__kernel void reassemble(__global int *in,
           __global int *out,
           __global int *zeros,
           __global int *ones,
           int k,
           int n)
{
  size_t idx = get_global_id(0);
  size_t tid = get_local_id(0);
  size_t dim = get_local_size(0);
  size_t gid = get_group_id(0);
  int t, index;

  if (idx < n) {
    t = in[idx];
    /*t = (v==((t>>k)&0x1));*/
    t = ((t>>k)&0x1);

    barrier(CLK_LOCAL_MEM_FENCE);

    if (t) {
      index = zeros[n-1] + ones[idx] - 1;
    } else {
      index = zeros[idx] - 1;
    }
    out[index] = in[idx];

    barrier(CLK_GLOBAL_MEM_FENCE);

    if (idx == 0) {
      __global int *tPtr = in;
      in = out;
      out = tPtr;
    }
  }
}