Real-Time Drainage Basin Modeling Using Concurrent Compute Shaders

This paper explores the use of graphics processing units (GPUs) for combined numerical analysis and visualization. Using an implementation of the O’Callaghan and Mark (1984) drainage accumulation algorithm, it explores the advantages and challenges that this environment brings to real-time, frame-by-frame production of spatial modeling products and their visualization. Modern GPUs employ massively-parallel processing architectures primarily configured for 3D rendering applications. However, the same environments also support general purpose numeric processing relevant to large scale simulations using languages like CUDA or OpenCL. Unfortunately, these languages do not interact directly with the ‘rendering pipeline,’ the GPU hardware responsible for transforming vertex, shape, patch, and fragment data into rendered pixels. In contrast, the most popular rendering APIs, OpenGL and Direct3D, have recently incorporated optional ‘compute shader’ processing stages that share the same data structures, memory models, and overall syntax used by the existing vertex, tessellation, geometry, and fragment shaders that feed the rendering pipeline. This addition allows developers to insert one or more purely numerical processing stages between graphics stages within the same API framework. Execution times for the implementation on a GPU with over 2,000 cores are analyzed to determine which operations are conducive to concurrent execution on the GPU as well as those processes and data access operations that lead to processing slowdowns.