PERFORMANCE OPTIMIZATION OF MATERIAL POINT METHOD SIMULATIONS ON GPU USING NVIDIA GVDB VOXELS

• Main Author: Christopher, Jonathan
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/42829
• Institution: Institut Teknologi Bandung

Material Point Method (MPM) is a method for realistically simulating various kinds of physical materials, recently used for both scientific work and the creation of movie visual effects. MPM-based simulations, especially of 3D objects and with millions of particles, are very computationally demanding. Parallelization on GPU has been proven effective to increase the performance of MPM; however, the relatively limited capacity of GPU memory requires the usage of memory-saving data structures for storing the MPM simulation grid data. We attempt to optimize the performance of MPM on the CUDA GPGPU platform, using NVIDIA's open source GVDB Voxels data structure for storing the MPM grid data and scattering with parallel sum reduction approach for parallelizing the particle-to-grid transfer step in MPM. Benchmarks show a speedup of around 85% for the particle-to-grid transfer step compared to previous parallelization approaches (naive scattering and gathering), which translates to a speedup of around 30% for a single MPM iteration. However, the performance of our MPM implementation has yet to match the performance of state-of-the-art MPM implementation that uses the GSPGrid data structure.