A Lightweight CUDA-Based Parallel Map Reprojection Method for Raster Datasets of Continental to Global Extent

Jing Li,Michael Finn,Marta Blanco Castano

doi:10.3390/ijgi6040092

Abstract

Geospatial transformations in the form of reprojection calculations for large datasets can be computationally intensive; as such, finding better, less expensive ways of achieving these computations is desired. In this paper, we report our efforts in developing a Compute Unified Device Architecture (CUDA)-based parallel algorithm to perform map reprojections for raster datasets on personal computers using Graphics Processing Units (GPUs). This algorithm has two unique features: a) an output-space-based parallel processing strategy to handle transformations more rigorously, and b) a chunk-based data decomposition method for projected space in conjunction with an on-the-fly data retrieval mechanism to avoid memory overflow. To demonstrate the performance of our CUDA-based map reprojection approaches, we have conducted tests between this method and the traditional serial version using the Central Processing Unit (CPU). The results show that speedup ratios range from 10 times to 100 times in all test scenarios. The lessons learned from the tests are summarized.

Highlights

Different map projections are useful for different purposes and applications, and as such, certain projections that minimize distortions of one or more aspects are often necessary
Since the last decade or so, NVIDIA’s Compute Unified Device Architecture (CUDA) platform was introduced as a lightweight and affordable alternative to the traditional parallel computing paradigm based on high-end supercomputers [2]
With CUDA, any personal computers equipped with NVIDIA Graphics Processing Units (GPUs) can be used to perform parallel processing, proving a cheaper option compared to supercomputing with CPU clusters

Summary

Introduction

Different map projections are useful for different purposes and applications (e.g., spatial analysis, data visualization), and as such, certain projections that minimize distortions of one or more aspects (e.g., area, angle) are often necessary. Computational challenges may rise when the data resolution is high, especially for projection or reprojection of raster datasets of continental to global extent (i.e., when the cartographic scale is small). Besides the methodological design and implementation, we conducted a series of tests to demonstrate the superior performance of GPU over CPU processes as well as the different data decomposition strategies with GPU components These assessments allowed us to explore the best practices of utilizing our CUDA algorithm, providing guidance for: (a) fine tuning the parallel implementation algorithm and (b) providing information to design algorithms in various computer architectures, including supercomputers and cloud computing.

Literature Review

Rigorous Raster Reprojection in a Serial Processing Manner

Test Environment

GPU Speedup Ratios

Findings

Conclusions