Struct AlphaLadderDevBuffers 

pub struct AlphaLadderDevBuffers {
    pub objective_dev: CudaSlice<f64>,
    pub status_dev: CudaSlice<u32>,
}

Device buffers for the fused alpha-ladder candidate-objective kernel.

candidate-objective mode: for each of the ALPHA_LADDER_LEN step sizes α_k the kernel evaluates η_trial_i = η_i + α_k · xδ_i, computes the per-row deviance, and atomically accumulates the sum into objective_dev[k]. Status flags are OR-accumulated into status_dev[k]. After a single memcpy_dtoh the host picks the first α that achieves deviance descent — no per-α kernel launch, no full row-output write.

Fields

objective_dev: CudaSlice<f64>

Device: summed deviance for each alpha step, length ALPHA_LADDER_LEN.

status_dev: CudaSlice<u32>

Device: OR-reduced status flags for each alpha step, length ALPHA_LADDER_LEN.

Implementations

impl AlphaLadderDevBuffers

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

Allocate the ladder device buffers on stream.

pub fn zero(&mut self, stream: &Arc<CudaStream>) -> Result<(), GpuError>

Zero all per-alpha accumulators in-place (call before each ladder launch).