audio_codec/

resampler.rs

//! Default polyphase FIR resampler (f32 core, exact integer timing).
//!
//! The filter is the classic Kaiser-windowed polyphase FIR (256 phases × 24
//! taps). The read position is tracked on an integer grid of
//! `input_rate * output_rate` ticks per second, so repeated 20 ms frames keep
//! a **constant, exact** output count (8 kHz → 48 kHz is always 960 samples,
//! never 959/961). The Q15 fixed-point core with the same timing lives in
//! [`crate::resampler_fixed`] for targets without an FPU.

use core::f64::consts::PI as PI_F64;

use fearless_simd::{Level, Simd, SimdBase, SimdFloat, dispatch, f32x8};

use super::{CodecError, Sample};

#[cfg(feature = "std")]
use super::PcmBuf;

/// Number of polyphase filter phases.
pub const NUM_PHASES: usize = 256;
/// Number of filter taps per phase.
pub const TAPS_PER_PHASE: usize = 24;
/// Required length of the caller-provided coefficient buffer (`NUM_PHASES * TAPS_PER_PHASE`).
pub const COEFFS_LEN: usize = NUM_PHASES * TAPS_PER_PHASE;

/// `f64::sin` polyfill that works in both std and no_std (via libm).
#[inline]
fn fsin(x: f64) -> f64 {
    #[cfg(feature = "std")]
    {
        f64::sin(x)
    }
    #[cfg(not(feature = "std"))]
    {
        libm::sin(x)
    }
}

/// `f64::sqrt` polyfill.
#[inline]
fn fsqrt(x: f64) -> f64 {
    #[cfg(feature = "std")]
    {
        f64::sqrt(x)
    }
    #[cfg(not(feature = "std"))]
    {
        libm::sqrt(x)
    }
}

pub struct Resampler<'a> {
    input_rate: usize,
    output_rate: usize,
    coeffs: &'a mut [f32],
    /// Last [`TAPS_PER_PHASE`] input samples, oldest first.
    history: [f32; TAPS_PER_PHASE],
    /// Ticks from the newest input sample to the next output sample, on the
    /// `input_rate * output_rate` grid (see [`Resampler::resample_into`]).
    pos: usize,
    /// SIMD level the filter loop is compiled for.
    level: Level,
}

fn bessel_i0(x: f64) -> f64 {
    let mut sum = 1.0_f64;
    let mut term = 1.0_f64;
    let x_sq = x * x * 0.25;

    for m in 1..=30 {
        term *= x_sq / (m * m) as f64;
        sum += term;
        if term < 1e-15 * sum {
            break;
        }
    }
    sum
}

fn kaiser_window(n: usize, n_total: usize, beta: f64) -> f64 {
    if n_total <= 1 {
        return 1.0;
    }
    let alpha = (n_total - 1) as f64 / 2.0;
    let x = (n as f64 - alpha) / alpha;
    let arg = beta * fsqrt(1.0 - x * x);
    bessel_i0(arg) / bessel_i0(beta)
}

impl<'a> Resampler<'a> {
    /// Create a new `Resampler`, writing polyphase filter coefficients into the
    /// caller-provided buffer.
    ///
    /// `coeffs.len()` must be at least [`COEFFS_LEN`] (= 6144 floats, ~24 KB).
    /// The buffer is held by the resampler for its entire lifetime; it is
    /// written once here and read on every subsequent `resample_into` call.
    pub fn new(
        input_rate: usize,
        output_rate: usize,
        coeffs: &'a mut [f32],
    ) -> Result<Self, CodecError> {
        if coeffs.len() < COEFFS_LEN {
            return Err(CodecError::BufferTooSmall);
        }
        if input_rate == 0 || output_rate == 0 {
            return Err(CodecError::InvalidInput);
        }
        // Guard the integer tick grid's `pos * NUM_PHASES` against overflow
        // (`pos < output_rate` in the hot loop, so this is the only product).
        if input_rate.checked_add(output_rate).is_none()
            || output_rate.checked_mul(NUM_PHASES).is_none()
        {
            return Err(CodecError::InvalidInput);
        }

        const KAISER_BETA: f64 = 7.0;

        let ratio = output_rate as f64 / input_rate as f64;
        let num_phases = NUM_PHASES;
        let taps_per_phase = TAPS_PER_PHASE;
        let filter_len = num_phases * taps_per_phase;

        let coeffs = &mut coeffs[..filter_len];

        let cutoff = if ratio < 1.0 {
            ratio * 0.5 * 0.95
        } else {
            0.5 * 0.95
        };

        let center = (taps_per_phase as f64 - 1.0) / 2.0;

        // Design the polyphase filter directly into the borrowed buffer.
        // We use a fixed-size scratch array on the stack (24 f64 = 192 bytes).
        let mut phase_coeffs: [f64; TAPS_PER_PHASE] = [0.0; TAPS_PER_PHASE];

        for p in 0..num_phases {
            let mut sum = 0.0_f64;

            for t in 0..taps_per_phase {
                let x = t as f64 - center - (p as f64 / num_phases as f64);

                let sinc_val = if x.abs() < 1e-10 {
                    2.0 * cutoff
                } else {
                    let x_pi = x * PI_F64;
                    fsin(x_pi * 2.0 * cutoff) / x_pi
                };

                let full_filter_idx = t * num_phases + p;
                let window = kaiser_window(full_filter_idx, filter_len, KAISER_BETA);

                phase_coeffs[t] = sinc_val * window;
                sum += phase_coeffs[t];
            }

            for t in 0..taps_per_phase {
                let normalized = (phase_coeffs[t] / sum) as f32;
                coeffs[p * taps_per_phase + t] = normalized;
            }
        }

        Ok(Self {
            input_rate,
            output_rate,
            coeffs,
            history: [0.0; TAPS_PER_PHASE],
            pos: 0,
            level: default_level(),
        })
    }

    pub fn input_rate(&self) -> usize {
        self.input_rate
    }

    pub fn output_rate(&self) -> usize {
        self.output_rate
    }

    /// 24-tap dot product as three 8-lane FMAs plus one horizontal sum.
    #[inline(always)]
    fn dot_product<S: Simd>(simd: S, a: &[f32; TAPS_PER_PHASE], b: &[f32; TAPS_PER_PHASE]) -> f32 {
        let a0 = f32x8::from_slice(simd, &a[0..8]);
        let a1 = f32x8::from_slice(simd, &a[8..16]);
        let a2 = f32x8::from_slice(simd, &a[16..24]);
        let b0 = f32x8::from_slice(simd, &b[0..8]);
        let b1 = f32x8::from_slice(simd, &b[8..16]);
        let b2 = f32x8::from_slice(simd, &b[16..24]);
        a0.mul_add(b0, a1.mul_add(b1, a2 * b2)).reduce_sum()
    }

    /// Resample `input` into the caller-provided `out` buffer.
    ///
    /// Returns the number of samples written. Returns
    /// [`CodecError::BufferTooSmall`] if `out` cannot hold the result; use
    /// [`Self::max_output_samples`] to size it.
    pub fn resample_into(
        &mut self,
        input: &[Sample],
        out: &mut [Sample],
    ) -> Result<usize, CodecError> {
        if self.input_rate == self.output_rate {
            if out.len() < input.len() {
                return Err(CodecError::BufferTooSmall);
            }
            out[..input.len()].copy_from_slice(input);
            return Ok(input.len());
        }

        // Exact number of outputs this call must produce (see the Q15 core for
        // the same grid): outputs sit `input_rate` ticks apart starting at
        // `pos`, and the call consumes `input.len() * output_rate` ticks.
        // Anything else means frame sizes drift across chunk boundaries (the
        // 0.4.x 959/961 bug); pinned by `debug_assert_eq!` below.
        let expected: u128 = {
            let ticks = (input.len() as u128) * (self.output_rate as u128);
            if ticks > self.pos as u128 {
                (ticks - self.pos as u128).div_ceil(self.input_rate as u128)
            } else {
                0
            }
        };

        // Dispatch once per call so the whole loop is compiled with the
        // selected target features and `dot_product` inlines into it.
        let level = self.level;
        let written = dispatch!(level, simd => self.resample_simd(simd, input, out))?;

        debug_assert_eq!(
            written as u128,
            expected,
            "resampler produced {written} samples, exact timing requires {expected} \
             for {} inputs ({} -> {}): output sizes would drift across chunk \
             boundaries",
            input.len(),
            self.input_rate,
            self.output_rate,
        );

        Ok(written)
    }

    #[inline(always)]
    fn resample_simd<S: Simd>(
        &mut self,
        simd: S,
        input: &[Sample],
        out: &mut [Sample],
    ) -> Result<usize, CodecError> {
        /// Input samples converted per block. Converting a whole block before
        /// filtering (instead of pushing one sample at a time into the
        /// history) avoids store-forwarding stalls: the wide window loads
        /// would otherwise hit a just-written narrow store on every output.
        const BLOCK: usize = 256;

        // Integer tick grid: one input sample spans `interval` ticks and one
        // output sample advances `step` ticks, so timing is exact and never
        // rounded — repeated 20 ms frames keep a constant output count.
        let (step, interval) = (self.input_rate, self.output_rate);
        // One output step in whole input samples plus a fraction of one, and
        // that fraction in phase units: `step_frac * 256 = dq * interval + dr`.
        // `new` checked that `interval * NUM_PHASES` fits.
        let (step_whole, step_frac) = (step / interval, step % interval);
        let (dq, dr) = (
            step_frac * NUM_PHASES / interval,
            step_frac * NUM_PHASES % interval,
        );

        let coeffs: &[[f32; TAPS_PER_PHASE]; NUM_PHASES] =
            self.coeffs.as_chunks().0[..NUM_PHASES].try_into().unwrap();
        // `[history | block]`: the window for an output at block index `i`
        // is `buf[i + 1..i + 1 + TAPS_PER_PHASE]`, i.e. the last
        // `TAPS_PER_PHASE` samples up to and including sample `i`.
        let mut buf = [0.0f32; TAPS_PER_PHASE + BLOCK];
        let mut written = 0usize;

        for chunk in input.chunks(BLOCK) {
            let n = chunk.len();
            buf[..TAPS_PER_PHASE].copy_from_slice(&self.history);
            for (dst, &src) in buf[TAPS_PER_PHASE..TAPS_PER_PHASE + n]
                .iter_mut()
                .zip(chunk)
            {
                *dst = src as f32;
            }

            // `pos` counts ticks from the block's first sample; an output is
            // due every `step` ticks before the block end.
            let pos = self.pos as u64;
            let end = n as u64 * interval as u64;
            let count = if pos < end {
                (end - pos).div_ceil(step as u64) as usize
            } else {
                0
            };
            let out_block = out
                .get_mut(written..written + count)
                .ok_or(CodecError::BufferTooSmall)?;

            // The output position, kept as the input sample it follows plus
            // the filter phase and the remainder below it:
            // `pos = idx * interval + frac`, `frac * 256 = phase * interval + rem`.
            // The phase is `frac * 256 / interval`; tracking it with carries
            // gives the same phase without a division per output.
            let mut idx = self.pos / interval;
            let frac = self.pos % interval;
            let mut phase = frac * NUM_PHASES / interval;
            let mut rem = frac * NUM_PHASES % interval;

            for slot in out_block {
                // `idx < n <= BLOCK` and `phase < NUM_PHASES`, so the masks never
                // change them; they let the compiler drop the bounds checks.
                let i = idx & (BLOCK - 1);
                let window = buf[i + 1..i + 1 + TAPS_PER_PHASE].try_into().unwrap();
                // `as` saturates to the i16 range.
                *slot = Self::dot_product(simd, &coeffs[phase & (NUM_PHASES - 1)], window) as i16;

                // Advance one output (`step` ticks).
                rem += dr;
                if rem >= interval {
                    rem -= interval;
                    phase += 1;
                }
                phase += dq;
                idx += step_whole + phase / NUM_PHASES;
                phase %= NUM_PHASES;
            }

            written += count;
            self.pos = (pos + count as u64 * step as u64 - end) as usize;
            self.history.copy_from_slice(&buf[n..n + TAPS_PER_PHASE]);
        }

        Ok(written)
    }

    /// Upper bound on the number of samples `resample_into` will produce for
    /// an input of `n_input` samples.
    pub fn max_output_samples(&self, n_input: usize) -> usize {
        if self.input_rate == self.output_rate {
            return n_input;
        }
        // Exact bound: outputs (`input_rate` ticks apart) that fit in `n_input`
        // input periods (`output_rate` ticks each); the most fit when `pos == 0`.
        let count = (n_input as u128 * self.output_rate as u128).div_ceil(self.input_rate as u128);
        count.min(usize::MAX as u128) as usize
    }

    /// Convenience wrapper that allocates a `Vec` and calls `resample_into`.
    #[cfg(feature = "std")]
    pub fn resample(&mut self, input: &[Sample]) -> PcmBuf {
        let max = self.max_output_samples(input.len());
        let mut out = vec![0i16; max];
        match self.resample_into(input, &mut out) {
            Ok(n) => {
                out.truncate(n);
                out
            }
            Err(_) => Vec::new(),
        }
    }

    pub fn reset(&mut self) {
        self.history.fill(0.0);
        self.pos = 0;
    }
}

/// Best SIMD level available: runtime-detected when possible (`std`, wasm32),
/// otherwise the level implied by the compile-time target features.
///
/// Only AVX-512 is downgraded, to AVX2: the kernel is 8 lanes wide either
/// way, but at the AVX-512 level LLVM also autovectorizes the buffer setup
/// and sample conversion with 512-bit instructions. On Ice Lake that was
/// 5-15% slower than AVX2 in 3 of 4 cases. Other levels are used as detected.
fn default_level() -> Level {
    let level = Level::try_detect().unwrap_or_else(Level::baseline);
    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    if level.as_avx512().is_some()
        && let Some(avx2) = level.as_avx2()
    {
        return avx2.level();
    }
    level
}

/// One-shot resampling convenience helper (allocates).
///
/// Only available with the `std` feature.
#[cfg(feature = "std")]
pub fn resample(input: &[Sample], input_sample_rate: u32, output_sample_rate: u32) -> PcmBuf {
    if input_sample_rate == output_sample_rate {
        return input.to_vec();
    }
    let mut coeffs = vec![0.0f32; COEFFS_LEN];
    let mut r = match Resampler::new(
        input_sample_rate as usize,
        output_sample_rate as usize,
        &mut coeffs,
    ) {
        Ok(r) => r,
        Err(_) => return Vec::new(),
    };
    r.resample(input)
}

/// Self-contained [`Resampler`] that owns its coefficient buffer (std only).
///
/// [`Resampler`] borrows a ~24 KB caller-provided coefficient slice, which
/// makes it awkward to store as a long-lived struct field (the buffer must
/// outlive the resampler). `BoxedResampler` heap-allocates the coefficients
/// once and keeps them alive for the resampler's whole lifetime, restoring
/// the pre-0.4 ergonomic `new(input_rate, output_rate)` + `resample(..)`
/// call shape for std users that hold resamplers in struct fields.
#[cfg(feature = "std")]
pub struct BoxedResampler {
    /// Heap allocation: the buffer address is stable for the box's lifetime
    /// and never reallocated, which is what the `Resampler<'static>` borrow
    /// below relies on. Field order matters for drop: `inner` (borrower)
    /// must drop before `coeffs` (borrowed).
    inner: Resampler<'static>,
    /// Kept alive only to back `inner`'s `'static` borrow; never read.
    #[allow(dead_code)]
    coeffs: Box<[f32]>,
}

#[cfg(feature = "std")]
impl BoxedResampler {
    /// Create a resampler that owns its polyphase filter coefficients.
    ///
    /// Fails only for zero rates (the coefficient buffer is always sized
    /// correctly internally).
    pub fn new(input_rate: usize, output_rate: usize) -> Result<Self, CodecError> {
        let mut coeffs = vec![0.0f32; COEFFS_LEN].into_boxed_slice();
        // SAFETY: `coeffs` is a heap box whose address cannot change while
        // the allocation lives. We extend the borrow to 'static solely to
        // store both the buffer and its borrower in the same struct; the
        // struct's field order drops `inner` before `coeffs`, and nothing
        // else can reach the buffer (it is moved into `Self` right after).
        let inner = unsafe {
            let ptr = coeffs.as_mut_ptr();
            let loan: &'static mut [f32] = core::slice::from_raw_parts_mut(ptr, coeffs.len());
            Resampler::new(input_rate, output_rate, loan)?
        };
        Ok(Self { inner, coeffs })
    }

    /// Convenience wrapper that allocates the output buffer.
    pub fn resample(&mut self, input: &[Sample]) -> PcmBuf {
        self.inner.resample(input)
    }

    /// Resample into a caller-provided buffer (no allocation).
    pub fn resample_into(
        &mut self,
        input: &[Sample],
        out: &mut [Sample],
    ) -> Result<usize, CodecError> {
        self.inner.resample_into(input, out)
    }

    /// Upper bound on the output sample count for `n_input` input samples.
    pub fn max_output_samples(&self, n_input: usize) -> usize {
        self.inner.max_output_samples(n_input)
    }

    /// Reset the resampling history (e.g. after a source discontinuity).
    pub fn reset(&mut self) {
        self.inner.reset();
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::f32::consts::PI as PI_F32;
    use std::time::Instant;

    fn new_resampler(input_rate: usize, output_rate: usize) -> Resampler<'static> {
        // Leak intentionally: tests are short-lived and we need a 'static reference.
        let coeffs: &'static mut [f32] = Box::leak(vec![0.0f32; COEFFS_LEN].into_boxed_slice());
        Resampler::new(input_rate, output_rate, coeffs).expect("resampler init")
    }

    /// The exact integer tick grid must keep the output count constant across
    /// repeated 20 ms frames (no 959/961 drift) — the whole point of the f32
    /// core having the same timing as the Q15 core.
    #[test]
    fn test_exact_sample_counts_for_repeated_20ms_blocks() {
        for input_rate in [8000, 16000, 44100, 48000] {
            for output_rate in [8000, 16000, 44100, 48000] {
                let mut r = new_resampler(input_rate, output_rate);
                let input = vec![1000; input_rate / 50];
                let mut output = vec![0; output_rate / 50];
                for block in 0..10 {
                    let n = r.resample_into(&input, &mut output).unwrap_or_else(|e| {
                        panic!("{input_rate} -> {output_rate}, block {block}: {e:?}")
                    });
                    assert_eq!(
                        n,
                        output.len(),
                        "{input_rate} -> {output_rate}, block {block}"
                    );
                }
            }
        }
    }

    /// 8 kHz -> 48 kHz (the phone-side to Opus path): every 20 ms frame must be
    /// exactly 960 samples, first block included.
    #[test]
    fn test_8k_to_48k_is_always_960() {
        let mut r = new_resampler(8000, 48000);
        for block in 0..10 {
            let out = r.resample(&vec![1000i16; 160]);
            assert_eq!(out.len(), 960, "block {block}");
        }
    }

    #[test]
    fn test_resample_8k_to_16k() {
        let mut resampler = new_resampler(8000, 16000);
        let input = vec![1000i16; 80];
        let output = resampler.resample(&input);
        assert!(output.len() >= 150 && output.len() <= 170);
        for &s in &output[48..output.len().saturating_sub(48)] {
            assert!((s - 1000).abs() < 100, "Value {} is too far from 1000", s);
        }
    }

    #[test]
    fn test_resample_16k_to_8k() {
        let mut resampler = new_resampler(16000, 8000);
        let input = vec![1000i16; 160];
        let output = resampler.resample(&input);
        assert!(output.len() >= 75 && output.len() <= 85);
        let skip = output.len() / 4;
        for &s in &output[skip..output.len() - skip] {
            assert!((s - 1000).abs() < 100, "Value {} is too far from 1000", s);
        }
    }

    #[test]
    fn test_frequency_response_downsample() {
        let mut resampler = new_resampler(16000, 8000);
        let freq = 2000.0_f32; // Well below 4kHz Nyquist
        let samples: Vec<i16> = (0..160)
            .map(|i| ((i as f32 * freq * 2.0 * PI_F32 / 16000.0).sin() * 10000.0) as i16)
            .collect();

        let output = resampler.resample(&samples);

        // Output should have similar amplitude (allowing for some attenuation)
        let input_rms: f32 = samples
            .iter()
            .map(|&s| (s as f32).powi(2))
            .sum::<f32>()
            .sqrt()
            / samples.len() as f32;
        let output_rms: f32 = output
            .iter()
            .map(|&s| (s as f32).powi(2))
            .sum::<f32>()
            .sqrt()
            / output.len() as f32;

        assert!(
            output_rms > input_rms * 0.7,
            "Too much attenuation: input_rms={}, output_rms={}",
            input_rms,
            output_rms
        );
    }

    #[test]
    fn test_aliasing_suppression() {
        let mut resampler = new_resampler(16000, 8000);
        let freq = 7000.0_f32; // Above 4kHz Nyquist of output
        let samples: Vec<i16> = (0..1600)
            .map(|i| ((i as f32 * freq * 2.0 * PI_F32 / 16000.0).sin() * 10000.0) as i16)
            .collect();

        let output = resampler.resample(&samples);

        let output_rms: f32 =
            (output.iter().map(|&s| (s as f32).powi(2)).sum::<f32>() / output.len() as f32).sqrt();
        let input_rms: f32 = 10000.0 / 1.414; // Expected RMS of sine wave with amplitude 10000

        assert!(
            output_rms < input_rms / 50.0,
            "Aliasing not sufficiently suppressed: output_rms={}",
            output_rms
        );
    }

    #[test]
    fn test_performance_48k_to_8k() {
        let mut resampler = new_resampler(48000, 8000);
        let input = vec![0i16; 48000];

        let start = Instant::now();
        let iterations = 100;
        for _ in 0..iterations {
            let _ = resampler.resample(&input);
            resampler.reset();
        }
        let duration = start.elapsed();
        let per_second = duration.as_secs_f64() / iterations as f64;
        println!(
            "Resampling 1s of 48kHz to 8kHz (24 taps) took: {:.4}ms",
            per_second * 1000.0
        );
        assert!(
            per_second < 0.1,
            "Performance regression: {}ms",
            per_second * 1000.0
        );
    }

    #[test]
    fn test_continuity_between_chunks() {
        let input_rate = 16000;
        let output_rate = 8000;

        let freq = 1000.0_f32;
        let total_samples = 3200;
        let input: Vec<i16> = (0..total_samples)
            .map(|i| ((i as f32 * freq * 2.0 * PI_F32 / input_rate as f32).sin() * 5000.0) as i16)
            .collect();

        let mut resampler1 = new_resampler(input_rate, output_rate);
        let output1 = resampler1.resample(&input);

        let mut resampler2 = new_resampler(input_rate, output_rate);
        let mid = input.len() / 2;
        let mut output2 = resampler2.resample(&input[..mid]);
        output2.extend_from_slice(&resampler2.resample(&input[mid..]));

        assert_eq!(output1.len(), output2.len(), "Output lengths differ");

        let max_diff: i16 = output1
            .iter()
            .zip(output2.iter())
            .map(|(a, b)| (a - b).abs())
            .max()
            .unwrap_or(0);

        assert!(
            max_diff < 100,
            "Large discontinuity between chunks: max_diff={}",
            max_diff
        );
    }

    /// `BoxedResampler` must produce byte-identical output to the borrowed
    /// `Resampler` fed the same coefficients, and survive being moved +
    /// reused across many calls (stable self-referential borrow).
    #[cfg(feature = "std")]
    #[test]
    fn test_boxed_resampler_matches_borrowed() {
        let input_rate = 48000usize;
        let output_rate = 8000usize;
        let input: Vec<i16> = (0..4800)
            .map(|i| ((i as f32 * 0.05).sin() * 8000.0) as i16)
            .collect();

        let expected = {
            let mut coeffs = vec![0.0f32; COEFFS_LEN];
            let mut borrowed = Resampler::new(input_rate, output_rate, &mut coeffs).unwrap();
            borrowed.resample(&input)
        };

        let mut boxed = BoxedResampler::new(input_rate, output_rate).unwrap();
        // Exact bound now (no `+1`): the integer tick grid never over-produces.
        let expected_max = (input.len() * output_rate).div_ceil(input_rate);
        assert_eq!(boxed.max_output_samples(input.len()), expected_max);
        let got = boxed.resample(&input);
        assert_eq!(
            expected, got,
            "BoxedResampler output must match borrowed Resampler"
        );

        // Reuse after move: the coefficient borrow must stay valid across
        // moves. The resampler is a streaming state machine — a second
        // `resample` of the same input keeps the tail of the previous run in
        // its history, so only the output LENGTH is stable across the move.
        let mut moved = boxed;
        let again = moved.resample(&input);
        assert_eq!(expected.len(), again.len());

        // reset() clears the history and phase, so a fresh `resample` of the
        // same input must reproduce the very first output exactly.
        moved.reset();
        let after_reset = moved.resample(&input);
        assert_eq!(expected, after_reset);

        assert!(
            BoxedResampler::new(0, 8000).is_err(),
            "zero input rate must error"
        );
        assert!(
            BoxedResampler::new(48000, 0).is_err(),
            "zero output rate must error"
        );
    }

    /// Streaming in chunks of any size (around the history length and the
    /// internal block size) keeps the running output count at exactly
    /// `ceil(consumed * out / in)` and the samples identical to one call.
    #[test]
    fn test_exact_timing_survives_unaligned_chunks_and_reset() {
        for (input_rate, output_rate) in [
            (16000, 48000),
            (48000, 16000),
            (44100, 48000),
            (48000, 44100),
        ] {
            let input: Vec<i16> = (0..2000).map(|i| ((i * 37 % 200) - 100) as i16).collect();
            let mut r = new_resampler(input_rate, output_rate);
            let expected = r.resample(&input);
            assert_eq!(
                expected.len(),
                (input.len() * output_rate).div_ceil(input_rate)
            );
            for chunk_size in [
                1, 2, 3, 5, 7, 11, 23, 24, 25, 47, 88, 160, 255, 256, 257, 319,
            ] {
                r.reset();
                let mut got = Vec::new();
                let mut consumed = 0;
                for chunk in input.chunks(chunk_size) {
                    got.extend(r.resample(chunk));
                    consumed += chunk.len();
                    assert_eq!(
                        got.len(),
                        (consumed * output_rate).div_ceil(input_rate),
                        "{input_rate} -> {output_rate}, chunk {chunk_size}, consumed {consumed}"
                    );
                    assert!(r.resample(&[]).is_empty());
                }
                assert_eq!(
                    got, expected,
                    "{input_rate} -> {output_rate}, chunk {chunk_size}"
                );
            }
        }
    }

    const RATIOS: [(usize, usize); 8] = [
        (8000, 16000),
        (16000, 8000),
        (48000, 8000),
        (8000, 48000),
        (44100, 48000),
        (48000, 44100),
        (8000, 22050),
        (22050, 16000),
    ];

    /// Deterministic test signals: sine + LCG noise, and a full-scale square
    /// wave whose filter overshoot exercises output saturation.
    fn test_signals(len: usize) -> [Vec<i16>; 2] {
        let mut seed = 0x1234_5678u32;
        let noisy: Vec<i16> = (0..len)
            .map(|i| {
                seed = seed.wrapping_mul(1_664_525).wrapping_add(1_013_904_223);
                let noise = ((seed >> 16) as i16 as f32) * 0.2;
                ((i as f32 * 0.07).sin() * 15000.0 + noise) as i16
            })
            .collect();
        let square: Vec<i16> = (0..len)
            .map(|i| {
                if (i / 10) % 2 == 0 {
                    i16::MAX
                } else {
                    i16::MIN
                }
            })
            .collect();
        [noisy, square]
    }

    /// Straightforward scalar implementation: per-sample shifting history,
    /// f64 accumulation, explicit clamp, and each output's position derived
    /// from its absolute index rather than a streaming accumulator.
    fn reference_resample(coeffs: &[f32], fin: usize, fout: usize, input: &[i16]) -> Vec<i16> {
        let mut history = [0.0f32; TAPS_PER_PHASE];
        let mut out = Vec::new();
        for (input_index, &sample) in input.iter().enumerate() {
            history.copy_within(1.., 0);
            history[TAPS_PER_PHASE - 1] = sample as f32;
            while out.len() * fin < (input_index + 1) * fout {
                let phase = (out.len() * fin % fout) * NUM_PHASES / fout;
                let c = &coeffs[phase * TAPS_PER_PHASE..(phase + 1) * TAPS_PER_PHASE];
                let acc: f64 = c
                    .iter()
                    .zip(&history)
                    .map(|(&c, &h)| c as f64 * h as f64)
                    .sum();
                out.push(acc.clamp(i16::MIN as f64, i16::MAX as f64) as i16);
            }
        }
        out
    }

    /// Every SIMD level reachable on this machine must match the scalar
    /// reference to within 1 LSB (FMA/summation order may flip truncation).
    #[test]
    fn test_simd_matches_scalar_reference() {
        let levels = [Level::baseline(), default_level(), Level::new()];
        for (from, to) in RATIOS {
            for input in test_signals(2000) {
                for level in levels {
                    let mut r = new_resampler(from, to);
                    r.level = level;
                    let expected = reference_resample(r.coeffs, from, to, &input);
                    let got = r.resample(&input);
                    assert_eq!(expected.len(), got.len(), "{from}->{to} {level:?}");
                    for (i, (&e, &g)) in expected.iter().zip(&got).enumerate() {
                        assert!(
                            (e as i32 - g as i32).abs() <= 1,
                            "{from}->{to} {level:?}: sample {i}: expected {e}, got {g}"
                        );
                    }
                }
            }
        }
    }

    /// Streaming in arbitrary chunk sizes (smaller than, around and larger
    /// than the history length and the internal block size) must be
    /// bit-identical to a single call.
    #[test]
    fn test_chunked_is_bit_identical() {
        let [input, _] = test_signals(3000);
        for (from, to) in RATIOS {
            let whole = new_resampler(from, to).resample(&input);
            for chunk in [1, 2, 7, 23, 24, 25, 47, 160, 255, 256, 257, 1000] {
                let mut r = new_resampler(from, to);
                let mut chunked = Vec::new();
                for part in input.chunks(chunk) {
                    chunked.extend_from_slice(&r.resample(part));
                }
                assert_eq!(whole, chunked, "{from}->{to} chunk={chunk}");
            }
        }
    }

    /// Constant input must come out at the same level for every ratio
    /// (each polyphase branch is normalized to unity DC gain).
    #[test]
    fn test_dc_gain_all_ratios() {
        for (from, to) in RATIOS {
            for level in [1000i16, -20000, i16::MAX] {
                let output = new_resampler(from, to).resample(&vec![level; 2000]);
                // Skip the filter warm-up (history starts zeroed).
                let warmup = (TAPS_PER_PHASE * to).div_ceil(from) + 1;
                for (i, &s) in output.iter().enumerate().skip(warmup) {
                    assert!(
                        (s as i32 - level as i32).abs() <= 2,
                        "{from}->{to}: sample {i} = {s}, expected ~{level}"
                    );
                }
            }
        }
    }

    /// Output never exceeds `max_output_samples`, for any chunk length and
    /// any carried-over position.
    #[test]
    fn test_max_output_samples_is_upper_bound() {
        for (from, to) in RATIOS {
            let mut r = new_resampler(from, to);
            let input = vec![100i16; 200];
            for n in 1..=200 {
                let mut out = vec![0i16; r.max_output_samples(n)];
                r.resample_into(&input[..n], &mut out)
                    .unwrap_or_else(|e| panic!("{from}->{to} n={n}: {e:?}"));
            }
        }
    }

    #[test]
    fn test_output_buffer_too_small() {
        let mut r = new_resampler(8000, 16000);
        let mut out = [0i16; 10];
        assert!(matches!(
            r.resample_into(&[0i16; 160], &mut out),
            Err(CodecError::BufferTooSmall)
        ));

        // Same-rate passthrough checks the size too.
        let mut r = new_resampler(8000, 8000);
        assert!(matches!(
            r.resample_into(&[0i16; 160], &mut out),
            Err(CodecError::BufferTooSmall)
        ));
    }

    #[test]
    fn test_same_rate_passthrough() {
        let [input, _] = test_signals(500);
        let mut r = new_resampler(16000, 16000);
        assert_eq!(r.max_output_samples(input.len()), input.len());
        assert_eq!(r.resample(&input), input);
    }

    #[test]
    fn test_new_rejects_bad_args() {
        let mut small = vec![0.0f32; COEFFS_LEN - 1];
        assert!(matches!(
            Resampler::new(8000, 16000, &mut small),
            Err(CodecError::BufferTooSmall)
        ));
        let mut coeffs = vec![0.0f32; COEFFS_LEN];
        assert!(matches!(
            Resampler::new(0, 16000, &mut coeffs),
            Err(CodecError::InvalidInput)
        ));
        assert!(matches!(
            Resampler::new(8000, 0, &mut coeffs),
            Err(CodecError::InvalidInput)
        ));
    }

    /// `reset` must restore the exact initial state, even mid-stream.
    #[test]
    fn test_reset_mid_stream() {
        let [input, _] = test_signals(1000);
        let mut r = new_resampler(44100, 48000);
        let fresh = r.resample(&input);
        // 13 samples leaves partial history and a non-zero `pos`.
        r.resample(&input[..13]);
        r.reset();
        assert_eq!(r.resample(&input), fresh);
    }
}