vector_ta/indicators/moving_averages/

gaussian.rs

#[cfg(all(feature = "python", feature = "cuda"))]
use crate::cuda::cuda_available;
#[cfg(all(feature = "python", feature = "cuda"))]
use crate::cuda::moving_averages::gaussian_wrapper::DeviceArrayF32Py;
#[cfg(all(feature = "python", feature = "cuda"))]
use crate::cuda::moving_averages::CudaGaussian;
use crate::utilities::data_loader::{source_type, Candles};
use crate::utilities::enums::Kernel;
use crate::utilities::helpers::{
    alloc_with_nan_prefix, detect_best_batch_kernel, detect_best_kernel, init_matrix_prefixes,
    make_uninit_matrix,
};
#[cfg(not(target_arch = "wasm32"))]
use rayon::prelude::*;
#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
use std::arch::x86_64::*;
use std::f64::consts::PI;
use std::mem::MaybeUninit;
use thiserror::Error;

const LANES_AVX512: usize = 8;
const LANES_AVX2: usize = 4;

impl<'a> AsRef<[f64]> for GaussianInput<'a> {
    #[inline(always)]
    fn as_ref(&self) -> &[f64] {
        match &self.data {
            GaussianData::Slice(slice) => slice,
            GaussianData::Candles { candles, source } => source_type(candles, source),
        }
    }
}

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

    #[cfg(not(all(target_arch = "wasm32", feature = "wasm")))]
    #[test]
    fn test_gaussian_into_matches_api() -> Result<(), Box<dyn std::error::Error>> {
        let mut data = Vec::with_capacity(256);

        for _ in 0..8 {
            data.push(f64::NAN);
        }

        for i in 0..(256 - 8) {
            let x = i as f64;
            data.push(x * 0.1 + (x * 0.07).sin());
        }

        let input = GaussianInput::from_slice(&data, GaussianParams::default());

        let base = gaussian(&input)?.values;

        let mut out = vec![0.0f64; data.len()];
        gaussian_into(&input, &mut out)?;

        assert_eq!(base.len(), out.len());

        fn eq_or_both_nan(a: f64, b: f64) -> bool {
            (a.is_nan() && b.is_nan()) || (a - b).abs() <= 1e-12
        }
        for i in 0..base.len() {
            assert!(
                eq_or_both_nan(base[i], out[i]),
                "mismatch at idx {}: base={}, out={}",
                i,
                base[i],
                out[i]
            );
        }
        Ok(())
    }
}

#[derive(Debug, Clone)]
pub enum GaussianData<'a> {
    Candles {
        candles: &'a Candles,
        source: &'a str,
    },
    Slice(&'a [f64]),
}

#[derive(Debug, Clone)]
pub struct GaussianOutput {
    pub values: Vec<f64>,
}

#[derive(Debug, Clone, Copy)]
#[cfg_attr(
    all(target_arch = "wasm32", feature = "wasm"),
    derive(Serialize, Deserialize)
)]
pub struct GaussianParams {
    pub period: Option<usize>,
    pub poles: Option<usize>,
}

impl Default for GaussianParams {
    fn default() -> Self {
        Self {
            period: Some(14),
            poles: Some(4),
        }
    }
}

#[derive(Debug, Clone)]
pub struct GaussianInput<'a> {
    pub data: GaussianData<'a>,
    pub params: GaussianParams,
}

impl<'a> GaussianInput<'a> {
    pub fn from_candles(c: &'a Candles, s: &'a str, p: GaussianParams) -> Self {
        Self {
            data: GaussianData::Candles {
                candles: c,
                source: s,
            },
            params: p,
        }
    }
    pub fn from_slice(sl: &'a [f64], p: GaussianParams) -> Self {
        Self {
            data: GaussianData::Slice(sl),
            params: p,
        }
    }
    pub fn with_default_candles(c: &'a Candles) -> Self {
        Self::from_candles(c, "close", GaussianParams::default())
    }
    pub fn get_period(&self) -> usize {
        self.params.period.unwrap_or(14)
    }
    pub fn get_poles(&self) -> usize {
        self.params.poles.unwrap_or(4)
    }
}

#[derive(Copy, Clone, Debug)]
pub struct GaussianBuilder {
    period: Option<usize>,
    poles: Option<usize>,
    kernel: Kernel,
}

impl Default for GaussianBuilder {
    fn default() -> Self {
        Self {
            period: None,
            poles: None,
            kernel: Kernel::Auto,
        }
    }
}

impl GaussianBuilder {
    pub fn new() -> Self {
        Self::default()
    }
    pub fn period(mut self, n: usize) -> Self {
        self.period = Some(n);
        self
    }
    pub fn poles(mut self, k: usize) -> Self {
        self.poles = Some(k);
        self
    }
    pub fn kernel(mut self, k: Kernel) -> Self {
        self.kernel = k;
        self
    }
    pub fn apply(self, c: &Candles) -> Result<GaussianOutput, GaussianError> {
        let p = GaussianParams {
            period: self.period,
            poles: self.poles,
        };
        let i = GaussianInput::from_candles(c, "close", p);
        gaussian_with_kernel(&i, self.kernel)
    }
    pub fn apply_slice(self, d: &[f64]) -> Result<GaussianOutput, GaussianError> {
        let p = GaussianParams {
            period: self.period,
            poles: self.poles,
        };
        let i = GaussianInput::from_slice(d, p);
        gaussian_with_kernel(&i, self.kernel)
    }
    pub fn into_stream(self) -> Result<GaussianStream, GaussianError> {
        let p = GaussianParams {
            period: self.period,
            poles: self.poles,
        };
        GaussianStream::try_new(p)
    }
}

#[derive(Debug, Error)]
pub enum GaussianError {
    #[error("gaussian: Input data slice is empty.")]
    EmptyInputData,
    #[error("gaussian: No data provided to Gaussian filter.")]
    NoData,
    #[error("gaussian: Invalid period: period = {period}, data length = {data_len}")]
    InvalidPeriod { period: usize, data_len: usize },
    #[error("gaussian: Invalid number of poles: expected 1..4, got {poles}")]
    InvalidPoles { poles: usize },
    #[error("gaussian: Period is longer than the data. period={period}, data_len={data_len}")]
    PeriodLongerThanData { period: usize, data_len: usize },
    #[error("gaussian: All values are NaN.")]
    AllValuesNaN,
    #[error("gaussian: Not enough valid data: needed = {needed}, valid = {valid}")]
    NotEnoughValidData { needed: usize, valid: usize },
    #[error("gaussian: Period too small. Period must be >= 2 for meaningful Gaussian filtering. Got period={period}")]
    DegeneratePeriod { period: usize },
    #[error("gaussian: Period of 1 causes degenerate filter (alpha=0). This produces constant zero output. Use period >= 2.")]
    PeriodOneDegenerate,
    #[error("gaussian: output length mismatch: expected {expected}, got {got}")]
    OutputLengthMismatch { expected: usize, got: usize },
    #[error("gaussian: invalid range: start={start}, end={end}, step={step}")]
    InvalidRange {
        start: usize,
        end: usize,
        step: usize,
    },
    #[error("gaussian: invalid kernel for batch API: {0:?}")]
    InvalidKernelForBatch(Kernel),
    #[error("gaussian: size overflow (rows={rows}, cols={cols})")]
    SizeOverflow { rows: usize, cols: usize },
}

#[inline]
pub fn gaussian(input: &GaussianInput) -> Result<GaussianOutput, GaussianError> {
    gaussian_with_kernel(input, Kernel::Auto)
}

#[inline(always)]
fn has_fma() -> bool {
    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    {
        std::is_x86_feature_detected!("fma")
    }

    #[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))]
    {
        false
    }
}

#[inline(always)]
pub fn gaussian_scalar(data: &[f64], period: usize, poles: usize, out: &mut [f64]) {
    debug_assert_eq!(data.len(), out.len());

    if has_fma() {
        unsafe { gaussian_scalar_fma(data, period, poles, out) }
    } else {
        gaussian_scalar_fallback(data, period, poles, out)
    }
}

#[cfg_attr(
    any(target_arch = "x86", target_arch = "x86_64"),
    target_feature(enable = "fma")
)]
unsafe fn gaussian_scalar_fma(data: &[f64], period: usize, poles: usize, out: &mut [f64]) {
    use core::f64::consts::PI;

    let beta = {
        let num = 1.0 - (2.0 * PI / period as f64).cos();
        let den = (2.0f64).powf(1.0 / poles as f64) - 1.0;
        num / den
    };
    let alpha = {
        let tmp = beta * beta + 2.0 * beta;
        -beta + tmp.sqrt()
    };

    match poles {
        1 => gaussian_poles1_fma(data, alpha, out),
        2 => gaussian_poles2_fma(data, alpha, out),
        3 => gaussian_poles3_fma(data, alpha, out),
        4 => gaussian_poles4_fma(data, alpha, out),
        _ => core::hint::unreachable_unchecked(),
    }
}

fn gaussian_scalar_fallback(data: &[f64], period: usize, poles: usize, out: &mut [f64]) {
    use core::f64::consts::PI;

    let beta = {
        let num = 1.0 - (2.0 * PI / period as f64).cos();
        let den = (2.0f64).powf(1.0 / poles as f64) - 1.0;
        num / den
    };
    let alpha = {
        let tmp = beta * beta + 2.0 * beta;
        -beta + tmp.sqrt()
    };

    unsafe {
        match poles {
            1 => gaussian_poles1_fma(data, alpha, out),
            2 => gaussian_poles2_fma(data, alpha, out),
            3 => gaussian_poles3_fma(data, alpha, out),
            4 => gaussian_poles4_fma(data, alpha, out),
            _ => core::hint::unreachable_unchecked(),
        }
    }
}

pub fn gaussian_with_kernel(
    input: &GaussianInput,
    kernel: Kernel,
) -> Result<GaussianOutput, GaussianError> {
    let data: &[f64] = match &input.data {
        GaussianData::Candles { candles, source } => source_type(candles, source),
        GaussianData::Slice(sl) => sl,
    };

    let len = data.len();
    let period = input.get_period();
    let poles = input.get_poles();

    if len == 0 {
        return Err(GaussianError::NoData);
    }

    if period == 1 {
        return Err(GaussianError::PeriodOneDegenerate);
    }

    if period < 2 {
        return Err(GaussianError::DegeneratePeriod { period });
    }

    if period > len {
        return Err(GaussianError::PeriodLongerThanData {
            period,
            data_len: len,
        });
    }

    if !(1..=4).contains(&poles) {
        return Err(GaussianError::InvalidPoles { poles });
    }

    let first_valid = data
        .iter()
        .position(|x| !x.is_nan())
        .ok_or(GaussianError::AllValuesNaN)?;
    if len - first_valid < period {
        return Err(GaussianError::NotEnoughValidData {
            needed: period,
            valid: len - first_valid,
        });
    }

    let chosen = match kernel {
        Kernel::Auto => Kernel::Scalar,
        other => other,
    };
    let warm = first_valid + period;
    let mut out = alloc_with_nan_prefix(len, warm);

    unsafe {
        match chosen {
            Kernel::Scalar | Kernel::ScalarBatch => gaussian_scalar(data, period, poles, &mut out),

            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx2 | Kernel::Avx2Batch => gaussian_avx2(data, period, poles, &mut out),
            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx512 | Kernel::Avx512Batch => gaussian_avx512(data, period, poles, &mut out),
            #[cfg(not(all(feature = "nightly-avx", target_arch = "x86_64")))]
            Kernel::Avx2 | Kernel::Avx2Batch | Kernel::Avx512 | Kernel::Avx512Batch => {
                gaussian_scalar(data, period, poles, &mut out)
            }

            Kernel::Auto => unreachable!(),
        }
    }

    Ok(GaussianOutput { values: out })
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline(always)]
#[allow(unused_variables)]
pub unsafe fn gaussian_avx2(data: &[f64], period: usize, poles: usize, out: &mut [f64]) {
    gaussian_scalar(data, period, poles, out);
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline(always)]
#[allow(unused_variables)]
pub unsafe fn gaussian_avx512(data: &[f64], period: usize, poles: usize, out: &mut [f64]) {
    gaussian_scalar(data, period, poles, out);
}

#[inline(always)]
unsafe fn gaussian_poles1_fma(inp: &[f64], alpha: f64, out: &mut [f64]) {
    let c0 = alpha;
    let c1 = 1.0 - alpha;

    let mut prev = 0.0;
    for i in 0..inp.len() {
        let x = *inp.get_unchecked(i);
        prev = c1.mul_add(prev, c0 * x);
        *out.get_unchecked_mut(i) = prev;
    }
}

#[inline(always)]
unsafe fn gaussian_poles2_fma(inp: &[f64], alpha: f64, out: &mut [f64]) {
    let a2 = alpha * alpha;
    let one = 1.0 - alpha;
    let c0 = a2;
    let c1 = 2.0 * one;
    let c2 = -(one * one);

    let mut prev1 = 0.0;
    let mut prev0 = 0.0;

    for i in 0..inp.len() {
        let x = *inp.get_unchecked(i);
        let y = c2.mul_add(prev0, c1.mul_add(prev1, c0 * x));
        prev0 = prev1;
        prev1 = y;
        *out.get_unchecked_mut(i) = y;
    }
}

#[inline(always)]
unsafe fn gaussian_poles3_fma(inp: &[f64], alpha: f64, out: &mut [f64]) {
    let a3 = alpha * alpha * alpha;
    let one = 1.0 - alpha;
    let one2 = one * one;

    let c0 = a3;
    let c1 = 3.0 * one;
    let c2 = -3.0 * one2;
    let c3 = one2 * one;

    let mut p2 = 0.0;
    let mut p1 = 0.0;
    let mut p0 = 0.0;

    for i in 0..inp.len() {
        let x = *inp.get_unchecked(i);
        let y = c3.mul_add(p0, c2.mul_add(p1, c1.mul_add(p2, c0 * x)));
        p0 = p1;
        p1 = p2;
        p2 = y;
        *out.get_unchecked_mut(i) = y;
    }
}

#[inline(always)]
unsafe fn gaussian_poles4_fma(inp: &[f64], alpha: f64, out: &mut [f64]) {
    let a4 = alpha * alpha * alpha * alpha;
    let one = 1.0 - alpha;
    let one2 = one * one;
    let one3 = one2 * one;

    let c0 = a4;
    let c1 = 4.0 * one;
    let c2 = -6.0 * one2;
    let c3 = 4.0 * one3;
    let c4 = -(one3 * one);

    let mut p3 = 0.0;
    let mut p2 = 0.0;
    let mut p1 = 0.0;
    let mut p0 = 0.0;

    for i in 0..inp.len() {
        let x = *inp.get_unchecked(i);
        let y = c4.mul_add(p0, c3.mul_add(p1, c2.mul_add(p2, c1.mul_add(p3, c0 * x))));
        p0 = p1;
        p1 = p2;
        p2 = p3;
        p3 = y;
        *out.get_unchecked_mut(i) = y;
    }
}

#[inline(always)]
fn gaussian_poles1(data: &[f64], n: usize, alpha: f64) -> Vec<f64> {
    let c0 = alpha;
    let c1 = 1.0 - alpha;
    let mut fil = vec![0.0; 1 + n];
    for i in 0..n {
        fil[i + 1] = c0 * data[i] + c1 * fil[i];
    }
    fil[1..1 + n].to_vec()
}
#[inline(always)]
fn gaussian_poles2(data: &[f64], n: usize, alpha: f64) -> Vec<f64> {
    let a2 = alpha * alpha;
    let one_a = 1.0 - alpha;
    let c0 = a2;
    let c1 = 2.0 * one_a;
    let c2 = -(one_a * one_a);
    let mut fil = vec![0.0; 2 + n];
    for i in 0..n {
        fil[i + 2] = c0 * data[i] + c1 * fil[i + 1] + c2 * fil[i];
    }
    fil[2..2 + n].to_vec()
}
#[inline(always)]
fn gaussian_poles3(data: &[f64], n: usize, alpha: f64) -> Vec<f64> {
    let a3 = alpha * alpha * alpha;
    let one_a = 1.0 - alpha;
    let one_a2 = one_a * one_a;
    let c0 = a3;
    let c1 = 3.0 * one_a;
    let c2 = -3.0 * one_a2;
    let c3 = one_a2 * one_a;
    let mut fil = vec![0.0; 3 + n];
    for i in 0..n {
        fil[i + 3] = c0 * data[i] + c1 * fil[i + 2] + c2 * fil[i + 1] + c3 * fil[i];
    }
    fil[3..3 + n].to_vec()
}
#[inline(always)]
fn gaussian_poles4(data: &[f64], n: usize, alpha: f64) -> Vec<f64> {
    let a4 = alpha * alpha * alpha * alpha;
    let one_a = 1.0 - alpha;
    let one_a2 = one_a * one_a;
    let one_a3 = one_a2 * one_a;
    let c0 = a4;
    let c1 = 4.0 * one_a;
    let c2 = -6.0 * one_a2;
    let c3 = 4.0 * one_a3;
    let c4 = -(one_a3 * one_a);
    let mut fil = vec![0.0; 4 + n];
    for i in 0..n {
        fil[i + 4] =
            c0 * data[i] + c1 * fil[i + 3] + c2 * fil[i + 2] + c3 * fil[i + 1] + c4 * fil[i];
    }
    fil[4..4 + n].to_vec()
}

#[derive(Debug, Clone)]
pub struct GaussianStream {
    period: usize,
    poles: u8,
    alpha: f64,
    one_minus: f64,

    c: [f64; 5],

    y: [f64; 4],

    idx: usize,
    init: bool,
}

impl GaussianStream {
    #[inline]
    pub fn try_new(params: GaussianParams) -> Result<Self, GaussianError> {
        let period = params.period.unwrap_or(14);
        let poles = params.poles.unwrap_or(4);

        if period == 1 {
            return Err(GaussianError::PeriodOneDegenerate);
        }
        if period < 2 {
            return Err(GaussianError::DegeneratePeriod { period });
        }
        if !(1..=4).contains(&poles) {
            return Err(GaussianError::InvalidPoles { poles });
        }

        let inv_n = 1.0 / (period as f64);
        let theta = core::f64::consts::PI * 2.0 * inv_n;
        let one_minus_cos = {
            let s = (0.5 * theta).sin();
            2.0 * (s * s)
        };

        #[inline(always)]
        fn denom_for_poles(k: usize) -> f64 {
            match k {
                1 => 1.0,
                2 => core::f64::consts::SQRT_2 - 1.0,
                3 => 0.259_921_049_894_873_2,
                4 => 0.189_207_115_002_721_06,
                _ => unreachable!(),
            }
        }
        let beta = one_minus_cos / denom_for_poles(poles);

        let alpha = if beta == 0.0 {
            0.0
        } else {
            let r = (beta * beta + 2.0 * beta).sqrt();
            (2.0 * beta) / (beta + r)
        };

        let one_minus = 1.0 - alpha;

        let a = alpha;
        let a2 = a * a;
        let a3 = a2 * a;
        let a4 = a2 * a2;
        let o = one_minus;
        let o2 = o * o;
        let o3 = o2 * o;
        let o4 = o2 * o2;

        let mut c = [0.0; 5];
        match poles {
            1 => {
                c[0] = a;
                c[1] = o;
            }
            2 => {
                c[0] = a2;
                c[1] = 2.0 * o;
                c[2] = -o2;
            }
            3 => {
                c[0] = a3;
                c[1] = 3.0 * o;
                c[2] = -3.0 * o2;
                c[3] = o3;
            }
            4 => {
                c[0] = a4;
                c[1] = 4.0 * o;
                c[2] = -6.0 * o2;
                c[3] = 4.0 * o3;
                c[4] = -o4;
            }
            _ => unreachable!(),
        }

        Ok(Self {
            period,
            poles: poles as u8,
            alpha,
            one_minus,
            c,
            y: [0.0; 4],
            idx: 0,
            init: true,
        })
    }

    #[inline(always)]
    pub fn update(&mut self, x: f64) -> f64 {
        let p = self.poles;
        let c = self.c;
        let mut y0 = self.y[0];
        let mut y1 = self.y[1];
        let mut y2 = self.y[2];
        let mut y3 = self.y[3];

        let y = match p {
            1 => self.alpha.mul_add(x, self.one_minus * y0),
            2 => c[2].mul_add(y1, c[1].mul_add(y0, c[0] * x)),
            3 => c[3].mul_add(y2, c[2].mul_add(y1, c[1].mul_add(y0, c[0] * x))),
            _ => c[4].mul_add(
                y3,
                c[3].mul_add(y2, c[2].mul_add(y1, c[1].mul_add(y0, c[0] * x))),
            ),
        };

        self.y[3] = y2;
        self.y[2] = y1;
        self.y[1] = y0;
        self.y[0] = y;

        self.idx = self.idx.wrapping_add(1);
        y
    }

    #[inline]
    pub fn update_many(&mut self, xs: &[f64], out: &mut [f64]) {
        debug_assert_eq!(xs.len(), out.len());
        for (i, &x) in xs.iter().enumerate() {
            out[i] = self.update(x);
        }
    }
}

#[derive(Clone, Debug)]
pub struct GaussianBatchRange {
    pub period: (usize, usize, usize),
    pub poles: (usize, usize, usize),
}

impl Default for GaussianBatchRange {
    fn default() -> Self {
        Self {
            period: (14, 263, 1),
            poles: (4, 4, 0),
        }
    }
}

#[derive(Clone, Debug, Default)]
pub struct GaussianBatchBuilder {
    range: GaussianBatchRange,
    kernel: Kernel,
}

impl GaussianBatchBuilder {
    pub fn new() -> Self {
        Self::default()
    }
    pub fn kernel(mut self, k: Kernel) -> Self {
        self.kernel = k;
        self
    }
    pub fn period_range(mut self, start: usize, end: usize, step: usize) -> Self {
        self.range.period = (start, end, step);
        self
    }
    pub fn period_static(mut self, p: usize) -> Self {
        self.range.period = (p, p, 0);
        self
    }
    pub fn poles_range(mut self, start: usize, end: usize, step: usize) -> Self {
        self.range.poles = (start, end, step);
        self
    }
    pub fn poles_static(mut self, p: usize) -> Self {
        self.range.poles = (p, p, 0);
        self
    }
    pub fn apply_slice(self, data: &[f64]) -> Result<GaussianBatchOutput, GaussianError> {
        gaussian_batch_with_kernel(data, &self.range, self.kernel)
    }
    pub fn apply_candles(
        self,
        c: &Candles,
        src: &str,
    ) -> Result<GaussianBatchOutput, GaussianError> {
        let slice = source_type(c, src);
        self.apply_slice(slice)
    }
    pub fn with_default_slice(
        data: &[f64],
        k: Kernel,
    ) -> Result<GaussianBatchOutput, GaussianError> {
        GaussianBatchBuilder::new().kernel(k).apply_slice(data)
    }
    pub fn with_default_candles(c: &Candles) -> Result<GaussianBatchOutput, GaussianError> {
        GaussianBatchBuilder::new()
            .kernel(Kernel::Auto)
            .apply_candles(c, "close")
    }
}

#[derive(Clone, Debug)]
pub struct GaussianBatchOutput {
    pub values: Vec<f64>,
    pub combos: Vec<GaussianParams>,
    pub rows: usize,
    pub cols: usize,
}

impl GaussianBatchOutput {
    pub fn row_for_params(&self, p: &GaussianParams) -> Option<usize> {
        self.combos.iter().position(|c| {
            c.period.unwrap_or(14) == p.period.unwrap_or(14)
                && c.poles.unwrap_or(4) == p.poles.unwrap_or(4)
        })
    }
    pub fn values_for(&self, p: &GaussianParams) -> Option<&[f64]> {
        self.row_for_params(p).map(|row| {
            let start = row * self.cols;
            &self.values[start..start + self.cols]
        })
    }
}

#[inline(always)]
fn expand_grid(r: &GaussianBatchRange) -> Result<Vec<GaussianParams>, GaussianError> {
    #[inline(always)]
    fn axis_usize((start, end, step): (usize, usize, usize)) -> Result<Vec<usize>, GaussianError> {
        if step == 0 || start == end {
            return Ok(vec![start]);
        }
        let (lo, hi) = if start <= end {
            (start, end)
        } else {
            (end, start)
        };
        let v: Vec<usize> = (lo..=hi).step_by(step).collect();
        if v.is_empty() {
            return Err(GaussianError::InvalidRange { start, end, step });
        }
        Ok(v)
    }
    let periods = axis_usize(r.period)?;
    let poles = axis_usize(r.poles)?;
    let mut out = Vec::with_capacity(periods.len().saturating_mul(poles.len()));
    for &p in &periods {
        for &k in &poles {
            out.push(GaussianParams {
                period: Some(p),
                poles: Some(k),
            });
        }
    }
    Ok(out)
}

pub fn gaussian_batch_with_kernel(
    data: &[f64],
    sweep: &GaussianBatchRange,
    k: Kernel,
) -> Result<GaussianBatchOutput, GaussianError> {
    let kernel = match k {
        Kernel::Auto => match detect_best_batch_kernel() {
            Kernel::Avx512Batch => Kernel::Avx2Batch,
            other => other,
        },
        other if other.is_batch() => other,
        other => return Err(GaussianError::InvalidKernelForBatch(other)),
    };
    let simd = match kernel {
        Kernel::Avx512Batch => Kernel::Avx512,
        Kernel::Avx2Batch => Kernel::Avx2,
        Kernel::ScalarBatch => Kernel::Scalar,
        _ => unreachable!(),
    };
    gaussian_batch_par_slice(data, sweep, simd)
}

#[inline(always)]
pub fn gaussian_batch_slice(
    data: &[f64],
    sweep: &GaussianBatchRange,
    kern: Kernel,
) -> Result<GaussianBatchOutput, GaussianError> {
    gaussian_batch_inner(data, sweep, kern, false)
}

#[inline(always)]
pub fn gaussian_batch_par_slice(
    data: &[f64],
    sweep: &GaussianBatchRange,
    kern: Kernel,
) -> Result<GaussianBatchOutput, GaussianError> {
    gaussian_batch_inner(data, sweep, kern, true)
}

#[inline]
pub fn gaussian_into_slice(
    dst: &mut [f64],
    input: &GaussianInput,
    kern: Kernel,
) -> Result<(), GaussianError> {
    let data = input.as_ref();

    if dst.len() != data.len() {
        return Err(GaussianError::OutputLengthMismatch {
            expected: data.len(),
            got: dst.len(),
        });
    }

    gaussian_with_kernel_into(input, kern, dst)
}

#[cfg(not(all(target_arch = "wasm32", feature = "wasm")))]
#[inline]
pub fn gaussian_into(input: &GaussianInput, out: &mut [f64]) -> Result<(), GaussianError> {
    let data = input.as_ref();
    if out.len() != data.len() {
        return Err(GaussianError::OutputLengthMismatch {
            expected: data.len(),
            got: out.len(),
        });
    }

    gaussian_with_kernel_into(input, Kernel::Auto, out)
}

#[inline(always)]
fn gaussian_with_kernel_into(
    input: &GaussianInput,
    kernel: Kernel,
    out: &mut [f64],
) -> Result<(), GaussianError> {
    let (data, period, poles, first, chosen) = gaussian_prepare(input, kernel)?;

    out[..first + period].fill(f64::NAN);

    gaussian_compute_into(data, period, poles, first, chosen, out);

    Ok(())
}

#[inline(always)]
fn gaussian_prepare<'a>(
    input: &'a GaussianInput,
    kernel: Kernel,
) -> Result<(&'a [f64], usize, usize, usize, Kernel), GaussianError> {
    let data: &[f64] = input.as_ref();
    let len = data.len();

    if len == 0 {
        return Err(GaussianError::NoData);
    }

    let first = data
        .iter()
        .position(|x| !x.is_nan())
        .ok_or(GaussianError::AllValuesNaN)?;

    let period = input.params.period.unwrap_or(14);
    let poles = input.params.poles.unwrap_or(4);

    if period == 1 {
        return Err(GaussianError::PeriodOneDegenerate);
    }

    if period < 2 {
        return Err(GaussianError::DegeneratePeriod { period });
    }

    if period > len {
        return Err(GaussianError::PeriodLongerThanData {
            period,
            data_len: len,
        });
    }

    if !(1..=4).contains(&poles) {
        return Err(GaussianError::InvalidPoles { poles });
    }

    if len - first < period {
        return Err(GaussianError::NotEnoughValidData {
            needed: period,
            valid: len - first,
        });
    }

    let chosen = match kernel {
        Kernel::Auto => Kernel::Scalar,
        other => other,
    };

    Ok((data, period, poles, first, chosen))
}

#[inline(always)]
fn gaussian_compute_into(
    data: &[f64],
    period: usize,
    poles: usize,
    first: usize,
    kernel: Kernel,
    out: &mut [f64],
) {
    unsafe {
        match kernel {
            Kernel::Scalar | Kernel::ScalarBatch => gaussian_scalar(data, period, poles, out),
            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx2 | Kernel::Avx2Batch => gaussian_avx2(data, period, poles, out),
            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx512 | Kernel::Avx512Batch => gaussian_avx512(data, period, poles, out),
            _ => unreachable!(),
        }
    }
}

#[inline(always)]
fn alpha_from(period: usize, poles: usize) -> f64 {
    let beta = {
        let numerator = 1.0 - (2.0 * PI / period as f64).cos();
        let denominator = (2.0_f64).powf(1.0 / poles as f64) - 1.0;
        numerator / denominator
    };
    let tmp = beta * beta + 2.0 * beta;
    -beta + tmp.sqrt()
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[target_feature(enable = "avx512f,fma")]
#[inline]
unsafe fn gaussian_rows8_avx512(
    data: &[f64],
    params: &[GaussianParams],
    out_rows: &mut [f64],
    cols: usize,
) {
    debug_assert_eq!(params.len(), LANES_AVX512);
    debug_assert_eq!(out_rows.len(), LANES_AVX512 * cols);

    let mut alpha_arr = [0.0f64; LANES_AVX512];
    let mut pole_arr = [0u32; LANES_AVX512];
    for (lane, prm) in params.iter().enumerate() {
        let p = prm.period.unwrap_or(14);
        let k = prm.poles.unwrap_or(4);
        alpha_arr[lane] = alpha_from(p, k);
        pole_arr[lane] = k as u32;
    }

    let alpha_v = _mm512_loadu_pd(alpha_arr.as_ptr());
    let one_minus_v = _mm512_sub_pd(_mm512_set1_pd(1.0), alpha_v);

    let mask_for = |stage: u32| -> __mmask8 {
        let mut m: u8 = 0;
        for lane in 0..LANES_AVX512 {
            if pole_arr[lane] > stage {
                m |= 1 << lane;
            }
        }
        m as __mmask8
    };
    let m0 = mask_for(0);
    let m1 = mask_for(1);
    let m2 = mask_for(2);
    let m3 = mask_for(3);

    let mut st0 = _mm512_setzero_pd();
    let mut st1 = _mm512_setzero_pd();
    let mut st2 = _mm512_setzero_pd();
    let mut st3 = _mm512_setzero_pd();

    for (t, &x_n) in data.iter().enumerate() {
        let x_vec = _mm512_set1_pd(x_n);

        let y0 = _mm512_fmadd_pd(alpha_v, x_vec, _mm512_mul_pd(one_minus_v, st0));
        st0 = _mm512_mask_mov_pd(st0, m0, y0);

        let y1 = _mm512_fmadd_pd(alpha_v, st0, _mm512_mul_pd(one_minus_v, st1));
        st1 = _mm512_mask_mov_pd(st1, m1, y1);

        let y2 = _mm512_fmadd_pd(alpha_v, st1, _mm512_mul_pd(one_minus_v, st2));
        st2 = _mm512_mask_mov_pd(st2, m2, y2);

        let y3 = _mm512_fmadd_pd(alpha_v, st2, _mm512_mul_pd(one_minus_v, st3));
        st3 = _mm512_mask_mov_pd(st3, m3, y3);

        let mut y = st0;
        y = _mm512_mask_mov_pd(y, m1, st1);
        y = _mm512_mask_mov_pd(y, m2, st2);
        y = _mm512_mask_mov_pd(y, m3, st3);

        let mut tmp = [0.0f64; LANES_AVX512];
        _mm512_storeu_pd(tmp.as_mut_ptr(), y);
        for lane in 0..LANES_AVX512 {
            out_rows[lane * cols + t] = tmp[lane];
        }
    }
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[target_feature(enable = "avx2,fma")]
#[inline]
unsafe fn gaussian_batch_tile_avx2(
    data: &[f64],
    combos: &[GaussianParams],
    out_mu: &mut [core::mem::MaybeUninit<f64>],
    cols: usize,
) {
    let out = core::slice::from_raw_parts_mut(out_mu.as_mut_ptr() as *mut f64, out_mu.len());

    let mut row = 0;
    while row + LANES_AVX2 <= combos.len() {
        gaussian_rows4_avx2(
            data,
            &combos[row..row + LANES_AVX2],
            &mut out[row * cols..(row + LANES_AVX2) * cols],
            cols,
        );
        row += LANES_AVX2;
    }
    for r in row..combos.len() {
        gaussian_row_scalar(data, &combos[r], &mut out[r * cols..(r + 1) * cols]);
    }
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[target_feature(enable = "avx2,fma")]
#[inline]
unsafe fn gaussian_rows4_avx2(
    data: &[f64],
    params: &[GaussianParams],
    out_rows: &mut [f64],
    cols: usize,
) {
    debug_assert_eq!(params.len(), LANES_AVX2);
    debug_assert_eq!(out_rows.len(), LANES_AVX2 * cols);

    let mut alpha_arr = [0.0f64; LANES_AVX2];
    let mut pole_arr = [0u32; LANES_AVX2];
    for (l, prm) in params.iter().enumerate() {
        alpha_arr[l] = alpha_from(prm.period.unwrap_or(14), prm.poles.unwrap_or(4));
        pole_arr[l] = prm.poles.unwrap_or(4) as u32;
    }
    let alpha_v = _mm256_loadu_pd(alpha_arr.as_ptr());
    let one_minus_v = _mm256_sub_pd(_mm256_set1_pd(1.0), alpha_v);

    let mut st0 = _mm256_setzero_pd();
    let mut st1 = _mm256_setzero_pd();
    let mut st2 = _mm256_setzero_pd();
    let mut st3 = _mm256_setzero_pd();

    let mut y0a = [0.0; LANES_AVX2];
    let mut y1a = [0.0; LANES_AVX2];
    let mut y2a = [0.0; LANES_AVX2];
    let mut y3a = [0.0; LANES_AVX2];

    for (t, &x_n) in data.iter().enumerate() {
        let x_v = _mm256_set1_pd(x_n);

        let y0_v = _mm256_fmadd_pd(alpha_v, x_v, _mm256_mul_pd(one_minus_v, st0));
        st0 = y0_v;

        let y1_v = _mm256_fmadd_pd(alpha_v, st0, _mm256_mul_pd(one_minus_v, st1));
        st1 = y1_v;

        let y2_v = _mm256_fmadd_pd(alpha_v, st1, _mm256_mul_pd(one_minus_v, st2));
        st2 = y2_v;

        let y3_v = _mm256_fmadd_pd(alpha_v, st2, _mm256_mul_pd(one_minus_v, st3));
        st3 = y3_v;

        _mm256_storeu_pd(y0a.as_mut_ptr(), y0_v);
        _mm256_storeu_pd(y1a.as_mut_ptr(), y1_v);
        _mm256_storeu_pd(y2a.as_mut_ptr(), y2_v);
        _mm256_storeu_pd(y3a.as_mut_ptr(), y3_v);

        for lane in 0..LANES_AVX2 {
            let final_y = match pole_arr[lane] {
                1 => y0a[lane],
                2 => y1a[lane],
                3 => y2a[lane],
                _ => y3a[lane],
            };
            out_rows[lane * cols + t] = final_y;
        }
    }
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[target_feature(enable = "avx512f,fma")]
#[inline]
unsafe fn gaussian_batch_tile_avx512(
    data: &[f64],
    combos: &[GaussianParams],
    out_mu: &mut [core::mem::MaybeUninit<f64>],
    cols: usize,
) {
    let out = core::slice::from_raw_parts_mut(out_mu.as_mut_ptr() as *mut f64, out_mu.len());

    let mut row = 0;
    while row + LANES_AVX512 <= combos.len() {
        gaussian_rows8_avx512(
            data,
            &combos[row..row + LANES_AVX512],
            &mut out[row * cols..(row + LANES_AVX512) * cols],
            cols,
        );
        row += LANES_AVX512;
    }

    for r in row..combos.len() {
        gaussian_row_scalar(data, &combos[r], &mut out[r * cols..(r + 1) * cols]);
    }
}

#[inline(always)]
fn gaussian_batch_inner(
    data: &[f64],
    sweep: &GaussianBatchRange,
    kern: Kernel,
    parallel: bool,
) -> Result<GaussianBatchOutput, GaussianError> {
    #[cfg(not(target_arch = "wasm32"))]
    use rayon::prelude::*;
    use std::{arch::is_x86_feature_detected, mem::MaybeUninit};

    let combos = expand_grid(sweep)?;
    if combos.is_empty() || data.is_empty() {
        return Err(GaussianError::NoData);
    }
    let cols = data.len();
    let rows = combos.len();

    let _total = rows
        .checked_mul(cols)
        .ok_or(GaussianError::SizeOverflow { rows, cols })?;

    let first_valid = data
        .iter()
        .position(|x| !x.is_nan())
        .ok_or(GaussianError::AllValuesNaN)?;

    for c in &combos {
        let period = c.period.unwrap_or(14);
        let poles = c.poles.unwrap_or(4);

        if period == 1 {
            return Err(GaussianError::PeriodOneDegenerate);
        }

        if period < 2 {
            return Err(GaussianError::DegeneratePeriod { period });
        }

        if period > cols {
            return Err(GaussianError::PeriodLongerThanData {
                period,
                data_len: cols,
            });
        }
        if !(1..=4).contains(&poles) {
            return Err(GaussianError::InvalidPoles { poles });
        }
        if cols - first_valid < period {
            return Err(GaussianError::NotEnoughValidData {
                needed: period,
                valid: cols - first_valid,
            });
        }
    }

    let warm: Vec<usize> = combos
        .iter()
        .map(|c| {
            let p = c.period.unwrap_or(14);
            (first_valid + p).min(cols)
        })
        .collect();

    let mut raw = make_uninit_matrix(rows, cols);
    init_matrix_prefixes(&mut raw, cols, &warm);

    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    let have_avx512 = cfg!(target_feature = "avx512f") && is_x86_feature_detected!("avx512f");
    #[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))]
    let have_avx512 = false;

    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    let have_avx2 = cfg!(target_feature = "avx2") && is_x86_feature_detected!("avx2");
    #[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))]
    let have_avx2 = false;

    let chosen = match kern {
        Kernel::Avx512 if have_avx512 => Kernel::Avx512,
        Kernel::Avx2 if have_avx2 => Kernel::Avx2,
        _ => Kernel::Scalar,
    };

    type RowRunner = unsafe fn(&[f64], &GaussianParams, &mut [f64]);

    let row_runner: RowRunner = match chosen {
        #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
        Kernel::Avx512 => gaussian_row_avx512,
        #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
        Kernel::Avx2 => gaussian_row_avx2,
        _ => gaussian_row_scalar,
    };

    #[inline(always)]
    unsafe fn compute_row(
        row_idx: usize,
        dst_mu: &mut [MaybeUninit<f64>],
        combos: &[GaussianParams],
        data: &[f64],
        cols: usize,
        runner: RowRunner,
    ) {
        let out = std::slice::from_raw_parts_mut(dst_mu.as_mut_ptr() as *mut f64, cols);
        runner(data, &combos[row_idx], out);
    }

    if parallel {
        #[cfg(not(target_arch = "wasm32"))]
        {
            match chosen {
                #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
                Kernel::Avx512 => {
                    let tiles = rows / LANES_AVX512;

                    raw.par_chunks_exact_mut(cols * LANES_AVX512)
                        .zip(combos.par_chunks_exact(LANES_AVX512))
                        .for_each(|(dst_blk, prm_blk)| unsafe {
                            gaussian_batch_tile_avx512(data, prm_blk, dst_blk, cols);
                        });

                    raw[tiles * cols * LANES_AVX512..]
                        .par_chunks_mut(cols)
                        .enumerate()
                        .for_each(|(i, dst)| unsafe {
                            compute_row(
                                tiles * LANES_AVX512 + i,
                                dst,
                                &combos,
                                data,
                                cols,
                                row_runner,
                            );
                        });
                }

                #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
                Kernel::Avx2 => {
                    let tiles = rows / LANES_AVX2;

                    raw.par_chunks_exact_mut(cols * LANES_AVX2)
                        .zip(combos.par_chunks_exact(LANES_AVX2))
                        .for_each(|(dst_blk, prm_blk)| unsafe {
                            gaussian_batch_tile_avx2(data, prm_blk, dst_blk, cols);
                        });

                    raw[tiles * cols * LANES_AVX2..]
                        .par_chunks_mut(cols)
                        .enumerate()
                        .for_each(|(i, dst)| unsafe {
                            compute_row(
                                tiles * LANES_AVX2 + i,
                                dst,
                                &combos,
                                data,
                                cols,
                                row_runner,
                            );
                        });
                }

                _ => {
                    raw.par_chunks_mut(cols)
                        .enumerate()
                        .for_each(|(row, dst)| unsafe {
                            compute_row(row, dst, &combos, data, cols, row_runner);
                        });
                }
            }
        }
        #[cfg(target_arch = "wasm32")]
        {
            for (row, dst) in raw.chunks_mut(cols).enumerate() {
                unsafe {
                    compute_row(row, dst, &combos, data, cols, row_runner);
                }
            }
        }
    } else {
        match chosen {
            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx512 => {
                let tiles = rows / LANES_AVX512;
                unsafe {
                    gaussian_batch_tile_avx512(
                        data,
                        &combos[..tiles * LANES_AVX512],
                        &mut raw[..tiles * cols * LANES_AVX512],
                        cols,
                    );
                }
                for (i, dst) in raw[tiles * cols * LANES_AVX512..]
                    .chunks_mut(cols)
                    .enumerate()
                {
                    unsafe {
                        compute_row(
                            tiles * LANES_AVX512 + i,
                            dst,
                            &combos,
                            data,
                            cols,
                            row_runner,
                        );
                    }
                }
            }

            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx2 => {
                let tiles = rows / LANES_AVX2;
                unsafe {
                    gaussian_batch_tile_avx2(
                        data,
                        &combos[..tiles * LANES_AVX2],
                        &mut raw[..tiles * cols * LANES_AVX2],
                        cols,
                    );
                }
                for (i, dst) in raw[tiles * cols * LANES_AVX2..]
                    .chunks_mut(cols)
                    .enumerate()
                {
                    unsafe {
                        compute_row(tiles * LANES_AVX2 + i, dst, &combos, data, cols, row_runner);
                    }
                }
            }

            _ => {
                for (row, dst) in raw.chunks_mut(cols).enumerate() {
                    unsafe {
                        compute_row(row, dst, &combos, data, cols, row_runner);
                    }
                }
            }
        }
    }

    let mut guard = core::mem::ManuallyDrop::new(raw);
    let values = unsafe {
        Vec::from_raw_parts(
            guard.as_mut_ptr() as *mut f64,
            guard.len(),
            guard.capacity(),
        )
    };
    Ok(GaussianBatchOutput {
        values,
        combos,
        rows,
        cols,
    })
}

#[inline(always)]
pub unsafe fn gaussian_row_scalar(data: &[f64], prm: &GaussianParams, out: &mut [f64]) {
    gaussian_scalar(data, prm.period.unwrap_or(14), prm.poles.unwrap_or(4), out);
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[target_feature(enable = "avx2,fma")]
#[inline]
pub unsafe fn gaussian_row_avx2(data: &[f64], prm: &GaussianParams, out: &mut [f64]) {
    gaussian_row_scalar(data, prm, out);
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[target_feature(enable = "avx512f,fma")]
#[inline]
pub unsafe fn gaussian_row_avx512(data: &[f64], prm: &GaussianParams, out: &mut [f64]) {
    gaussian_row_scalar(data, prm, out);
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::skip_if_unsupported;
    use crate::utilities::data_loader::read_candles_from_csv;
    use proptest::prelude::*;

    fn check_gaussian_partial_params(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);
        let file_path = "src/data/2018-09-01-2024-Bitfinex_Spot-4h.csv";
        let candles = read_candles_from_csv(file_path)?;
        let default_params = GaussianParams {
            period: None,
            poles: None,
        };
        let input = GaussianInput::from_candles(&candles, "close", default_params);
        let output = gaussian_with_kernel(&input, kernel)?;
        assert_eq!(output.values.len(), candles.close.len());
        Ok(())
    }

    fn check_gaussian_accuracy(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);
        let file_path = "src/data/2018-09-01-2024-Bitfinex_Spot-4h.csv";
        let candles = read_candles_from_csv(file_path)?;
        let params = GaussianParams {
            period: Some(14),
            poles: Some(4),
        };
        let input = GaussianInput::from_candles(&candles, "close", params);
        let result = gaussian_with_kernel(&input, kernel)?;
        let expected_last_five = [
            59221.90637814869,
            59236.15215167245,
            59207.10087088464,
            59178.48276885589,
            59085.36983209433,
        ];
        let start = result.values.len().saturating_sub(5);
        for (i, &val) in result.values[start..].iter().enumerate() {
            let diff = (val - expected_last_five[i]).abs();
            assert!(
                diff < 1e-4,
                "[{}] Gaussian {:?} mismatch at idx {}: got {}, expected {}",
                test_name,
                kernel,
                i,
                val,
                expected_last_five[i]
            );
        }
        Ok(())
    }

    fn check_gaussian_default_candles(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);
        let file_path = "src/data/2018-09-01-2024-Bitfinex_Spot-4h.csv";
        let candles = read_candles_from_csv(file_path)?;
        let input = GaussianInput::with_default_candles(&candles);
        match input.data {
            GaussianData::Candles { source, .. } => assert_eq!(source, "close"),
            _ => panic!("Expected GaussianData::Candles"),
        }
        let output = gaussian_with_kernel(&input, kernel)?;
        assert_eq!(output.values.len(), candles.close.len());
        Ok(())
    }

    fn check_gaussian_zero_period(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);
        let data = [1.0, 2.0, 3.0];
        let params = GaussianParams {
            period: Some(0),
            poles: Some(2),
        };
        let input = GaussianInput::from_slice(&data, params);
        let res = gaussian_with_kernel(&input, kernel);
        assert!(
            matches!(res, Err(GaussianError::DegeneratePeriod { .. })),
            "[{test_name}] expected DegeneratePeriod error for period=0"
        );
        Ok(())
    }

    fn check_gaussian_empty_input(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);
        let empty: [f64; 0] = [];
        let input = GaussianInput::from_slice(&empty, GaussianParams::default());
        let res = gaussian_with_kernel(&input, kernel);
        assert!(
            matches!(res, Err(GaussianError::NoData)),
            "[{test_name}] expected NoData error"
        );
        Ok(())
    }

    fn check_gaussian_invalid_poles(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);
        let data = [1.0, 2.0, 3.0, 4.0];
        let params = GaussianParams {
            period: Some(2),
            poles: Some(5),
        };
        let input = GaussianInput::from_slice(&data, params);
        let res = gaussian_with_kernel(&input, kernel);
        assert!(
            matches!(res, Err(GaussianError::InvalidPoles { .. })),
            "[{test_name}] expected InvalidPoles error"
        );
        Ok(())
    }

    fn check_gaussian_all_nan(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);
        let data = [f64::NAN; 5];
        let params = GaussianParams {
            period: Some(3),
            poles: None,
        };
        let input = GaussianInput::from_slice(&data, params);
        let res = gaussian_with_kernel(&input, kernel);
        assert!(
            matches!(res, Err(GaussianError::AllValuesNaN)),
            "[{test_name}] expected AllValuesNaN error"
        );
        Ok(())
    }

    fn check_gaussian_period_exceeds_length(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);
        let data_small = [10.0, 20.0, 30.0];
        let params = GaussianParams {
            period: Some(10),
            poles: None,
        };
        let input = GaussianInput::from_slice(&data_small, params);
        let res = gaussian_with_kernel(&input, kernel);
        assert!(
            res.is_err(),
            "[{}] Gaussian should fail with period exceeding length",
            test_name
        );
        Ok(())
    }

    fn check_gaussian_very_small_dataset(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);
        let single_point = [42.0];
        let params = GaussianParams {
            period: Some(14),
            poles: None,
        };
        let input = GaussianInput::from_slice(&single_point, params);
        let res = gaussian_with_kernel(&input, kernel);
        assert!(
            res.is_err(),
            "[{}] Gaussian should fail with insufficient data",
            test_name
        );
        Ok(())
    }

    fn check_gaussian_reinput(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);
        let file_path = "src/data/2018-09-01-2024-Bitfinex_Spot-4h.csv";
        let candles = read_candles_from_csv(file_path)?;
        let first_params = GaussianParams {
            period: Some(14),
            poles: Some(4),
        };
        let first_input = GaussianInput::from_candles(&candles, "close", first_params);
        let first_result = gaussian_with_kernel(&first_input, kernel)?;
        assert_eq!(first_result.values.len(), candles.close.len());
        let second_params = GaussianParams {
            period: Some(7),
            poles: Some(2),
        };
        let second_input = GaussianInput::from_slice(&first_result.values, second_params);
        let second_result = gaussian_with_kernel(&second_input, kernel)?;
        assert_eq!(second_result.values.len(), first_result.values.len());

        for i in 10..second_result.values.len() {
            assert!(
                !second_result.values[i].is_nan(),
                "NaN found at index {}",
                i
            );
        }
        Ok(())
    }

    fn check_gaussian_nan_handling(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);
        let file_path = "src/data/2018-09-01-2024-Bitfinex_Spot-4h.csv";
        let candles = read_candles_from_csv(file_path)?;
        let input = GaussianInput::from_candles(&candles, "close", GaussianParams::default());
        let res = gaussian_with_kernel(&input, kernel)?;
        assert_eq!(res.values.len(), candles.close.len());

        let skip = input.params.poles.unwrap_or(4);
        for val in res.values.iter().skip(skip) {
            assert!(
                val.is_finite(),
                "[{}] Gaussian output should be finite once settled.",
                test_name
            );
        }
        Ok(())
    }

    fn check_gaussian_streaming(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);
        let file_path = "src/data/2018-09-01-2024-Bitfinex_Spot-4h.csv";
        let candles = read_candles_from_csv(file_path)?;
        let period = 14;
        let poles = 4;
        let input = GaussianInput::from_candles(
            &candles,
            "close",
            GaussianParams {
                period: Some(period),
                poles: Some(poles),
            },
        );
        let batch_output = gaussian_with_kernel(&input, kernel)?.values;
        let mut stream = GaussianStream::try_new(GaussianParams {
            period: Some(period),
            poles: Some(poles),
        })?;
        let mut stream_values = Vec::with_capacity(candles.close.len());
        for &price in &candles.close {
            stream_values.push(stream.update(price));
        }
        assert_eq!(batch_output.len(), stream_values.len());
        let skip = poles;
        for (i, (&b, &s)) in batch_output
            .iter()
            .zip(stream_values.iter())
            .enumerate()
            .skip(skip)
        {
            if b.is_nan() && s.is_nan() {
                continue;
            }
            let diff = (b - s).abs();
            assert!(
                diff < 1e-9,
                "[{}] Gaussian streaming f64 mismatch at idx {}: batch={}, stream={}, diff={}",
                test_name,
                i,
                b,
                s,
                diff
            );
        }
        Ok(())
    }

    fn check_gaussian_property(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);

        let strat = (2usize..=50).prop_flat_map(|period| {
            (
                prop::collection::vec(
                    (-1e6f64..1e6f64).prop_filter("finite", |x| x.is_finite()),
                    period..400,
                ),
                Just(period),
                1usize..=4,
            )
        });

        proptest::test_runner::TestRunner::default()
            .run(&strat, |(data, period, poles)| {
                let params = GaussianParams {
                    period: Some(period),
                    poles: Some(poles),
                };
                let input = GaussianInput::from_slice(&data, params);

                let GaussianOutput { values: out } = gaussian_with_kernel(&input, kernel).unwrap();
                let GaussianOutput { values: ref_out } =
                    gaussian_with_kernel(&input, Kernel::Scalar).unwrap();

                prop_assert_eq!(out.len(), data.len());

                let first_valid = data.iter().position(|x| !x.is_nan()).unwrap_or(data.len());

                let expected_warmup = first_valid + period;

                for i in 0..first_valid {
                    prop_assert!(
                        out[i].is_nan(),
                        "idx {}: expected NaN for NaN input, got {}",
                        i,
                        out[i]
                    );
                }

                for i in first_valid..data.len() {
                    if data[i].is_finite() && !data[first_valid..=i].iter().any(|x| x.is_nan()) {
                        prop_assert!(
                            out[i].is_finite(),
                            "idx {}: expected finite value, got {}",
                            i,
                            out[i]
                        );
                    }
                }

                let stability_point = first_valid + period * 2;
                if period > 1 && stability_point + 20 < data.len() {
                    let window_start = stability_point;
                    let window_end = (stability_point + 50).min(data.len());

                    let input_slice = &data[window_start..window_end];
                    let output_slice = &out[window_start..window_end];

                    if input_slice.len() > 10
                        && input_slice.iter().all(|x| x.is_finite())
                        && output_slice.iter().all(|x| x.is_finite())
                    {
                        let input_mean: f64 =
                            input_slice.iter().sum::<f64>() / input_slice.len() as f64;
                        let output_mean: f64 =
                            output_slice.iter().sum::<f64>() / output_slice.len() as f64;

                        let input_var: f64 = input_slice
                            .iter()
                            .map(|x| (x - input_mean).powi(2))
                            .sum::<f64>()
                            / input_slice.len() as f64;
                        let output_var: f64 = output_slice
                            .iter()
                            .map(|x| (x - output_mean).powi(2))
                            .sum::<f64>()
                            / output_slice.len() as f64;

                        if input_var > 1e-10 {
                            prop_assert!(
								output_var <= input_var + 1e-15,
								"Gaussian filter MUST reduce variance: input_var={}, output_var={}, period={}, poles={}",
								input_var, output_var, period, poles
							);
                        }
                    }
                }

                if period < 2 {
                    prop_assert!(
                        false,
                        "Test should not generate period < 2, but got period={}",
                        period
                    );
                }

                let stability_check = first_valid + period * 3;
                if data[first_valid..]
                    .windows(2)
                    .all(|w| (w[0] - w[1]).abs() < 1e-10)
                    && stability_check < data.len()
                {
                    let constant_val = data[first_valid];
                    for i in stability_check..data.len() {
                        if out[i].is_finite() {
                            prop_assert!(
								(out[i] - constant_val).abs() <= 1e-12,
								"constant input MUST produce constant output: idx={}, expected={}, got={}, diff={}",
								i, constant_val, out[i], (out[i] - constant_val).abs()
							);
                        }
                    }
                }

                if kernel != Kernel::Scalar {
                    for i in first_valid..data.len() {
                        if out[i].is_finite() && ref_out[i].is_finite() {
                            let y_bits = out[i].to_bits();
                            let r_bits = ref_out[i].to_bits();
                            let diff_bits = if y_bits > r_bits {
                                y_bits - r_bits
                            } else {
                                r_bits - y_bits
                            };

                            prop_assert!(
								diff_bits <= 10 || (out[i] - ref_out[i]).abs() < 1e-14,
								"kernel consistency failed at idx {}: {:?}={}, Scalar={}, diff_bits={}, abs_diff={}",
								i, kernel, out[i], ref_out[i], diff_bits, (out[i] - ref_out[i]).abs()
							);
                        }
                    }
                }

                if stability_point + 30 < data.len() {
                    match poles {
                        1 => {
                            let check_start = (first_valid + period * 2).min(data.len() / 2);
                            let check_end = data.len();

                            if check_end > check_start + 10 {
                                let input_slice = &data[check_start..check_end];
                                let output_slice = &out[check_start..check_end];

                                if input_slice.iter().all(|x| x.is_finite())
                                    && output_slice.iter().all(|x| x.is_finite())
                                {
                                    let input_mean =
                                        input_slice.iter().sum::<f64>() / input_slice.len() as f64;
                                    let output_mean = output_slice.iter().sum::<f64>()
                                        / output_slice.len() as f64;

                                    let input_var = input_slice
                                        .iter()
                                        .map(|x| (x - input_mean).powi(2))
                                        .sum::<f64>()
                                        / input_slice.len() as f64;
                                    let output_var = output_slice
                                        .iter()
                                        .map(|x| (x - output_mean).powi(2))
                                        .sum::<f64>()
                                        / output_slice.len() as f64;

                                    if input_var > 1e-10 {
                                        prop_assert!(
											output_var <= input_var,
											"1-pole filter should reduce variance: input_var={}, output_var={}",
											input_var, output_var
										);
                                    }
                                }
                            }
                        }
                        2 | 3 | 4 => {
                            let params_1pole = GaussianParams {
                                period: Some(period),
                                poles: Some(1),
                            };
                            let input_1pole = GaussianInput::from_slice(&data, params_1pole);
                            if let Ok(GaussianOutput { values: out_1pole }) =
                                gaussian_with_kernel(&input_1pole, kernel)
                            {
                                let window_start = stability_point;
                                let window_end = (stability_point + 30).min(data.len() - 1);

                                if window_end > window_start + 5 {
                                    let accel_multi: f64 = (window_start + 2..window_end)
                                        .map(|i| {
                                            if out[i - 2].is_finite()
                                                && out[i - 1].is_finite()
                                                && out[i].is_finite()
                                            {
                                                ((out[i] - out[i - 1]) - (out[i - 1] - out[i - 2]))
                                                    .abs()
                                            } else {
                                                0.0
                                            }
                                        })
                                        .sum::<f64>();

                                    let accel_1pole: f64 = (window_start + 2..window_end)
                                        .map(|i| {
                                            if out_1pole[i - 2].is_finite()
                                                && out_1pole[i - 1].is_finite()
                                                && out_1pole[i].is_finite()
                                            {
                                                ((out_1pole[i] - out_1pole[i - 1])
                                                    - (out_1pole[i - 1] - out_1pole[i - 2]))
                                                    .abs()
                                            } else {
                                                0.0
                                            }
                                        })
                                        .sum::<f64>();

                                    let non_zero_count =
                                        data.iter().filter(|&&x| x.abs() > 1e-10).count();
                                    if accel_1pole > 1e-10
                                        && accel_multi > 1e-10
                                        && non_zero_count > 5
                                    {
                                        prop_assert!(
											accel_multi <= accel_1pole * 1.1,
											"{}-pole filter should be smoother than 1-pole: accel_{}pole={}, accel_1pole={}",
											poles, poles, accel_multi, accel_1pole
										);
                                    }
                                }
                            }
                        }
                        _ => {}
                    }
                }

                Ok(())
            })
            .unwrap();

        Ok(())
    }

    macro_rules! generate_all_gaussian_tests {
        ($($test_fn:ident),*) => {
            paste::paste! {
                $(
                    #[test]
                    fn [<$test_fn _scalar_f64>]() {
                        let _ = $test_fn(stringify!([<$test_fn _scalar_f64>]), Kernel::Scalar);
                    }
                    #[test]
                    fn [<$test_fn _avx2_f64>]() {
                        let _ = $test_fn(stringify!([<$test_fn _avx2_f64>]), Kernel::Avx2);
                    }
                    #[test]
                    fn [<$test_fn _avx512_f64>]() {
                        let _ = $test_fn(stringify!([<$test_fn _avx512_f64>]), Kernel::Avx512);
                    }
                )*
            }
        }
    }

    fn check_gaussian_edge_cases(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);

        {
            let data = vec![1.0, 2.0, 3.0, 4.0, 5.0];
            let params = GaussianParams {
                period: Some(1),
                poles: Some(1),
            };
            let input = GaussianInput::from_slice(&data, params);
            let result = gaussian_with_kernel(&input, kernel);

            assert!(
                result.is_err(),
                "[{}] Period=1 should return error, but got Ok",
                test_name
            );

            if let Err(e) = result {
                match e {
                    GaussianError::PeriodOneDegenerate => {}
                    _ => panic!("[{}] Wrong error type for period=1: {:?}", test_name, e),
                }
            }
        }

        {
            let data = vec![1.0, 2.0, 3.0, 4.0, 5.0];
            let params = GaussianParams {
                period: Some(data.len()),
                poles: Some(2),
            };
            let input = GaussianInput::from_slice(&data, params);
            let result = gaussian_with_kernel(&input, kernel)?;
            assert_eq!(result.values.len(), data.len());

            for i in 0..data.len() {
                assert!(
                    result.values[i].is_finite() || result.values[i].is_nan(),
                    "[{}] Output should be finite or NaN at index {}, got {}",
                    test_name,
                    i,
                    result.values[i]
                );
            }
        }

        {
            let data = vec![42.0];
            let params = GaussianParams {
                period: Some(1),
                poles: Some(1),
            };
            let input = GaussianInput::from_slice(&data, params);
            let result = gaussian_with_kernel(&input, kernel);

            assert!(
                result.is_err(),
                "[{}] Single data point with period=1 should return error",
                test_name
            );
        }

        {
            let data = vec![0.0; 10];
            let params = GaussianParams {
                period: Some(3),
                poles: Some(2),
            };
            let input = GaussianInput::from_slice(&data, params);
            let result = gaussian_with_kernel(&input, kernel)?;

            for i in 3..result.values.len() {
                assert!(
                    result.values[i].abs() < 1e-15 || result.values[i].is_nan(),
                    "[{}] All-zero input should produce zero output at index {}: got {}",
                    test_name,
                    i,
                    result.values[i]
                );
            }
        }

        {
            let data = vec![1.0, 2.0, 3.0];
            let params = GaussianParams {
                period: Some(5),
                poles: Some(2),
            };
            let input = GaussianInput::from_slice(&data, params);
            let result = gaussian_with_kernel(&input, kernel);

            assert!(
                result.is_err(),
                "[{}] Period > data.len() should return error, but got Ok",
                test_name
            );

            if let Err(e) = result {
                match e {
                    GaussianError::InvalidPeriod { .. }
                    | GaussianError::PeriodLongerThanData { .. } => {}
                    _ => panic!(
                        "[{}] Wrong error type for period > data.len(): {:?}",
                        test_name, e
                    ),
                }
            }
        }

        {
            let data = vec![1.0, -1.0, 1.0, -1.0, 1.0, -1.0, 1.0, -1.0];
            let params = GaussianParams {
                period: Some(2),
                poles: Some(4),
            };
            let input = GaussianInput::from_slice(&data, params);
            let result = gaussian_with_kernel(&input, kernel)?;

            let start = 2;
            if result.values.len() > start + 2 {
                let slice = &result.values[start..];
                let valid_values: Vec<f64> =
                    slice.iter().filter(|x| x.is_finite()).copied().collect();

                if valid_values.len() > 2 {
                    let mean = valid_values.iter().sum::<f64>() / valid_values.len() as f64;
                    let variance = valid_values.iter().map(|x| (x - mean).powi(2)).sum::<f64>()
                        / valid_values.len() as f64;

                    assert!(
                        variance < 0.6,
                        "[{}] 4-pole filter should smooth alternating input: variance={}",
                        test_name,
                        variance
                    );
                }
            }
        }

        Ok(())
    }

    #[cfg(debug_assertions)]
    fn check_gaussian_no_poison(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test_name);

        let file_path = "src/data/2018-09-01-2024-Bitfinex_Spot-4h.csv";
        let candles = read_candles_from_csv(file_path)?;

        let test_cases = vec![
            GaussianParams {
                period: Some(14),
                poles: Some(4),
            },
            GaussianParams {
                period: Some(10),
                poles: Some(1),
            },
            GaussianParams {
                period: Some(30),
                poles: Some(2),
            },
            GaussianParams {
                period: Some(20),
                poles: Some(3),
            },
            GaussianParams {
                period: Some(50),
                poles: Some(4),
            },
            GaussianParams {
                period: Some(5),
                poles: Some(1),
            },
            GaussianParams {
                period: Some(100),
                poles: Some(4),
            },
            GaussianParams {
                period: None,
                poles: None,
            },
        ];

        for params in test_cases {
            let input = GaussianInput::from_candles(&candles, "close", params);
            let output = gaussian_with_kernel(&input, kernel)?;

            for (i, &val) in output.values.iter().enumerate() {
                if val.is_nan() {
                    continue;
                }

                let bits = val.to_bits();

                if bits == 0x11111111_11111111 {
                    panic!(
                        "[{}] Found alloc_with_nan_prefix poison value {} (0x{:016X}) at index {} \
                         with params period={:?}, poles={:?}",
                        test_name, val, bits, i, params.period, params.poles
                    );
                }

                if bits == 0x22222222_22222222 {
                    panic!(
                        "[{}] Found init_matrix_prefixes poison value {} (0x{:016X}) at index {} \
                         with params period={:?}, poles={:?}",
                        test_name, val, bits, i, params.period, params.poles
                    );
                }

                if bits == 0x33333333_33333333 {
                    panic!(
                        "[{}] Found make_uninit_matrix poison value {} (0x{:016X}) at index {} \
                         with params period={:?}, poles={:?}",
                        test_name, val, bits, i, params.period, params.poles
                    );
                }
            }
        }

        Ok(())
    }

    #[cfg(not(debug_assertions))]
    fn check_gaussian_no_poison(
        _test_name: &str,
        _kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        Ok(())
    }

    generate_all_gaussian_tests!(
        check_gaussian_partial_params,
        check_gaussian_accuracy,
        check_gaussian_default_candles,
        check_gaussian_zero_period,
        check_gaussian_period_exceeds_length,
        check_gaussian_very_small_dataset,
        check_gaussian_empty_input,
        check_gaussian_invalid_poles,
        check_gaussian_all_nan,
        check_gaussian_reinput,
        check_gaussian_nan_handling,
        check_gaussian_streaming,
        check_gaussian_property,
        check_gaussian_edge_cases,
        check_gaussian_no_poison
    );

    fn check_batch_default_row(
        test: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test);
        let file = "src/data/2018-09-01-2024-Bitfinex_Spot-4h.csv";
        let c = read_candles_from_csv(file)?;
        let output = GaussianBatchBuilder::new()
            .kernel(kernel)
            .apply_candles(&c, "close")?;
        let def = GaussianParams::default();
        let row = output.values_for(&def).expect("default row missing");
        assert_eq!(row.len(), c.close.len());
        let expected = [
            59221.90637814869,
            59236.15215167245,
            59207.10087088464,
            59178.48276885589,
            59085.36983209433,
        ];
        let start = row.len() - 5;
        for (i, &v) in row[start..].iter().enumerate() {
            assert!(
                (v - expected[i]).abs() < 1e-4,
                "[{test}] default-row mismatch at idx {i}: {v} vs {expected:?}"
            );
        }
        Ok(())
    }

    macro_rules! gen_batch_tests {
        ($fn_name:ident) => {
            paste::paste! {
                #[test] fn [<$fn_name _scalar>]()      {
                    let _ = $fn_name(stringify!([<$fn_name _scalar>]), Kernel::ScalarBatch);
                }
                #[test] fn [<$fn_name _avx2>]()        {
                    let _ = $fn_name(stringify!([<$fn_name _avx2>]), Kernel::Avx2Batch);
                }
                #[test] fn [<$fn_name _avx512>]()      {
                    let _ = $fn_name(stringify!([<$fn_name _avx512>]), Kernel::Avx512Batch);
                }
                #[test] fn [<$fn_name _auto_detect>]() {
                    let _ = $fn_name(stringify!([<$fn_name _auto_detect>]), Kernel::Auto);
                }
            }
        };
    }

    #[cfg(debug_assertions)]
    fn check_batch_no_poison(test: &str, kernel: Kernel) -> Result<(), Box<dyn std::error::Error>> {
        skip_if_unsupported!(kernel, test);

        let file = "src/data/2018-09-01-2024-Bitfinex_Spot-4h.csv";
        let c = read_candles_from_csv(file)?;

        let batch_configs = vec![
            ((10, 30, 10), (1, 4, 1)),
            ((5, 5, 0), (1, 1, 0)),
            ((100, 120, 20), (2, 4, 2)),
            ((7, 21, 7), (1, 3, 2)),
            ((15, 45, 15), (1, 4, 3)),
            ((3, 12, 3), (1, 2, 1)),
        ];

        for ((p_start, p_end, p_step), (poles_start, poles_end, poles_step)) in batch_configs {
            let output = GaussianBatchBuilder::new()
                .kernel(kernel)
                .period_range(p_start, p_end, p_step)
                .poles_range(poles_start, poles_end, poles_step)
                .apply_candles(&c, "close")?;

            for (idx, &val) in output.values.iter().enumerate() {
                if val.is_nan() {
                    continue;
                }

                let bits = val.to_bits();
                let row = idx / output.cols;
                let col = idx % output.cols;
                let combo = &output.combos[row];

                if bits == 0x11111111_11111111 {
                    panic!(
						"[{}] Found alloc_with_nan_prefix poison value {} (0x{:016X}) at row {} col {} \
                         (flat index {}) with params period={:?}, poles={:?}",
						test, val, bits, row, col, idx, combo.period, combo.poles
					);
                }

                if bits == 0x22222222_22222222 {
                    panic!(
						"[{}] Found init_matrix_prefixes poison value {} (0x{:016X}) at row {} col {} \
                         (flat index {}) with params period={:?}, poles={:?}",
						test, val, bits, row, col, idx, combo.period, combo.poles
					);
                }

                if bits == 0x33333333_33333333 {
                    panic!(
						"[{}] Found make_uninit_matrix poison value {} (0x{:016X}) at row {} col {} \
                         (flat index {}) with params period={:?}, poles={:?}",
						test, val, bits, row, col, idx, combo.period, combo.poles
					);
                }
            }
        }

        Ok(())
    }

    #[cfg(not(debug_assertions))]
    fn check_batch_no_poison(
        _test: &str,
        _kernel: Kernel,
    ) -> Result<(), Box<dyn std::error::Error>> {
        Ok(())
    }

    gen_batch_tests!(check_batch_default_row);
    gen_batch_tests!(check_batch_no_poison);
}

#[cfg(feature = "python")]
use crate::utilities::kernel_validation::validate_kernel;
#[cfg(feature = "python")]
use numpy::{IntoPyArray, PyArray1, PyArrayMethods, PyReadonlyArray1};
#[cfg(feature = "python")]
use pyo3::exceptions::PyValueError;
#[cfg(feature = "python")]
use pyo3::prelude::*;
#[cfg(feature = "python")]
use pyo3::types::PyDict;

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
use serde::{Deserialize, Serialize};
#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
use wasm_bindgen::prelude::*;

#[inline(always)]
fn gaussian_batch_inner_into(
    data: &[f64],
    sweep: &GaussianBatchRange,
    kern: Kernel,
    parallel: bool,
    out: &mut [f64],
) -> Result<Vec<GaussianParams>, GaussianError> {
    let combos = expand_grid(sweep)?;
    if combos.is_empty() || data.is_empty() {
        return Err(GaussianError::NoData);
    }
    let cols = data.len();
    let rows = combos.len();
    let expected = rows
        .checked_mul(cols)
        .ok_or(GaussianError::SizeOverflow { rows, cols })?;
    if out.len() != expected {
        return Err(GaussianError::OutputLengthMismatch {
            expected,
            got: out.len(),
        });
    }

    let first_valid = data
        .iter()
        .position(|x| !x.is_nan())
        .ok_or(GaussianError::AllValuesNaN)?;

    for c in &combos {
        let period = c.period.unwrap_or(14);
        let poles = c.poles.unwrap_or(4);

        if period == 1 {
            return Err(GaussianError::PeriodOneDegenerate);
        }

        if period < 2 {
            return Err(GaussianError::DegeneratePeriod { period });
        }

        if period > cols {
            return Err(GaussianError::PeriodLongerThanData {
                period,
                data_len: cols,
            });
        }
        if !(1..=4).contains(&poles) {
            return Err(GaussianError::InvalidPoles { poles });
        }
        if cols - first_valid < period {
            return Err(GaussianError::NotEnoughValidData {
                needed: period,
                valid: cols - first_valid,
            });
        }
    }

    let warm: Vec<usize> = combos
        .iter()
        .map(|c| {
            let w = first_valid + c.period.unwrap_or(14);
            w.min(cols)
        })
        .collect();

    let raw = unsafe {
        std::slice::from_raw_parts_mut(out.as_mut_ptr() as *mut MaybeUninit<f64>, out.len())
    };

    unsafe { init_matrix_prefixes(raw, cols, &warm) };

    let chosen = match kern {
        Kernel::Auto => Kernel::Scalar,
        other => other,
    };

    type RowRunner = unsafe fn(&[f64], &GaussianParams, &mut [f64]);
    let row_runner: RowRunner = match chosen {
        #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
        Kernel::Avx512 => gaussian_row_avx512,
        #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
        Kernel::Avx2 => gaussian_row_avx2,
        _ => gaussian_row_scalar,
    };

    let do_row = |row: usize, dst_mu: &mut [MaybeUninit<f64>]| unsafe {
        let dst = std::slice::from_raw_parts_mut(dst_mu.as_mut_ptr() as *mut f64, dst_mu.len());
        row_runner(data, &combos[row], dst);
    };

    if parallel {
        #[cfg(not(target_arch = "wasm32"))]
        {
            raw.par_chunks_mut(cols)
                .enumerate()
                .for_each(|(row, slice)| do_row(row, slice));
        }
        #[cfg(target_arch = "wasm32")]
        {
            for (row, slice) in raw.chunks_mut(cols).enumerate() {
                do_row(row, slice);
            }
        }
    } else {
        for (row, slice) in raw.chunks_mut(cols).enumerate() {
            do_row(row, slice);
        }
    }

    Ok(combos)
}

#[cfg(feature = "python")]
#[pyfunction(name = "gaussian")]
#[pyo3(signature = (data, period, poles, kernel=None))]
pub fn gaussian_py<'py>(
    py: Python<'py>,
    data: PyReadonlyArray1<'py, f64>,
    period: usize,
    poles: usize,
    kernel: Option<&str>,
) -> PyResult<Bound<'py, PyArray1<f64>>> {
    use numpy::{IntoPyArray, PyArray1, PyArrayMethods};

    let slice_in = data.as_slice()?;
    let kern = validate_kernel(kernel, false)?;

    let params = GaussianParams {
        period: Some(period),
        poles: Some(poles),
    };
    let gaussian_in = GaussianInput::from_slice(slice_in, params);

    let result_vec: Vec<f64> = py
        .allow_threads(|| gaussian_with_kernel(&gaussian_in, kern).map(|o| o.values))
        .map_err(|e| PyValueError::new_err(e.to_string()))?;

    Ok(result_vec.into_pyarray(py))
}

#[cfg(feature = "python")]
#[pyclass(name = "GaussianStream")]
pub struct GaussianStreamPy {
    stream: GaussianStream,
}

#[cfg(feature = "python")]
#[pymethods]
impl GaussianStreamPy {
    #[new]
    fn new(period: usize, poles: usize) -> PyResult<Self> {
        let params = GaussianParams {
            period: Some(period),
            poles: Some(poles),
        };
        let stream =
            GaussianStream::try_new(params).map_err(|e| PyValueError::new_err(e.to_string()))?;
        Ok(GaussianStreamPy { stream })
    }

    fn update(&mut self, value: f64) -> Option<f64> {
        Some(self.stream.update(value))
    }
}

#[cfg(feature = "python")]
#[pyfunction(name = "gaussian_batch")]
#[pyo3(signature = (data, period_range, poles_range, kernel=None))]
pub fn gaussian_batch_py<'py>(
    py: Python<'py>,
    data: PyReadonlyArray1<'py, f64>,
    period_range: (usize, usize, usize),
    poles_range: (usize, usize, usize),
    kernel: Option<&str>,
) -> PyResult<Bound<'py, PyDict>> {
    use numpy::{IntoPyArray, PyArray1, PyArrayMethods};
    use pyo3::types::PyDict;

    let slice_in = data.as_slice()?;
    let kern = validate_kernel(kernel, true)?;

    let sweep = GaussianBatchRange {
        period: period_range,
        poles: poles_range,
    };

    let combos = expand_grid(&sweep).map_err(|e| PyValueError::new_err(e.to_string()))?;
    let rows = combos.len();
    let cols = slice_in.len();

    let out_arr = unsafe { PyArray1::<f64>::new(py, [rows * cols], false) };
    let slice_out = unsafe { out_arr.as_slice_mut()? };

    let combos = py
        .allow_threads(|| {
            let kernel = match kern {
                Kernel::Auto => match detect_best_batch_kernel() {
                    Kernel::Avx512Batch => Kernel::Avx2Batch,
                    other => other,
                },
                k => k,
            };
            let simd = match kernel {
                Kernel::Avx512Batch => Kernel::Avx512,
                Kernel::Avx2Batch => Kernel::Avx2,
                Kernel::ScalarBatch => Kernel::Scalar,
                _ => unreachable!(),
            };
            gaussian_batch_inner_into(slice_in, &sweep, simd, true, slice_out)
        })
        .map_err(|e| PyValueError::new_err(e.to_string()))?;

    let dict = PyDict::new(py);
    dict.set_item("values", out_arr.reshape((rows, cols))?)?;
    dict.set_item(
        "periods",
        combos
            .iter()
            .map(|p| p.period.unwrap() as u64)
            .collect::<Vec<_>>()
            .into_pyarray(py),
    )?;
    dict.set_item(
        "poles",
        combos
            .iter()
            .map(|p| p.poles.unwrap_or(4) as u64)
            .collect::<Vec<_>>()
            .into_pyarray(py),
    )?;

    Ok(dict)
}

#[cfg(all(feature = "python", feature = "cuda"))]
#[pyfunction(name = "gaussian_cuda_batch_dev")]
#[pyo3(signature = (data_f32, period_range, poles_range, device_id=0))]
pub fn gaussian_cuda_batch_dev_py(
    py: Python<'_>,
    data_f32: numpy::PyReadonlyArray1<'_, f32>,
    period_range: (usize, usize, usize),
    poles_range: (usize, usize, usize),
    device_id: usize,
) -> PyResult<DeviceArrayF32Py> {
    if !cuda_available() {
        return Err(PyValueError::new_err("CUDA not available"));
    }

    let slice = data_f32.as_slice()?;
    let sweep = GaussianBatchRange {
        period: period_range,
        poles: poles_range,
    };

    let cuda = CudaGaussian::new(device_id).map_err(|e| PyValueError::new_err(e.to_string()))?;
    let stream = cuda.stream_handle();
    let dev_id = cuda.device_id();
    let ctx_guard = cuda.context_arc();
    let inner = py
        .allow_threads(|| cuda.gaussian_batch_dev(slice, &sweep))
        .map_err(|e| PyValueError::new_err(e.to_string()))?;

    Ok(DeviceArrayF32Py::new_from_rust(
        inner, stream, ctx_guard, dev_id,
    ))
}

#[cfg(all(feature = "python", feature = "cuda"))]
#[pyfunction(name = "gaussian_cuda_many_series_one_param_dev")]
#[pyo3(signature = (prices_tm_f32, period, poles, device_id=0))]
pub fn gaussian_cuda_many_series_one_param_dev_py(
    py: Python<'_>,
    prices_tm_f32: numpy::PyReadonlyArray2<'_, f32>,
    period: usize,
    poles: usize,
    device_id: usize,
) -> PyResult<DeviceArrayF32Py> {
    use numpy::PyUntypedArrayMethods;

    if !cuda_available() {
        return Err(PyValueError::new_err("CUDA not available"));
    }

    let shape = prices_tm_f32.shape();
    let rows = shape[0];
    let cols = shape[1];

    let flat = prices_tm_f32.as_slice()?;
    let params = GaussianParams {
        period: Some(period),
        poles: Some(poles),
    };

    let cuda = CudaGaussian::new(device_id).map_err(|e| PyValueError::new_err(e.to_string()))?;
    let stream = cuda.stream_handle();
    let dev_id = cuda.device_id();
    let ctx_guard = cuda.context_arc();
    let inner = py
        .allow_threads(|| {
            cuda.gaussian_many_series_one_param_time_major_dev(flat, cols, rows, &params)
        })
        .map_err(|e| PyValueError::new_err(e.to_string()))?;

    Ok(DeviceArrayF32Py::new_from_rust(
        inner, stream, ctx_guard, dev_id,
    ))
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen]
pub fn gaussian_js(data: &[f64], period: usize, poles: usize) -> Result<Vec<f64>, JsValue> {
    let params = GaussianParams {
        period: Some(period),
        poles: Some(poles),
    };
    let input = GaussianInput::from_slice(data, params);

    let mut output = vec![0.0; data.len()];

    gaussian_into_slice(&mut output, &input, Kernel::Auto)
        .map_err(|e| JsValue::from_str(&e.to_string()))?;

    Ok(output)
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen]
pub fn gaussian_into(
    in_ptr: *const f64,
    out_ptr: *mut f64,
    len: usize,
    period: usize,
    poles: usize,
) -> Result<(), JsValue> {
    if in_ptr.is_null() || out_ptr.is_null() {
        return Err(JsValue::from_str("null pointer passed to gaussian_into"));
    }

    unsafe {
        let data = std::slice::from_raw_parts(in_ptr, len);

        if period == 1 {
            return Err(JsValue::from_str("Period of 1 causes degenerate filter"));
        }
        if period < 2 {
            return Err(JsValue::from_str(&format!(
                "Period must be >= 2, got {}",
                period
            )));
        }
        if period > len {
            return Err(JsValue::from_str(&format!(
                "Period {} is longer than data length {}",
                period, len
            )));
        }

        if !(1..=4).contains(&poles) {
            return Err(JsValue::from_str("Invalid poles (must be 1-4)"));
        }

        let params = GaussianParams {
            period: Some(period),
            poles: Some(poles),
        };
        let input = GaussianInput::from_slice(data, params);

        if in_ptr == out_ptr {
            let mut temp = vec![0.0; len];
            gaussian_into_slice(&mut temp, &input, Kernel::Auto)
                .map_err(|e| JsValue::from_str(&e.to_string()))?;

            let out = std::slice::from_raw_parts_mut(out_ptr, len);
            out.copy_from_slice(&temp);
        } else {
            let out = std::slice::from_raw_parts_mut(out_ptr, len);
            gaussian_into_slice(out, &input, Kernel::Auto)
                .map_err(|e| JsValue::from_str(&e.to_string()))?;
        }

        Ok(())
    }
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen]
pub fn gaussian_alloc(len: usize) -> *mut f64 {
    let mut vec = Vec::<f64>::with_capacity(len);
    let ptr = vec.as_mut_ptr();
    std::mem::forget(vec);
    ptr
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen]
pub fn gaussian_free(ptr: *mut f64, len: usize) {
    if !ptr.is_null() {
        unsafe {
            let _ = Vec::from_raw_parts(ptr, len, len);
        }
    }
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen]
pub fn gaussian_batch_js(
    data: &[f64],
    period_start: usize,
    period_end: usize,
    period_step: usize,
    poles_start: usize,
    poles_end: usize,
    poles_step: usize,
) -> Result<Vec<f64>, JsValue> {
    let sweep = GaussianBatchRange {
        period: (period_start, period_end, period_step),
        poles: (poles_start, poles_end, poles_step),
    };

    gaussian_batch_inner(data, &sweep, Kernel::Auto, false)
        .map(|output| output.values)
        .map_err(|e| JsValue::from_str(&e.to_string()))
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen]
pub fn gaussian_batch_metadata_js(
    period_start: usize,
    period_end: usize,
    period_step: usize,
    poles_start: usize,
    poles_end: usize,
    poles_step: usize,
) -> Result<Vec<f64>, JsValue> {
    let sweep = GaussianBatchRange {
        period: (period_start, period_end, period_step),
        poles: (poles_start, poles_end, poles_step),
    };

    let combos = expand_grid(&sweep).map_err(|e| JsValue::from_str(&e.to_string()))?;
    let metadata: Vec<f64> = combos
        .iter()
        .flat_map(|combo| {
            vec![
                combo.period.unwrap_or(14) as f64,
                combo.poles.unwrap_or(4) as f64,
            ]
        })
        .collect();

    Ok(metadata)
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[derive(Serialize, Deserialize)]
pub struct GaussianBatchConfig {
    pub period_range: (usize, usize, usize),
    pub poles_range: (usize, usize, usize),
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[derive(Serialize, Deserialize)]
pub struct GaussianBatchJsOutput {
    pub values: Vec<f64>,
    pub combos: Vec<GaussianParams>,
    pub rows: usize,
    pub cols: usize,
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen(js_name = gaussian_batch)]
pub fn gaussian_batch_unified_js(data: &[f64], config: JsValue) -> Result<JsValue, JsValue> {
    let config: GaussianBatchConfig = serde_wasm_bindgen::from_value(config)
        .map_err(|e| JsValue::from_str(&format!("Invalid config: {}", e)))?;

    let sweep = GaussianBatchRange {
        period: config.period_range,
        poles: config.poles_range,
    };

    let output = gaussian_batch_inner(data, &sweep, Kernel::Auto, false)
        .map_err(|e| JsValue::from_str(&e.to_string()))?;

    let js_output = GaussianBatchJsOutput {
        values: output.values,
        combos: output.combos,
        rows: output.rows,
        cols: output.cols,
    };

    serde_wasm_bindgen::to_value(&js_output)
        .map_err(|e| JsValue::from_str(&format!("Serialization error: {}", e)))
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen]
pub fn gaussian_batch_into(
    in_ptr: *const f64,
    out_ptr: *mut f64,
    len: usize,
    period_start: usize,
    period_end: usize,
    period_step: usize,
    poles_start: usize,
    poles_end: usize,
    poles_step: usize,
) -> Result<usize, JsValue> {
    if in_ptr.is_null() || out_ptr.is_null() {
        return Err(JsValue::from_str(
            "null pointer passed to gaussian_batch_into",
        ));
    }

    unsafe {
        let data = std::slice::from_raw_parts(in_ptr, len);

        let sweep = GaussianBatchRange {
            period: (period_start, period_end, period_step),
            poles: (poles_start, poles_end, poles_step),
        };

        let combos = expand_grid(&sweep).map_err(|e| JsValue::from_str(&e.to_string()))?;
        let rows = combos.len();
        let cols = len;

        let out = std::slice::from_raw_parts_mut(out_ptr, rows * cols);

        gaussian_batch_inner_into(data, &sweep, Kernel::Auto, false, out)
            .map_err(|e| JsValue::from_str(&e.to_string()))?;

        Ok(rows)
    }
}