Simplifying High-Performance Geospatial Computing on GPGPUs Using Parallel Primitives: A Case Study of Quadtree Constructions on Large-Scale Geospatial Rasters

The increasingly available Graphics Processing Units (GPU) hardware resources and the emerging General Purpose computing on GPU (GPGPU) technologies provide an alternative and complementary solution to existing cluster based high-performance geospatial computing. However, the complexities of the unique GPGPU hardware architectures and the steep learning curve of GPGPU programming have imposed signficant technical challenges on the geospatial computing community to develop efficient parallel geospatial data structures and algorithms that can make full use of the hardware capabilities to solve ever growing large and complex real world geospatial problems. In this study, we propose a practical approach to simplifying highperformance geospatial computing on GPGPUs by using parallel primitives. We take a case study of quadtree construction on large-scale geospatial rasters to demonstrate the effectiveness of the proposed approach. Comparing the proposed parallel primitives based implementation with a naïve CUDA implementation, a signficant reduction on coding complexity and a 10X speedup have been achieved. We believe that GPGPU based software development using generic parallel primitives can be a first step towards developing geospatial-specific and more efficient parallel primitives for high-performance geospatial computing in both personal and cluster computing environments and boost the performance of geospatial cyberinfrastructure.