Struct SolveRowBuffers 

pub struct SolveRowBuffers {
    pub grad_eta: CudaSlice<f64>,
    pub w_solver: CudaSlice<f64>,
    pub deviance: CudaSlice<f64>,
    pub status: CudaSlice<u32>,
    pub n: usize,
}

Device-resident per-row output buffers for the solve-row mode.

Allocates only the four fields the PIRLS solver reads on every Newton iteration: grad_eta (score for Xᵀg RHS), w_solver (working weight for XᵀWX assembly), deviance (per-row deviance for convergence check), and status (diagnostic flags OR-reduced to a single host u32). Written by launch_solve_row_on_stream; used instead of RowOutputDevBuffers during the hot inner loop to reduce device memory and kernel store traffic.

Fields

grad_eta: CudaSlice<f64>

∂ℓ/∂η_i — score for Xᵀg RHS formation.

w_solver: CudaSlice<f64>

Stabilised Hessian weight — fed to XᵀWX assembly.

deviance: CudaSlice<f64>

Per-row deviance contribution — summed for convergence check.

status: CudaSlice<u32>

Bitmask flags OR-reduced to detect numerical issues.

n: usize

Implementations

impl SolveRowBuffers

pub fn allocate(stream: &Arc<CudaStream>, n: usize) -> Result<Self, GpuError>

Allocate the four solve-row output buffers (length n) on stream.