← ArchivePaper2020

CUDA Deformers for Model Reduction

GPU CUDA implementation of reduced-coordinate deformable-body simulation with optimized memory layout for real-time character deformation.

Open the paper ↗Trace in the lineage graph

Abstract

Real-time deformable object simulation is important in interactive applications such as games and virtual reality. One common approach to achieve speed is to employ model reduction, a technique whereby the equations of motion of a deformable object are projected to a suitable low-dimensional space. Improving the real-time performance of model-reduced systems has been the subject of much research. While modern GPUs play an important role in real-time simulation and parallel computing, existing model reduction systems typically utilize CPUs and seldom employ GPUs. We give a method to efficiently employ GPUs for vertex position computation in model-reduced simulations. Our CUDA-based algorithm gives a substantial speedup compared to a CPU implementation, thanks to our system architecture that employs a memory layout friendly to GPU memory, reduces the communication between the CPU and GPU, and enables the CPU and GPU to work in parallel.

How to read this

Category: Method / systems: a GPU implementation for reduced deformable simulation
Contributions: Gives a CUDA method to efficiently compute vertex positions in model-reduced deformable simulations on the GPU
Designs a system architecture with a GPU-friendly memory layout that reduces CPU-GPU communication and lets CPU and GPU run in parallel
Reports a substantial speedup over a CPU implementation for real-time deformable bodies
Context: Builds on model-reduction for deformable simulation, in the lineage of Fulton et al.'s 'Latent-space Dynamics for Reduced Deformable Simulation', addressing the underused GPU side of such systems.Builds on: Latent-space Dynamics for Reduced Deformable Simulation
Correctness: Assumes a model-reduced (low-dimensional subspace) formulation is acceptable for the target objects; the contribution is an engineering speedup measured against a CPU baseline, so gains depend on mesh resolution, network/basis size, and the chosen reduction quality rather than improved physical accuracy.
Clarity: Implementation-focused; a first pass conveys the architecture and where the speedup comes from, with details in the memory-layout discussion.
How to read it: Focus on the memory layout and CPU-GPU work split; read closely only if you are porting reduced simulation to CUDA, otherwise a first pass suffices.

Builds on

Latent-space Dynamics for Reduced Deformable Simulation 2019

Built upon by

Nothing yet.

Related work

Pose-Space Subspace Dynamics 2016 / SIGGRAPH
Compression and Direct Manipulation of Complex Blendshape Models 2011 / SIGGRAPH Asia
FEM Simulation of 3D Deformable Solids: A Practitioner's Guide to Theory, Discretization and Model Reduction 2012 / Course
Stable Spaces for Real-time Clothing 2010 / SIGGRAPH

Keywords

model reduction gpu acceleration cuda modal matrix deformable object simulation real-time simulation vertex displacement memory optimization thread utilization plant dynamics

This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →