CUDA-Accelerated Soft Body Simulation

University of Pennsylvania, CIS 565: GPU Programming and Architecture, Final Project

• Gehan Zheng (LinkedIn, personal website), Hanting Xu

Click here for documentation (CIS5650 Final Project version)

Overview

This project is a CUDA-accelerated soft body simulation framework originally developed as a final project for CIS 5650: GPU Programming and Architecture at Upenn.

The goal of this project is to explore GPU-based physics simulation by building a lightweight, extensible simulation framework with minimal external dependencies. The system is designed to support rapid experimentation with different:

• physical models,
• numerical solvers,
• GPU-accelerated linear algebra pipelines.

---

Features

• Linear Solvers

  • Sparse Cholesky Prefactorization W/ Approximate Minimum Degree Ordering
  • Dense Cholesky Prefactorization
  • Jacobi Solver (Naive)
  • Cholesky Decomposition
  • Preconditioned Conjugate Gradient
    • Incomplete Cholesky Preconditioner
    • Jacobi Preconditioner

• FEM

  • Projective Dynamics
    • Direct (Cusolver's Cholesky)
    • Chebyshev Acceleration
  • Explicit Euler
  • Incremental Potential Contact (IPC)
    • Barrier
    • Frictional Contact
    • Dirichlet Boundary Conditions
      • Equality Constraints
    • Materials
      • Corotational
      • Neo-Hookean

• Collision Detection

  • Real-Time BVH Construction
  • Continuous Collision Detection (CCD)

Dependencies

System Requirements

• Operating System

  • Windows
  • Linux

• CUDA Toolkit ≥ 12.0 (cublas, cusolver required)

• CMake ≥ 3.18

• OpenGL

Third-Party Libraries

The following libraries are included directly in the project:

• OpenGL
• ImGui
• GLFW
• Eigen
• spdlog
• Catch2

External tools:

• CUDA Toolkit
• CMake

---

Configuration

Environment Configuration

The full runtime configuration is specified in context.json. This file defines simulation contexts, solver settings, and physical parameters.

---

Scene Configuration

The framework supports multiple simulation contexts, each representing an independent scene. A context may contain:

• one or more soft bodies,
• rigid bodies,

Each context can be configured independently with physical parameters such as time step size, gravity, damping coefficients. Contexts can be switched at runtime.

---

Solver Configuration

Solver behavior is controlled on a per-context basis.

• Single-precision (float)

  • Uses the Projective Dynamics (PD) solver

• Double-precision (double)

  • Uses the Incremental Potential Contact (IPC) solver

Only parameters relevant to the active solver are applied.

Notes on Solver Usage

• The PD solver supports interactive object dragging.
• IPC is not recommended for scenes with a large number of degrees of freedom; for large vertex counts, careful parameter tuning is required, otherwise the simulation may fail to converge and pause.
• For large-scale systems, Cholesky-based solvers can become prohibitively slow; PCG with a Jacobi preconditioner is recommended instead.
• Linear solvers can be switched via ImGui before simulation starts.

Screenshots

showcase.mp4