_LINEAR;


struct input {
  int filter_width;
  int filter_height;
  int width;
  int height;
};

TYPE_FLT read_data(image2d_t input_frame, int x, int z)
{
  int2 gid = (int2)(x, z);
  uint4 pixel = read_imageui(input_frame, sampler, gid);
  TYPE_INT temp =  (TYPE_INT)((TYPE_INT)pixel.x & MASK) | 
                    (TYPE_INT)(((TYPE_INT)pixel.y & MASK) << SHIFT) |
                    (TYPE_INT)(((TYPE_INT)pixel.z & MASK) << (SHIFT * 2)) | 
                    (TYPE_INT)(((TYPE_INT)pixel.w & MASK) << (SHIFT * 3));
  TYPE_FLT raw_data = *((TYPE_FLT*)(&temp));
  return raw_data;
}


kernel void Enhance(read_only image2d_t input_frame, read_write image2d_t output_frame, global TYPE_FLT* filter, struct input params) 
{
  int2 gid = (int2)(get_global_id(0), get_global_id(1));
  TYPE_FLT input = read_data(input_frame, gid.x, gid.y);
  TYPE_FLT output_data = 0;

  int p = params.filter_height;
  int q = params.filter_width;

  int lag1 = (p - 1) / 2;
  int lag2 = (q - 1) / 2;

  if(gid.x >= lag2 && gid.x < params.width - lag2 && gid.y >= lag1 && gid.y < params.height - lag1)
  {
    for(int i = -lag1; i <= lag1 ; i++)
    {
      for(int j = -lag2; j <= lag2 ; j++)
      {
          if(gid.x == 947 && gid.y == 1)
          {
            printf("input: %f; ", output_data);
          }

        output_data += read_data(input_frame, gid.x + j, gid.y + i) * filter[(i + lag1) * q + (j + lag2)];
      }
    }
  }
  else
  {
    output_data = input;
  }

  if(output_data < 0)
    output_data = 0;

  if(output_data > 255)
    output_data = 255;

  
  TYPE_INT out = *((TYPE_INT*)(&output_data));
  uint4 pixel;
  pixel.x = (TYPE_INT)(out & MASK);
  pixel.y = (TYPE_INT)((out >> SHIFT) & MASK);
  pixel.z = (TYPE_INT)((out >> (SHIFT *2)) & MASK);
  pixel.w = (TYPE_INT)((out >> (SHIFT * 3)) & MASK);

  write_imageui(output_frame, gid, pixel);
}