The increasing demand for high-speed, reliable wireless communication has driven the adoption of massive multiple-input multiple-output (MIMO) systems, which leverage a large number of antennas to enhance performance. Despite their potential, massive MIMO systems introduce substantial computational challenges, for instance in uplink detection. This thesis addresses these challenges by exploring software acceleration techniques to optimize the performance of massive MIMO systems.

In the context of uplink detection, the study focuses on linear detection algorithms, such as zero-forcing (ZF) and minimum mean square error (MMSE) and employs graphics processing units (GPUs) to accelerate matrix operations. Techniques such as block Cholesky and QR decompositions were implemented to reduce computational overhead. The results demonstrate significant reductions in execution time and improvements in scalability, achieving notable speedups while balancing precision and performance trade-offs.

Additionally, the thesis investigates the application of convolutional neural networks (CNNs) for channel state information (CSI)-based positioning. By optimizing CNN architectures and employing pruning techniques, the study enhances localization accuracy while minimizing computational requirements. These advancements enable precise positioning in resource-constrained environments, supporting advanced applications in 5G and beyond.

The thesis also proposes future work directions, emphasizing the potential of hardware-based implementations using the dataflow model of computation within the compute abstraction layer (CAL) framework. By modelling algorithms as actor networks with explicit data dependencies, CAL facilitates efficient and scalable hardware designs, particularly for field-programmable gate arrays (FPGAs). This approach offers a promising pathway to address the increasing computational demands of next-generation massive MIMO systems.