How 99 Lines of Taichi Code Enable High‑Performance CG Simulations

Taichi Overview

Taichi is a domain‑specific language for high‑performance computer graphics, installable via pip install taichi and used with import taichi as ti. It achieves up to 188× speedup over TensorFlow and 13.4× over PyTorch while reducing code size to about one‑tenth of traditional implementations.

Material Point Method (MPM) and MLS‑MPM

MPM is a widely used method for simulating continuous media in visual effects (e.g., Disney's Frozen ) but historically required weeks of compute time for short scenes. Hu Yuanming introduced Moving Least Squares MPM (MLS‑MPM), which unifies the affine velocity field from AIPC with deformation‑gradient updates via moving‑least‑squares interpolation. MLS‑MPM doubles the simulation speed and enables new phenomena such as material cutting and rigid‑body coupling.

The reference implementation consists of 88 lines of C++ and is available at https://github.com/yuanming-hu/taichi_mpm.

ChainQueen Differentiable Physics Engine

ChainQueen is a differentiable hybrid Lagrangian‑Eulerian simulator built on MLS‑MPM. It provides high‑precision forward simulation and gradient computation, allowing gradient‑based optimization of neural‑network controllers without reinforcement learning.

DiffTaichi: Differentiable Programming

DiffTaichi extends Taichi with automatic differentiation. The project includes ten physical simulators (elastic objects, rigid bodies, fluids, light refraction, multi‑object scenes, etc.) that can be benchmarked on a standard laptop. Most models run in 2–3 hours without a GPU; the differentiable elastic object simulator uses only 110 lines of Taichi code compared to ~490 lines of hand‑written CUDA while achieving comparable performance. Overall, Taichi outperforms TensorFlow (188×) and PyTorch (13.4×) in speed.

Technical Motivation

Computer graphics lacked a universal, easy‑to‑use tool comparable to TensorFlow in deep learning, creating a high entry barrier. By embedding Taichi in Python, leveraging existing IDEs and libraries (NumPy, Matplotlib), and distributing it as a pip‑installable package, high‑performance graphics programming becomes accessible across platforms.

Key references:

ArXiv paper: https://arxiv.org/abs/1910.00935

Taichi repository: https://github.com/yuanming-hu/taichi

MLS‑MPM demo repository: https://github.com/yuanming-hu/taichi_mpm