Features

  • Write CUDA kernel in 100% Python syntax
  • Runtime JIT compilation
  • Rich vector math library
  • Mesh processing and queries
  • AD
  • Interop with PyTorch, JAX
  • USD import/export
import warp as wp


@wp.kernel
def integrate(
    p: wp.array(dtype=wp.vec3),
    v: wp.array(dtype=wp.vec3),
    f: wp.array(dtype=wp.vec3),
    m: wp.array(dtype=float),
):
    # thread id
    tid = wp.tid()
    # Semi-implicit Euler step
    v[tid] = v[tid] + (f[tid] * m[tid] + wp.vec3(0.0, -9.8, 0.0)) * dt
    x[tid] = x[tid] + v[tid] * dt


# kernel launch
wp.launch(integrate, dim=1024, inputs=[x, v, f, ...], device="cuda:0")

Python GPU Ecosystem

Warp Python Modules

  • warp.core: math, geometry, vector library
  • warp.sim: real-time simulation for robotic control, rigid/soft bodies, particles, cloth, URDF/MJCF/UsdPhysics
  • warp.fem: PDE
  • warp.llm

Data Model

  • Host/Device memory managed through wp.array type
  • Builtin spatial math types: vec2, vec3, vec4, mat22, mat33, mat44, quad, transform
  • Support for all common array protocols: __array_interface__, __cuda_array_interface__, __dlpack__
  • 0-copy interop with PyTorch, JAX

Compilation Pipeline

Execution Model

  • Kernels 启动 N 维线程块 (N ≤ 4)
  • 纯 SIMT 模型
    • No shared memory
    • No warp-level primitives

Custom CUDA Code

调用原始 CUDA 代码的好处：

  • 能用共享内存
  • Fine grained synchronization
  • Cooperative operations

Mesh, Hash Grid, Sparse Grid

Warp Sim

Integrators:

  • Symplectic Euler (semi-implicit)
  • XPBD (implicit)
  • Featherstone: which provides more stable simulation of articulated rigid body dynamics in generalized coordinates
  • ctypes
  • NVRTC (Runtime Compilation)
  • NVCC
  • PTX (Parallel Thread Execution, IR)
  • CUBIN