Optimization for viewshed analysis on GPU

Different algorithms have been raised for viewshed analysis and measures were taken to get the compromise between performance and accuracy. The most accurate and standard algorithm is still the basic interpolation method, though its time cost is high. However, the development of Graphic Processing Unit (GPU) enables us to acquire high performance with normal PC, especially when the Compute Unified Device Architecture (CUDA) is put forward by NVIDIA for general purpose computing. In this paper, we will analyze the feasibility to map the basic interpolation method into GPU application and give our approach to achieve this goal. Further, we will introduce two critical measures in this approach: one is how to assign the data into different memory spaces on GPU according to their different access characteristics; the other is how to regularize the computing instructions and minimize branch parts in the procedure. At most, nearly 70 times speedup is reached in the experiment compared with the basic interpolation method on CPU.