Crate oxicuda_sparse 

OxiCUDA Sparse – GPU-Accelerated Sparse Matrix Operations

This crate provides GPU-accelerated sparse matrix operations, serving as a pure Rust equivalent to NVIDIA’s cuSPARSE library.

Sparse formats

Multiple storage formats are supported via the format module:

• CsrMatrix – Compressed Sparse Row (primary format)
• CscMatrix – Compressed Sparse Column
• CooMatrix – Coordinate (triplet)
• BsrMatrix – Block Sparse Row
• EllMatrix – ELLPACK
• HybMatrix – HYB (Hybrid ELL+COO)

Operations

• ops::spmv – Sparse matrix-vector multiply (y = alpha*A*x + beta*y)
• ops::spmm – Sparse-dense matrix multiply (C = alpha*A*B + beta*C)
• ops::spgemm – Sparse-sparse matrix multiply (C = A*B)
• ops::sptrsv – Sparse triangular solve (L*x = b or U*x = b)
• ops::sddmm – Sampled Dense-Dense Matrix Multiply

Preconditioners

• preconditioner::ilu0 – Incomplete LU(0) for general systems
• preconditioner::ic0 – Incomplete Cholesky(0) for SPD systems
• preconditioner::ic_k – Incomplete Cholesky with level-of-fill (IC(k))
• preconditioner::amg – Smoothed-aggregation algebraic multigrid (setup, V-cycle, and a standalone solver)

Eigensolvers

• eig::lobpcg – LOBPCG for the smallest eigenpairs of a sparse SPD matrix, with an optional diagonal (Jacobi) preconditioner
• eig::shift_invert – shift-invert power iteration for the eigenvalue closest to a target shift (interior eigenvalues)

Compatibility shims

• compat::cusparse_compat – a cuSPARSE-flavoured drop-in surface (cusparse_spmv / cusparse_spgemm / cusparse_spsv) over the host CSR type, for porting cuSPARSE code with minimal churn

These advanced solvers operate on the host-resident host_csr::HostCsr representation, which can be lifted from a CsrMatrix via HostCsr::from_gpu.

Example

use oxicuda_sparse::prelude::*;

Re-exports

pub use error::SparseError;

pub use error::SparseResult;

pub use format::BsrMatrix;

pub use format::CooMatrix;

pub use format::CscMatrix;

pub use format::Csr5Matrix;

pub use format::CsrMatrix;

pub use format::EllMatrix;

pub use format::HybMatrix;

pub use format::HybPartition;

pub use format::HybStatistics;

pub use handle::SparseHandle;

Modules

compat

Drop-in compatibility shims mirroring vendor sparse-library APIs.

eig

Sparse eigenvalue solvers.

error

Error types for OxiCUDA Sparse operations.

format

Sparse matrix storage formats.

handle

Sparse handle management.

host_csr

Host-resident CSR matrix and shared CPU kernels for advanced solvers.

ops

Sparse matrix operations.

preconditioner

Incomplete factorization preconditioners for iterative solvers.

prelude

Prelude for convenient imports.