oxicuda-sparse

GPU-accelerated sparse matrix operations -- pure Rust cuSPARSE equivalent.

Part of the OxiCUDA project.

Overview

oxicuda-sparse provides a full suite of GPU-accelerated sparse matrix operations implemented entirely in Rust, targeting feature parity with NVIDIA's cuSPARSE. It supports five standard sparse storage formats with efficient conversions between them, and offers all the key sparse linear algebra primitives needed for scientific computing and machine learning workloads.

The SpMV implementation includes multiple algorithm variants -- scalar, vector, and adaptive -- allowing the autotuner or user to select the best strategy for a given sparsity pattern. SpGEMM uses a two-phase symbolic + numeric approach to minimize memory allocation overhead. SpTRSV employs level-set parallelism for efficient triangular solves on the GPU.

Preconditioners (ILU(0) and IC(0)) are included for use with iterative solvers in oxicuda-solver. The SparseHandle integrates with the BLAS handle from oxicuda-blas for seamless interoperability.

Sparse Formats

Format	Type	Description
`CsrMatrix`	CSR	Compressed Sparse Row (primary)
`CscMatrix`	CSC	Compressed Sparse Column
`CooMatrix`	COO	Coordinate (triplet) format
`BsrMatrix`	BSR	Block Sparse Row
`EllMatrix`	ELL	ELLPACK (fixed entries per row)

Format conversions are available in the format::convert module (CSR to/from CSC, COO, BSR, ELL).

Operations

Operation	Function	Description
SpMV	`ops::spmv`	Sparse matrix-vector multiply: y = alphaAx + beta*y
SpMM	`ops::spmm`	Sparse-dense matrix multiply: C = alphaAB + beta*C
SpGEMM	`ops::spgemm`	Sparse-sparse multiply: C = A*B (two-phase)
SpTRSV	`ops::sptrsv`	Sparse triangular solve: Lx = b or Ux = b
SDDMM	`ops::sddmm`	Sampled Dense-Dense Matrix Multiply

Preconditioners

ILU(0) -- Incomplete LU factorization (zero fill-in) for general systems
IC(0) -- Incomplete Cholesky factorization for symmetric positive definite systems

Quick Start

use oxicuda_sparse::prelude::*;

// Create a CSR matrix from raw arrays
let row_ptr = vec![0u32, 2, 4, 6];
let col_idx = vec![0u32, 1, 0, 1, 0, 1];
let values  = vec![1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0];
let csr = CsrMatrix::new(3, 2, row_ptr, col_idx, values);

// With a GPU context:
// let handle = SparseHandle::new(&ctx)?;
// spmv(&handle, 1.0, &csr_gpu, &x, 0.0, &mut y, SpMVAlgo::Vector)?;

Feature Flags

Feature	Description
`f16`	Half-precision (fp16) sparse support

Status

Metric	Value
Version	0.2.0
Tests passing	406
Release date	2026-06-16

oxicuda-sparse 0.2.0