in the decoding of later layers in the first iteration.
Hence, for these layers, the a posteriori messages computed
from the previous layer are the v-to-c messages of the
current layer. Starting from the second decoding iteration, the
c-to-v messages of the same layer from the previous iteration
need to be subtracted from the a posteriori messages to generate
the v-to-c messages according to (3). Storing the c-to-v messages
of each layer requires a RAM A, which is significantly
larger than a RAM E when is not small. A RAM A also occupies
substantially more area than path constructors. Accordingly,
we propose to store the sorted v-to-c messages instead
of c-to-v messages. When c-to-v messages are needed in the
decoding of the same layer in the next decoding iteration, they
are recovered from the sorting results using Path Constructor-II,
which also has copies of the architecture in Fig. 4.
RAM E blocks are required to store the results and intermediate
data of the sorting. The allocation of these RAM E blocks will
be detailed in Section IV-B. The two multiplexors in the bottom
right corner of Fig. 5 are used to pass proper sorted messages to
the two sets of path constructors.