torchpcl¶

torchpcl implements common point-cloud processing routines for PyTorch, in particular fast nearest-neighbor search and registration. It strives to provide an easy-to-use, yet performant developer experience.

The API is not final yet and might change at any time.

Installation¶

We recommend uv for installation. Binary wheels from tagged releases are available through the wheel index. Select the index matching your installed PyTorch and CUDA build; uv will then select the wheel for the current Python and machine architecture.

Add the selected wheel index to your project's pyproject.toml:

[project]
dependencies = ["torchpcl"]

[tool.uv]
find-links = [
  "https://schmiddo.github.io/torchpcl/wheels/torch2.14.0-cu130/",
]

For example, for PyTorch 2.14.0 with CUDA 13.0:

uv pip install --no-deps torchpcl \
  --find-links https://schmiddo.github.io/torchpcl/wheels/torch2.14.0-cu130/

Alternatively, install the newest version from source. The cuBQL headers are the only Git submodule required by the build:

git clone --recurse-submodules https://github.com/Schmiddo/torchpcl.git
cd torchpcl
uv sync

Ensure that CUDA_HOME points to your CUDA toolkit location. Use TORCHPCL_WITH_CUDA=1 to force a GPU-enabled build and TORCHPCL_WITH_CUDA=0 for a CPU-only build.

Documentation¶

• Read the packed batching guide for the point-cloud data model, devices, dtypes, and gradient behavior.
• Consult the API reference for the available operations and public Python interface.
• See the repository README for end-to-end examples and development instructions.