motion_estimation.cl

__constant sampler_t sampler =
CLK_NORMALIZED_COORDS_FALSE
| CLK_ADDRESS_CLAMP_TO_EDGE
| CLK_FILTER_NEAREST;

const int BLOCKSIZE = 16;
const int ME_RANGE = 32;
const int ME_RANGE_x2 = 64;

__kernel void motion_est(	
  read_only image2d_t reference,
  read_only image2d_t block,
  global int *output) {
	const int2 pos = { get_global_id(0), get_global_id(1) };
  const int2 blkpos = pos/ME_RANGE_x2;
  const int2 searchpos = (pos%ME_RANGE_x2)-ME_RANGE;
	unsigned int sum = 0;
	for (int y = searchpos.s1; y < searchpos.s1+BLOCKSIZE; y++) {
		for (int x = searchpos.s0; x <searchpos.s0+BLOCKSIZE; x++) {
      int4 pixel_ref = convert_int4(read_imagef(reference, sampler, blkpos*BLOCKSIZE + (int2)(x, y))*255);
      int4 pixel = convert_int4(read_imagef(block, sampler, blkpos*BLOCKSIZE + (int2)(x, y))*255);
			sum += abs_diff(pixel_ref.x, pixel.x);
		}
	}
  output[get_global_id(0) +  get_global_id(1)*((832/BLOCKSIZE)*ME_RANGE_x2)] = sum;
}