Module solver 

cuSOLVER-backed dense solver kernels for the GPU HAL.

This module owns CUDA solver functionality that is shared by GPU linear algebra dispatch and higher-level solver code. CPU solves do not live behind these entry points: unavailable CUDA support is reported as an error.

Structs

RefinementOutcome

Outcome reported by iterative_refinement_cholesky_solve.

Functions

check_deferred_potrf_info

Download the POTRF deferred info scalar and return an error if non-zero.

check_deferred_potrs_info

Download the POTRS deferred info scalar and return an error if non-zero.

cholesky_logdet_from_col_major

cholesky_lower_gpu

cholesky_solve_gpu

cholesky_solve_only_gpu

Solution-only mixed-precision solve: like cholesky_solve_gpu but skips the redundant fp64 POTRF when the fp32 + refinement path succeeds, since the caller does not consume the log-determinant. This is the path that delivers the full mixed-precision speedup (expensive O(p³) factor stays fp32) for the PIRLS Newton direction solve, where the logdet is discarded. The solution is full fp64 accuracy via iterative refinement.

context_and_stream

iterative_refinement_cholesky_solve

Solve A x = b with fp32 Cholesky factorization + fp64-residual iterative refinement, automatically falling back to fp64 when the policy rejects the attempt or when the fp32 path fails / diverges.

pinned_htod

potrf_in_place

potrf_in_place_reuse

POTRF factorization using pre-allocated workspace and info buffers.

potrf_query_lwork

Query the cuSOLVER POTRF workspace size for a p×p matrix.

potrs_in_place

potrs_in_place_reuse

POTRS triangular solve using a pre-allocated info buffer.

solver_backend_status