Effective feature learning from real-world point cloud scenes for semantic segmentation remains a challenging task, particularly due to the complexity and scale of the data. To address this, we propose a high-performance semantic segmentation network that aggregates global and local features for real point cloud processing. The architecture builds on the hierarchical feature extraction principles of PointNet++ and introduces a density-adaptive local neighbourhood feature extraction layer, which dynamically adjusts grouping scales to better handle non-uniformly distributed point clouds. To further enhance performance, spatial attention modules are incorporated into both the encoder and decoder, enabling the learning of global inter-point correlations through a self-attention mechanism while residual connections mitigate gradient vanishing issues. Additionally, an improved spatial encoding strategy explicitly captures spatial structures and optimizes network parameters. Designed with scalability and parallelism in mind, this network architecture is well-suited for deployment in high-performance cloud computing environments. Experiments on benchmark datasets S3DIS and Semantic3D demonstrate significant improvements over the baseline PointNet++, with mIoU and overall accuracy increasing by 16.9 and 7.4 percentage points respectively. This work lays the foundation for leveraging emerging cloud-based quantum and high-performance computing platforms for large-scale 3D semantic understanding tasks.