<p>This work addresses large-scale 3D point-cloud analysis, targeting scene-level semantic segmentation as the primary task and using shape classification for secondary validation. Real-world point clouds are unordered, sparse, and strongly non-uniform. Most point‑cloud segmentation methods suffer from weak pre‑attention geometry modeling, static attention with poor adaptivity to density and structure, and the quadratic time–memory cost of self‑attention that hinders scene‑scale deployment. We propose GDPNet, a hybrid GNN-Transformer that follows the principle of front-loaded geometry, global attention allocation, and dual-path efficiency. An anisotropic GNN is inserted between the MLP and attention to aggregate adjacency-graph geometry before global modeling. Grouped vector attention is augmented with learnable positional multipliers/biases and a density-contrast term to modulate weights directly, elevating neighborhood adaptivity to weight-level adaptivity so that aggregation adjusts across dense/sparse and flat/high-curvature regions. For efficiency, Performer linearizes attention on the operator side, while GridPool applies regular grid partitioning with intra-cell pooling on the data side to reduce the active point count; together they lower time and memory. Compared with Point Transformer v3 (PTv3), GDPNet achieves higher performance with + 4.5 mIoU and + 5.0 mAcc on SemanticKITTI, + 0.8 mIoU and + 1.2 mAcc on S3DIS (Area 5), and + 0.4 mAcc and + 0.4 OA on ModelNet40. Ablation studies indicate that combining front-loaded geometric priors, weight modulation jointly conditioned on position and density, and dual-path efficiency sharpens boundaries, improves global consistency, and provides near-linear scaling suitable for large scenes and resource-constrained deployment.</p>