Struct CudaMegakernelPlanCache 

pub struct CudaMegakernelPlanCache { /* private fields */ }

Bounded LRU cache for CUDA megakernel topology plans.

Implementations

impl CudaMegakernelPlanCache

pub fn new() -> Self

Create a cache with the default production entry bound.

pub fn with_max_entries(max_entries: usize) -> Self

Create a cache with an explicit entry bound.

pub fn get_or_insert_with( &mut self, key: CudaMegakernelPlanCacheKey, build: impl FnOnce() -> CudaMegakernelTopologyDecision, ) -> Result<CudaMegakernelCachedPlan, CudaMegakernelMemoryError>

Return a cached plan or insert a newly selected topology decision.

pub fn get_or_select_topology( &mut self, graph_layout_hash: u64, analysis_kind: CudaMegakernelAnalysisKind, device: CudaMegakernelDeviceKey, sample: CudaMegakernelScheduleSample, graph: CudaMegakernelGraphShape, memory: CudaMegakernelMemoryBudget, launch_overhead_ns: f64, fusion_pressure: f64, ) -> Result<CudaMegakernelCachedPlan, CudaMegakernelMemoryError>

Return a cached topology plan or select and cache one from the current CUDA telemetry sample.

This is the hot-path convenience API: callers provide stable graph, analysis, device, and telemetry inputs, while the cache owns the pressure bucketing needed to avoid stale sparse/dense decisions.

pub fn get_or_plan_execution( &mut self, graph_layout_hash: u64, analysis_kind: CudaMegakernelAnalysisKind, device: CudaMegakernelDeviceKey, sample: CudaMegakernelScheduleSample, graph: CudaMegakernelGraphShape, bytes_per_node: u64, bytes_per_edge: u64, frontier_bytes: u64, scratch_bytes: u64, output_bytes: u64, budget_bytes: u64, launch_overhead_ns: f64, fusion_pressure: f64, ) -> Result<CudaMegakernelExecutionPlan, CudaMegakernelMemoryError>

Return a cache-backed, memory-validated CUDA megakernel execution plan.

The cache key uses sparse-plan memory pressure because sparse is the lower-bound resident footprint shared by every topology. A cache hit reuses the prior topology decision, then this method validates the exact current dense/fused/sparse byte budget before returning a launchable plan. If the cached non-sparse topology no longer fits, the method downgrades to sparse only after proving the sparse plan fits.

pub fn stats(&self) -> CudaMegakernelPlanCacheStats

Return cache counters.

pub fn clear(&mut self)

Drop every cached plan and preserve counters for observability.

Trait Implementations

impl Debug for CudaMegakernelPlanCache

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

Formats the value using the given formatter.

impl Default for CudaMegakernelPlanCache

fn default() -> Self

Returns the “default value” for a type.