high performance computing on graphics processing units: hgpu.org

This study targets at speeding up polygon rasterization in large-scale geospatial datasets by utilizing massively parallel General Purpose Graphics Processing Units (GPGPU) computing for efficient spatial indexing and analysis based on a dynamically integrated vector-raster data model. As the first step, we have designed and implemented a parallelization schema for moderately large polygons using the Compute Unified Device Architecture (CUDA). Experiment results on 41,768 real world geospatial polygons with vertex numbers between 64 and 1024, which are selected among a total of 717,057 polygons with 1,199,799 rings in the experiment dataset, show that our implementation can speed up the computation of intersection points among polygon edges and scan lines by more than 20 times on a Nvidia C2050 GPU card. Extending the design and implementation to support polygons with arbitrarily large numbers of vertices by extensively using efficient sorting is discussed. The paper also reports the design and implementation of a profile quadtree to better understand the data and the distributions of its parallel computing tasks, in addition to help select polygon groups for experiments.