A sparse-memory-encoding GPU-MPM framework for large-scale simulations of granular flows

Abstract

The Material Point Method (MPM) is increasingly recognized as an effective tool for simulating complex granular flows. While GPU computing has been widely used in MPM applications for large-scale problems, its heavy reliance on contiguous memory distribution can significantly hinder efficiency and limit simulation capabilities due to memory capacity constraints. This study presents a sparse-memory-encoding framework that incorporates advanced algorithms to address these limitations in large-scale simulations. We introduce a novel algorithm for atomic-free dual mapping between material points and nodes, in conjunction with warp-wise particle-to-grid mappings organized within a block-cell-material point hierarchy. Moreover, the framework features an efficient memory shift algorithm that optimizes memory usage for material properties. This optimization enables the seamless integration of commonly used material constitutive models, including elastic, elastoplastic, and hyper-plastic models, as well as various iteration schemes such as “update stress first”, “update stress last”, and “modified update stress last” within a cohesive framework. Furthermore, the framework accommodates incorporating diverse boundary conditions, such as Dirichlet, Neumann, and arbitrary-shaped rigid body contact, thus broadening its applicability to real-world engineering challenges, including landslides. The framework can effectively and efficiently handle large-scale, high-fidelity simulations of granular flows.