Real-Time Snow Simulation - Integrating Weather Data and Cloud Rendering

Abstract

Over the last decade the NTNU Heterogenous Computing Labratory (HPC-Lab) at the Norwegian University of Science and Technology (NTNU) has had master students work- ing on a real-time snow simulator. Evolving from a complex and highly parallell Central Processing Unit (CPU) smoke simulation, the simulator now covers models for wind sim- ulation, snow particle physics, and avalanche prediction all computed in parallel on both CPU and Graphics Processing Unit (GPU) using Open Computing Language (OpenCL) or Compute Unified Device Architecture (CUDA). The resulting simulation is rendered using modern techniques and Open Graphics Library (OpenGL). In this thesis, the wind and in-air snow simulation are improved by removing and re- ducing simplifications and assumptions. By extending control of the boundary conditions and initial distributions, simulations can be run with external context such as real-world weather information. The simulation boundary conditions are extended to support interpolation between ground-truth points, and a simple, yet novel approach is introduced for selecting interpo- lation weights. The interpolation cost is shown to be less than 2ms more than for constant values in the worst case, for a frame time of 26ms, and at less than 1ms of additional load- ing time. The neighborhood calculation time is smaller than the time to load ground thruth points from disk. Snow precipitation rates across the simulation domain are controllable from animated data-sets or procedural functions, allowing the use of radar imaging as a source of simulation data. The additional cost of the rejection sampling is shown to be negligible for reasonable configuration values. To help visualize the precipitation rates cloud rendering is introduced to the simula- tor. While previous work focused on terrain and snow particles, the sky was simplified. Animated clouds are shown to be a computationally costly, but affordable, addition to the simulator, improving the quality of the rendered images. The run-time cost of high quality clouds is shown to a frame time increase of less than 50% at typical camera positions, and even last-generation mid-range GPUs maintain frame-rates of over 30HZ. High-end consumer GPUs maintain over 30HZ even in worst-case scenarios and almost 60HZ in normal use. Finally, the choice of Pseudo-Random Number Generator (PRNG) on both the CPU and GPU are re-evaluated. A change is made necessary to support varied precipitation rates, but also improves the statistical properties of the simulation. The performance cost of the improvement is shown to be negligible at less than a 5% increase in run-time of the CUDA version of the snow particle update kernel.