Migrating from CUDSS.jl
The handle API keeps the names of the parameters and phases of CUDSS.jl; the functions are renamed after Julia conventions.
| CUDSS.jl | SparseDirectSolver.jl |
|---|---|
CudssSolver(A, structure, view) | DirectSolver(A, structure, view) |
cudss(phase, solver, x, b) | execute!(phase, solver, x, b) |
cudss_set(solver, name, value) | setparam!(solver, name, value) |
cudss_get(solver, name) | getparam(solver, name) (or getparam! into a buffer) |
cudss_update(solver, A) | update!(solver, A); with CUDA.jl ≥ 6.5 both packages export update!, so write using SparseDirectSolver: update! |
CudssMatrix (sparse), CudssMatrix (dense), CudssConfig | CSR, MatrixDescriptor, Options |
CudssData | no public counterpart: the data object lives inside the DirectSolver |
cholesky, ldlt, lu on CuSparseMatrixCSR | the same functions, see LinearAlgebra interface |
Structures ("G", "S", "H", "SPD", "HPD"), views ('L', 'U', 'F'), phase strings and parameter strings are identical, so a port is mostly a renaming:
# CUDSS.jl
solver = CudssSolver(A_gpu, "S", 'L')
cudss_set(solver, "pivot_threshold", 0.01)
cudss("analysis", solver, x_gpu, b_gpu)
cudss("factorization", solver, x_gpu, b_gpu)
inertia = cudss_get(solver, "inertia")
cudss("solve", solver, x_gpu, b_gpu)
# SparseDirectSolver.jl
solver = DirectSolver(A_gpu, "S", 'L')
setparam!(solver, "pivot_threshold", 0.01)
execute!("analysis", solver, x_gpu, b_gpu)
execute!("factorization", solver, x_gpu, b_gpu)
inertia = getparam(solver, "inertia")
execute!("solve", solver, x_gpu, b_gpu)Differences
- cuDSS parameters that are not implemented are not silently dropped: the multi-device parameters raise a
NotSupportedError, the hybrid memory and execution modes warn once and are ignored; see How unimplemented parameters are handled. "factorization_alg"and"solve_alg"keep their spellings but select this package's own algorithms (seeSparseDirectSolver.FactorizationAlgandSparseDirectSolver.SolveAlg).- Global pivoting (
pivot_type'C'/'R') is not supported: pivots are chosen inside each front, and failed pivots are perturbed rather than delayed. - Extensions beyond cuDSS:
pivot_sign,pivot_stats,ir_mode = "fgmres",pivot_pairs,amalgamation,schedule(EXTRA_PARAMETERS). - The matrix and the vectors may live on any KernelAbstractions backend, not only CUDA.