Efficient Spatially Sparse Computation
As a data-oriented programming language, Taichi decouples computation from data organization. For example, you can freely switch between arrays of structures (AOS) and structures of arrays (SOA), or between multi-level pointer arrays and simple dense arrays. Taichi has native support for sparse data structures, and the Taichi compiler effectively simplifies data structure accesses. This allows users to compose data organization components into complex hierarchical and sparse structures. The Taichi compiler optimizes data access.
We have developed 10 different differentiable physical simulators using Taichi, for deep learning and robotics tasks. Thanks to the built-in reverse-mode automatic differentiation system, most of these differentiable simulators are developed within only 2 hours. Accurate gradients from these differentiable simulators make controller optimization orders of magnitude faster than reinforcement learning.