lac/tests/latency.rs

//! Per-frame latency distribution and heap footprint on real audio.
//!
//! Answers three questions the aggregate benchmark leaves fuzzy:
//!
//! 1. **What's the tail of encode latency?** The mean in a bench says
//!    nothing about how bad the worst frames are. A realtime system needs
//!    P99 below the frame period; otherwise one slow frame blows the
//!    entire deadline. We report P50/P95/P99/max on real speech and music.
//! 2. **What's decode speed in isolation?** The MCU test bundles
//!    decode+mix+encode; this test measures decode alone so we know how
//!    cheap the receive path actually is.
//! 3. **Peak heap per frame?** Important if LAC is embedded alongside a
//!    heavier codec (LVC video) — we want to know how much transient
//!    allocation each audio frame costs. We wrap the global allocator
//!    with a simple counter for the duration of the test.
//!
//! Run with `cargo test --test latency --release -- --nocapture`.
//! Tests serialise themselves via a process-wide mutex so the
//! tracking-allocator counters stay coherent even under `cargo test`'s
//! default multi-threaded runner. `--test-threads=1` is no longer
//! required for correctness but still recommended for clean,
//! in-order console output.
//!
//! # Measurement stability
//!
//! For stable P99 numbers, pin the harness to a fixed core and disable
//! frequency scaling before running:
//!
//! ```text
//! sudo cpupower frequency-set -g performance
//! taskset -c 0 cargo test --test latency --release -- --nocapture
//! ```
//!
//! On a noisy CI runner the P99 values include scheduler jitter and can
//! overstate real-world cost by 2-5×. The P99 hard-deadline asserts
//! below use the frame period as the ceiling, which is still a wide
//! safety margin (~40× headroom in steady state) so jitter alone won't
//! flake the suite.

use std::alloc::{GlobalAlloc, Layout, System};
use std::path::{Path, PathBuf};
use std::sync::Mutex;
use std::sync::atomic::{AtomicUsize, Ordering};
use std::time::{Duration, Instant};

use hound::WavReader;
use lac::{decode_frame, encode_frame};

const CORPUS_DIR: &str = "corpus";

// ── Tracking allocator ──────────────────────────────────────────────────────

/// Global allocator wrapper tracking current and peak bytes outstanding.
/// Counts are process-global; the per-test `_lock(&TEST_MUTEX)` guard
/// (`MEASUREMENT_LOCK.lock()` at the top of each test body) serialises
/// access so concurrent test threads don't corrupt each other's
/// measurements.
struct TrackingAllocator;

/// Serialises latency tests so the process-global tracking-allocator
/// counters stay coherent under multi-threaded `cargo test`. Each test
/// takes the lock at entry and holds it for the whole measurement
/// window. This is only about allocator-counter coherence; a panic
/// inside a test section will still release the mutex via unwind, so
/// the `PoisonError` path intentionally ignores poison.
static MEASUREMENT_LOCK: Mutex<()> = Mutex::new(());

/// Cumulative bytes currently allocated from the tracked allocator. Updated
/// on every alloc/dealloc; reset to 0 between measurements via
/// `reset_peak`.
static CURRENT_BYTES: AtomicUsize = AtomicUsize::new(0);

/// Peak of `CURRENT_BYTES` observed since the last `reset_peak`.
static PEAK_BYTES: AtomicUsize = AtomicUsize::new(0);

/// Cumulative count of `alloc` calls since the last `reset_peak`. Tracked
/// separately from bytes because a regression can keep peak-bytes flat
/// (same sized buffers, different provenance) while multiplying the call
/// count — e.g. a refactor that replaces one reused `Vec` with a fresh
/// `Vec::new()` per frame.
static CALL_COUNT: AtomicUsize = AtomicUsize::new(0);

unsafe impl GlobalAlloc for TrackingAllocator {
    unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
        let ptr = unsafe { System.alloc(layout) };
        if !ptr.is_null() {
            let new = CURRENT_BYTES.fetch_add(layout.size(), Ordering::Relaxed) + layout.size();
            // `fetch_max` updates the peak only if `new` exceeds the stored
            // value; cheap, wait-free, sufficient for single-threaded tests.
            PEAK_BYTES.fetch_max(new, Ordering::Relaxed);
            CALL_COUNT.fetch_add(1, Ordering::Relaxed);
        }
        ptr
    }

    unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
        unsafe { System.dealloc(ptr, layout) };
        CURRENT_BYTES.fetch_sub(layout.size(), Ordering::Relaxed);
    }
}

#[global_allocator]
static ALLOC: TrackingAllocator = TrackingAllocator;

/// Reset the peak to the current allocation level so the next measurement
/// window starts fresh. Also clears the per-window allocation counter.
/// Call once before the code under test.
fn reset_peak() {
    PEAK_BYTES.store(CURRENT_BYTES.load(Ordering::Relaxed), Ordering::Relaxed);
    CALL_COUNT.store(0, Ordering::Relaxed);
}

/// Read the maximum bytes outstanding since the last `reset_peak`, minus
/// the current baseline — i.e., the peak *transient* heap the code under
/// test held. Ignores allocations that were still live at reset time
/// (test scaffolding, pre-loaded corpus).
fn peak_delta_since_reset() -> usize {
    let peak = PEAK_BYTES.load(Ordering::Relaxed);
    let baseline = CURRENT_BYTES.load(Ordering::Relaxed);
    peak.saturating_sub(baseline)
}

/// Read the number of `alloc` calls made since the last `reset_peak`.
fn call_count_since_reset() -> usize {
    CALL_COUNT.load(Ordering::Relaxed)
}

// ── Corpus loading ──────────────────────────────────────────────────────────

fn corpus(name: &str) -> PathBuf {
    Path::new(CORPUS_DIR).join(name)
}

fn load_mono(path: &Path) -> Option<Vec<i32>> {
    let mut reader = WavReader::open(path).ok()?;
    let spec = reader.spec();
    if spec.sample_format != hound::SampleFormat::Int
        || spec.channels != 1
        || spec.bits_per_sample > 24
    {
        return None;
    }
    let samples: Result<Vec<i32>, _> = reader.samples::<i32>().collect();
    samples.ok()
}

macro_rules! require {
    ($path:expr) => {
        if !$path.exists() {
            eprintln!("skipping: corpus file not found: {}", $path.display());
            return;
        }
    };
}

// ── Latency harness ─────────────────────────────────────────────────────────

/// Distribution summary computed from a sorted `Vec<Duration>`.
struct Dist {
    count: usize,
    p50: Duration,
    p95: Duration,
    p99: Duration,
    max: Duration,
    mean: Duration,
}

fn dist_of(mut samples: Vec<Duration>) -> Dist {
    samples.sort();
    let count = samples.len();
    let p = |frac: f64| samples[((count as f64 - 1.0) * frac).round() as usize];
    let total: Duration = samples.iter().sum();
    Dist {
        count,
        p50: p(0.50),
        p95: p(0.95),
        p99: p(0.99),
        max: *samples.last().unwrap(),
        mean: total / count as u32,
    }
}

fn fmt_us(d: Duration) -> String {
    format!("{:.1}µs", d.as_nanos() as f64 / 1000.0)
}

/// Measure encode and decode latency distributions over every
/// `frame_size`-sample chunk in `samples`. Also reports peak per-frame
/// heap allocation (encoder side only — the decoder footprint is
/// measured separately in the per-phase breakdown below).
///
/// `sample_rate` is used to convert the frame-sample count into a
/// real-time frame period, which gates the P99 assertion below.
fn run_latency(name: &str, samples: &[i32], frame_size: usize, sample_rate: u32) {
    // Serialise against every other latency-test thread: the tracking
    // allocator is a single pair of atomic counters. The
    // `.unwrap_or_else` handles `PoisonError` so one failed test can't
    // wedge the rest of the suite.
    let _guard = MEASUREMENT_LOCK
        .lock()
        .unwrap_or_else(|poisoned| poisoned.into_inner());
    let frames: Vec<&[i32]> = samples.chunks_exact(frame_size).collect();
    // Warm-up: first few encodes allocate LLVM+allocator arenas that
    // don't reflect steady-state behaviour. Skip them in the measurement.
    let warmup = 32.min(frames.len());
    for f in &frames[..warmup] {
        std::hint::black_box(encode_frame(f));
    }

    // ── Encode latency and transient heap ────────────────────────────
    let mut encode_times = Vec::with_capacity(frames.len() - warmup);
    reset_peak();
    let mut peak_encode_bytes = 0usize;
    let mut peak_encode_allocs = 0usize;
    let mut encoded_bytes_total = 0usize;
    let mut encoded_frames: Vec<Vec<u8>> = Vec::with_capacity(frames.len() - warmup);
    for f in &frames[warmup..] {
        reset_peak();
        let t = Instant::now();
        let encoded = encode_frame(f);
        encode_times.push(t.elapsed());
        encoded_bytes_total += encoded.len();
        peak_encode_bytes = peak_encode_bytes.max(peak_delta_since_reset());
        peak_encode_allocs = peak_encode_allocs.max(call_count_since_reset());
        encoded_frames.push(encoded);
    }

    // ── Decode latency and transient heap ────────────────────────────
    let mut decode_times = Vec::with_capacity(encoded_frames.len());
    let mut peak_decode_bytes = 0usize;
    let mut peak_decode_allocs = 0usize;
    for ef in &encoded_frames {
        reset_peak();
        let t = Instant::now();
        let _samples = decode_frame(ef).expect("decode");
        decode_times.push(t.elapsed());
        peak_decode_bytes = peak_decode_bytes.max(peak_delta_since_reset());
        peak_decode_allocs = peak_decode_allocs.max(call_count_since_reset());
    }

    let enc = dist_of(encode_times);
    let dec = dist_of(decode_times);
    // Frame period = frame_size / sample_rate, expressed in nanoseconds
    // as an integer so the subsequent P99 comparison is exact (no float
    // epsilon). Example: 320 samples at 16 kHz → 20_000_000 ns = 20 ms.
    let frame_period =
        Duration::from_nanos((frame_size as u64 * 1_000_000_000) / sample_rate as u64);

    eprintln!();
    eprintln!("== {name} ({frame_size}-sample frames @ {sample_rate} Hz) ==");
    eprintln!(
        "  encode latency:  p50={}  p95={}  p99={}  max={}  mean={}",
        fmt_us(enc.p50),
        fmt_us(enc.p95),
        fmt_us(enc.p99),
        fmt_us(enc.max),
        fmt_us(enc.mean)
    );
    eprintln!(
        "                    headroom at p99 vs frame period ({:.1}ms): {:.1}×",
        frame_period.as_micros() as f64 / 1000.0,
        frame_period.as_nanos() as f64 / enc.p99.as_nanos() as f64,
    );
    eprintln!(
        "  decode latency:  p50={}  p95={}  p99={}  max={}  mean={}",
        fmt_us(dec.p50),
        fmt_us(dec.p95),
        fmt_us(dec.p99),
        fmt_us(dec.max),
        fmt_us(dec.mean)
    );
    eprintln!(
        "  peak heap / frame:    encode={}B  decode={}B",
        peak_encode_bytes, peak_decode_bytes
    );
    eprintln!(
        "  peak allocs / frame:  encode={}  decode={}",
        peak_encode_allocs, peak_decode_allocs
    );
    eprintln!(
        "  throughput:  encoded_frames={}  total_encoded_bytes={}  avg_bytes/frame={:.1}",
        enc.count,
        encoded_bytes_total,
        encoded_bytes_total as f64 / enc.count as f64
    );

    // Real-time invariant: P99 per-frame cost must stay below the frame
    // period. Steady-state headroom is ~40× so a CI runner with heavy
    // scheduler jitter still passes comfortably; a 40× regression
    // (encoder bug, allocator hot-path change) trips this assert.
    assert!(
        enc.p99 < frame_period,
        "encode P99 {} exceeds frame period {} — real-time deadline missed",
        fmt_us(enc.p99),
        fmt_us(frame_period),
    );
    assert!(
        dec.p99 < frame_period,
        "decode P99 {} exceeds frame period {} — real-time deadline missed",
        fmt_us(dec.p99),
        fmt_us(frame_period),
    );
}

// ── Tests ───────────────────────────────────────────────────────────────────

#[test]
fn latency_headset_speech_320() {
    // 320 samples @ 16 kHz = 20 ms frame — standard voice-chat period.
    let path = corpus("ES2002a.Headset-0.wav");
    require!(path);
    let samples = load_mono(&path).expect("load");
    // Cap to ~60 s of audio so the test doesn't dominate CI time.
    let cap = (16_000 * 60).min(samples.len());
    run_latency("headset_speech", &samples[..cap], 320, 16_000);
}

#[test]
fn latency_headset_speech_160() {
    // 160 samples @ 16 kHz = 10 ms frame — tighter latency mode used by
    // WebRTC and similar real-time systems.
    let path = corpus("ES2002a.Headset-0.wav");
    require!(path);
    let samples = load_mono(&path).expect("load");
    let cap = (16_000 * 60).min(samples.len());
    run_latency("headset_speech_10ms", &samples[..cap], 160, 16_000);
}

#[test]
fn latency_headset_speech_480() {
    // 480 samples at 16 kHz = 30 ms frame. The same sample count at
    // 48 kHz is WebRTC's 10 ms full-band frame; since the codec only
    // cares about frame sample count (not sample rate) this exercises
    // the same search-grid shape that a 48 kHz WebRTC deployment would
    // hit. 480 = 2^5 · 3 · 5, so partition orders 0..=5 are valid; 6
    // and 7 are not, which differs from the 2048-sample dense case.
    let path = corpus("ES2002a.Headset-0.wav");
    require!(path);
    let samples = load_mono(&path).expect("load");
    let cap = (16_000 * 60).min(samples.len());
    run_latency("headset_speech_480", &samples[..cap], 480, 16_000);
}

#[test]
fn latency_headset_speech_prime() {
    // 503 is prime, so only `partition_order = 0` divides it — the
    // encoder skips the partition search entirely and emits a single
    // Rice partition. Covers a code path that power-of-two and
    // smooth-composite frame sizes never reach.
    let path = corpus("ES2002a.Headset-0.wav");
    require!(path);
    let samples = load_mono(&path).expect("load");
    let cap = (16_000 * 60).min(samples.len());
    run_latency("headset_speech_prime503", &samples[..cap], 503, 16_000);
}

#[test]
fn latency_mixed_meeting_320() {
    let path = corpus("ES2002a.Mix-Headset.wav");
    require!(path);
    let samples = load_mono(&path).expect("load");
    let cap = (16_000 * 60).min(samples.len());
    run_latency("mixed_meeting", &samples[..cap], 320, 16_000);
}

#[test]
fn latency_array_speech_320() {
    // Distant mic — residuals are noisier, so encode cost per frame
    // typically rises. Useful to confirm P99 doesn't blow up on less
    // predictable content.
    let path = corpus("ES2002a.Array1-01.wav");
    require!(path);
    let samples = load_mono(&path).expect("load");
    let cap = (16_000 * 60).min(samples.len());
    run_latency("array_speech", &samples[..cap], 320, 16_000);
}