What is Quadrants?

What is Quadrants?#

2026-11-02

3 min read time

Applies to Linux

Quadrants is a multi-platform compiler for physics simulation. You write kernels in Python and Quadrants lowers them to parallel code for AMD ROCm (HIP), Vulkan (SPIR-V), Apple Metal, and x86 and ARM64 CPUs. Genesis uses it for the numeric kernels inside its solvers, collision code, and linear algebra.

The project was forked from Taichi Lang and is now an independent compiler. It does not aim to stay compatible with upstream Taichi Lang, as development is led by Genesis AI. This ROCm edition is the AMD-Ecosystem/quadrants branch. The user guide can be found at genesis-embodied-ai.github.io/quadrants.

Why Quadrants?#

Quadrants exists so a physics engine can keep one kernel language and still run on the machines researchers actually have:

  • One kernel source, several backends: A function marked @qd.kernel or @qd.func is the source of the AMDGPU, Metal, Vulkan, and CPU programs. Switching backends does not mean rewriting the solver.

  • Built for large batched simulation: The compiler’s layout types, cache, and GPU graphs are aimed at the launch and memory patterns of a robotics simulator, not at a general graphics sandbox.

  • Room to extend the math: Autodiff, subgroup operations, and register-resident tiles are available to kernel authors, which is how a rigid solver can implement its own linear algebra instead of calling a separate library for every step.

  • A small surface: GUI, mobile, and several older graphics backends were removed so the compiler that Genesis ships is the one that is still maintained.

Features and use cases#

Quadrants provides the following features:

  • Python kernel language: qd.init selects the architecture. Fields, ndarrays, and qd.Tensor hold simulation state. Dataclass structs pass through qd.func with no extra kernel-runtime cost. BufferView is a bounds-checked window onto an ndarray.

  • AMDGPU, Metal, Vulkan, and CPU backends: The ROCm 10.1 image enables AMDGPU and leaves CUDA and Vulkan off. The same source can target the other architectures in a build that turns them on.

  • Launch and compile speed: Fastcache skips front-end parsing when a kernel has already been compiled. GPU graphs capture a sequence of kernels so a step is one launch. perf_dispatch times alternate implementations and keeps the faster one.

  • SIMT building blocks: Subgroup shuffle and reductions cover AMDGPU, Metal, and Vulkan. Tile types keep 16x16 and 32x32 matrices in registers for blocked factorizations.

  • Autodiff: Reverse-mode differentiation, including dynamic loops, plus forward mode, custom gradients, and a tape.

  • PyTorch interop: to_torch(copy=False) and to_numpy(copy=False) use DLPack on CPU, AMDGPU, and Metal, and a torch tensor can be passed into a kernel.

  • A Python interpreter for kernels: qd.init(qd.python) runs a kernel as ordinary Python so it can be stepped through in a debugger.

Quadrants is commonly used in the following scenarios:

  • As the Genesis compiler: Rigid, deformable, and particle solvers in Genesis are Quadrants kernels. The ROCm Genesis image is a Quadrants wheel with Genesis installed on top.

  • Standalone kernel development: pip install quadrants, or the ROCm image quadrants-release-rocm10.1, is enough to compile and run kernels without Genesis.

  • Portable physics code: A prototype written once can be checked on a CPU, then on an AMD GPU, by changing the architecture passed to qd.init.

  • Differentiable simulation: Reverse-mode autodiff through a kernel is how a caller gets gradients without hand-writing an adjoint.

Contents