Struct PartitionedReducer 

pub struct PartitionedReducer { /* private fields */ }

Performs memory-efficient reductions by splitting input into fixed-size chunks and accumulating partial results.

Implementations

impl PartitionedReducer

pub fn new(config: PartitionConfig) -> Self

Create a new reducer with the given configuration.

pub fn reduce_all(&mut self, data: &[f64]) -> Result<f64, PartitionedError>

Reduce all elements of a flat 1-D slice to a single scalar.

pub fn reduce_axis( &mut self, data: &[f64], shape: &[usize], axis: usize, ) -> Result<(Vec<f64>, Vec<usize>), PartitionedError>

Reduce along a single axis of an N-dimensional tensor.

data is the row-major flat representation of a tensor with the given shape. The returned value is the flat representation of the reduced tensor together with its shape (the axis dimension is removed).

pub fn log_sum_exp(&self, data: &[f64]) -> Result<f64, PartitionedError>

Numerically stable log-sum-exp: log(Σ exp(x_i)).

Uses the max-subtraction trick to avoid overflow.

pub fn stats(&self) -> &PartitionedStats

Return the accumulated statistics since the last reset.

pub fn reset_stats(&mut self)

Reset accumulated statistics.