ndarray

ndarray brings NumPy-style n-dimensional arrays to Rust: a generic ArrayBase type supporting owned, shared, and view-based storage, multidimensional slicing and broadcasting, element-wise and linear-algebra operations, and optional BLAS/rayon backends for high-performance numerics. It is the foundation numerical and scientific-computing crates in the Rust ecosystem build on when they need array-shaped data with NumPy-like ergonomics rather than plain Vecs or nested arrays.

What You Get

• A generic ArrayBase type parameterized over storage (owned, shared, view) and dimensionality (1D through N-D, fixed or dynamic)
• NumPy-style slicing, broadcasting, and axis operations via the s![] macro and Axis API
• Element-wise arithmetic operator overloads and reduction/aggregation methods (sum, mean, fold, etc.)
• Optional BLAS backend integration (via the blas feature and cblas-sys) for accelerated matrix multiplication
• Optional Rayon integration for data-parallel iteration over array elements and axes

Common Use Cases

• Implementing numerical algorithms (linear algebra, signal processing, simulations) that need n-dimensional array manipulation in Rust
• Building machine-learning or scientific-computing crates that require NumPy-like array semantics without a Python dependency
• Working with zero-copy array views over externally-owned buffers (e.g. memory-mapped data or FFI buffers)
• Parallelizing array computations across cores via the optional Rayon integration

Under The Hood

Architecture - The crate centers on ArrayBase<S, D>, generic over a storage trait S (owned Vec-backed, Rc-shared, or borrowed view) and a dimension type D, with dozens of impl_*.rs modules (impl_1d, impl_2d, impl_dyn, impl_raw_views, impl_cow, etc.) implementing dimension- and storage-specific behavior against that one shared type, so the same array API works uniformly whether the data is owned, shared, or borrowed. Tech Stack - Pure Rust (edition 2021, Rust 1.87+), depending on num-traits/num-complex/num-integer for numeric abstractions and matrixmultiply for pure-Rust matrix multiplication by default, with optional rayon for parallelism and cblas-sys/libc behind a blas feature flag for calling out to an external BLAS implementation; it’s structured as a Cargo workspace with a companion ndarray-rand crate. Code Quality - The tests/ directory has dedicated suites per concern (array-construct, assign, azip, broadcast, complex, dimension, format), benches/ cover performance-sensitive paths, and proptest-regressions/ captures property-test-discovered regressions, reflecting a codebase that takes numeric-correctness edge cases seriously. API Design - The API deliberately mirrors NumPy’s mental model (axes, broadcasting, slicing macros) to lower the switching cost for developers coming from Python’s scientific stack, though the type-level encoding of storage and dimensionality means generic code over arrays can require verbose trait bounds compared to NumPy’s dynamically-typed arrays.