baracuda-cusolver

Safe Rust wrappers for NVIDIA cuSOLVER — dense and sparse linear-system solvers, eigensolvers, and matrix factorizations on the GPU. Covers cuSolverDn, cuSolverSp, cuSolverRf, and cuSolverMg.

Generic over scalar type (f32 / f64 / Complex32 / Complex64) where the math allows.

Coverage

• Dense (cuSolverDn):
  • LU: getrf + getrs
  • QR: geqrf + orgqr + ormqr
  • Cholesky: potrf + potrs + potri
  • SVD (gesvd, gesvdj, batched gesvdj)
  • Eigensolvers: syevd, syevj, batched syevj
  • Least-squares: gels
  • Generic 64-bit X* family (Xgetrf, Xgetrs, Xgeqrf, Xpotrf, Xpotrs, Xsyevd) for very large matrices.
• Sparse (cuSolverSp): csrlsvchol, csrlsvqr.
• Refactor (cuSolverRf): full Rf API.
• Multi-GPU (cuSolverMg): getrf, potrf, syevd.

Workspace handling

Like cuBLASLt and cuSPARSE, cuSOLVER requires the caller to provide device + host scratch sized via a bufferSize query. baracuda-cusolver exposes both queries and the actual operation as separate calls so you can amortize the allocation across many problems of the same shape.

Pairs with baracuda-cusolver-sys for the raw FFI surface.

Part of the baracuda workspace.

License

Dual MIT / Apache-2.0.