explicitly disable cudss superpanels, which seems to make factorization more efficient in large problems

FenglongSong · FenglongSong · commit d61bd2b543e4 · 2026-05-27T14:13:36.000+02:00
diff --git a/cupiqp/sparse/sparse_direct_solver.py b/cupiqp/sparse/sparse_direct_solver.py
@@ -136,6 +136,7 @@ def __init__(self, matrix: Union[csr_matrix, UniformBatchedCsrMatrix], use_deter
                 ubatch_size.dtype.itemsize,
             )
         self._cudss_solver.plan_config.reordering_algorithm = DirectSolverAlgType.ALG_DEFAULT
+        self._cudss_solver.plan_config.use_superpanels = 0
         self._cudss_solver.solution_config.ir_num_steps = 0  # NOTE: iterative refinement steps, to be tuned
         # cudss has IR_TOL, but not implemented yet according to https://docs.nvidia.com/cuda/cudss/types.html#c.cudssConfigParam_t.CUDSS_CONFIG_IR_TOL
         # self._cudss_solver.plan_config.pivot_type = cudss_bindings.PivotType.PIVOT_COL

Original file line number	Diff line number	Diff line change
`@@ -136,6 +136,7 @@ def __init__(self, matrix: Union[csr_matrix, UniformBatchedCsrMatrix], use_deter`
`136`	`136`	`ubatch_size.dtype.itemsize,`
`137`	`137`	`)`
`138`	`138`	`self._cudss_solver.plan_config.reordering_algorithm = DirectSolverAlgType.ALG_DEFAULT`
	`139`	`+ self._cudss_solver.plan_config.use_superpanels = 0`
`139`	`140`	`self._cudss_solver.solution_config.ir_num_steps = 0 # NOTE: iterative refinement steps, to be tuned`
`140`	`141`	`# cudss has IR_TOL, but not implemented yet according to https://docs.nvidia.com/cuda/cudss/types.html#c.cudssConfigParam_t.CUDSS_CONFIG_IR_TOL`
`141`	`142`	`# self._cudss_solver.plan_config.pivot_type = cudss_bindings.PivotType.PIVOT_COL`