unsigned-float 0.3.0

//! Bulk SIMD operations for the unsigned floating-point storage formats.
//!
//! The scalar newtypes intentionally retain their precise, portable
//! promote/compute/demote contract. SIMD is therefore exposed as an opt-in
//! slice API: storage lanes are widened into native `f32` (`Uf16`) or `f64`
//! (`Uf32`) vectors, computed there, and rounded back to the exact scalar
//! encoding. Normal finite lanes take the register-only bit-expansion path;
//! subnormals, infinities, NaNs, overflows, and negative results fall back to
//! the scalar conversion for bit-for-bit compatibility.
//!
//! This feature uses nightly `portable_simd`. The chosen lane widths follow
//! the portable `axnn-cpu` tier: 128-bit F32/F64 vectors by default and
//! 256-bit vectors when the crate is compiled with AVX2 or AVX-512 enabled.

use core::fmt;
use core::simd::prelude::*;

use crate::{Uf16, Uf16E5M11, Uf16E6M10, Uf32};

/// F32 lanes processed by the UF16 path in one vector operation.
#[cfg(any(target_feature = "avx2", target_feature = "avx512f"))]
pub const UF16_LANES: usize = 8;
/// F32 lanes processed by the UF16 path in one vector operation.
#[cfg(not(any(target_feature = "avx2", target_feature = "avx512f")))]
pub const UF16_LANES: usize = 4;

/// F64 lanes processed by the UF32 path in one vector operation.
#[cfg(any(target_feature = "avx2", target_feature = "avx512f"))]
pub const UF32_LANES: usize = 4;
/// F64 lanes processed by the UF32 path in one vector operation.
#[cfg(not(any(target_feature = "avx2", target_feature = "avx512f")))]
pub const UF32_LANES: usize = 2;

/// Invalid source/destination slice relationship passed to a bulk operation.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum SimdError {
    /// Two input planes that must be elementwise aligned have different sizes.
    InputLengthMismatch { left: usize, right: usize },
    /// The output plane does not have one element for each input element.
    OutputLengthMismatch { input: usize, output: usize },
}

impl fmt::Display for SimdError {
    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            Self::InputLengthMismatch { left, right } => {
                write!(formatter, "SIMD input lengths differ ({left} and {right})")
            }
            Self::OutputLengthMismatch { input, output } => {
                write!(
                    formatter,
                    "SIMD output length is {output}, expected {input}"
                )
            }
        }
    }
}

trait Uf16Layout: Copy {
    const EXPONENT_BITS: u32;
    const MANTISSA_BITS: u32;
    /// Native F32 exponent corresponding to an encoded exponent of zero.
    const F32_EXPONENT_BIAS: u32;

    fn from_bits(bits: u16) -> Self;
    fn to_bits(self) -> u16;
    fn from_f32(value: f32) -> Self;
    fn to_f32(self) -> f32;
}

impl Uf16Layout for Uf16E5M11 {
    const EXPONENT_BITS: u32 = 5;
    const MANTISSA_BITS: u32 = 11;
    const F32_EXPONENT_BIAS: u32 = 112;

    fn from_bits(bits: u16) -> Self {
        Self::from_bits(bits)
    }

    fn to_bits(self) -> u16 {
        self.to_bits()
    }

    fn from_f32(value: f32) -> Self {
        Self::from_f32(value)
    }

    fn to_f32(self) -> f32 {
        self.to_f32()
    }
}

impl Uf16Layout for Uf16E6M10 {
    const EXPONENT_BITS: u32 = 6;
    const MANTISSA_BITS: u32 = 10;
    const F32_EXPONENT_BIAS: u32 = 96;

    fn from_bits(bits: u16) -> Self {
        Self::from_bits(bits)
    }

    fn to_bits(self) -> u16 {
        self.to_bits()
    }

    fn from_f32(value: f32) -> Self {
        Self::from_f32(value)
    }

    fn to_f32(self) -> f32 {
        self.to_f32()
    }
}

fn output_len(input: usize, output: usize) -> Result<(), SimdError> {
    if input == output {
        Ok(())
    } else {
        Err(SimdError::OutputLengthMismatch { input, output })
    }
}

fn binary_len(left: usize, right: usize, output: usize) -> Result<(), SimdError> {
    if left != right {
        return Err(SimdError::InputLengthMismatch { left, right });
    }
    output_len(left, output)
}

fn uf16_max_exponent<T: Uf16Layout>() -> u32 {
    (1 << T::EXPONENT_BITS) - 1
}

fn can_decode_uf16<T: Uf16Layout>(value: T) -> bool {
    let exponent = (value.to_bits() as u32 >> T::MANTISSA_BITS) & uf16_max_exponent::<T>();
    exponent != 0 && exponent != uf16_max_exponent::<T>()
}

fn can_encode_uf16<T: Uf16Layout>(value: f32) -> bool {
    let bits = value.to_bits();
    let exponent = (bits >> 23) & 0xff;
    let max_normal = uf16_max_exponent::<T>() - 1;
    bits >> 31 == 0
        && exponent > T::F32_EXPONENT_BIAS
        // Keep the largest normal bin scalar: a round carry there can become
        // infinity, while every lane accepted here stays normal after RNE.
        && exponent < T::F32_EXPONENT_BIAS + max_normal
}

fn decode_uf16_fast<T: Uf16Layout>(src: &[T]) -> Simd<f32, UF16_LANES> {
    debug_assert_eq!(src.len(), UF16_LANES);
    debug_assert!(src.iter().copied().all(can_decode_uf16::<T>));
    let raw = Simd::<u32, UF16_LANES>::from_array(core::array::from_fn(|lane| {
        src[lane].to_bits() as u32
    }));
    let bits =
        (raw << Simd::splat(23 - T::MANTISSA_BITS)) + Simd::splat(T::F32_EXPONENT_BIAS << 23);
    Simd::<f32, UF16_LANES>::from_bits(bits)
}

fn encode_uf16_fast<T: Uf16Layout>(src: Simd<f32, UF16_LANES>, dst: &mut [T]) {
    debug_assert_eq!(dst.len(), UF16_LANES);
    debug_assert!(src.to_array().into_iter().all(can_encode_uf16::<T>));
    let bits = src.to_bits();
    let fraction = bits & Simd::splat(0x007f_ffff_u32);
    let drop = 23 - T::MANTISSA_BITS;
    let mantissa = fraction >> Simd::splat(drop);
    let discarded = fraction & Simd::splat((1_u32 << drop) - 1);
    // Round-to-nearest-even without a per-lane comparison.
    let rounding =
        (discarded + Simd::splat((1_u32 << (drop - 1)) - 1) + (mantissa & Simd::splat(1)))
            >> Simd::splat(drop);
    let rounded = mantissa + rounding;
    let carry = rounded >> Simd::splat(T::MANTISSA_BITS);
    let exponent = (bits >> Simd::splat(23)) - Simd::splat(T::F32_EXPONENT_BIAS) + carry;
    let raw = (exponent << Simd::splat(T::MANTISSA_BITS))
        | (rounded & Simd::splat((1_u32 << T::MANTISSA_BITS) - 1));
    for (lane, bits) in raw.to_array().into_iter().enumerate() {
        dst[lane] = T::from_bits(bits as u16);
    }
}

fn decode_uf16<T: Uf16Layout>(src: &[T], dst: &mut [f32]) -> Result<(), SimdError> {
    output_len(src.len(), dst.len())?;
    let vector_end = src.len() / UF16_LANES * UF16_LANES;
    for offset in (0..vector_end).step_by(UF16_LANES) {
        let input = &src[offset..offset + UF16_LANES];
        if input.iter().copied().all(can_decode_uf16::<T>) {
            decode_uf16_fast(input).copy_to_slice(&mut dst[offset..]);
        } else {
            for lane in 0..UF16_LANES {
                dst[offset + lane] = input[lane].to_f32();
            }
        }
    }
    for index in vector_end..src.len() {
        dst[index] = src[index].to_f32();
    }
    Ok(())
}

fn encode_uf16<T: Uf16Layout>(src: &[f32], dst: &mut [T]) -> Result<(), SimdError> {
    output_len(src.len(), dst.len())?;
    let vector_end = src.len() / UF16_LANES * UF16_LANES;
    for offset in (0..vector_end).step_by(UF16_LANES) {
        let input = Simd::<f32, UF16_LANES>::from_slice(&src[offset..]);
        if input.to_array().into_iter().all(can_encode_uf16::<T>) {
            encode_uf16_fast(input, &mut dst[offset..offset + UF16_LANES]);
        } else {
            for lane in 0..UF16_LANES {
                dst[offset + lane] = T::from_f32(src[offset + lane]);
            }
        }
    }
    for index in vector_end..src.len() {
        dst[index] = T::from_f32(src[index]);
    }
    Ok(())
}

fn binary_uf16<T: Uf16Layout>(
    left: &[T],
    right: &[T],
    output: &mut [T],
    vector: impl Fn(Simd<f32, UF16_LANES>, Simd<f32, UF16_LANES>) -> Simd<f32, UF16_LANES>,
    scalar: impl Fn(f32, f32) -> f32,
) -> Result<(), SimdError> {
    binary_len(left.len(), right.len(), output.len())?;
    let vector_end = left.len() / UF16_LANES * UF16_LANES;
    for offset in (0..vector_end).step_by(UF16_LANES) {
        let lhs = &left[offset..offset + UF16_LANES];
        let rhs = &right[offset..offset + UF16_LANES];
        if lhs.iter().copied().all(can_decode_uf16::<T>)
            && rhs.iter().copied().all(can_decode_uf16::<T>)
        {
            let result = vector(decode_uf16_fast(lhs), decode_uf16_fast(rhs));
            if result.to_array().into_iter().all(can_encode_uf16::<T>) {
                encode_uf16_fast(result, &mut output[offset..offset + UF16_LANES]);
                continue;
            }
        }
        for lane in 0..UF16_LANES {
            output[offset + lane] = T::from_f32(scalar(lhs[lane].to_f32(), rhs[lane].to_f32()));
        }
    }
    for index in vector_end..left.len() {
        output[index] = T::from_f32(scalar(left[index].to_f32(), right[index].to_f32()));
    }
    Ok(())
}

/// Decode packed [`Uf16`] lanes to F32. Normal finite blocks use SIMD bit expansion.
pub fn decode_uf16_to_f32(src: &[Uf16], dst: &mut [f32]) -> Result<(), SimdError> {
    decode_uf16(src, dst)
}

/// Encode F32 lanes as [`Uf16`] with the scalar constructor's exact RNE behavior.
pub fn encode_f32_to_uf16(src: &[f32], dst: &mut [Uf16]) -> Result<(), SimdError> {
    encode_uf16(src, dst)
}

/// Decode packed [`Uf16E6M10`] lanes to F32.
pub fn decode_uf16e6m10_to_f32(src: &[Uf16E6M10], dst: &mut [f32]) -> Result<(), SimdError> {
    decode_uf16(src, dst)
}

/// Encode F32 lanes as [`Uf16E6M10`] with scalar-equivalent rounding.
pub fn encode_f32_to_uf16e6m10(src: &[f32], dst: &mut [Uf16E6M10]) -> Result<(), SimdError> {
    encode_uf16(src, dst)
}

/// Elementwise [`Uf16`] addition through native F32 SIMD lanes.
pub fn add_uf16(left: &[Uf16], right: &[Uf16], output: &mut [Uf16]) -> Result<(), SimdError> {
    binary_uf16(
        left,
        right,
        output,
        |left, right| left + right,
        |left, right| left + right,
    )
}

/// Elementwise [`Uf16`] subtraction through native F32 SIMD lanes.
pub fn sub_uf16(left: &[Uf16], right: &[Uf16], output: &mut [Uf16]) -> Result<(), SimdError> {
    binary_uf16(
        left,
        right,
        output,
        |left, right| left - right,
        |left, right| left - right,
    )
}

/// Elementwise [`Uf16`] multiplication through native F32 SIMD lanes.
pub fn mul_uf16(left: &[Uf16], right: &[Uf16], output: &mut [Uf16]) -> Result<(), SimdError> {
    binary_uf16(
        left,
        right,
        output,
        |left, right| left * right,
        |left, right| left * right,
    )
}

/// Elementwise [`Uf16`] division through native F32 SIMD lanes.
pub fn div_uf16(left: &[Uf16], right: &[Uf16], output: &mut [Uf16]) -> Result<(), SimdError> {
    binary_uf16(
        left,
        right,
        output,
        |left, right| left / right,
        |left, right| left / right,
    )
}

/// Elementwise [`Uf16E6M10`] addition through native F32 SIMD lanes.
pub fn add_uf16e6m10(
    left: &[Uf16E6M10],
    right: &[Uf16E6M10],
    output: &mut [Uf16E6M10],
) -> Result<(), SimdError> {
    binary_uf16(
        left,
        right,
        output,
        |left, right| left + right,
        |left, right| left + right,
    )
}

/// Elementwise [`Uf16E6M10`] subtraction through native F32 SIMD lanes.
pub fn sub_uf16e6m10(
    left: &[Uf16E6M10],
    right: &[Uf16E6M10],
    output: &mut [Uf16E6M10],
) -> Result<(), SimdError> {
    binary_uf16(
        left,
        right,
        output,
        |left, right| left - right,
        |left, right| left - right,
    )
}

/// Elementwise [`Uf16E6M10`] multiplication through native F32 SIMD lanes.
pub fn mul_uf16e6m10(
    left: &[Uf16E6M10],
    right: &[Uf16E6M10],
    output: &mut [Uf16E6M10],
) -> Result<(), SimdError> {
    binary_uf16(
        left,
        right,
        output,
        |left, right| left * right,
        |left, right| left * right,
    )
}

/// Elementwise [`Uf16E6M10`] division through native F32 SIMD lanes.
pub fn div_uf16e6m10(
    left: &[Uf16E6M10],
    right: &[Uf16E6M10],
    output: &mut [Uf16E6M10],
) -> Result<(), SimdError> {
    binary_uf16(
        left,
        right,
        output,
        |left, right| left / right,
        |left, right| left / right,
    )
}

fn can_decode_uf32(value: Uf32) -> bool {
    let exponent = value.to_bits() >> 24;
    exponent != 0 && exponent != 0xff
}

fn can_encode_uf32(value: f64) -> bool {
    let bits = value.to_bits();
    let exponent = (bits >> 52) & 0x7ff;
    // An encoded exponent of 254 is kept scalar because rounding may carry to infinity.
    bits >> 63 == 0 && exponent > 896 && exponent < 1150
}

fn decode_uf32_fast(src: &[Uf32]) -> Simd<f64, UF32_LANES> {
    debug_assert_eq!(src.len(), UF32_LANES);
    debug_assert!(src.iter().copied().all(can_decode_uf32));
    let raw = Simd::<u64, UF32_LANES>::from_array(core::array::from_fn(|lane| {
        src[lane].to_bits() as u64
    }));
    let bits = (raw << Simd::splat(28)) + Simd::splat(896_u64 << 52);
    Simd::<f64, UF32_LANES>::from_bits(bits)
}

fn encode_uf32_fast(src: Simd<f64, UF32_LANES>, dst: &mut [Uf32]) {
    debug_assert_eq!(dst.len(), UF32_LANES);
    debug_assert!(src.to_array().into_iter().all(can_encode_uf32));
    let bits = src.to_bits();
    let fraction = bits & Simd::splat(0x000f_ffff_ffff_ffff_u64);
    let mantissa = fraction >> Simd::splat(28);
    let discarded = fraction & Simd::splat((1_u64 << 28) - 1);
    let rounding = (discarded + Simd::splat((1_u64 << 27) - 1) + (mantissa & Simd::splat(1)))
        >> Simd::splat(28);
    let rounded = mantissa + rounding;
    let carry = rounded >> Simd::splat(24);
    let exponent = (bits >> Simd::splat(52)) - Simd::splat(896_u64) + carry;
    let raw = (exponent << Simd::splat(24)) | (rounded & Simd::splat(0x00ff_ffff_u64));
    for (lane, bits) in raw.to_array().into_iter().enumerate() {
        dst[lane] = Uf32::from_bits(bits as u32);
    }
}

/// Decode packed [`Uf32`] lanes to F64. Normal finite blocks use SIMD bit expansion.
pub fn decode_uf32_to_f64(src: &[Uf32], dst: &mut [f64]) -> Result<(), SimdError> {
    output_len(src.len(), dst.len())?;
    let vector_end = src.len() / UF32_LANES * UF32_LANES;
    for offset in (0..vector_end).step_by(UF32_LANES) {
        let input = &src[offset..offset + UF32_LANES];
        if input.iter().copied().all(can_decode_uf32) {
            decode_uf32_fast(input).copy_to_slice(&mut dst[offset..]);
        } else {
            for lane in 0..UF32_LANES {
                dst[offset + lane] = input[lane].to_f64();
            }
        }
    }
    for index in vector_end..src.len() {
        dst[index] = src[index].to_f64();
    }
    Ok(())
}

/// Encode F64 lanes as [`Uf32`] with the scalar constructor's exact RNE behavior.
pub fn encode_f64_to_uf32(src: &[f64], dst: &mut [Uf32]) -> Result<(), SimdError> {
    output_len(src.len(), dst.len())?;
    let vector_end = src.len() / UF32_LANES * UF32_LANES;
    for offset in (0..vector_end).step_by(UF32_LANES) {
        let input = Simd::<f64, UF32_LANES>::from_slice(&src[offset..]);
        if input.to_array().into_iter().all(can_encode_uf32) {
            encode_uf32_fast(input, &mut dst[offset..offset + UF32_LANES]);
        } else {
            for lane in 0..UF32_LANES {
                dst[offset + lane] = Uf32::from_f64(src[offset + lane]);
            }
        }
    }
    for index in vector_end..src.len() {
        dst[index] = Uf32::from_f64(src[index]);
    }
    Ok(())
}

fn binary_uf32(
    left: &[Uf32],
    right: &[Uf32],
    output: &mut [Uf32],
    vector: impl Fn(Simd<f64, UF32_LANES>, Simd<f64, UF32_LANES>) -> Simd<f64, UF32_LANES>,
    scalar: impl Fn(f64, f64) -> f64,
) -> Result<(), SimdError> {
    binary_len(left.len(), right.len(), output.len())?;
    let vector_end = left.len() / UF32_LANES * UF32_LANES;
    for offset in (0..vector_end).step_by(UF32_LANES) {
        let lhs = &left[offset..offset + UF32_LANES];
        let rhs = &right[offset..offset + UF32_LANES];
        if lhs.iter().copied().all(can_decode_uf32) && rhs.iter().copied().all(can_decode_uf32) {
            let result = vector(decode_uf32_fast(lhs), decode_uf32_fast(rhs));
            if result.to_array().into_iter().all(can_encode_uf32) {
                encode_uf32_fast(result, &mut output[offset..offset + UF32_LANES]);
                continue;
            }
        }
        for lane in 0..UF32_LANES {
            output[offset + lane] = Uf32::from_f64(scalar(lhs[lane].to_f64(), rhs[lane].to_f64()));
        }
    }
    for index in vector_end..left.len() {
        output[index] = Uf32::from_f64(scalar(left[index].to_f64(), right[index].to_f64()));
    }
    Ok(())
}

/// Elementwise [`Uf32`] addition through native F64 SIMD lanes.
pub fn add_uf32(left: &[Uf32], right: &[Uf32], output: &mut [Uf32]) -> Result<(), SimdError> {
    binary_uf32(
        left,
        right,
        output,
        |left, right| left + right,
        |left, right| left + right,
    )
}

/// Elementwise [`Uf32`] subtraction through native F64 SIMD lanes.
pub fn sub_uf32(left: &[Uf32], right: &[Uf32], output: &mut [Uf32]) -> Result<(), SimdError> {
    binary_uf32(
        left,
        right,
        output,
        |left, right| left - right,
        |left, right| left - right,
    )
}

/// Elementwise [`Uf32`] multiplication through native F64 SIMD lanes.
pub fn mul_uf32(left: &[Uf32], right: &[Uf32], output: &mut [Uf32]) -> Result<(), SimdError> {
    binary_uf32(
        left,
        right,
        output,
        |left, right| left * right,
        |left, right| left * right,
    )
}

/// Elementwise [`Uf32`] division through native F64 SIMD lanes.
pub fn div_uf32(left: &[Uf32], right: &[Uf32], output: &mut [Uf32]) -> Result<(), SimdError> {
    binary_uf32(
        left,
        right,
        output,
        |left, right| left / right,
        |left, right| left / right,
    )
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::vec;
    use std::vec::Vec;

    fn lcg(state: &mut u64) -> u64 {
        *state = state
            .wrapping_mul(6_364_136_223_846_793_005)
            .wrapping_add(1_442_695_040_888_963_407);
        *state
    }

    fn verify_uf16_conversions<T: Uf16Layout>() {
        let source: Vec<T> = (u16::MIN..=u16::MAX).map(T::from_bits).collect();
        let mut decoded = vec![0.0; source.len()];
        decode_uf16(&source, &mut decoded).unwrap();
        for (value, actual) in source.iter().copied().zip(decoded) {
            assert_eq!(actual.to_bits(), value.to_f32().to_bits());
        }

        let mut state = 0x6f_1d_5eed_u64;
        let mut input = vec![0.0; 32_771];
        input[0] = 0.0;
        input[1] = -0.0;
        input[2] = f32::INFINITY;
        input[3] = f32::NEG_INFINITY;
        input[4] = f32::NAN;
        for value in &mut input[5..] {
            *value = f32::from_bits(lcg(&mut state) as u32);
        }
        let mut encoded = vec![T::from_bits(0); input.len()];
        encode_uf16(&input, &mut encoded).unwrap();
        for (value, actual) in input.into_iter().zip(encoded) {
            assert_eq!(actual.to_bits(), T::from_f32(value).to_bits());
        }
    }

    fn verify_uf16_binary<T: Uf16Layout>() {
        let mut state = 0x9a_6d_ef_41_u64;
        let left: Vec<T> = (0..(UF16_LANES * 19 + 3))
            .map(|_| T::from_bits(lcg(&mut state) as u16))
            .collect();
        let right: Vec<T> = (0..left.len())
            .map(|_| T::from_bits(lcg(&mut state) as u16))
            .collect();
        let mut output = vec![T::from_bits(0); left.len()];

        binary_uf16(
            &left,
            &right,
            &mut output,
            |left, right| left + right,
            |left, right| left + right,
        )
        .unwrap();
        for ((left, right), actual) in left.iter().zip(&right).zip(&output) {
            assert_eq!(
                actual.to_bits(),
                T::from_f32(left.to_f32() + right.to_f32()).to_bits()
            );
        }

        binary_uf16(
            &left,
            &right,
            &mut output,
            |left, right| left - right,
            |left, right| left - right,
        )
        .unwrap();
        for ((left, right), actual) in left.iter().zip(&right).zip(&output) {
            assert_eq!(
                actual.to_bits(),
                T::from_f32(left.to_f32() - right.to_f32()).to_bits()
            );
        }

        binary_uf16(
            &left,
            &right,
            &mut output,
            |left, right| left * right,
            |left, right| left * right,
        )
        .unwrap();
        for ((left, right), actual) in left.iter().zip(&right).zip(&output) {
            assert_eq!(
                actual.to_bits(),
                T::from_f32(left.to_f32() * right.to_f32()).to_bits()
            );
        }

        binary_uf16(
            &left,
            &right,
            &mut output,
            |left, right| left / right,
            |left, right| left / right,
        )
        .unwrap();
        for ((left, right), actual) in left.iter().zip(&right).zip(&output) {
            assert_eq!(
                actual.to_bits(),
                T::from_f32(left.to_f32() / right.to_f32()).to_bits()
            );
        }
    }

    #[test]
    fn uf16e5m11_bulk_paths_are_bit_exact() {
        verify_uf16_conversions::<Uf16>();
        verify_uf16_binary::<Uf16>();
    }

    #[test]
    fn uf16e6m10_bulk_paths_are_bit_exact() {
        verify_uf16_conversions::<Uf16E6M10>();
        verify_uf16_binary::<Uf16E6M10>();
    }

    #[test]
    fn uf32_bulk_paths_are_bit_exact() {
        let mut state = 0x03_2d_99_ef_u64;
        let mut source = vec![Uf32::ZERO, Uf32::MIN_POSITIVE, Uf32::INFINITY, Uf32::NAN];
        source.extend((0..32_767).map(|_| Uf32::from_bits(lcg(&mut state) as u32)));
        let mut decoded = vec![0.0; source.len()];
        decode_uf32_to_f64(&source, &mut decoded).unwrap();
        for (value, actual) in source.iter().copied().zip(decoded) {
            assert_eq!(actual.to_bits(), value.to_f64().to_bits());
        }

        let mut encoded_input = vec![0.0; 32_771];
        encoded_input[0] = 0.0;
        encoded_input[1] = -0.0;
        encoded_input[2] = f64::INFINITY;
        encoded_input[3] = f64::NAN;
        for value in &mut encoded_input[4..] {
            *value = f64::from_bits(lcg(&mut state));
        }
        let mut encoded = vec![Uf32::ZERO; encoded_input.len()];
        encode_f64_to_uf32(&encoded_input, &mut encoded).unwrap();
        for (value, actual) in encoded_input.into_iter().zip(encoded) {
            assert_eq!(actual.to_bits(), Uf32::from_f64(value).to_bits());
        }

        let left: Vec<Uf32> = (0..(UF32_LANES * 19 + 1))
            .map(|_| Uf32::from_bits(lcg(&mut state) as u32))
            .collect();
        let right: Vec<Uf32> = (0..left.len())
            .map(|_| Uf32::from_bits(lcg(&mut state) as u32))
            .collect();
        let mut output = vec![Uf32::ZERO; left.len()];
        macro_rules! assert_uf32_binary {
            ($vector:expr, $scalar:expr) => {{
                binary_uf32(&left, &right, &mut output, $vector, $scalar).unwrap();
                for ((left, right), actual) in left.iter().zip(&right).zip(&output) {
                    assert_eq!(
                        actual.to_bits(),
                        Uf32::from_f64($scalar(left.to_f64(), right.to_f64())).to_bits()
                    );
                }
            }};
        }
        assert_uf32_binary!(|left, right| left + right, |left: f64, right: f64| left
            + right);
        assert_uf32_binary!(|left, right| left - right, |left: f64, right: f64| left
            - right);
        assert_uf32_binary!(|left, right| left * right, |left: f64, right: f64| left
            * right);
        assert_uf32_binary!(|left, right| left / right, |left: f64, right: f64| left
            / right);
    }

    #[test]
    fn bulk_operations_reject_mismatched_planes() {
        let input = [Uf16::ONE; 2];
        let other = [Uf16::ONE; 1];
        let mut output = [Uf16::ZERO; 2];
        assert_eq!(
            add_uf16(&input, &other, &mut output),
            Err(SimdError::InputLengthMismatch { left: 2, right: 1 })
        );
        let mut short = [0.0; 1];
        assert_eq!(
            decode_uf16_to_f32(&input, &mut short),
            Err(SimdError::OutputLengthMismatch {
                input: 2,
                output: 1
            })
        );
    }
}