The ntLDPC_G98042 (17664,14592) IP Core is defined in IEEE 802.3ca-2020, it is used by ITU-T G.9804.2-09.2021 standard document and it is based on an implementation of QC-LDPC Quasi-Cyclic LDPC Codes. These LDPC codes are based on block-structured LDPC codes with circular block matrices. The entire parity check matrix can be partitioned into an array of block matrices; each block matrix is either a zero matrix or a right cyclic shift of an identity matrix. The parity check matrix designed in this way can be conveniently represented by a base matrix represented by cyclic shifts. The main advantage of this feature is that they offer high throughput at low implementation complexity.
The ntLDPCΕ_G98042 encoder IP implements a 256-bit parallel systematic LDPC encoder. The Generator LDPC Matrix is calculated off-line, compressed and stored in ROM. It is partitioned to 12 layers and each layer, when multiplied by the 14592 payload block, produces 256 parity bits. The multiplier architecture may be parameterized before synthesis to generate multiple multiplier instances [1:4,6], in order to effectively process multiple layers in parallel and improve the IP throughput rate. Shortened blocks are supported with granularity of 128-bit boundaries and 384 or 512 parity bits puncturing is also optionally supported.
The ntLDPCD_G98042 decoder IP Core may optionally implement one of two approximations of the log-domain LDPC iterative decoding algorithm (Belief propagation) known as either Layered Offset Min-Sum Algorithm (OMS) or Layered Lambda-min Algorithm (LMIN). Selecting between the two algorithms presents a decoding performance vs. system resources utilization trade-off. The OMS algorithm is chosen for this implementation, given the high code rate of the Parity Check Matrix (PCM).
The ntLDPCD_G98042 decoder IP implements a 256-bit parallel systematic LDPC layered decoder. Each layer corresponds to Z=256 expanded rows of the original LDPC matrix. Each layer element corresponds to the active ZxZ shifted identity sub-matrices within the layer. Each layer element is shifted accordingly and processed by the parallel decoding datapath unit, in order to update the layers’ LLR estimates and extrinsic information iteratively until the required number of decoding iterations has been run.
The decoder IP also features a powerful optional syndrome check early termination (ET) criterion, to maintain identical error correction performance, while significantly increasing its throughput rate and/or reducing hardware cost. Additionally it reports how many decoding iterations have been performed when ET is activated, for system performance observation and calibration purposes. A top level architecture deployment wrapper allows to expand the parallelism degree of the decoder before synthesis, effec-tively implementing a trade-off between utilized area and throughput rate. Finally a simple, yet robust, flow control handshaking mechanism is included in both IPs, which is used to communicate the IPs availability to adjacent system components at 128-bit parallel bus interface. This logic is easily portable into any communication protocol, like AXI4 stream IF.
