Low-Density Parity-Check Code Decoder Design and Error Characterization on an FPGA Based Framework

Abstract

Low-Density Parity-Check (LDPC) codes have gained popularity in communication systems and standards due to their capacity approaching error correction performance. Among all the hard-decision based LDPC decoders, Gallager B (GaB), due to simplicity of its operations, poses as the most hardware friendly algorithm and an attractive solution for meeting the high-throughput demand in communication systems. However, GaB sufferers from poor error correction performance. In this work, we first propose a resource efficient GaB hardware architecture that delivers the best throughput while using fewest Field Programmable Gate Array (FPGA) resources with respect to the state of the art comparable LDPC decoding algorithms. We then introduce a Probabilistic GaB (PGaB) algorithm that disturbs the decisions made during the decoding iterations randomly with a probability value determined based on experimental studies. We achieve up to four orders of magnitude better error correction performance than the GaB with a 3.4% improvement in normalized throughput performance. PGaB requires around 40% less energy than GaB as the probabilistic execution results with reducing the average iteration count by up to 62% compared to the GaB. We also show that our PGaB consistently results with an improvement in maximum operational clock rate compared to the state of the art implementations. In this dissertation, we also present a high throughput FPGA based framework to accelerate error characterization of the LDPC codes. Our flexible framework allows the end user adjust the simulation parameters and rapidly study various LDPC codes and decoders. We first show that the connection intensive bipartite graph based LDPC decoder hardware architecture creates routing stress for longer codewords that are utilized in today's communications systems and standards. We address this problem by partitioning each processing element (PE) in the bipartite graph in such a way that the inputs of a PE are evenly distributed over its partitions. This allows depopulating the Loo Up Table (LUT) resources utilized for the decoder architecture by spreading the logic across the FPGA. We show that even though LUT usage increases, critical path delay reduces with the depopulation. More importantly, with the depopulation technique an unroutable design becomes routable, which allows longer codewords to be mapped on the FPGA. We then conduct two experiments on error correction performance analysis for the GaB and PGaB algorithms, demonstrate our framework's ability to reach a resolution level that is not attainable with general purpose processor (GPP) based simulations, which reduces the time scale of simulations to 24 hours from an estimated 199 years. We finally conduct the first study on identifying all possible codewords that are not correctable by the GaB for the case where a codeword has four errors. We reduce the time scale of this simulation that requires processing 117 billion codewords to 4 hours and 38 minutes with our framework from an estimated 7800 days on a single GPP.