![]() |
Eigen-Contrib
5.0.1-dev+master
|
#include <contrib/Eigen/src/GPU/GpuSparseSolverBase.h>
Pass-through configuration for the cuDSS-backed sparse direct solvers (SparseLLT, SparseLDLT, SparseLU). Fields left at their defaults keep the corresponding cuDSS default, which is tuned for performance rather than maximum robustness. Non-default fields require cuDSS >= 0.8, whose cudssReorderingAlg_t etc. name the algorithms with stable values; earlier versions declare no algorithm enumerator and accept only a default config, which EIGEN_HAS_CUDSS_SOLVER_CONFIG reports.
Public Member Functions | |
| bool | isDefault () const |
Public Attributes | |
| bool | hybridExecute |
| bool | hybridMemory |
| int64_t | hybridMemoryDeviceLimit |
| double | pivotEpsilon |
| double | pivotThreshold |
| int | refinementSteps |
| double | refinementTolerance |
|
inline |
True when every field keeps the cuDSS default.
| bool Eigen::gpu::SparseSolverConfig::hybridExecute |
Split factorization/solve work between host and device. cuDSS requires this before analysis, which it changes — set it before analyzePattern().
| bool Eigen::gpu::SparseSolverConfig::hybridMemory |
Let factor data spill to host memory when device memory is insufficient. Changes what analyzePattern() builds — set it before that phase.
| int64_t Eigen::gpu::SparseSolverConfig::hybridMemoryDeviceLimit |
Device-memory budget in bytes for hybridMemory; negative keeps the cuDSS default heuristic, which is what cuDSS itself spells as -1. Zero is a budget of its own and is passed through.
| double Eigen::gpu::SparseSolverConfig::pivotEpsilon |
Perturbation applied to near-zero pivots; negative keeps the cuDSS default.
| double Eigen::gpu::SparseSolverConfig::pivotThreshold |
Pivot admissibility threshold in [0, 1]; negative keeps the cuDSS default.
| int Eigen::gpu::SparseSolverConfig::refinementSteps |
Iterative-refinement steps during solve(); negative keeps the cuDSS default (0).
| double Eigen::gpu::SparseSolverConfig::refinementTolerance |
Iterative-refinement stopping tolerance; negative keeps the cuDSS default.