audio_samples 1.0.5

use bytemuck::NoUninit;
use ndarray::ScalarOperand;
use num_traits::{FromPrimitive, Num, NumCast, One, ToBytes, Zero};
use serde::{Deserialize, Serialize};

use crate::repr::SampleType;
use crate::{AudioSamples, I24};
use std::fmt::{Debug, Display};
use std::ops::{Add, AddAssign, Div, DivAssign, Mul, MulAssign, Rem, RemAssign, Sub, SubAssign};

/// Performs a raw numeric cast from type `S` into `Self`.
///
/// Unlike [`ConvertTo`] and [`ConvertFrom`], casts apply standard numeric
/// conversion rules equivalent to an `as` expression — no audio-aware
/// scaling, rounding, or saturation is applied. This trait is used
/// internally to move values between numeric types for arithmetic or
/// indexing purposes.
///
/// For audio-meaningful conversions (e.g. `i16` → `f32` in the range
/// `[-1.0, 1.0]`), use [`ConvertTo`] or [`ConvertFrom`] instead.
pub trait CastFrom<S>: Sized {
    /// Casts a value of type `S` into `Self` using raw numeric conversion.
    ///
    /// # Arguments
    ///
    /// – `value` — the source value to cast.
    ///
    /// # Returns
    ///
    /// The cast value as `Self`. No audio-aware scaling is applied.
    fn cast_from(value: S) -> Self;
}

/// Performs a raw numeric cast from `Self` into type `T`.
///
/// This is the directional complement of [`CastFrom`]. A blanket
/// implementation delegates to `T::cast_from(self)`, so this trait does
/// not need to be implemented directly.
///
/// For audio-meaningful conversions, use [`ConvertTo`] instead.
pub trait CastInto<T>: Sized
where
    Self: CastFrom<T>,
{
    /// Casts `self` into type `T` using raw numeric conversion.
    ///
    /// # Returns
    ///
    /// The cast value as `T`. No audio-aware scaling is applied.
    fn cast_into(self) -> T;
}

/// Convenience supertrait asserting that a type can be raw-cast into all
/// supported audio sample types: `u8`, `i16`, [`I24`](i24::I24), `i32`, `f32`, and
/// `f64`.
///
/// A type satisfies this bound automatically when it implements [`CastInto`]
/// for all six target types; no manual implementation is required.
pub trait Castable:
    CastInto<u8> + CastInto<i16> + CastInto<I24> + CastInto<i32> + CastInto<f32> + CastInto<f64>
{
}
impl<T> Castable for T where
    T: CastInto<u8> + CastInto<i16> + CastInto<I24> + CastInto<i32> + CastInto<f32> + CastInto<f64>
{
}

mod sealed {
    pub trait Sealed {}
}
use sealed::Sealed;

/// Zero-sized marker type parameterised by byte width `N`.
///
/// Used together with [`SupportedByteSize`] as a compile-time guard on
/// [`AudioSample::as_bytes`] to restrict byte serialisation to widths that
/// are valid for the supported sample types: 1, 2, 3, 4, and 8.
#[non_exhaustive]
pub struct SampleByteSize<const N: usize>;

/// Sealed marker trait indicating that byte width `N` is a supported audio
/// sample size.
///
/// Implemented for [`SampleByteSize<N>`] where `N` is 1, 2, 3, 4, or 8,
/// corresponding to the byte widths of `u8`, `i16`, `I24`, `i32`/`f32`,
/// and `f64` respectively.
///
/// This trait is sealed and cannot be implemented outside this crate.
pub trait SupportedByteSize: Sealed {}

impl<const N: usize> Sealed for SampleByteSize<N> {}

impl SupportedByteSize for SampleByteSize<1> {}
impl SupportedByteSize for SampleByteSize<2> {}
impl SupportedByteSize for SampleByteSize<3> {}
impl SupportedByteSize for SampleByteSize<4> {}
impl SupportedByteSize for SampleByteSize<8> {}

/// Core trait for all audio sample types.
///
/// `AudioSample` is the foundational constraint that every audio sample
/// numeric type in this crate must satisfy. It provides a uniform interface
/// for arithmetic, conversions, byte serialisation, and interoperability
/// with `ndarray` scalar operations.
///
/// ## Supported Types
///
/// | Type    | Representation              | Bit depth |
/// |---------|-----------------------------|-----------|
/// | `u8`    | Unsigned PCM (silence = 128)| 8         |
/// | `i16`   | Signed integer PCM          | 16        |
/// | [`I24`](i24::I24) | Signed integer PCM (24-bit) | 24        |
/// | `i32`   | Signed integer PCM          | 32        |
/// | `f32`   | Normalised float            | 32        |
/// | `f64`   | Normalised float            | 64        |
///
/// For `u8`, the silence level is 128 (mid-scale unsigned PCM convention).
/// For `f32` and `f64`, `MAX` and `MIN` are `1.0` and `-1.0`.
///
/// # Safety
///
/// All implementors must satisfy [`bytemuck::NoUninit`], guaranteeing that
/// the byte representation contains no uninitialised or padding bytes. This
/// is required for safe byte-level serialisation via [`AudioSample::to_bytes`]
/// and [`AudioSample::as_bytes`].
///
/// ## Examples
///
/// ```
/// use audio_samples::AudioSample;
///
/// let sample: f32 = 0.5;
/// let bytes = sample.to_bytes();
/// assert_eq!(bytes.len(), 4); // f32 is 4 bytes
///
/// // Raw numeric cast (not audio-normalised)
/// let raw: f64 = sample.as_float();
/// assert_eq!(raw, 0.5_f64);
/// ```
pub trait AudioSample:
    // Standard library traits
    Copy
    + Sized
    + Default
    + Display
    + Debug
    + Sync
    + Send
    + PartialEq
    + PartialOrd
    + Add<Output = Self>
    + AddAssign<Self>
    + Sub<Output = Self>
    + SubAssign<Self>
    + Mul<Output = Self>
    + MulAssign<Self>
    + Div<Output = Self>
    + DivAssign<Self>
    + Rem<Output = Self>
    + RemAssign<Self>
    + Into<Self>
    + From<Self>
    + ToString

    // External crate traits
    + NoUninit // bytemuck trait to ensure no uninitialized bytes
    + Num // num-traits trait for numeric operations
    + One // num-traits trait for 1 value
    + Zero // num-traits trait for 0 value
    + ToBytes // num-traits trait for byte conversion
    + Serialize // serde trait for serialization
    + Deserialize<'static> //serde trait for deserialisation // Need to make these optional. 
    + FromPrimitive // num-traits trait for conversion from primitive types
    + NumCast // num-traits trait for casting between numeric types
    + ScalarOperand // ndarray trait for scalar operations

    // Library-specific traits. Most of which are below.
    // They define how to convert between types depending on the context.
    // Sometimes we are dealing with audio samples and float representations between -1.0 and 1.0, sometimes we are dealing with raw integer representations that we need to cast to floats for specific operations, but not -1.0 to 1.0, for various operations.
    + ConvertTo<Self> // "I can convert to myself" trait
    + ConvertTo<u8> // "I can convert to u8" trait
    + ConvertTo<i16> // "I can convert to i16" trait
    + ConvertTo<I24> // "I can convert to I24" trait
    + ConvertTo<i32> // "I can convert to i32" trait
    + ConvertTo<f32> // "I can convert to f32" trait
    + ConvertTo<f64> // "I can convert to f64" trait
    + CastFrom<usize> // "I can cast from a  usize"
    + Castable // "I can be cast into supported types"
{
    /// Returns this sample value unchanged.
    ///
    /// This is an identity method provided for API uniformity. It is useful
    /// in generic contexts where a value must be consumed through a trait
    /// interface but should pass through unmodified.
    ///
    /// # Returns
    ///
    /// `self` unchanged.
    ///
    /// ## Examples
    ///
    /// ```
    /// use audio_samples::AudioSample;
    ///
    /// let s: f32 = 0.75;
    /// assert_eq!(s.into_inner(), 0.75_f32);
    /// ```
    #[inline]
    #[must_use]
    fn into_inner(self) -> Self {
        self
    }

    /// Clamps `self` to the closed interval `[min, max]`.
    ///
    /// This is a per-type hook used internally by [`clip`][crate::AudioProcessing::clip]
    /// to select the most efficient clamping implementation.  Integer types
    /// use [`Ord::clamp`]; floating-point types use their inherent `.clamp()`
    /// method which the compiler maps to `VMAXPS`/`VMINPS` in vectorised
    /// loops.  The default implementation uses explicit comparisons and is
    /// correct for all types.
    ///
    /// # Arguments
    /// - `min` — lower bound (inclusive).
    /// - `max` — upper bound (inclusive).
    ///
    /// # Returns
    /// `min` if `self < min`, `max` if `self > max`, otherwise `self`.
    #[inline]
    #[must_use]
    fn clamp_to(self, min: Self, max: Self) -> Self {
        if self < min { min } else if self > max { max } else { self }
    }

    /// Returns the absolute-value maximum of a slice using a type-specific
    /// fast path when available, or `None` to signal fallback to the generic
    /// scalar path.
    ///
    /// The default returns `None`.  Types with hardware-specific
    /// implementations (e.g. `f32` on x86-64 with AVX2) override this.
    #[doc(hidden)]
    #[inline]
    fn avx2_abs_max(_slice: &[Self]) -> Option<Self> { None }

    /// Converts this sample into a byte vector in native-endian order.
    ///
    /// Each sample is serialised to its native byte representation using
    /// the platform's endianness. The returned vector has length
    /// `Self::BYTES as usize`.
    ///
    /// For converting a whole slice at once, prefer
    /// [`AudioSample::slice_to_bytes`].
    ///
    /// # Returns
    ///
    /// A `Vec<u8>` of length `Self::BYTES as usize` containing the
    /// native-endian byte representation of this sample.
    ///
    /// ## Examples
    ///
    /// ```
    /// use audio_samples::AudioSample;
    ///
    /// let sample: i16 = 256;
    /// let bytes = sample.to_bytes();
    /// assert_eq!(bytes.len(), 2);
    /// ```
    #[inline]
    #[must_use]
    fn to_bytes(self) -> Vec<u8> {
        self.to_ne_bytes().as_ref().to_vec()
    }

    /// Converts this sample into a fixed-size byte array in native-endian order.
    ///
    /// The const generic parameter `N` must equal the byte size of `Self`
    /// (e.g. `N = 4` for `f32` or `i32`). Supported widths — 1, 2, 3, 4,
    /// and 8 — are enforced at compile time via the [`SupportedByteSize`]
    /// bound.
    ///
    /// # Arguments
    ///
    /// – `N` (const generic) — the number of bytes in the output array.
    ///   Must match `Self::BYTES`; mismatches are compile errors.
    ///
    /// # Returns
    ///
    /// A `[u8; N]` containing the native-endian byte representation of
    /// this sample.
    ///
    /// ## Examples
    ///
    /// ```
    /// use audio_samples::AudioSample;
    ///
    /// let sample: i16 = 0x0100_i16;
    /// let bytes: [u8; 2] = sample.as_bytes::<2>();
    /// assert_eq!(bytes.len(), 2);
    /// ```
    #[inline]
    fn as_bytes<const N: usize>(&self) -> [u8; N]
        where SampleByteSize<N>: SupportedByteSize,
    {
        let bytes_ref = self.to_ne_bytes();
        let bytes_slice: &[u8] = bytes_ref.as_ref();
        let mut result = [0u8; N];
        result.copy_from_slice(bytes_slice);
        result
    }

    #[inline]
    /// Converts a slice of samples into a byte vector in native-endian order.
    ///
    /// Uses [`bytemuck`] to reinterpret the slice as bytes. This avoids
    /// element-wise copying when the alignment allows it.
    ///
    /// # Arguments
    ///
    /// – `samples` — a slice of samples to serialise.
    ///
    /// # Returns
    ///
    /// A `Vec<u8>` of length `samples.len() * Self::BYTES as usize`.
    ///
    /// ## Examples
    ///
    /// ```
    /// use audio_samples::AudioSample;
    ///
    /// let samples = [1_i16, 2, 3];
    /// let bytes = i16::slice_to_bytes(&samples);
    /// assert_eq!(bytes.len(), 6); // 3 samples × 2 bytes each
    /// ```
    fn slice_to_bytes(samples: &[Self]) -> Vec<u8> {
        Vec::from(bytemuck::cast_slice(samples))
    }

    #[inline]
    /// Casts this sample value to `f64` without audio-aware scaling.
    ///
    /// This is a raw numeric cast equivalent to `self as f64`. No
    /// normalisation into `[-1.0, 1.0]` is applied. For integer sample
    /// types, the raw integer magnitude is preserved.
    ///
    /// For audio-aware conversion that scales integer samples into the
    /// floating-point range, use [`ConvertTo::<f64>::convert_to`] instead.
    ///
    /// # Returns
    ///
    /// The sample value cast to `f64`, preserving the raw numeric magnitude.
    ///
    /// ## Examples
    ///
    /// ```
    /// use audio_samples::AudioSample;
    ///
    /// // Raw cast: the integer value 32767 becomes 32767.0, not 1.0.
    /// let sample: i16 = 32767;
    /// assert_eq!(sample.as_float(), 32767.0_f64);
    ///
    /// // For float types the value is unchanged.
    /// let f: f32 = 0.5;
    /// assert_eq!(f.as_float(), 0.5_f64);
    /// ```
    fn as_float(self) -> f64
    {
        self.cast_into()
    }

    /// Maximum representable amplitude value for this sample type.
    ///
    /// For integer types this is the type's integer maximum (e.g. `32767`
    /// for `i16`, `255` for `u8`). For float types (`f32`, `f64`) this
    /// is `1.0`.
    const MAX: Self;

    /// Minimum representable amplitude value for this sample type.
    ///
    /// For integer types this is the type's integer minimum (e.g. `-32768`
    /// for `i16`, `0` for `u8`). For float types this is `-1.0`.
    const MIN: Self;

    /// Bit depth of this sample type (e.g. `8` for `u8`, `16` for `i16`).
    const BITS: u8;

    /// Byte width of this sample type, derived as `BITS / 8`.
    const BYTES: u32 = Self::BITS as u32 / 8;

    /// Human-readable label for this sample type, used for display and
    /// plotting. Examples: `"u8"`, `"i16"`, `"I24"`, `"f32"`, `"f64"`.
    const LABEL: &'static str;

    /// Enum variant identifying this sample type at runtime.
    ///
    /// Corresponds to a variant of [`SampleType`].
    const SAMPLE_TYPE: SampleType;
}

/// Supertrait combining [`AudioSample`] with full bidirectional conversion
/// support across all standard sample types.
///
/// A type satisfies `StandardSample` automatically when it implements
/// [`AudioSample`] together with [`CastInto<f64>`], [`CastFrom<f64>`], and
/// [`ConvertFrom`] for every standard sample type (`u8`, `i16`, [`I24`](i24::I24),
/// `i32`, `f32`, `f64`). No manual implementation is required.
///
/// The supported standard sample types are `u8`, `i16`, [`I24`](i24::I24), `i32`,
/// `f32`, and `f64`. Most audio processing operations in this crate are
/// generic over `T: StandardSample`.
///
/// Prefer this bound over [`AudioSample`] when bidirectional conversions
/// between sample types are required.
pub trait StandardSample:
    AudioSample
    + CastInto<f64>
    + CastFrom<f64>
    + ConvertFrom<Self>
    + ConvertFrom<u8>
    + ConvertFrom<i16>
    + ConvertFrom<I24>
    + ConvertFrom<i32>
    + ConvertFrom<f32>
    + ConvertFrom<f64>
    + Castable
{
}

impl<T> StandardSample for T where
    T: AudioSample
        + CastInto<f64>
        + CastFrom<f64>
        + ConvertFrom<Self>
        + ConvertFrom<u8>
        + ConvertFrom<i16>
        + ConvertFrom<I24>
        + ConvertFrom<i32>
        + ConvertFrom<f32>
        + ConvertFrom<f64>
{
}

/// Trait for converting one sample type to another with audio-aware scaling.
///
/// `ConvertTo` performs conversions that are intended for audio sample values rather than raw
/// numeric casts.
///
/// ## Conversion Behavior
/// - **Integer ↔ Integer**: PCM-style bit-depth scaling (e.g. `i16::MAX` maps to `0x7FFF0000i32`)
/// - **Integer ↔ Float**: normalized scaling into $[-1.0, 1.0]$ using asymmetric endpoints
/// - **Float ↔ Integer**: clamp to $[-1.0, 1.0]$, then scale, round, and saturate
/// - **I24 Special Handling**: conversions treat `I24` as a 24-bit signed PCM integer
///
/// ## Example
/// ```rust
/// use audio_samples::ConvertTo;
///
/// let sample_i16: i16 = 16384; // approximately half-scale
/// let sample_f32: f32 = sample_i16.convert_to();
/// assert!((sample_f32 - 0.5).abs() < 1e-4);
///
/// let sample_i32: i32 = sample_i16.convert_to();
/// assert_eq!(sample_i32, 0x4000_0000);
/// ```
pub trait ConvertTo<Dst> {
    /// Converts this sample into `Dst` using audio-aware scaling.
    ///
    /// # Returns
    ///
    /// The converted sample as `Dst`. Conversion semantics follow the rules
    /// documented on [`ConvertTo`]: integer-to-float converts to
    /// `[-1.0, 1.0]`; float-to-integer clamps, scales, and rounds;
    /// integer-to-integer shifts bit depth with saturation.
    fn convert_to(self) -> Dst;
}

/// Audio-aware conversion from a source sample type into `Self`.
///
/// This is the blanket-implemented inverse of [`ConvertTo`]. Implementing
/// `ConvertFrom<Src>` for `Dst` automatically provides `ConvertTo<Dst>`
/// for `Src` via the blanket implementation.
///
/// Conversion semantics are the same as documented on [`ConvertTo`]:
/// integer ↔ float conversions apply asymmetric normalised scaling;
/// integer ↔ integer conversions use PCM-style bit-depth shifting with
/// saturation; `u8` uses mid-scale unsigned PCM conventions.
pub trait ConvertFrom<Src> {
    /// Converts a sample of type `Src` into `Self`.
    ///
    /// # Arguments
    ///
    /// – `source` — the source sample to convert.
    ///
    /// # Returns
    ///
    /// The converted sample as `Self`.
    fn convert_from(source: Src) -> Self;
}

impl<Src, Dst> ConvertTo<Dst> for Src
where
    Dst: ConvertFrom<Src>,
{
    #[inline]
    fn convert_to(self) -> Dst {
        Dst::convert_from(self)
    }
}

// Identity
macro_rules! impl_identity_conversion {
    ($ty:ty) => {
        impl ConvertFrom<$ty> for $ty {
            #[inline]
            fn convert_from(source: $ty) -> Self {
                source
            }
        }
    };
}

// Integer -> Integer, saturating (with bit-shift scaling if needed)
macro_rules! impl_int_to_int_conversion {
    ($from:ty, $to:ty) => {
        impl ConvertFrom<$from> for $to {
            #[inline]
            fn convert_from(source: $from) -> Self {
                let from_bits = <$from>::BITS as i32;
                let to_bits = <$to>::BITS as i32;

                let v = <i128 as From<$from>>::from(source);
                let scaled = if from_bits < to_bits {
                    v << (to_bits - from_bits)
                } else if from_bits > to_bits {
                    v >> (from_bits - to_bits)
                } else {
                    v
                };

                let min = <i128 as From<$to>>::from(<$to>::MIN);
                let max = <i128 as From<$to>>::from(<$to>::MAX);

                if scaled < min {
                    <$to>::MIN
                } else if scaled > max {
                    <$to>::MAX
                } else {
                    scaled as $to
                }
            }
        }
    };
}

// I24 -> Integer (any standard integer), saturating
macro_rules! impl_i24_to_int {
    ($to:ty) => {
        impl ConvertFrom<I24> for $to {
            #[inline]
            fn convert_from(source: I24) -> Self {
                let to_bits = <$to>::BITS as i32;
                let v = <i128 as From<i32>>::from(source.to_i32());

                let shift = to_bits - 24;
                let scaled = if shift >= 0 {
                    v << shift
                } else {
                    v >> (-shift)
                };

                let min = <i128 as From<$to>>::from(<$to>::MIN);
                let max = <i128 as From<$to>>::from(<$to>::MAX);
                if scaled < min {
                    <$to>::MIN
                } else if scaled > max {
                    <$to>::MAX
                } else {
                    scaled as $to
                }
            }
        }
    };
}

// Integer -> I24, saturating
macro_rules! impl_int_to_i24 {
    ($from:ty) => {
        impl ConvertFrom<$from> for I24 {
            #[inline]
            fn convert_from(source: $from) -> Self {
                let from_bits = <$from>::BITS as i32;
                let v = <i128 as From<$from>>::from(source);

                let shift = 24 - from_bits;
                let scaled = if shift >= 0 {
                    v << shift
                } else {
                    v >> (-shift)
                };

                let min = <i128 as From<i32>>::from(I24::MIN.to_i32());
                let max = <i128 as From<i32>>::from(I24::MAX.to_i32());
                let clamped = if scaled < min {
                    min as i32
                } else if scaled > max {
                    max as i32
                } else {
                    scaled as i32
                };

                I24::saturating_from_i32(clamped)
            }
        }
    };
}

// I24 -> Float (normalised)
macro_rules! impl_i24_to_float {
    ($to:ty) => {
        impl ConvertFrom<I24> for $to {
            #[inline]
            #[allow(clippy::cast_lossless)]
            fn convert_from(source: I24) -> Self {
                let v = source.to_i32() as $to;
                let max = I24::MAX.to_i32() as $to;
                let min = I24::MIN.to_i32() as $to;
                if v < 0.0 { v / -min } else { v / max }
            }
        }
    };
}

// Integer -> Float (normalised)
macro_rules! impl_int_to_float {
    ($from:ty, $to:ty) => {
        impl ConvertFrom<$from> for $to {
            #[inline]
            fn convert_from(source: $from) -> Self {
                let v = source;
                if v < 0 {
                    (v as $to) / (-(<$from>::MIN as $to))
                } else {
                    (v as $to) / (<$from>::MAX as $to)
                }
            }
        }
    };
}

// Float -> Integer (clamp + scale + truncating cast)
//
// Uses IEEE max/min (f32::max / f32::min) instead of clamp so that NaN
// inputs produce a defined finite value: max(NaN, -1.0) = -1.0 per IEEE
// 754-2008 maxNum semantics (returns the non-NaN operand). This lets LLVM
// prove the intermediate value is finite, eliminating the NaN-handling
// branches that Rust's saturating `as i32` would otherwise generate.
// Without those branches LLVM can emit vcvttps2dq + vpackssdw (vectorised
// f32→i32→i16) instead of the scalar vcvttss2si path.
//
// Semantic differences from a fully-IEEE-correct implementation:
// - NaN input → −MAX (e.g. −32767) instead of 0 or ±MAX.
// - Uses truncation rather than round-to-nearest (error ≤ 1 LSB, inaudible).
// - −1.0 maps to −MAX (e.g. −32767 for i16), not MIN (−32768).  The one-
//   code-unit asymmetry at full negative scale is inaudible in practice.
// Used for i16 (and similarly small integer types) where $to::MAX is exactly
// representable in f32/f64 (32767 < 2^23). `to_int_unchecked` removes the
// NaN/overflow branches that `as i32` would otherwise emit, letting LLVM
// vectorise the loop with vcvttps2dq + vpackssdw.
macro_rules! impl_float_to_int {
    ($from:ty, $to:ty) => {
        impl ConvertFrom<$from> for $to {
            #[inline]
            fn convert_from(source: $from) -> Self {
                // IEEE max/min: NaN input → -1.0 (non-NaN). v ∈ [-1.0, 1.0].
                let v = source.max(-1.0).min(1.0);
                // scaled ∈ [−MAX, MAX]. For i16, MAX = 32767 which is exactly
                // representable in f32, so scaled is strictly within i32 range.
                let scaled = v * (<$to>::MAX as $from);
                // SAFETY: scaled is finite and |scaled| ≤ $to::MAX ≤ 32767,
                // which fits comfortably in i32. No NaN/overflow is possible.
                let as_i32: i32 = unsafe { scaled.to_int_unchecked() };
                as_i32 as $to
            }
        }
    };
}

// Used for i32 where i32::MAX = 2147483647 rounds up to 2147483648.0 in f32
// (not exactly representable). Using `to_int_unchecked` there would be UB for
// source = 1.0, so we use the saturating `as $to` cast instead.
macro_rules! impl_float_to_large_int {
    ($from:ty, $to:ty) => {
        impl ConvertFrom<$from> for $to {
            #[inline]
            fn convert_from(source: $from) -> Self {
                let v = source.max(-1.0).min(1.0);
                let scaled = v * (<$to>::MAX as $from);
                // Saturating cast handles the case where $to::MAX as $from
                // rounds up (e.g. i32::MAX → 2147483648.0f32).
                scaled.clamp(<$to>::MIN as $from, <$to>::MAX as $from) as $to
            }
        }
    };
}

// Float -> I24 (clamp + scale + round + saturating)
macro_rules! impl_float_to_i24 {
    ($from:ty) => {
        impl ConvertFrom<$from> for I24 {
            #[inline]
            fn convert_from(source: $from) -> Self {
                let v = source.clamp(-1.0, 1.0);
                let scaled = if v < 0.0 {
                    v * (-(I24::MIN.to_i32() as $from))
                } else {
                    v * (I24::MAX.to_i32() as $from)
                };
                let rounded = scaled.round();
                let clamped = if rounded < (I24::MIN.to_i32() as $from) {
                    I24::MIN.to_i32()
                } else if rounded > (I24::MAX.to_i32() as $from) {
                    I24::MAX.to_i32()
                } else {
                    rounded as i32
                };
                I24::saturating_from_i32(clamped)
            }
        }
    };
}

// Float -> Float
macro_rules! impl_float_to_float {
    ($from:ty, $to:ty) => {
        impl ConvertFrom<$from> for $to {
            #[inline]
            fn convert_from(source: $from) -> Self {
                source as $to
            }
        }
    };
}

// ========================
// u8 <-> Signed Integer / I24 / Float
// ========================
//
// Conventions:
// - u8 is unsigned PCM with 128 as zero.
// - Centred value c = (u8 as i32) - 128  in [-128, 127].
// - Asymmetric scaling:
//     negative side uses divisor 128 (so 0 maps to -1.0 exactly)
//     positive side uses divisor 127 (so 255 maps to +1.0 exactly)
//
// - For u8 -> signed-int/I24: scale c into destination integer range using the
//   same asymmetric idea (negative uses abs(min), positive uses max).
// - For signed-int/I24 -> u8: invert scaling, round-to-nearest, clamp to [0,255].
//
// Notes:
// - All intermediate arithmetic is done in i128 to avoid overflow for i32::MIN abs.
// - Rounding is symmetric "nearest" for integer division (positive numerators only).

#[inline]
const fn div_round_nearest_i128(num: i128, den: i128) -> i128 {
    // Assumes den > 0 and num >= 0
    (num + (den / 2)) / den
}

// u8 -> signed integer (i16/i32), saturating + asymmetric scaling
macro_rules! impl_u8_to_int {
    ($to:ty) => {
        impl ConvertFrom<u8> for $to {
            #[inline]
            #[allow(clippy::cast_lossless)]
            fn convert_from(source: u8) -> Self {
                let c: i128 = (source as i128) - 128; // [-128, 127]

                let scaled: i128 = if c < 0 {
                    // Map -128 -> MIN exactly
                    // c is negative, abs range is 128
                    c * (-(<$to>::MIN as i128)) / 128
                } else {
                    // Map +127 -> MAX exactly
                    c * (<$to>::MAX as i128) / 127
                };

                // Should already be in-range, but keep the same saturating style.
                let min = <$to>::MIN as i128;
                let max = <$to>::MAX as i128;
                if scaled < min {
                    <$to>::MIN
                } else if scaled > max {
                    <$to>::MAX
                } else {
                    scaled as $to
                }
            }
        }
    };
}

// signed integer -> u8, saturating + asymmetric scaling + rounding
macro_rules! impl_int_to_u8 {
    ($from:ty) => {
        impl ConvertFrom<$from> for u8 {
            #[inline]
            #[allow(clippy::cast_lossless)]
            fn convert_from(source: $from) -> Self {
                let v: i128 = source as i128;

                let out_i128: i128 = if v < 0 {
                    // v in [MIN, -1]
                    let mag = (-v) as i128; // positive
                    let den = (-(<$from>::MIN as i128)); // abs(min), e.g. 32768 for i16
                    let scaled = div_round_nearest_i128(mag * 128, den); // 0..128
                    128 - scaled
                } else {
                    // v in [0, MAX]
                    let den = (<$from>::MAX as i128);
                    let scaled = div_round_nearest_i128(v * 127, den); // 0..127
                    128 + scaled
                };

                // Clamp to [0, 255]
                if out_i128 < 0 {
                    0
                } else if out_i128 > 255 {
                    255
                } else {
                    out_i128 as u8
                }
            }
        }
    };
}

// u8 -> I24 (asymmetric scaling), saturating
macro_rules! impl_u8_to_i24 {
    () => {
        impl ConvertFrom<u8> for I24 {
            #[inline]
            fn convert_from(source: u8) -> Self {
                let c: i128 = (<i128 as From<u8>>::from(source)) - 128; // [-128, 127]
                let min = <i128 as From<i32>>::from(I24::MIN.to_i32());
                let max = <i128 as From<i32>>::from(I24::MAX.to_i32());

                let scaled: i128 = if c < 0 {
                    c * (-min) / 128
                } else {
                    c * max / 127
                };

                // Clamp into i24 range, then saturating construct.
                let clamped: i32 = if scaled < min {
                    min as i32
                } else if scaled > max {
                    max as i32
                } else {
                    scaled as i32
                };

                I24::saturating_from_i32(clamped)
            }
        }
    };
}

// I24 -> u8 (invert asymmetric scaling), saturating + rounding
macro_rules! impl_i24_to_u8 {
    () => {
        impl ConvertFrom<I24> for u8 {
            #[inline]
            fn convert_from(source: I24) -> Self {
                let v: i128 = <i128 as From<i32>>::from(source.to_i32());
                let min = <i128 as From<i32>>::from(I24::MIN.to_i32()); // negative
                let max = <i128 as From<i32>>::from(I24::MAX.to_i32()); // positive

                let out_i128: i128 = if v < 0 {
                    let mag = <i128 as From<i128>>::from(-v);
                    let den = <i128 as From<i128>>::from(-min); // abs(min) = 8388608
                    let scaled = div_round_nearest_i128(mag * 128, den); // 0..128
                    128 - scaled
                } else {
                    let den = <i128 as From<i128>>::from(max); // 8388607
                    let scaled = div_round_nearest_i128(v * 127, den); // 0..127
                    128 + scaled
                };

                if out_i128 < 0 {
                    0
                } else if out_i128 > 255 {
                    255
                } else {
                    out_i128 as u8
                }
            }
        }
    };
}

// u8 -> float (normalised, asymmetric endpoints)
macro_rules! impl_u8_to_float {
    ($to:ty) => {
        impl ConvertFrom<u8> for $to {
            #[inline]
            fn convert_from(source: u8) -> Self {
                let c: i32 = (<i32 as From<u8>>::from(source)) - 128; // [-128, 127]
                let v = c as $to;
                if c < 0 {
                    v / (128.0 as $to)
                } else {
                    v / (127.0 as $to)
                }
            }
        }
    };
}

// float -> u8 (clamp + asymmetric scale + round + saturate)
macro_rules! impl_float_to_u8 {
    ($from:ty) => {
        impl ConvertFrom<$from> for u8 {
            #[inline]
            fn convert_from(source: $from) -> Self {
                let v = source.clamp(-1.0, 1.0);

                // Convert float to centred integer c in [-128, 127] with asymmetric scaling.
                // Negative maps to [-128, 0], positive maps to [0, 127].
                let c: i128 = if v < 0.0 {
                    // -1.0 -> -128 exactly
                    (v * (128.0 as $from)).round() as i128
                } else {
                    // +1.0 -> +127 exactly
                    (v * (127.0 as $from)).round() as i128
                };

                let out = 128i128 + c;

                if out < 0 {
                    0
                } else if out > 255 {
                    255
                } else {
                    out as u8
                }
            }
        }
    };
}

// ========================
// u8 Identity
// ========================

impl_identity_conversion!(u8);

// ========================
// u8 <-> Integer
// ========================

impl_u8_to_int!(i16);
impl_u8_to_int!(i32);

impl_int_to_u8!(i16);
impl_int_to_u8!(i32);

// ========================
// u8 <-> I24
// ========================

impl_u8_to_i24!();
impl_i24_to_u8!();

// ========================
// u8 <-> Float
// ========================

impl_u8_to_float!(f32);
impl_u8_to_float!(f64);

impl_float_to_u8!(f32);
impl_float_to_u8!(f64);

// ========================
// Identity
// ========================

impl_identity_conversion!(i16);
impl_identity_conversion!(I24);
impl_identity_conversion!(i32);
impl_identity_conversion!(f32);
impl_identity_conversion!(f64);

// ========================
// Integer <-> Integer (Saturating, No Normalisation)
// ========================

impl_int_to_int_conversion!(i16, i32);
impl_int_to_int_conversion!(i32, i16);

// ========================
// I24 <-> Integer
// ========================

impl_i24_to_int!(i16);
impl_i24_to_int!(i32);

impl_int_to_i24!(i16);
impl_int_to_i24!(i32);

// ========================
// Integer -> Float (Normalised +- 1.0)
// ========================

impl_int_to_float!(i16, f32);
impl_int_to_float!(i16, f64);

impl_int_to_float!(i32, f32);
impl_int_to_float!(i32, f64);

// ========================
// I24 -> Float (Normalised +- 1.0)
// ========================

impl_i24_to_float!(f32);
impl_i24_to_float!(f64);

// ========================
// Float -> Integer (Clamped, Rounded, Saturating)
// ========================

impl_float_to_int!(f32, i16);
impl_float_to_int!(f64, i16);

impl_float_to_large_int!(f32, i32);
impl_float_to_large_int!(f64, i32);

// ========================
// Float -> I24 (Clamped, Rounded, Saturating)
// ========================

impl_float_to_i24!(f32);
impl_float_to_i24!(f64);

// ========================
// Float <-> Float
// ========================

impl_float_to_float!(f32, f64);
impl_float_to_float!(f64, f32);

// ========================
// AVX2 fast path for f32 abs-max scan
// ========================

/// Computes the maximum absolute value of a contiguous f32 slice using AVX2.
///
/// Processes 8 f32 samples per iteration using `_mm256_andnot_ps` to mask
/// the sign bit (branchless abs) and `_mm256_max_ps` to track the running
/// maximum across 8 independent lanes.  A horizontal reduction folds the
/// lanes to a single scalar at the end.
///
/// # Safety
/// Caller must ensure that the AVX2 feature is available
/// (e.g. via `is_x86_feature_detected!("avx2")`).
#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2")]
unsafe fn abs_max_f32_avx2(slice: &[f32]) -> f32 {
    use std::arch::x86_64::{
        _mm_cvtss_f32, _mm_max_ps, _mm_max_ss, _mm_movehl_ps, _mm_shuffle_ps, _mm256_andnot_ps,
        _mm256_castps256_ps128, _mm256_extractf128_ps, _mm256_loadu_ps, _mm256_max_ps,
        _mm256_set1_ps, _mm256_setzero_ps,
    };

    // Bit-mask that strips the IEEE 754 sign bit from every lane.
    let sign_mask = _mm256_set1_ps(-0.0_f32);
    // Eight independent max accumulators — no sequential dependency.
    let mut max_v = _mm256_setzero_ps();

    let chunks = slice.len() / 8;
    let ptr = slice.as_ptr();

    for i in 0..chunks {
        // SAFETY: i * 8 + 7 < slice.len() because i < chunks = slice.len() / 8.
        unsafe {
            let v = _mm256_loadu_ps(ptr.add(i * 8));
            let abs_v = _mm256_andnot_ps(sign_mask, v); // abs via bitmask, no branch
            max_v = _mm256_max_ps(max_v, abs_v);
        }
    }

    // Horizontal reduction: fold 8 f32 lanes down to one scalar.
    let hi128 = _mm256_extractf128_ps(max_v, 1); // lanes [4..7]
    let lo128 = _mm256_castps256_ps128(max_v); // lanes [0..3]
    let max128 = _mm_max_ps(lo128, hi128);
    let shuf = _mm_movehl_ps(max128, max128);
    let max64 = _mm_max_ps(max128, shuf);
    let shuf2 = _mm_shuffle_ps(max64, max64, 0x1);
    let max32 = _mm_max_ss(max64, shuf2);
    let mut result = _mm_cvtss_f32(max32);

    // Scalar tail for the remaining < 8 samples.
    for &x in &slice[chunks * 8..] {
        let ax = x.abs();
        if ax > result {
            result = ax;
        }
    }
    result
}

// ========================
// AudioSample Implementations
// ========================
impl AudioSample for u8 {
    const MAX: Self = Self::MAX;
    const MIN: Self = Self::MIN;
    const BITS: Self = 8;
    const LABEL: &'static str = "u8";
    const SAMPLE_TYPE: SampleType = SampleType::U8;
    #[inline]
    fn clamp_to(self, min: Self, max: Self) -> Self {
        Ord::clamp(self, min, max)
    }
}

impl AudioSample for i16 {
    const MAX: Self = Self::MAX;
    const MIN: Self = Self::MIN;
    const BITS: u8 = 16;
    const LABEL: &'static str = "i16";
    const SAMPLE_TYPE: SampleType = SampleType::I16;
    #[inline]
    fn clamp_to(self, min: Self, max: Self) -> Self {
        Ord::clamp(self, min, max)
    }
}

impl AudioSample for I24 {
    #[inline]
    fn slice_to_bytes(samples: &[Self]) -> Vec<u8> {
        Self::write_i24s_ne(samples)
    }

    const MAX: Self = Self::MAX;
    const MIN: Self = Self::MIN;
    const BITS: u8 = 24;
    const LABEL: &'static str = "I24";
    const SAMPLE_TYPE: SampleType = SampleType::I24;
    #[inline]
    fn clamp_to(self, min: Self, max: Self) -> Self {
        Ord::clamp(self, min, max)
    }
}

impl AudioSample for i32 {
    const MAX: Self = Self::MAX;
    const MIN: Self = Self::MIN;
    const BITS: u8 = 32;
    const LABEL: &'static str = "i32";
    const SAMPLE_TYPE: SampleType = SampleType::I32;
    #[inline]
    fn clamp_to(self, min: Self, max: Self) -> Self {
        Ord::clamp(self, min, max)
    }
}

impl AudioSample for f32 {
    const MAX: Self = 1.0;
    const MIN: Self = -1.0;
    const BITS: u8 = 32;
    const LABEL: &'static str = "f32";
    const SAMPLE_TYPE: SampleType = SampleType::F32;
    /// Uses `f32::clamp` — compiles to `VMAXSS`/`VMINSS` (and `VMAXPS`/`VMINPS` in
    /// vectorised loops), matching what C achieves with `-ffast-math`.
    #[inline]
    fn clamp_to(self, min: Self, max: Self) -> Self {
        self.clamp(min, max)
    }

    #[inline]
    fn avx2_abs_max(slice: &[Self]) -> Option<Self> {
        #[cfg(target_arch = "x86_64")]
        if is_x86_feature_detected!("avx2") {
            // safety: caller must ensure AVX2 is available, which we check with is_x86_feature_detected.
            return Some(unsafe { abs_max_f32_avx2(slice) });
        }
        None
    }
}

impl AudioSample for f64 {
    const MAX: Self = 1.0;
    const MIN: Self = -1.0;
    const BITS: u8 = 64;
    const LABEL: &'static str = "f64";
    const SAMPLE_TYPE: SampleType = SampleType::F64;
    /// Uses `f64::clamp` — compiles to `VMAXSD`/`VMINSD` in vectorised loops.
    #[inline]
    fn clamp_to(self, min: Self, max: Self) -> Self {
        self.clamp(min, max)
    }
}

// ========================
// Generate All Conversions
// ========================

macro_rules! impl_cast_from {
    ($src:ty => [$($dst:ty),+]) => {
        $(
            impl CastFrom<$src> for $dst {
                #[inline]
                fn cast_from(value: $src) -> Self {
                    value as $dst
                }
            }
        )+
    };
}

impl_cast_from!(u8 => [u8, i16, i32, f32, f64]);
impl_cast_from!(i16 => [u8, i16, i32, f32, f64]);
impl_cast_from!(i32 => [u8, i16, i32, f32, f64]);
impl_cast_from!(f64 => [u8, i16, i32, f32, f64]);
impl_cast_from!(f32 => [u8, i16, i32, f32, f64]);

/// Macro to implement the `CastFrom` trait for multiple type pairs
macro_rules! impl_cast_from_i24 {
    // Simple direct casts (no clamping or special logic)
    ($src:ty => $dst:ty) => {
        impl CastFrom<$src> for $dst {
            #[inline]
            #[allow(clippy::cast_lossless)]
            fn cast_from(value: $src) -> Self {
                value as $dst
            }
        }
    };

    // Clamped casts for usize -> integer
    (clamp_usize $src:ty => $dst:ty, $max:expr) => {
        impl CastFrom<$src> for $dst {
            #[inline]
            fn cast_from(value: $src) -> Self {
                if value > $max as $src {
                    $max
                } else {
                    value as $dst
                }
            }
        }
    };

    // usize -> I24 with clamping and try_from_i32
    (usize_to_i24 $src:ty => $dst:ty) => {
        impl CastFrom<$src> for $dst {
            #[inline]
            fn cast_from(value: $src) -> Self {
                if value > I24::MAX.to_i32() as $src {
                    I24::MAX
                } else {
                    match I24::try_from_i32(value as i32) {
                        Some(x) => x,
                        None => I24::MIN,
                    }
                }
            }
        }
    };

    // I24 -> primitive
    (i24_to_primitive $src:ty => $dst:ty) => {
        impl CastFrom<$src> for $dst {
            #[inline]
            #[allow(clippy::cast_lossless)]
            fn cast_from(value: $src) -> Self {
                value.to_i32() as $dst
            }
        }
    };

    // primitive -> I24
    (primitive_to_i24 $src:ty => $dst:ty) => {
        impl CastFrom<$src> for $dst {
            #[inline]
            #[allow(clippy::cast_lossless)]
            fn cast_from(value: $src) -> Self {
                I24::try_from_i32(value as i32).unwrap_or(I24::MIN)
            }
        }
    };

    // identity
    (identity $t:ty) => {
        impl CastFrom<$t> for $t {
            #[inline]
            fn cast_from(value: $t) -> Self {
                value
            }
        }
    };
}

// usize to primitives
impl_cast_from_i24!(clamp_usize usize => u8, u8::MAX);
impl_cast_from_i24!(clamp_usize usize => i16, i16::MAX);
impl_cast_from_i24!(usize_to_i24 usize => I24);
impl_cast_from_i24!(clamp_usize usize => i32, i32::MAX);
impl_cast_from_i24!(usize => f32);
impl_cast_from_i24!(usize => f64);

// I24 to primitives
impl_cast_from_i24!(i24_to_primitive I24 => u8);
impl_cast_from_i24!(i24_to_primitive I24 => i16);
impl_cast_from_i24!(identity I24);
impl_cast_from_i24!(i24_to_primitive I24 => i32);
impl_cast_from_i24!(i24_to_primitive I24 => f32);
impl_cast_from_i24!(i24_to_primitive I24 => f64);

// primitives to I24
impl_cast_from_i24!(primitive_to_i24 u8 => I24);
impl_cast_from_i24!(primitive_to_i24 i16 => I24);
impl_cast_from_i24!(primitive_to_i24 i32 => I24);
impl_cast_from_i24!(primitive_to_i24 f32 => I24);
impl_cast_from_i24!(primitive_to_i24 f64 => I24);

macro_rules! impl_cast_into {
    ($src:ty => [$($dst:ty),+]) => {
        $(
            impl CastInto<$dst> for $src {
                #[inline]
                fn cast_into(self) -> $dst {
                    <$dst>::cast_from(self)
                }
            }
        )+
    };
}
impl_cast_into!(u8 => [u8, i16, i32, f32, f64]);
impl_cast_into!(i16 => [u8, i16, i32, f32, f64]);
impl_cast_into!(i32 => [u8, i16, i32, f32, f64]);
impl_cast_into!(f64 => [u8, i16, i32, f32, f64]);
impl_cast_into!(f32 => [u8, i16, i32, f32, f64]);

/// Macro to implement the `CastInto` trait for multiple type pairs
macro_rules! impl_cast_into_i24 {
    // I24 -> primitive (via to_i32)
    (i24_to_primitive $src:ty => $dst:ty) => {
        impl CastInto<$dst> for $src {
            #[inline]
            fn cast_into(self) -> $dst {
                self.to_i32() as $dst
            }
        }
    };

    // primitive -> I24 (via CastFrom)
    (primitive_to_i24 $src:ty => $dst:ty) => {
        impl CastInto<$dst> for $src {
            #[inline]
            fn cast_into(self) -> $dst {
                <$dst as CastFrom<$src>>::cast_from(self)
            }
        }
    };

    // identity
    (identity $t:ty) => {
        impl CastInto<$t> for $t {
            #[inline]
            fn cast_into(self) -> $t {
                self
            }
        }
    };
}
// I24 to primitives

impl_cast_into_i24!(i24_to_primitive I24 => u8);
impl_cast_into_i24!(i24_to_primitive I24 => i16);
impl_cast_into_i24!(identity I24);
impl_cast_into_i24!(i24_to_primitive I24 => i32);
impl_cast_into_i24!(i24_to_primitive I24 => f32);
impl_cast_into_i24!(i24_to_primitive I24 => f64);

// primitives to I24
impl_cast_into_i24!(primitive_to_i24 u8 => I24);
impl_cast_into_i24!(primitive_to_i24 i16 => I24);
impl_cast_into_i24!(primitive_to_i24 i32 => I24);
impl_cast_into_i24!(primitive_to_i24 f32 => I24);
impl_cast_into_i24!(primitive_to_i24 f64 => I24);

/// Audio sample conversion and casting operations for [`AudioSamples`].
///
/// This trait defines the public conversion surface for transforming an
/// [`AudioSamples`] value from one sample representation to another.
///
/// ## Purpose
///
/// Audio data is commonly represented using both integer PCM formats and
/// floating-point formats. This trait provides two explicit conversion modes:
///
/// - *Audio-aware conversion* for interpreting numeric values as audio samples,
///   applying the appropriate scaling and clamping when moving between integer
///   and floating-point representations.
/// - *Raw numeric casting* for transforming values using standard numeric rules
///   without audio-specific scaling.
///
/// The two modes are intentionally distinct and must be selected explicitly by
/// the caller.
///
/// ## Behavioural Guarantees
///
/// - All operations return a new owned [`AudioSamples`] value.
/// - Sample rate, channel structure, and sample ordering are preserved.
/// - The source audio is not modified.
/// - Conversions are total and do not return `Result`.
///
/// When converting from floating-point to fixed-width integer formats, values
/// outside the representable range are clamped.
///
/// ## Assumptions
///
/// The conversion behaviour is defined by the conversion traits implemented for
/// the involved sample types. This trait is implemented for `AudioSamples<T>`
/// where those conversions are available.
pub trait AudioTypeConversion: Sized {
    /// The source sample type of the audio being converted.
    type Sample: StandardSample;

    /// Converts the audio to a different sample type using audio-aware scaling.
    ///
    /// Performs an audio-aware conversion from `Self::Sample` to `O`. The
    /// amplitude meaning of each sample is preserved: integer-to-float
    /// conversions normalise into `[-1.0, 1.0]`; float-to-integer conversions
    /// clamp, scale, and round; integer-to-integer conversions shift bit depth
    /// with saturation. For `u8`, the mid-scale unsigned PCM convention is
    /// applied.
    ///
    /// The source audio is not modified; a new owned value is returned.
    ///
    /// # Returns
    ///
    /// A new owned [`AudioSamples<'static, O>`] with samples converted from
    /// `Self::Sample` to `O`. Sample rate, channel count, and temporal
    /// ordering are preserved.
    ///
    /// ## Examples
    ///
    /// ```
    /// use audio_samples::{AudioSamples, sample_rate, AudioTypeConversion};
    /// use ndarray::array;
    ///
    /// let audio = AudioSamples::new_mono(
    ///     array![0i16, i16::MAX],
    ///     sample_rate!(44100),
    /// ).unwrap();
    ///
    /// let float_audio = audio.to_format::<f32>();
    /// let samples = float_audio.as_mono().unwrap();
    /// assert_eq!(samples[0], 0.0_f32);
    /// assert!((samples[1] - 1.0_f32).abs() < 1e-4);
    /// ```
    fn to_format<O>(&self) -> AudioSamples<'static, O>
    where
        Self::Sample: ConvertTo<O> + ConvertFrom<O>,
        O: StandardSample;

    /// Converts the audio to a different sample type, consuming the source.
    ///
    /// This is the consuming counterpart to [`AudioTypeConversion::to_format`].
    /// It performs the same audio-aware conversion, but takes ownership of the
    /// input value. Prefer this method over `to_format` when the source is no
    /// longer needed, to avoid an unnecessary clone.
    ///
    /// # Returns
    ///
    /// A new owned [`AudioSamples<'static, O>`] with samples converted from
    /// `Self::Sample` to `O`. Sample rate, channel count, and temporal
    /// ordering are preserved.
    ///
    /// ## Examples
    ///
    /// ```
    /// use audio_samples::{AudioSamples, AudioTypeConversion, sample_rate};
    /// use ndarray::array;
    ///
    /// let audio = AudioSamples::new_mono(array![0i16, i16::MAX], sample_rate!(44100)).unwrap();
    /// let audio_f32: AudioSamples<'static, f32> = audio.to_type::<f32>();
    /// assert_eq!(audio_f32[0], 0.0);
    /// assert!((audio_f32[1] - 1.0).abs() < 1e-4);
    /// ```
    fn to_type<O>(self) -> AudioSamples<'static, O>
    where
        Self::Sample: ConvertTo<O> + ConvertFrom<O>,
        O: StandardSample;

    /// Casts the audio to a different sample type without audio-aware scaling.
    ///
    /// Performs a raw numeric cast from `Self::Sample` to `O`, equivalent to
    /// an `as` cast applied element-wise. No normalisation, clamping, or
    /// bit-depth scaling is applied. Integer values are preserved as their
    /// raw numeric magnitude.
    ///
    /// Use [`AudioTypeConversion::to_format`] when amplitude meaning must be
    /// preserved across sample types.
    ///
    /// # Returns
    ///
    /// A new owned [`AudioSamples<'static, O>`] containing raw-cast samples.
    /// The source audio is unchanged. Sample rate, channel count, and temporal
    /// ordering are preserved.
    ///
    /// ## Examples
    ///
    /// ```
    /// use audio_samples::{AudioSamples, AudioTypeConversion, sample_rate};
    /// use ndarray::array;
    ///
    /// // Raw cast: i16 value 1000 becomes f32 value 1000.0, not 0.030518...
    /// let audio = AudioSamples::new_mono(array![1000i16, -500], sample_rate!(44100)).unwrap();
    /// let raw = audio.cast_as::<f32>();
    /// assert_eq!(raw[0], 1000.0_f32);
    /// assert_eq!(raw[1], -500.0_f32);
    /// ```
    fn cast_as<O>(&self) -> AudioSamples<'static, O>
    where
        Self::Sample: CastInto<O> + ConvertTo<O>,
        O: StandardSample;

    /// Casts the audio to a different sample type without audio-aware scaling,
    /// consuming the source.
    ///
    /// This is the consuming counterpart to [`AudioTypeConversion::cast_as`].
    /// It performs the same raw numeric cast but takes ownership of the input.
    ///
    /// # Returns
    ///
    /// A new owned [`AudioSamples<'static, O>`] containing raw-cast samples.
    /// Sample rate, channel count, and ordering are preserved.
    ///
    /// ## Examples
    ///
    /// ```
    /// use audio_samples::{AudioSamples, AudioTypeConversion, sample_rate};
    /// use ndarray::array;
    ///
    /// let audio = AudioSamples::new_mono(array![255u8, 128u8, 0u8], sample_rate!(44100)).unwrap();
    /// let raw: AudioSamples<'static, i16> = audio.cast_to::<i16>();
    /// assert_eq!(raw[0], 255);
    /// assert_eq!(raw[1], 128);
    /// assert_eq!(raw[2], 0);
    /// ```
    fn cast_to<O>(self) -> AudioSamples<'static, O>
    where
        Self::Sample: CastInto<O> + ConvertTo<O>,
        O: StandardSample;

    /// Casts the audio to `f64` without audio-aware scaling.
    ///
    /// Each sample value is raw-cast to `f64` as a number, preserving its
    /// numeric magnitude without normalisation. For integer sample types,
    /// the integer value (e.g. `32767`) is cast directly to `f64`, not
    /// scaled to `1.0`.
    ///
    /// For audio-aware conversion into `[-1.0, 1.0]`, use
    /// [`AudioTypeConversion::as_f64`] instead.
    ///
    /// # Returns
    ///
    /// A new owned [`AudioSamples<'static, f64>`] with raw-cast sample values.
    #[inline]
    fn cast_as_f64(&self) -> AudioSamples<'static, f64> {
        self.cast_as::<f64>()
    }

    /// Converts the audio to `f64` using audio-aware scaling.
    ///
    /// Integer sample values are normalised into `[-1.0, 1.0]` according to
    /// the [`ConvertTo`] rules for the source type. For `u8` audio, value
    /// `128` maps to `0.0`. For float sources the values are widened unchanged.
    ///
    /// # Returns
    ///
    /// A new owned [`AudioSamples<'static, f64>`] with normalised sample values.
    #[inline]
    fn as_float(&self) -> AudioSamples<'static, f64> {
        self.to_format::<f64>()
    }

    /// Converts the audio to `f64` using audio-aware scaling.
    ///
    /// Equivalent to [`AudioTypeConversion::as_float`]. Integer sample values
    /// are normalised into `[-1.0, 1.0]`.
    ///
    /// # Returns
    ///
    /// A new owned [`AudioSamples<'static, f64>`] with normalised sample values.
    #[inline]
    fn as_f64(&self) -> AudioSamples<'static, f64> {
        self.to_format::<f64>()
    }

    /// Converts the audio to `f32` using audio-aware scaling.
    ///
    /// Integer sample values are normalised into `[-1.0, 1.0]`. For `u8`
    /// audio, value `128` maps to `0.0`.
    ///
    /// # Returns
    ///
    /// A new owned [`AudioSamples<'static, f32>`] with normalised sample values.
    #[inline]
    fn as_f32(&self) -> AudioSamples<'static, f32> {
        self.to_format::<f32>()
    }

    /// Converts the audio to 32-bit signed integer PCM using audio-aware
    /// scaling.
    ///
    /// Float samples in `[-1.0, 1.0]` are scaled and rounded to the `i32`
    /// range with saturation. For `u8` audio, mid-scale value `128` maps to
    /// `0`.
    ///
    /// # Returns
    ///
    /// A new owned [`AudioSamples<'static, i32>`] in signed 32-bit PCM format.
    #[inline]
    fn as_i32(&self) -> AudioSamples<'static, i32> {
        self.to_format::<i32>()
    }

    /// Converts the audio to 16-bit signed integer PCM using audio-aware
    /// scaling.
    ///
    /// Float samples in `[-1.0, 1.0]` are scaled and rounded to the `i16`
    /// range with saturation. For `u8` audio, mid-scale value `128` maps to
    /// `0`.
    ///
    /// # Returns
    ///
    /// A new owned [`AudioSamples<'static, i16>`] in signed 16-bit PCM format.
    #[inline]
    fn as_i16(&self) -> AudioSamples<'static, i16> {
        self.to_format::<i16>()
    }

    /// Converts the audio to 24-bit signed integer PCM using audio-aware
    /// scaling.
    ///
    /// Float samples in `[-1.0, 1.0]` are scaled and rounded to the [`I24`](i24::I24)
    /// range with saturation. For `u8` audio, mid-scale value `128` maps to
    /// `0`.
    ///
    /// # Returns
    ///
    /// A new owned [`AudioSamples<'static, I24>`] in signed 24-bit PCM format.
    #[inline]
    fn as_i24(&self) -> AudioSamples<'static, I24> {
        self.to_format::<I24>()
    }

    /// Converts the audio to 8-bit unsigned PCM format using audio-aware
    /// scaling.
    ///
    /// Float samples in `[-1.0, 1.0]` and signed integer samples are scaled
    /// to the `u8` range with mid-scale silence at `128`. Negative full-scale
    /// maps to `0`; positive full-scale maps to `255`.
    ///
    /// # Returns
    ///
    /// A new owned [`AudioSamples<'static, u8>`] in unsigned 8-bit PCM format.
    #[inline]
    fn as_u8(&self) -> AudioSamples<'static, u8> {
        self.to_format::<u8>()
    }
}

#[cfg(test)]
mod conversion_tests {
    use super::*;

    use i24::i24;

    macro_rules! assert_approx_eq {
        ($left:expr, $right:expr, $tolerance:expr) => {
            assert!(
                ($left - $right).abs() < $tolerance,
                "assertion failed: `{} ≈ {}` (tolerance: {})",
                $left,
                $right,
                $tolerance
            );
        };
    }

    #[test]
    fn u8_tests() {
        let zero: u8 = 0;
        let mid: u8 = 128;
        let max: u8 = 255;
        let neg_one_f32: f32 = -1.0;
        let zero_f32: f32 = 0.0;
        let one_f32: f32 = 1.0;

        let zero_to_neg_one: f32 = zero.convert_to();
        let mid_to_zero: f32 = mid.convert_to();
        let max_to_one: f32 = max.convert_to();
        assert_approx_eq!(zero_to_neg_one as f64, -1.0, 1e-10);
        assert_approx_eq!(mid_to_zero as f64, 0.0, 1e-10);
        assert_approx_eq!(max_to_one as f64, 1.0, 1e-10);

        let neg_one_to_u8: u8 = neg_one_f32.convert_to();
        let zero_to_u8: u8 = zero_f32.convert_to();
        let one_to_u8: u8 = one_f32.convert_to();

        assert_eq!(neg_one_to_u8, 0);
        assert_eq!(zero_to_u8, 128);
        assert_eq!(one_to_u8, 255);
    }

    // Edge cases for i16 conversions
    #[test]
    fn i16_edge_cases() {
        // Test minimum value
        let min_i16: i16 = i16::MIN;
        let min_i16_to_f32: f32 = min_i16.convert_to();
        // Use higher epsilon for floating point comparison
        assert_approx_eq!(min_i16_to_f32 as f64, -1.0, 1e-5);

        let min_i16_to_i32: i32 = min_i16.convert_to();
        assert_eq!(min_i16_to_i32, i32::MIN);

        let min_i16_to_i24: I24 = min_i16.convert_to();
        let expected_i24_min = i24!(i32::MIN >> 8);
        assert_eq!(min_i16_to_i24.to_i32(), expected_i24_min.to_i32());

        // Test maximum value
        let max_i16: i16 = i16::MAX;
        let max_i16_to_f32: f32 = max_i16.convert_to();
        assert_approx_eq!(max_i16_to_f32 as f64, 1.0, 1e-4);

        let max_i16_to_i32: i32 = max_i16.convert_to();
        assert_eq!(max_i16_to_i32, 0x7FFF0000);

        // Test zero
        let zero_i16: i16 = 0;
        let zero_i16_to_f32: f32 = zero_i16.convert_to();
        assert_approx_eq!(zero_i16_to_f32 as f64, 0.0, 1e-10);

        let zero_i16_to_i32: i32 = zero_i16.convert_to();
        assert_eq!(zero_i16_to_i32, 0);

        let zero_i16_to_i24: I24 = zero_i16.convert_to();
        assert_eq!(zero_i16_to_i24.to_i32(), 0);

        // Test mid-range positive
        let half_max_i16: i16 = i16::MAX / 2;
        let half_max_i16_to_f32: f32 = half_max_i16.convert_to();
        // Use higher epsilon for floating point comparison of half values
        assert_approx_eq!(half_max_i16_to_f32 as f64, 0.5, 1e-4);

        let half_max_i16_to_i32: i32 = half_max_i16.convert_to();
        assert_eq!(half_max_i16_to_i32, 0x3FFF0000);

        // Test mid-range negative
        let half_min_i16: i16 = i16::MIN / 2;
        let half_min_i16_to_f32: f32 = half_min_i16.convert_to();
        assert_approx_eq!(half_min_i16_to_f32 as f64, -0.5, 1e-4);

        // let half_min_i16_to_i32: i32 = half_min_i16.convert_to();
        // assert_eq!(half_min_i16_to_i32, 0xC0010000); // i16::MIN/2 == -16384
    }

    // Edge cases for i32 conversions
    #[test]
    fn i32_edge_cases() {
        // Test minimum value
        let min_i32: i32 = i32::MIN;
        let min_i32_to_f32: f32 = min_i32.convert_to();
        assert_approx_eq!(min_i32_to_f32 as f64, -1.0, 1e-6);

        let min_i32_to_f64: f64 = min_i32.convert_to();
        assert_approx_eq!(min_i32_to_f64, -1.0, 1e-12);

        let min_i32_to_i16: i16 = min_i32.convert_to();
        assert_eq!(min_i32_to_i16, i16::MIN);

        // Test maximum value
        let max_i32: i32 = i32::MAX;
        let max_i32_to_f32: f32 = max_i32.convert_to();
        assert_approx_eq!(max_i32_to_f32 as f64, 1.0, 1e-6);

        let max_i32_to_f64: f64 = max_i32.convert_to();
        assert_approx_eq!(max_i32_to_f64, 1.0, 1e-12);

        let max_i32_to_i16: i16 = max_i32.convert_to();
        assert_eq!(max_i32_to_i16, i16::MAX);

        // Test zero
        let zero_i32: i32 = 0;
        let zero_i32_to_f32: f32 = zero_i32.convert_to();
        assert_approx_eq!(zero_i32_to_f32 as f64, 0.0, 1e-10);

        let zero_i32_to_f64: f64 = zero_i32.convert_to();
        assert_approx_eq!(zero_i32_to_f64, 0.0, 1e-12);

        let zero_i32_to_i16: i16 = zero_i32.convert_to();
        assert_eq!(zero_i32_to_i16, 0);

        // Test quarter-range values
        let quarter_max_i32: i32 = i32::MAX / 4;
        let quarter_max_i32_to_f32: f32 = quarter_max_i32.convert_to();
        assert_approx_eq!(quarter_max_i32_to_f32 as f64, 0.25, 1e-6);

        let quarter_min_i32: i32 = i32::MIN / 4;
        let quarter_min_i32_to_f32: f32 = quarter_min_i32.convert_to();
        assert_approx_eq!(quarter_min_i32_to_f32 as f64, -0.25, 1e-6);
    }

    // Edge cases for f32 conversions
    #[test]
    fn f32_edge_cases() {
        // Test -1.0 (minimum valid value)
        let min_f32: f32 = -1.0;
        let min_f32_to_i16: i16 = min_f32.convert_to();
        // For exact -1.0, we can get -32767 due to rounding in the implementation
        // This is acceptable since it's only 1 bit off from the true min
        assert!(
            min_f32_to_i16 == i16::MIN || min_f32_to_i16 == -32767,
            "Expected either -32768 or -32767, got {}",
            min_f32_to_i16
        );

        let min_f32_to_i32: i32 = min_f32.convert_to();
        assert!(
            min_f32_to_i32 == i32::MIN || min_f32_to_i32 == -2147483647,
            "Expected either i32::MIN or -2147483647, got {}",
            min_f32_to_i32
        );

        let min_f32_to_i24: I24 = min_f32.convert_to();
        let expected_i24 = I24::MIN;
        let diff = (min_f32_to_i24.to_i32() - expected_i24.to_i32()).abs();
        assert!(diff <= 1, "I24 values differ by more than 1, {}", diff);

        // Test 1.0 (maximum valid value)
        let max_f32: f32 = 1.0;
        let max_f32_to_i16: i16 = max_f32.convert_to();
        println!("DEBUG: f32 -> i16 conversion for 1.0");
        println!(
            "Input: {}, Output: {}, Expected: {}",
            max_f32,
            max_f32_to_i16,
            i16::MAX
        );
        assert_eq!(max_f32_to_i16, i16::MAX);

        let max_f32_to_i32: i32 = max_f32.convert_to();
        println!("DEBUG: f32 -> i32 conversion for 1.0");
        println!(
            "Input: {}, Output: {}, Expected: {}",
            max_f32,
            max_f32_to_i32,
            i32::MAX
        );
        assert_eq!(max_f32_to_i32, i32::MAX);

        // Test 0.0
        let zero_f32: f32 = 0.0;
        let zero_f32_to_i16: i16 = zero_f32.convert_to();
        println!("DEBUG: f32 -> i16 conversion for 0.0");
        println!(
            "Input: {}, Output: {}, Expected: 0",
            zero_f32, zero_f32_to_i16
        );
        assert_eq!(zero_f32_to_i16, 0);

        let zero_f32_to_i32: i32 = zero_f32.convert_to();
        println!("DEBUG: f32 -> i32 conversion for 0.0");
        println!(
            "Input: {}, Output: {}, Expected: 0",
            zero_f32, zero_f32_to_i32
        );
        assert_eq!(zero_f32_to_i32, 0);

        let zero_f32_to_i24: I24 = zero_f32.convert_to();
        println!("DEBUG: f32 -> I24 conversion for 0.0");
        println!(
            "Input: {}, Output: {} (i32 value), Expected: 0",
            zero_f32,
            zero_f32_to_i24.to_i32()
        );
        assert_eq!(zero_f32_to_i24.to_i32(), 0);

        // Test clamping of out-of-range values
        let large_f32: f32 = 2.0;
        let large_f32_to_i16: i16 = large_f32.convert_to();
        assert_eq!(large_f32_to_i16, i16::MAX);

        let neg_large_f32: f32 = -2.0;
        let neg_large_f32_to_i16: i16 = neg_large_f32.convert_to();
        assert!(
            neg_large_f32_to_i16 == i16::MIN || neg_large_f32_to_i16 == -32767,
            "Expected either -32768 or -32767, got {}",
            neg_large_f32_to_i16
        );

        let large_f32_to_i32: i32 = large_f32.convert_to();
        assert_eq!(large_f32_to_i32, i32::MAX);

        let neg_large_f32_to_i32: i32 = neg_large_f32.convert_to();
        assert!(
            neg_large_f32_to_i32 == i32::MIN || neg_large_f32_to_i32 == -2147483647,
            "Expected either i32::MIN or -2147483647, got {}",
            neg_large_f32_to_i32
        );

        // Test small values
        let small_value: f32 = 1.0e-6;
        let small_value_to_i16: i16 = small_value.convert_to();
        assert_eq!(small_value_to_i16, 0);

        let small_value_to_i32: i32 = small_value.convert_to();
        assert_eq!(small_value_to_i32, 2147); // 1.0e-6 * 2147483647 rounded to nearest

        // Test values near 0.5
        let half_f32: f32 = 0.5;
        let half_f32_to_i16: i16 = half_f32.convert_to();
        assert_eq!(half_f32_to_i16, 16383); // 0.5 * 32767 = 16383.5, truncated to 16383

        let neg_half_f32: f32 = -0.5;
        let neg_half_f32_to_i16: i16 = neg_half_f32.convert_to();
        assert_eq!(neg_half_f32_to_i16, -16383); // -0.5 * 32767 = -16383.5, truncated toward zero
    }

    // Edge cases for f64 conversions
    #[test]
    fn f64_edge_cases() {
        // Test -1.0 (minimum valid value)
        let min_f64: f64 = -1.0;
        let min_f64_to_i16: i16 = min_f64.convert_to();

        println!("DEBUG: f64 -> i16 conversion for -1.0");
        println!(
            "Input: {}, Output: {}, Expected: {} or {}",
            min_f64,
            min_f64_to_i16,
            i16::MIN,
            -32767
        );

        // Due to rounding in the implementation, sometimes -1.0 can convert to -32767
        // This is acceptable since it's only 1 bit off from the true min
        assert!(
            min_f64_to_i16 == i16::MIN || min_f64_to_i16 == -32767,
            "Expected either -32768 or -32767, got {}",
            min_f64_to_i16
        );

        let min_f64_to_i32: i32 = min_f64.convert_to();

        println!("DEBUG: f64 -> i32 conversion for -1.0");
        println!(
            "Input: {}, Output: {}, Expected: {} or {}",
            min_f64,
            min_f64_to_i32,
            i32::MIN,
            -2147483647
        );

        assert!(
            min_f64_to_i32 == i32::MIN || min_f64_to_i32 == -2147483647,
            "Expected either i32::MIN or -2147483647, got {}",
            min_f64_to_i32
        );

        let min_f64_to_f32: f32 = min_f64.convert_to();

        println!("DEBUG: f64 -> f32 conversion for -1.0");
        println!(
            "Input: {}, Output: {}, Expected: -1.0",
            min_f64, min_f64_to_f32
        );

        assert_approx_eq!(min_f64_to_f32 as f64, -1.0, 1e-6);

        // Test 1.0 (maximum valid value)
        let max_f64: f64 = 1.0;
        let max_f64_to_i16: i16 = max_f64.convert_to();
        assert_eq!(max_f64_to_i16, i16::MAX);

        let max_f64_to_i32: i32 = max_f64.convert_to();
        assert_eq!(max_f64_to_i32, i32::MAX);

        let max_f64_to_f32: f32 = max_f64.convert_to();
        assert_approx_eq!(max_f64_to_f32 as f64, 1.0, 1e-6);

        // Test 0.0
        let zero_f64: f64 = 0.0;
        let zero_f64_to_i16: i16 = zero_f64.convert_to();
        assert_eq!(zero_f64_to_i16, 0);

        let zero_f64_to_i32: i32 = zero_f64.convert_to();
        assert_eq!(zero_f64_to_i32, 0);

        let zero_f64_to_f32: f32 = zero_f64.convert_to();
        assert_approx_eq!(zero_f64_to_f32 as f64, 0.0, 1e-10);

        // Test clamping of out-of-range values
        let large_f64: f64 = 2.0;
        let large_f64_to_i16: i16 = large_f64.convert_to();
        assert_eq!(large_f64_to_i16, i16::MAX);

        let neg_large_f64: f64 = -2.0;
        let neg_large_f64_to_i16: i16 = neg_large_f64.convert_to();
        assert!(
            neg_large_f64_to_i16 == i16::MIN || neg_large_f64_to_i16 == -32767,
            "Expected either -32768 or -32767, got {}",
            neg_large_f64_to_i16
        );

        // Test very small values
        let tiny_value: f64 = 1.0e-12;
        let tiny_value_to_i16: i16 = tiny_value.convert_to();
        assert_eq!(tiny_value_to_i16, 0);

        let tiny_value_to_i32: i32 = tiny_value.convert_to();
        assert_eq!(tiny_value_to_i32, 0);

        let tiny_value_to_f32: f32 = tiny_value.convert_to();
        assert_approx_eq!(tiny_value_to_f32 as f64, 0.0, 1e-10);
    }

    // Tests for I24 conversions
    #[test]
    fn i24_conversion_tests() {
        // Create an I24 with a known value
        let i24_value = i24!(4660 << 8); //  So converting back to i16 gives 4660
        println!(
            "DEBUG: Created I24 value from 4660 << 8 = {}",
            i24_value.to_i32()
        );

        // Test I24 to i16
        let i24_to_i16: i16 = i24_value.convert_to();
        let expected_i16 = 0x1234_i16;
        println!("DEBUG: I24 -> i16 conversion");
        println!(
            "I24 (as i32): {}, i16: {}, Expected: {}",
            i24_value.to_i32(),
            i24_to_i16,
            expected_i16
        );
        assert_eq!(i24_to_i16, expected_i16);

        // Test I24 to f32
        let i24_to_f32: f32 = i24_value.convert_to();
        let expected_f32 = (0x123456 as f32) / (I24::MAX.to_i32() as f32);
        println!("DEBUG: I24 -> f32 conversion");
        println!(
            "I24 (as i32): {}, f32: {}, Expected: {}",
            i24_value.to_i32(),
            i24_to_f32,
            expected_f32
        );
        // Print the difference to help debug
        println!("DEBUG: Difference: {}", (i24_to_f32 - expected_f32).abs());
        assert_approx_eq!(i24_to_f32 as f64, expected_f32 as f64, 1e-4);

        // Test I24 to f64
        let i24_to_f64: f64 = i24_value.convert_to();
        let expected_f64 = (0x123456 as f64) / (I24::MAX.to_i32() as f64);
        println!("DEBUG: I24 -> f64 conversion");
        println!(
            "I24 (as i32): {}, f64: {}, Expected: {}",
            i24_value.to_i32(),
            i24_to_f64,
            expected_f64
        );
        // Print the difference to help debug
        println!("DEBUG: Difference: {}", (i24_to_f64 - expected_f64).abs());
        assert_approx_eq!(i24_to_f64, expected_f64, 1e-4);
    }

    // Tests for convert_from functionality
    #[test]
    fn convert_from_tests() {
        // Test i16::convert_from with different source types
        let f32_source: f32 = 0.5;
        let i16_result: i16 = i16::convert_from(f32_source);
        assert_eq!(i16_result, 16383); // 0.5 * 32767 = 16383.5, truncated to 16383

        let i32_source: i32 = 65536;
        let i16_result: i16 = i16::convert_from(i32_source);
        assert_eq!(i16_result, 1); // 65536 >> 16 = 1

        // Test f32::convert_from with different source types
        let i16_source: i16 = 16384;
        let f32_result: f32 = f32::convert_from(i16_source);
        assert_approx_eq!(f32_result as f64, 0.5, 1e-4);

        let i32_source: i32 = i32::MAX / 2;
        let f32_result: f32 = f32::convert_from(i32_source);
        assert_approx_eq!(f32_result as f64, 0.5, 1e-4);

        // Test I24::convert_from
        let i16_source: i16 = 4660; // 0x1234
        let i24_result: I24 = I24::convert_from(i16_source);
        assert_eq!(i24_result.to_i32(), 4660 << 8); // Should be shifted left by 8 bits

        // Test with zero values
        let zero_f32: f32 = 0.0;
        let zero_i16: i16 = i16::convert_from(zero_f32);
        assert_eq!(zero_i16, 0);

        let zero_i16_source: i16 = 0;
        let zero_f32_result: f32 = f32::convert_from(zero_i16_source);
        assert_approx_eq!(zero_f32_result as f64, 0.0, 1e-10);
    }

    // Tests for round trip conversions
    #[test]
    fn round_trip_conversions() {
        // i16 -> f32 -> i16
        for sample in [-32768, -16384, 0, 16384, 32767].iter() {
            let original = *sample;
            let intermediate: f32 = original.convert_to();
            let round_tripped: i16 = intermediate.convert_to();

            println!("DEBUG: i16->f32->i16 conversion");
            println!(
                "Original i16: {}, f32: {}, Round trip i16: {}",
                original, intermediate, round_tripped
            );

            assert!(
                (original - round_tripped).abs() <= 1,
                "Expected {}, got {}",
                original,
                round_tripped
            );
        }

        // i32 -> f32 -> i32 (will lose precision)
        for &sample in &[i32::MIN, i32::MIN / 2, 0, i32::MAX / 2, i32::MAX] {
            let original = sample;
            let intermediate: f32 = original.convert_to();
            let round_tripped: i32 = intermediate.convert_to();

            // Special case for extreme values
            if original == i32::MIN {
                // Allow off-by-one for MIN value
                assert!(
                    round_tripped == i32::MIN || round_tripped == -2147483647,
                    "Expected either i32::MIN or -2147483647, got {}",
                    round_tripped
                );
            } else if original == i32::MAX || original == 0 {
                assert_eq!(
                    original, round_tripped,
                    "Failed in i32->f32->i32 with extreme value {}",
                    original
                );
            } else {
                // For other values, we expect close but not exact due to precision
                let ratio = (round_tripped as f64) / (original as f64);
                assert!(
                    ratio > 0.999 && ratio < 1.001,
                    "Round trip error too large: {} -> {}",
                    original,
                    round_tripped
                );
            }
        }

        // f32 -> i16 -> f32
        for &sample in &[-1.0, -0.5, 0.0, 0.5, 1.0] {
            let original: f32 = sample;
            let intermediate: i16 = original.convert_to();
            let round_tripped: f32 = intermediate.convert_to();

            // For all values, we check approximately but with a more generous epsilon
            assert_approx_eq!(original as f64, round_tripped as f64, 1e-4);
        }

        // i16 -> I24 -> i16
        for &sample in &[i16::MIN, -16384, 0, 16384, i16::MAX] {
            let original = sample;
            let intermediate: I24 = original.convert_to();
            let round_tripped: i16 = intermediate.convert_to();

            // For extreme negative values, allow 1-bit difference
            if original == i16::MIN {
                assert!(
                    round_tripped == i16::MIN || round_tripped == -32767,
                    "Expected either -32768 or -32767, got {}",
                    round_tripped
                );
            } else {
                assert_eq!(
                    original, round_tripped,
                    "Failed in i16->I24->i16 with value {}",
                    original
                );
            }
        }
    }
}