taichi-dev/taichi

Taichi Lang lets you write fast numerical code that runs on your GPU, but without leaving Python. Normally if you want code to use a graphics card for heavy math, physics simulations, rendering, AI, image processing, you have to drop into a separate low-level language. Taichi sits inside Python: you mark certain functions with a decorator, and a just-in-time compiler turns them into optimized machine code that runs on the GPU or, if no GPU is available, on the CPU instead. The way it works is that you write almost-normal Python inside functions tagged with @ti.kernel. When Taichi sees one of these, its JIT compiler (built on frameworks like LLVM) converts the function to parallel instructions for whichever backend is available. Data goes into containers called SNodes that can describe dense or sparse multi-dimensional fields, which is useful for simulations where most of the grid is empty. The README's hello-world example draws an animated Julia-set fractal in about 30 lines, computing every pixel in parallel inside a single kernel. Taichi integrates with NumPy and PyTorch, and the project also ships a cross-platform 3D visualizer, differentiable programming, and experimental quantized computation. You would use it when plain Python or NumPy is too slow for real-time physical simulation, computer graphics, robotics, vision, visual effects, or general number crunching, and you do not want to maintain a separate C++ or CUDA codebase. It installs with pip, runs on Windows, Linux, and macOS, and the same code can target CUDA, Vulkan, OpenGL 4.3+, Apple Metal, plain CPUs, or experimentally WebAssembly.