Research And Implementation Of A GPU-based PBD Simulation Algorithm

As computing power of graphics processing unit(GPU)continues to increase,the simulation algorithms in computer graphics have become the research hotspot.The physically-based simulation of mechanical are widely used in a variety of simulation systems,including games,movies,medical and many other fields.The simulation algorithm is complex to solve and it is difficult to maintain a stable frame rate in the real-time interactive system.In addition,there are some limitations in special material model simulation.Position Based Dynamics(PBD)defines constraints based on the model material properties and the simulation model is directly controlled through these constraint parameters.The calculation efficiency of PDB is higher than that of physically-based simulation and it is easy to control.It is imperative to apply GPU acceleration technology to PBD simulation algorithm.However,the PBD algorithm has a lot of constraint dependencies thus it must be forced to be serializable.In this paper,we propose a new GPU acceleration strategy to realize the real-time disassembly of constraint dependencies in the simulation process so that the dynamic constraints of the PBD algorithm can be parallelized.The main contents of this paper are as follows:1.We achieve the specific constraints PBD simulation for flexible body model by converting Finite Element Method(FEM)to corresponding PBD constraints.FEM constraints make almost no difference between PBD simulation and physically-based simulation computation results,but the computational efficiency is much higher.We employ cloth simulation experiments to compare the differences of various types of constraints.2.We propose a PBD simulation algorithm based on GPU acceleration technology.We design the GPU acceleration schemes for two typical iterative solvers in PBD algorithm separately:Gauss-Seidel method and Jacobi method.We experiment the two GPU acceleration schemes in complicated cloth collision scenario to compare their speedup ratio and convergence.3.We propose a hybrid PBD constraint packet parallel scheme which balances the speedup ratio and convergence of the GPU acceleration algorithm.Existing hybrid method applies Bron-Kerbosch algorithm to search the cliques in constraint topological graph and disassembles the constraint dependencies.The cliques search is an NP problem which can't support the PBD simulation algorithm with a variable constraint topology.To address this problem,we define a "vertex constraint set" to approximate the cliques which is grouped according to the order of the "vertex constraint set".It promotes the computational efficiency of existing hybrid methods and enables the methods to handle model simulation of variable constraint topology,e.g.,model with fracture and tear behaviors.Experiments are applied to validate that our hybrid PBD constraint packet parallel scheme has strong stability for all kinds of complex model scenarios.