fastfields-dlpack

fastfields-dlpack is the low-level core of the fastfields project. It imports as fastfields.dlpack and exposes the field operators as raw, in-place calls that work directly on any array — NumPy, PyTorch or CuPy — sharing memory with zero copies.

Most people don't use this package directly. The friendly, array-returning wrappers — fastfields.numpy, fastfields.torch, fastfields.cupy and the unified fastfields.auto — are built on top of it and are what you normally want. Reach for fastfields.dlpack when you want the thinnest possible layer and are happy to manage output buffers yourself.

Install

pip install fastfields-dlpack \
    --extra-index-url https://fastfields.github.io/whl/cpu/

Use it

Functions write results in place or into a pre-allocated output you pass in:

import numpy as np
import fastfields.dlpack as ff

x = np.array([0, np.inf, np.inf, 0, np.inf], dtype=np.float32)
ff.dt_euclidean(x)          # in-place squared Euclidean distance transform

What's inside

The same operation families as the higher-level packages, in their raw in-place form: distance transforms (dt_euclidean, dt_l1, the dt_spline_* and dt_mesh point distances), positive-definite linear algebra (sym_matvec, sym_addmatvec_, sym_solve, sym_invert, …), and resampling (resample, restriction, spline_coeff). The Spline and Bound enums document the integer order/boundary arguments.

See the API reference for full signatures and options.

On CUDA

The same calls work on CuPy and PyTorch CUDA arrays and take the stream to submit on. They are asynchronous, so the caller must keep every array alive until that stream is synchronized — see CUDA usage for the full contract.