Struct ResidentDesignGram 

pub struct ResidentDesignGram { /* private fields */ }

#1017 Phase 3: a device-resident design matrix for repeated Xᵀ·diag(w)·X Gram evaluations that uploads X to the device ONCE.

The per-call try_fast_xt_diag_x re-uploads the full n×p X on every call. The SAE / IRLS inner loop holds X fixed and rebuilds the Gram once per Newton/PIRLS weight update, so the repeated H2D of X is pure waste — measured on an A100 (#1412) it makes the XtWX GEMM ~98% of the pipeline at <20% device utilisation (the device is starved by staging, not arithmetic). This handle uploads X once at construction; each Self::gram crosses only the n-vector w H2D and the p×p Gram D2H, so the per-Gram transfer shrinks by a factor of p.

Admission keys on the same work-based DispatchOp::XtDiagX gate as the per-call path (so it engages exactly when the Gram is GPU-profitable) and the numerics are bit-identical to try_fast_xt_diag_x on the same device (same cublasDdgmm row-scale + gemm reduction order). On a non-CUDA host, a below-threshold shape, or any device failure, Self::try_new returns None and the caller keeps its CPU/per-call path — residency never changes the result, only where (and how often) X is staged.

Implementations

impl ResidentDesignGram

pub fn try_new(x: ArrayView2<'_, f64>) -> Option<Self>

Upload x (n×p) to the device once. Returns None when CUDA is unavailable, the shape is below the GPU Gram threshold, or the upload fails.

pub fn gram(&self, w: ArrayView1<'_, f64>) -> Option<Array2<f64>>

Compute Xᵀ·diag(w)·X reusing the resident X. w must have one entry per design row. Returns None on a shape mismatch or device failure.

pub fn solve_normal_equations( &self, w: ArrayView1<'_, f64>, rhs: ArrayView1<'_, f64>, ridge: f64, ) -> Option<Array1<f64>>

Solve the penalized normal equations (Xᵀ·diag(w)·X + ridge·I)·β = rhs with the Gram, its Cholesky factor, and the RHS all kept DEVICE-RESIDENT — only w (n), rhs (p), and the solution β (p) cross the bus.

This is the #1017 Phase-3 fix for the next ceiling after Self::gram: the bare Gram still pays a p×p D2H (134 MB at p=4096), but the SAE/IRLS inner step only needs β, so chaining row-scale→GEMM→POTRF→TRSM on-device and returning only the p-vector removes that transfer entirely. Returns None on a shape mismatch, a non-PD Gram, or any device failure — the caller then runs the CPU normal-equations solve. The numerics match a host Cholesky((XᵀWX+ridge·I)) solve up to IEEE-754 reduction order.

pub fn dims(&self) -> (usize, usize)

(n, p) of the resident design.