Struct Tuner 

pub struct Tuner { /* private fields */ }

Workgroup-size autotuner.

Implementations

impl Tuner

pub fn new(adapter_fp: &str, mode: Mode) -> Self

Build a new tuner for the adapter fingerprinted as adapter_fp.

pub fn cache_path_for_adapter(adapter_fp: &str) -> PathBuf

Cache file path for a given adapter fingerprint.

pub fn candidates_for(&self, max_invocations: u32) -> Vec<u32>

Candidate workgroup sizes bounded by max_invocations.

pub const fn default_workgroup_size() -> [u32; 3]

Default workgroup size used without cache data.

pub const fn mode(&self) -> Mode

Mode this tuner is running in.

pub fn resolve(&self, program_fp: &str) -> [u32; 3]

Resolve the workgroup size for a program key.

pub fn resolve_key(&self, key: &TunerProgramKey) -> [u32; 3]

Resolve the workgroup size for a typed program/static-shape key.

pub fn record_decision(&mut self, program_fp: impl Into<String>, size: [u32; 3])

Record a sweep outcome in memory.

pub fn record_key_decision(&mut self, key: TunerProgramKey, size: [u32; 3])

Record a sweep outcome for a typed key.

pub fn best_of<T: BackendTimer>( &self, program: &Program, candidates: impl IntoIterator<Item = [u32; 3]>, timer: &mut T, ) -> Result<Option<TuningMeasurement>, T::Error>

Measure candidate sizes and choose the fastest one.

Errors

Returns a backend timing error from BackendTimer.

pub fn best_of_natural_gradient<T: BackendTimer>( &self, program: &Program, candidates: impl IntoIterator<Item = [u32; 3]>, timer: &mut T, fisher_inv_sqrt_q16: &[u32], policy: NaturalGradientPolicy, ) -> Result<Result<NaturalGradientTuningStep, NaturalGradientTuningError>, T::Error>

Measure candidates, then choose the next probe with a Fisher-preconditioned natural-gradient policy.

This is the concrete runtime handoff for VYRE_AUTOTUNER=natural. It reuses the same backend timer as Self::best_of, records every measured candidate, and feeds those measurements into NaturalGradientPolicy. The returned step includes both the raw fastest measurement and the Fisher-directed next candidate.

Errors

Returns backend timing errors from BackendTimer or policy errors from NaturalGradientPolicy.

pub fn natural_gradient_step( &self, measurements: &[TuningMeasurement], fisher_inv_sqrt_q16: &[u32], policy: NaturalGradientPolicy, ) -> Result<NaturalGradientTuningStep, NaturalGradientTuningError>

Convert measured candidates into a Fisher-preconditioned next probe.

This keeps the best-of-N timing hook compatible while giving CUDA and other GPU backends a richer update rule than “pick the current fastest sample forever.” Backends can feed fisher_inv_sqrt_q16 from the primitive-backed natural-gradient self-substrate path.

Errors

Returns NaturalGradientTuningError when the policy input is malformed.

pub fn persist(&self) -> Result<(), String>

Write the cache to disk.

Errors

Returns the structured error from TunerCache::save.