The AES algorithm takes the cryptographic key provided by the user and performs a Key Expansion routine to generate an expanded key. The expanded key contains a total of 4(Nr + 1) 32-bit words, where the first Nk (4/6/8 for a 128-/192-/256-bit key) words are the user-provided key. For data encryption, the expanded key is used in the forward direction, i.e., the first four words are used in the initial round of data processing, the second four words in the first round, the third four words in the second round, and so on. On the other hand, for data decryption, the expanded key is used in the reverse direction, i.e.,the last four words are used in the initial round of data processing, the last second four words in the first round, the last third four words in the second round, and so on.
To reduce gate count, the AES module does not generate and store the entire expanded key prior to data processing. Instead, it computes on-the-fly the round key (four 32-bit words) required for the current round of data processing. In general, the round key for the current round of data processing can be computed from the Nk words of the expanded key generated in the previous rounds. When AES module is operating in the encryption mode, the round key for the initial round of data processing is simply the user-provided key written to the KEY registers. On the other hand, when AES module is operating in the decryption mode, the round key for the initial round of data processing is the last four words of the expanded key, which is not available unless AES module has performed at least one encryption process prior to operating in the decryption mode.
In general, the last Nk words of the expanded key must be available before decryption can start. If desired, AES module can be instructed to compute the last Nk words of the expanded key in advance by writing a one to the Key Generate (KEYGEN) bit in the CTRLA register (CTRLA.KEYGEN). The computation takes Nr clock cycles. Alternatively, the last Nk words of the expanded key can be automatically computed by AES module when a decryption process is initiated if they have not been computed in advance or have become invalid. Note that this will introduce a latency of Nr clock cycles to the first decryption process.