oxiblas-matrix

Matrix types and views for OxiBLAS

Overview

oxiblas-matrix provides efficient, type-safe matrix types and views for the OxiBLAS library. It implements zero-copy views, lazy evaluation, and various matrix storage formats (dense, banded, packed, symmetric, triangular).

Features

Core Matrix Types

• Mat<T> - Owned, column-major dense matrix
• MatRef<'a, T> - Immutable view into a matrix
• MatMut<'a, T> - Mutable view into a matrix
• DiagRef<'a, T> - Diagonal matrix view

Specialized Storage

• BandedMatrix<T> - Banded matrix storage (efficient for tridiagonal, pentadiagonal, etc.)
• PackedMatrix<T> - Packed storage for symmetric/triangular matrices (saves 50% memory)
• SymmetricMatrix<T> - Symmetric matrix with upper/lower storage
• TriangularMatrix<T> - Triangular matrix (upper/lower with unit/non-unit diagonal)

Advanced Features

• Lazy evaluation - LazyMat<T> for deferred computation and expression templates
• Copy-on-write - CowMat<T> for efficient memory management
• Memory-mapped matrices - MmapMat<T> for large out-of-core matrices (with mmap feature)
• Nalgebra compatibility - Convert to/from nalgebra types (with nalgebra feature)

Safety & Performance

• Zero-copy views - No allocation for submatrices
• Compile-time size checks - Where possible using const generics
• SIMD-friendly layout - Proper alignment for vectorization
• Cache-aware - Column-major (Fortran) order for BLAS compatibility
• Prefetching support - Manual prefetch hints for performance-critical code

Installation

Add this to your Cargo.toml:

[dependencies]
oxiblas-matrix = "0.1"

# With serialization
oxiblas-matrix = { version = "0.1", features = ["serde"] }

# With memory-mapped matrices
oxiblas-matrix = { version = "0.1", features = ["mmap"] }

# With nalgebra interop
oxiblas-matrix = { version = "0.1", features = ["nalgebra"] }

# All features
oxiblas-matrix = { version = "0.1", features = ["serde", "mmap", "nalgebra"] }

Usage

Creating Matrices

use oxiblas_matrix::Mat;

// From dimensions (initialized to zero)
let a = Mat::<f64>::zeros(3, 3);

// From a slice (column-major order)
let data = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0];
let b = Mat::from_slice(2, 3, &data);

// From rows (more intuitive for initialization)
let c = Mat::from_rows(&[
    &[1.0, 2.0, 3.0],
    &[4.0, 5.0, 6.0],
]);

// Identity matrix
let id = Mat::<f64>::identity(4);

// Diagonal matrix
let diag = Mat::from_diag(&[1.0, 2.0, 3.0, 4.0]);

Matrix Access

use oxiblas_matrix::Mat;

let mut a = Mat::from_rows(&[
    &[1.0, 2.0, 3.0],
    &[4.0, 5.0, 6.0],
]);

// Element access (row, col)
let val = a[(0, 1)];  // 2.0
a[(1, 2)] = 9.0;      // Modify element

// Row/column access
let row = a.row(0);   // [1.0, 2.0, 3.0]
let col = a.col(1);   // [2.0, 5.0]

// Dimensions
let (nrows, ncols) = (a.nrows(), a.ncols());

Matrix Views

use oxiblas_matrix::Mat;

let a = Mat::from_rows(&[
    &[1.0, 2.0, 3.0, 4.0],
    &[5.0, 6.0, 7.0, 8.0],
    &[9.0, 10.0, 11.0, 12.0],
]);

// Create immutable view
let view = a.as_ref();

// Submatrix view (zero-copy!)
let sub = view.submatrix(0, 0, 2, 2);  // Top-left 2×2 block

// Transpose view (zero-copy!)
let at = view.t();

// Mutable view
let mut b = a.clone();
let mut_view = b.as_mut();
mut_view[(0, 0)] = 99.0;

Banded Matrices

use oxiblas_matrix::BandedMatrix;

// Tridiagonal matrix (1 subdiagonal, 1 superdiagonal)
let mut band = BandedMatrix::<f64>::new(5, 5, 1, 1);

// Set diagonal
for i in 0..5 {
    band.set(i, i, 2.0);
}

// Set super/subdiagonals
for i in 0..4 {
    band.set(i, i + 1, -1.0);  // Superdiagonal
    band.set(i + 1, i, -1.0);  // Subdiagonal
}

// Use in BLAS operations
// (Much more efficient than full dense storage)

Packed Matrices

use oxiblas_matrix::{PackedMatrix, Uplo};

// Symmetric matrix in packed storage (only upper triangle stored)
let mut packed = PackedMatrix::<f64>::new(4, Uplo::Upper);

// Set elements (only upper triangle)
packed.set(0, 0, 1.0);
packed.set(0, 1, 2.0);
packed.set(0, 2, 3.0);
// ... symmetric elements accessed via symmetry

// Saves 50% memory for large symmetric matrices!

Lazy Evaluation

use oxiblas_matrix::{Mat, LazyMat};

let a = Mat::from_rows(&[[1.0, 2.0], [3.0, 4.0]]);
let b = Mat::from_rows(&[[5.0, 6.0], [7.0, 8.0]]);

// Build expression tree without computation
let lazy = LazyMat::add(
    LazyMat::from_mat(a.as_ref()),
    LazyMat::from_mat(b.as_ref()),
);

// Evaluate when needed
let result = lazy.eval();  // Now computes a + b

Memory-Mapped Matrices

#[cfg(feature = "mmap")]
{
    use oxiblas_matrix::MmapMat;
    use std::path::Path;

    // Create or open memory-mapped matrix
    let path = Path::new("large_matrix.bin");
    let mmap = MmapMat::<f64>::create(path, 10000, 10000)?;

    // Access like regular matrix (backed by disk)
    let val = mmap[(100, 200)];

    // Efficient for matrices larger than RAM
}

Nalgebra Interop

#[cfg(feature = "nalgebra")]
{
    use oxiblas_matrix::Mat;
    use nalgebra as na;

    // From nalgebra
    let na_mat = na::DMatrix::<f64>::zeros(3, 3);
    let ox_mat = Mat::from_nalgebra(&na_mat);

    // To nalgebra
    let back = ox_mat.to_nalgebra();
}

Matrix Storage Formats

Dense (Column-Major)

Standard BLAS-compatible storage:

Matrix:     Storage:
[a b c]     [a d b e c f]
[d e f]

• Memory: m × n elements
• Cache: Column-wise access is cache-friendly
• Use for: General matrices

Banded

Stores only diagonals:

Matrix:          Storage (banded):
[a b 0 0]        [0 a b c d]
[e f g 0]        [a b c d 0]
[0 h i j]        [b c d 0 0]
[0 0 k l]

• Memory: (kl + ku + 1) × n where kl=lower bandwidth, ku=upper bandwidth
• Cache: Excellent for narrow banded systems
• Use for: Tridiagonal, pentadiagonal, sparse banded systems

Packed (Symmetric/Triangular)

Stores only upper or lower triangle:

Symmetric:       Packed storage:
[a b c]          [a b c d e f]
[b d e]          (only upper triangle)
[c e f]

• Memory: n × (n + 1) / 2 (50% savings)
• Cache: Compact storage
• Use for: Symmetric or triangular matrices

Performance Characteristics

Zero-Copy Design

All view types are zero-copy:

use oxiblas_matrix::Mat;

let a = Mat::from_rows(&[
    &[1.0, 2.0, 3.0, 4.0],
    &[5.0, 6.0, 7.0, 8.0],
]);

// No allocation - just pointer arithmetic
let sub1 = a.as_ref().submatrix(0, 0, 2, 2);
let sub2 = a.as_ref().submatrix(0, 2, 2, 2);
let transposed = a.as_ref().t();

// All views reference the same underlying data!

Feature Flags

Safety

• Bounds checking - All indexing operations check bounds in debug mode
• No unsafe in public API - Unsafe code is internal and carefully audited
• Lifetime safety - Views cannot outlive their underlying data
• Thread safety - Send and Sync where appropriate

Architecture

oxiblas-matrix/
├── mat.rs              # Owned Mat<T> type
├── mat_ref.rs          # Immutable MatRef<'a, T>
├── mat_mut.rs          # Mutable MatMut<'a, T>
├── banded.rs           # BandedMatrix<T>
├── packed.rs           # PackedMatrix<T>
├── symmetric.rs        # SymmetricMatrix<T>
├── triangular.rs       # TriangularMatrix<T>
├── lazy.rs             # LazyMat<T> expression templates
├── cow.rs              # CowMat<T> copy-on-write
├── mmap.rs             # MmapMat<T> (feature-gated)
├── nalgebra_compat.rs  # Nalgebra conversions (feature-gated)
├── ops.rs              # Operator overloads
└── prefetch.rs         # Prefetch utilities

Examples

See the examples directory in the main repository:

• matrix_basics.rs - Creating and manipulating matrices
• matrix_views.rs - Zero-copy submatrix views
• banded_matrices.rs - Working with banded storage
• mmap_matrices.rs - Large out-of-core matrices (requires mmap feature)

Testing

# Run all tests
cargo test --package oxiblas-matrix

# With all features
cargo test --package oxiblas-matrix --all-features

# Property-based tests (QuickCheck)
cargo test --package oxiblas-matrix -- --ignored

Related Crates

• oxiblas-core - Core traits and SIMD abstractions
• oxiblas-blas - BLAS operations on matrices
• oxiblas-lapack - LAPACK decompositions
• oxiblas - Meta-crate with unified API

Comparison with Other Libraries

License

Licensed under either of:

• Apache License, Version 2.0 (LICENSE-APACHE)
• MIT license (LICENSE-MIT)

at your option.

Contributing

Contributions are welcome! Areas of interest:

1. Sparse matrix views - Integration with oxiblas-sparse
2. Block matrix - Hierarchical block structure
3. Strided matrices - Non-contiguous storage patterns
4. GPU support - CUDA/ROCm memory views

References

• BLAS Standard - Matrix storage conventions
• LAPACK Documentation - Packed and banded formats
• ndarray - General-purpose array library
• nalgebra - Linear algebra library