Module sparse 

Sparse tensor support for numr

This module provides sparse tensor formats and operations for efficient storage and computation on tensors with mostly zero elements.

Sparse Formats

Three standard sparse formats are supported:

• COO (Coordinate): Stores (row, col, value) triplets. Best for construction and format conversion. O(nnz) storage.

• CSR (Compressed Sparse Row): Row-major compressed format. Best for row slicing and matrix-vector multiplication (SpMV). O(nnz + nrows) storage.

• CSC (Compressed Sparse Column): Column-major compressed format. Best for column slicing and some linear algebra operations. O(nnz + ncols) storage.

Usage

// Create COO tensor from triplets
let rows = vec![0, 1, 2];
let cols = vec![1, 0, 2];
let values = vec![1.0f32, 2.0, 3.0];
let shape = [3, 3];

let coo = SparseTensor::<CpuRuntime>::from_coo_slices(
    &rows, &cols, &values, shape, &device
)?;

// Convert to CSR for efficient SpMV
let csr = coo.to_csr()?;

// Sparse matrix-vector multiplication
let x = Tensor::<CpuRuntime>::from_slice(&[1.0f32, 2.0, 3.0], &[3], &device);
// let sparse_ops = CpuRuntime::default_client(&device);
// let y = sparse_ops.spmv(&csr, &x)?;  // y = A * x

When to Use Sparse

Sparse formats are beneficial when:

• Sparsity > 90% (less than 10% non-zeros)
• Memory is constrained
• Operations preserve sparsity (SpMV, element-wise on non-zeros)

Dense is usually faster when:

• Sparsity < 50%
• Operations produce dense results
• Using GPU with high memory bandwidth

Re-exports

pub use crate::ops::traits::Sparse24Ops;

pub use structured::Sparse24Tensor;

Modules

structured

2:4 Structured sparsity format

Structs

CooData

COO (Coordinate) sparse matrix data

CscData

CSC (Compressed Sparse Column) sparse matrix data

CsrData

CSR (Compressed Sparse Row) sparse matrix data

Enums

NormType

Norm type for computing row/column norms

SparseFormat

Sparse matrix storage format

SparseTensor

Sparse tensor with runtime-selected storage format

Traits

SparseOps

Trait for sparse tensor operations

SparseScaling

Sparse matrix scaling and equilibration operations

SparseStorage

Trait for sparse storage backends