def uncompress_rle(data, nbits):
  data_uncompressed, i = [], 0
  maskval, maskcnt = (1 << nbits) - 1, (0xFF >> nbits >> 1) << nbits
  while i < len(data):
    v = data[i]
    mv = v & maskval
    c = (v & maskcnt) >> nbits
    if v & 0b10000000:
      i += 1
      c |= data[i] << (7 - nbits)
    data_uncompressed.extend([mv] * c)
    i += 1
  return bytearray(data_uncompressed)

def pack(data, nbits):
  data_packed, npacks = [], 8 // nbits
  ipack, val_packed = 0, 0
  for k in range(len(data)):
    val_packed |= data[k] << (nbits * ipack)
    ipack = (ipack + 1) % npacks
    if not ipack or k + 1 >= len(data):
      data_packed.append(val_packed)
      val_packed = 0
  return bytearray(data_packed)

def compress_rle(data, nbits):
  size = len(data)
  maskcol = (1 << nbits) - 1
  maskcnt = 0xFF >> nbits >> 1
  rle = []
  i = 0
  val_cnt = 0
  while i <= size:
    val_cour = None
    if i < size: val_cour = data[i]
    if i==0: val_prec = val_cour
    val_cnt += val_prec == val_cour
    if val_prec != val_cour or i==size:
      while val_cnt:
        vim = val_prec
        if val_cnt <= maskcnt:
          ic = val_cnt
          vim |= ic << nbits
        else:
          ic = min(val_cnt, (1<<(15-nbits))-1)
          rle.append(vim | ((ic & maskcnt) << nbits) | (1<<7))
          vim = ic >> (7-nbits)
        val_cnt -= ic
        rle.append(vim)
      val_cnt = 1
      val_prec = val_cour
    i += 1
  return bytearray(rle)

def compress_rle_img_cols(nbits, data, w, h):
  cols = []
  for i in range(w):
    cols.append(compress_rle(data[i*h:(i+1)*h], nbits))
  return tuple(cols)