Struct PrefetchPredictor 

pub struct PrefetchPredictor {
    compute_durations: VecDeque<u64>,
    loading_durations: VecDeque<u64>,
    max_history: usize,
    prefetch_lead_ns: u64,
}

Tracks phase durations to predict when to trigger prefetch.

Fields

compute_durations: VecDeque<u64>

Recent GpuCompute phase durations (ns).

loading_durations: VecDeque<u64>

Recent DataLoading phase durations (ns).

max_history: usize

Maximum history size.

prefetch_lead_ns: u64

How far before predicted compute end to trigger prefetch (ns).

Implementations

impl PrefetchPredictor

pub fn new(max_history: usize, prefetch_lead_ns: u64) -> Self

pub fn record_compute(&mut self, duration_ns: u64)

Record a completed GpuCompute phase duration.

pub fn record_loading(&mut self, duration_ns: u64)

Record a completed DataLoading phase duration.

pub fn predicted_compute_ns(&self) -> Option<u64>

Predict the next GpuCompute duration using exponential moving average.

pub fn predicted_loading_ns(&self) -> Option<u64>

Predict the next DataLoading duration.

pub fn should_prefetch(&self, elapsed_compute_ns: u64) -> bool

Should we trigger prefetch now? Call this periodically during GpuCompute phase with elapsed time.

pub fn overlap_efficiency(&self) -> f64

Calculate the overlap efficiency. 1.0 = perfect overlap (data ready exactly when GPU finishes) <1.0 = GPU had to wait for data

1.0 = data was ready before GPU finished (ideal)