Module gpu_cuda 

Optional, off-by-default, NVIDIA-only CUDA acceleration for the Hessian-vector products at the heart of the Newton-CG inner loop, calling the pure-Rust oxicuda-* library crates directly.

This module is ADDITIVE and entirely feature-gated behind the cuda feature (off by default). It does not replace or modify the existing gpu subsystem (src/gpu/) or the wgpu Newton-CG / CG / L-BFGS paths (src/unconstrained/{newton_gpu,cg_gpu,lbfgs_gpu}.rs). Every public function is f64-native — there is no silent f32 downcast at the boundary, which is the advantage of the oxicuda CUDA path over the f32 wgpu path.

Each entry point degrades safely when no NVIDIA device is present: cuda_is_available never panics, and the compute functions return an crate::error::OptimizeError::ComputationError rather than aborting.

Library-call pattern

The inner CG loop of the Newton-CG subsystem performs, per iteration, a Hessian-vector product H · p (a dense matrix×vector multiply, O(n²)) that dominates the per-iteration cost for large n. This module hands that product off to the pre-built, battle-tested oxicuda-blas GEMM rather than authoring any bespoke kernels, mirroring the proven matrix×vector layout of the Phase-1 interpolate reference (scirs2-interpolate/src/gpu_cuda.rs::cuda_eval_gemm).

Scoping note

This is a standalone additive function, deliberately NOT wired into minimize_newton_cg / newton.rs — exactly like the Phase-1 crates kept their cuda_* functions standalone.

Functions

cuda_hessian_vector_product

Compute the Hessian-vector product H · v on the GPU in f64.

cuda_is_available

Returns true iff the CUDA driver initializes and at least one NVIDIA device is visible. Never panics.