Struct DynamicOrder2Accumulator 

pub struct DynamicOrder2Accumulator { /* private fields */ }

Allocation-minimal runtime-width lowering of an additive order-two program.

This is the dynamic-width sibling of MappedOrder2Accumulator. A caller first reduces its mathematical row program to a scalar value plus N independent composed sources. This accumulator then performs the universal chain rule

g_i  = sum_t (f_t' q_t,i)
H_ij = sum_t (f_t'' q_t,i q_t,j + f_t' q_t,ij)

in one gradient pass and one upper-triangular Hessian pass. It owns exactly the two buffers that become the returned gradient and Hessian; no per-term derivative buffer or jet temporary is allocated. N is const-generic so a small fixed source set is visible to LLVM and can be unrolled.

Implementations

impl DynamicOrder2Accumulator

pub fn from_composed_sum<T: DynamicOrder2Term, const N: usize>( dimension: usize, value: f64, terms: &[T; N], ) -> Self

Lower a fused additive source set at runtime primary width dimension.

pub fn into_channels(self) -> (f64, Vec<f64>, Vec<f64>)

Finish the lowering as (value, gradient, row-major symmetric Hessian).

Trait Implementations

impl Debug for DynamicOrder2Accumulator

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

Formats the value using the given formatter.