GPU Implementation and Optimization of a Flexible MAP Decoder for Synchronization Correction

TL;DR

This paper introduces an optimized GPU-based MAP decoder that achieves significant speedups and supports a wide range of code sizes and conditions, enabling practical decoding of larger and more challenging codes.

Contribution

The paper presents a dynamic, highly efficient GPU implementation of a flexible MAP decoder with reduced memory usage and broad applicability across various code sizes and channel conditions.

Findings

Decoding speedups of over 100x on mid-range GPUs

Supports a wide range of code sizes and channel conditions

Reduced memory variant maintains error correction performance

Abstract

In this paper we present an optimized parallel implementation of a flexible MAP decoder for synchronization error correcting codes, supporting a very wide range of code sizes and channel conditions. On mid-range GPUs we demonstrate decoding speedups of more than two orders of magnitude over a CPU implementation of the same optimized algorithm, and more than an order of magnitude over our earlier GPU implementation. The prominent challenge is to maintain high parallelization efficiency over a wide range of code sizes and channel conditions, and different execution hardware. We ensure this with a dynamic strategy for choosing parallel execution parameters at run-time. We also present a variant that trades off some decoding speed for significantly reduced memory requirement, with no loss to the decoder's error correction performance. The increased throughput of our implementation and its ability to work with less memory allow us to analyse larger codes and poorer channel conditions, and makes practical use of such codes more feasible.