GPU Cloth Simulation

About the Cloth Simulation

The goal of this project was to create an NVIDIA CUDA kernel for a particle-based cloth simulation using Verlet integration. We got a framework that took care of rendering the cloth and "only" had to implement the Verlet integration using CUDA kernels.

How the cloth is held together

The cloth is held together by three different kinds of springs, called structural, shear and flexion. These springs connect certain particles and are used to calculate how the cloth should deform and move.

The simulation kernel does the following, starting one thread for each particle:

• Launch the simulation kernel, and update the particle positions using Verlet integration: x(t+dt) = (1 + damp) * x(t) - damp * x(t-dt) + a * dt².
• The acceleration a should contain gravity, and should also consider the current wind direction and strength given by wind_x, wind_y and wind_z.
• Make sure that fixed particles do not change their position.
• Repeatedly, fixUpIterations times, launch a kernel that enforces the spring constraints and obstacle collisions, again starting one thread per particle.
• For each particle, iterate over all particles connected via springs and compute the movement that would be required to restore each spring to its resting position.
• Based on the mass ratio between the particle pairs (and information about particles being fixed), compute the part of the movement that would have to be carried out by the current particle.
• Sum these movements up and update the particle position with a fraction of that movement (fixUpPercent of the full movement).
• Iterate over all spherical obstacles and determine whether the particle's updated position would be inside any of them. If it is, move the particle out of the sphere along the sphere normal.