The low resolution version of the paper is available here. A higher resolution paper is available on arxiv.
YASPS is a symbolic differentiation framework that tackles the simple problem: How do we write energies on meshes with collisions?
For example, a simple collision between two points is easy to write when each point can freely move around.
But what if some points can move around freely (it has 3 DoF), and the others are controlled by some rigid body (they all share the same 6 DoF).
What then? Do we write out all 4 combinations?
- Free-Free
- Free-Rigid
- Rigid-Rigid
- Rigid-Free
And what if there are 3 ways to control a points, what if I now write an energy with 4 points, which leads to 3^4=81 combinations?
YASPS says: no, just use the JOIN and UNION opeartor that YASPS introduces, and you only need to write the energy as one function.
And with those two operators, we can easily write simulations like those:
| Cage Deformation | Contact with Friction | Repulsive Curve |
|---|---|---|
 |
 |
 |
YASPS generates all the code on GPU, and is very fast, so fast that we directly compared to hand optimized GPU simulation framework (table 1 in the paper), and we are still faster.
As long as you can write the energy correctly, you can trust YASPS to do the differentiation efficiently.
YASPS already includes a conjugate gradient solver. It directly works with the Hessian and the gradient that YASPS computes, and gives you the solution of
Go to ./yasps directory and use the install script to install the package
cd yasps
./install.shAll the examples used in the paper are in the examples directory.