__kernel void sample(__global uint* global_A, __global uint* global_B,
                     uint nPoints, uint nQueries) {
  //for (uint i = 0; i < nPoints; i++) {
  //  printf("%u ", global_A[i]);
  //}
  printf("\n\n");
  for (uint loopQueryId = 0; loopQueryId < nQueries; loopQueryId++) {
    global_A[0] = 0;

    uint maximum_depth = nPoints;
    for (uint depth = 0; depth <= maximum_depth; depth++) {
      for (uint i = 0; i < nPoints; i++) {
        global_B[i] = global_A[i] + 1;
        printf("%u ", global_B[i]);
      }
      printf("\n");
      __global uint* temp = global_A;
      global_A = global_B;
      global_B = temp;
    }
    printf("\n");
  }
}