Struct KernelSelector 

pub struct KernelSelector { /* private fields */ }

Picks a kernel implementation from measured execution times.

Calibration runs blocks of consecutive calls of one implementation so that worker wake-up and cache warm-up are charged the way steady execution would see them. The first calls of every block, which pay for moving the state between cores, are discarded. After the configured number of rounds, the implementation with the lower median call time is kept; the partitioned kernel must win by a margin, since the sequential kernel needs no worker threads.

A measured decision is checked again after a number of calls that doubles with every check, so a choice made while the host was busy, or before the simulation reached its steady state, does not stay wrong for the whole run. Checks become rare quickly and cost a few dozen calls each.

Implementations

impl KernelSelector

pub fn new(selection: KernelSelection) -> Self

pub fn diagnostic_id(&self) -> &str

Process-local identity preserved across moves, but not across clones.

pub fn next(&self) -> (KernelChoice, bool)

The implementation to run next, and whether its duration should be reported with Self::record.

pub fn decided(&self) -> Option<KernelChoice>

The current decision, or None while measuring.

pub fn advance(&mut self, calls: u64)

Account for calls executions of the decided implementation; a measured decision is checked again when its interval has elapsed.

pub fn record(&mut self, choice: KernelChoice, nanos: u64)

Report the duration of a call started after Self::next asked for timing.

Trait Implementations

impl Clone for KernelSelector

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for KernelSelector

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

Formats the value using the given formatter.