Arraymancer is a Nim-based N-dimensional array (tensor) library designed for high-performance numerical computing and deep learning. It offers an ergonomic syntax inspired by NumPy and PyTorch, targeting researchers and developers who need a fast, portable solution for CPU, CUDA, and OpenCL backends.

How It Works

Arraymancer leverages Nim's fast compilation and metaprogramming capabilities to provide a high-level, Python-like API with C-level performance. It supports multiple backends (CPU with OpenMP, CUDA, OpenCL) and allows for custom BLAS/LAPACK library integration. The library includes automatic differentiation for deep learning tasks, enabling the definition and training of neural networks with a concise syntax.

Quick Start & Requirements

• Install: nimble install arraymancer or nimble install arraymancer@#head for the development version.
• Prerequisites: A BLAS and LAPACK library (e.g., OpenBLAS, MKL, Apple Accelerate). CUDA and CuDNN are required for GPU acceleration.
• Documentation: Arraymancer official documentation (Note: Documentation may only be updated for release versions; check the examples folder for the latest features).

Highlighted Details

• Supports tensors up to 6 dimensions.
• Includes deep learning primitives like Conv2D, MaxPool2D, Linear, and GRULayer.
• Offers features like broadcasting, explicit slicing, reshaping, and file I/O (.csv, .npy, HDF5).
• Can compile to a self-contained binary with minimal dependencies.

Maintenance & Community

The project's last update mentioned in the README was July 2019 (v0.5.1). Further community engagement channels are not explicitly listed.

Licensing & Compatibility

The README does not explicitly state a license. Given the lack of a LICENSE file or explicit mention, users should exercise caution regarding commercial use and closed-source linking.

Limitations & Caveats

The deep learning features are noted as unstable and subject to interface changes. CUDA and OpenCL tensor implementations are less feature-complete than CPU tensors, with limitations on data types and operations like iteration or slicing mutations. The project's development activity appears to have slowed significantly since 2019.