//! Encode/decode throughput benchmarks.
//!
//! Uses the nightly `test` crate harness (`#[bench]`), not criterion. Run
//! with `cargo bench`. Results are wall-clock nanoseconds per iteration; to
//! convert to samples-per-second, divide the frame sample count by the
//! reported ns/iter and multiply by 10⁹.
//!
//! Representative sizes exercise the encoder's exhaustive search behaviour
//! at different `partition_order` ceilings — 256 and 1024 are power-of-two
//! frames (all seven partition orders available); 960 and 2880 mimic
//! Opus/WebRTC frame sizes (only some partition orders divide evenly, so
//! the inner search is sparser).

#![feature(test)]

extern crate test;

use lac::{decode_frame, encode_frame};
use test::Bencher;

// ── Synthetic-signal benches ────────────────────────────────────────────────
//
// These benches drive the encoder without any WAV I/O overhead, so the
// measurement is pure codec work. Four signal shapes cover the space:
//
// - `silence`: order-0 short-circuit path (skips Levinson entirely).
// - `multi_sine`: maximally LPC-friendly — order search converges fast.
// - `pseudo_speech`: AR(2) resonance on LFSR excitation, which is the
//   textbook model of the vocal tract. Residuals are near-Laplacian,
//   exercising the Rice k-search at realistic speech statistics.
// - `filtered_noise`: LFSR through a biquad low-pass. Broad-spectrum
//   content with no strong tonal structure — the hard case for LPC and
//   a reasonable proxy for music-like workload.

fn silence(n: usize) -> Vec<i32> {
    vec![0i32; n]
}

fn multi_sine(n: usize) -> Vec<i32> {
    // Three superimposed sinusoids with incommensurate frequencies, scaled
    // to the 24-bit range. Picks up realistic LPC workload without needing
    // to read a WAV file in the bench body.
    (0..n)
        .map(|i| {
            let t = i as f64;
            let a = (t * 0.11).sin() * 3_000_000.0;
            let b = (t * 0.27).sin() * 1_500_000.0;
            let c = (t * 0.43).sin() * 750_000.0;
            (a + b + c) as i32
        })
        .collect()
}

/// 32-bit Galois LFSR. Deterministic (seeded) pseudo-random i32 sequence
/// in approximately `±2^19` — one-eighth of full 24-bit scale, chosen to
/// leave headroom for AR(2) resonance gain without clipping.
fn lfsr_noise(n: usize, seed: u32) -> Vec<i32> {
    // Non-zero seed required: the LFSR would otherwise lock at zero.
    let mut state = if seed == 0 { 0xACE1_ACE1 } else { seed };
    (0..n)
        .map(|_| {
            // Maximal-length 32-bit Galois polynomial x^32 + x^22 + x^2 + x + 1
            // (tap mask 0x8020_0003). Period = 2^32 − 1, which is comfortably
            // larger than any bench frame size.
            let lsb = state & 1;
            state >>= 1;
            if lsb != 0 {
                state ^= 0x8020_0003;
            }
            // Sign-extend via `as i32`, arithmetic shift narrows to ~±2^19.
            (state as i32) >> 12
        })
        .collect()
}

/// AR(2) pseudo-speech: LFSR excitation filtered through a single formant
/// resonance at ~700 Hz / 16 kHz with bandwidth ~100 Hz. The pole pair
/// gives speech-shaped spectral envelope; residuals are near-Laplacian,
/// matching the statistical profile of real vowel segments. This is the
/// content class LPC + Rice is designed for, so it stresses the inner
/// encoder loops more realistically than multi_sine (which converges
/// instantly) or white noise (which doesn't benefit from LPC).
fn pseudo_speech(n: usize) -> Vec<i32> {
    // Q14 coefficients for a pole at r·e^{±jθ} with r = 0.9806, θ = 2π·700/16000.
    // a1 = 2r·cos(θ) · 2^14 ≈ 30933
    // a2 = −r²      · 2^14 ≈ −15751
    // The implied real-valued poles live inside the unit circle so the
    // recursion is stable for unbounded input lengths.
    const A1_Q14: i64 = 30_933;
    const A2_Q14: i64 = -15_751;
    let excitation = lfsr_noise(n, 0x5EED);
    let mut out = Vec::with_capacity(n);
    let mut y1: i64 = 0;
    let mut y2: i64 = 0;
    for &e in &excitation {
        // Resonance gain at the peak is ~1/(1−r) ≈ 50, so y magnitudes
        // reach ~2^19 · 50 ≈ 2^25. i64 intermediates prevent the
        // multiply-accumulate from overflowing; the final clamp keeps
        // the output inside the codec's 24-bit input contract.
        let sum = A1_Q14 * y1 + A2_Q14 * y2;
        // Round-to-nearest for the Q14 → integer demotion.
        let ar = (sum + (1 << 13)) >> 14;
        let y = ar + e as i64;
        let clamped = y.clamp(-((1 << 23) - 1), (1 << 23) - 1);
        out.push(clamped as i32);
        y2 = y1;
        y1 = clamped;
    }
    out
}

/// Broadband low-passed noise — LFSR excitation through a simple
/// single-pole IIR low-pass (pole at 0.9 in Q15). Covers the content
/// class where LPC cannot predict efficiently: residuals retain most
/// of the source entropy, so the Rice coder dominates the output bit
/// budget.
fn filtered_noise(n: usize) -> Vec<i32> {
    const POLE_Q15: i64 = 29_491; // round(0.9 · 2^15)
    // (1 − pole) in Q15 is the DC gain correction keeping output
    // magnitude near the excitation range.
    const ONE_MINUS_POLE_Q15: i64 = (1 << 15) - POLE_Q15;
    let excitation = lfsr_noise(n, 0xFEED);
    let mut out = Vec::with_capacity(n);
    let mut y: i64 = 0;
    for &e in &excitation {
        // y[n] = pole·y[n−1] + (1−pole)·x[n], all in Q15.
        let sum = POLE_Q15 * y + ONE_MINUS_POLE_Q15 * e as i64;
        y = (sum + (1 << 14)) >> 15;
        out.push(y.clamp(-((1 << 23) - 1), (1 << 23) - 1) as i32);
    }
    out
}

macro_rules! encode_bench {
    ($name:ident, $signal:ident, $size:expr) => {
        #[bench]
        fn $name(b: &mut Bencher) {
            let samples = $signal($size);
            b.iter(|| encode_frame(test::black_box(&samples)));
        }
    };
}

macro_rules! decode_bench {
    ($name:ident, $signal:ident, $size:expr) => {
        #[bench]
        fn $name(b: &mut Bencher) {
            let samples = $signal($size);
            let encoded = encode_frame(&samples);
            b.iter(|| decode_frame(test::black_box(&encoded)).unwrap());
        }
    };
}

encode_bench!(encode_silence_960, silence, 960);
encode_bench!(encode_silence_4096, silence, 4096);
encode_bench!(encode_sine_256, multi_sine, 256);
encode_bench!(encode_sine_960, multi_sine, 960);
encode_bench!(encode_sine_1024, multi_sine, 1024);
encode_bench!(encode_sine_2048, multi_sine, 2048);
encode_bench!(encode_sine_2880, multi_sine, 2880);
encode_bench!(encode_sine_4096, multi_sine, 4096);
encode_bench!(encode_speech_320, pseudo_speech, 320);
encode_bench!(encode_speech_960, pseudo_speech, 960);
encode_bench!(encode_speech_2048, pseudo_speech, 2048);
encode_bench!(encode_music_960, filtered_noise, 960);
encode_bench!(encode_music_2048, filtered_noise, 2048);
encode_bench!(encode_music_4096, filtered_noise, 4096);

decode_bench!(decode_silence_4096, silence, 4096);
decode_bench!(decode_sine_960, multi_sine, 960);
decode_bench!(decode_sine_4096, multi_sine, 4096);
decode_bench!(decode_speech_960, pseudo_speech, 960);
decode_bench!(decode_music_2048, filtered_noise, 2048);

// ── SIMD dot-product kernel isolation benches ───────────────────────────────
//
// The encode benches above measure end-to-end encode cost, which is
// dominated by per-frame allocations, Rice k-search, and the order
// loop. When evaluating a SIMD kernel change these confound the signal.
// These benches exercise just the kernel at realistic LPC orders so a
// change to `compute_residuals` shows up directly.

macro_rules! compute_residuals_bench {
    ($name:ident, $order:expr, $len:expr) => {
        #[bench]
        fn $name(b: &mut Bencher) {
            let samples = multi_sine($len);
            let coeffs: Vec<i16> = (0..$order)
                .map(|i| ((i as i16) * 711).wrapping_sub(100))
                .collect();
            b.iter(|| {
                lac::compute_residuals(test::black_box(&samples), test::black_box(&coeffs), 1)
            });
        }
    };
}

compute_residuals_bench!(compute_residuals_order_4_n320, 4, 320);
compute_residuals_bench!(compute_residuals_order_8_n320, 8, 320);
compute_residuals_bench!(compute_residuals_order_16_n320, 16, 320);
compute_residuals_bench!(compute_residuals_order_32_n320, 32, 320);
compute_residuals_bench!(compute_residuals_order_8_n960, 8, 960);
compute_residuals_bench!(compute_residuals_order_32_n960, 32, 960);
compute_residuals_bench!(compute_residuals_order_32_n4096, 32, 4096);