Struct DeviceResidentRowHess 

pub struct DeviceResidentRowHess { /* private fields */ }

Device-resident state produced by [launch_bms_flex_row_kernel_device_resident] and consumed by [launch_bms_flex_row_hvp] / [launch_bms_flex_row_diagonal].

Owns the row-Hessian + design slices on-device so the host can issue many HVPs against the same β snapshot without round-tripping 626 MB through host RAM. Drop releases the device memory back to the CUDA runtime. Per-row Hessian storage layout on the device. The build path is free to emit either, and the Hv / diag kernels read whichever the storage says.

Charter (Block 9 Phase 4): packed-upper halves the DRAM footprint of the n × r² cache (per-row r*(r+1)/2 doubles instead of r²), at the cost of a single per-entry index conversion in the kernel. The benchmark decides whether the packed path becomes the default for large-scale fits (r = 20 → 210 vs 400 doubles per row, ~47.5% smaller). The numerics are bit-equal because each H_i is symmetric by construction (the row kernel emits a symmetric block by construction — see the symmetric scratch-write loop in bms_flex_row_kernel’s shared-memory finaliser).

Trait Implementations

impl Debug for DeviceResidentRowHess

Available on Linux only.

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.