vector_ta/indicators/moving_averages/

frama.rs

#[cfg(all(feature = "python", feature = "cuda"))]
use crate::cuda::moving_averages::CudaFrama;
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(all(feature = "nightly-avx", target_arch = "x86_64"))]
use core::arch::x86_64::*;
#[cfg(not(target_arch = "wasm32"))]
use rayon::prelude::*;
use std::collections::VecDeque;
use std::convert::AsRef;
use std::error::Error;
#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
use std::hint::unlikely;
use std::mem::{swap, MaybeUninit};
use thiserror::Error;

impl<'a> AsRef<[f64]> for FramaInput<'a> {
    #[inline(always)]
    fn as_ref(&self) -> &[f64] {
        match &self.data {
            FramaData::Candles { candles } => candles.select_candle_field("close").unwrap(),
            FramaData::Slices { close, .. } => close,
        }
    }
}

#[inline(always)]
unsafe fn seed_sma(close: &[f64], first: usize, win: usize, out: &mut [f64]) {
    let mut sum = 0.0;
    for k in 0..win {
        sum += *close.get_unchecked(first + k);
    }
    *out.get_unchecked_mut(first + win - 1) = sum / win as f64;
}

#[derive(Debug, Clone)]
pub enum FramaData<'a> {
    Candles {
        candles: &'a Candles,
    },
    Slices {
        high: &'a [f64],
        low: &'a [f64],
        close: &'a [f64],
    },
}

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

#[derive(Debug, Clone)]
#[cfg_attr(
    all(target_arch = "wasm32", feature = "wasm"),
    derive(Serialize, Deserialize)
)]
pub struct FramaParams {
    pub window: Option<usize>,
    pub sc: Option<usize>,
    pub fc: Option<usize>,
}

impl Default for FramaParams {
    fn default() -> Self {
        Self {
            window: Some(10),
            sc: Some(300),
            fc: Some(1),
        }
    }
}

#[derive(Debug, Clone)]
pub struct FramaInput<'a> {
    pub data: FramaData<'a>,
    pub params: FramaParams,
}

impl<'a> FramaInput<'a> {
    #[inline]
    pub fn from_candles(candles: &'a Candles, params: FramaParams) -> Self {
        Self {
            data: FramaData::Candles { candles },
            params,
        }
    }
    #[inline]
    pub fn from_slices(
        high: &'a [f64],
        low: &'a [f64],
        close: &'a [f64],
        params: FramaParams,
    ) -> Self {
        Self {
            data: FramaData::Slices { high, low, close },
            params,
        }
    }
    #[inline]
    pub fn with_default_candles(candles: &'a Candles) -> Self {
        Self::from_candles(candles, FramaParams::default())
    }
    #[inline]
    pub fn get_window(&self) -> usize {
        self.params.window.unwrap_or(10)
    }
    #[inline]
    pub fn get_sc(&self) -> usize {
        self.params.sc.unwrap_or(300)
    }
    #[inline]
    pub fn get_fc(&self) -> usize {
        self.params.fc.unwrap_or(1)
    }

    #[inline]
    pub fn slices(&self) -> (&'a [f64], &'a [f64], &'a [f64]) {
        match &self.data {
            FramaData::Candles { candles } => (
                candles.select_candle_field("high").unwrap(),
                candles.select_candle_field("low").unwrap(),
                candles.select_candle_field("close").unwrap(),
            ),
            FramaData::Slices { high, low, close } => (*high, *low, *close),
        }
    }
}

#[derive(Copy, Clone, Debug)]
pub struct FramaBuilder {
    window: Option<usize>,
    sc: Option<usize>,
    fc: Option<usize>,
    kernel: Kernel,
}

impl Default for FramaBuilder {
    fn default() -> Self {
        Self {
            window: None,
            sc: None,
            fc: None,
            kernel: Kernel::Auto,
        }
    }
}
impl FramaBuilder {
    #[inline(always)]
    pub fn new() -> Self {
        Self::default()
    }
    #[inline(always)]
    pub fn window(mut self, n: usize) -> Self {
        self.window = Some(n);
        self
    }
    #[inline(always)]
    pub fn sc(mut self, x: usize) -> Self {
        self.sc = Some(x);
        self
    }
    #[inline(always)]
    pub fn fc(mut self, x: usize) -> Self {
        self.fc = Some(x);
        self
    }
    #[inline(always)]
    pub fn kernel(mut self, k: Kernel) -> Self {
        self.kernel = k;
        self
    }
    #[inline(always)]
    pub fn apply(self, c: &Candles) -> Result<FramaOutput, FramaError> {
        let p = FramaParams {
            window: self.window,
            sc: self.sc,
            fc: self.fc,
        };
        let i = FramaInput::from_candles(c, p);
        frama_with_kernel(&i, self.kernel)
    }
    #[inline(always)]
    pub fn apply_slices(
        self,
        high: &[f64],
        low: &[f64],
        close: &[f64],
    ) -> Result<FramaOutput, FramaError> {
        let p = FramaParams {
            window: self.window,
            sc: self.sc,
            fc: self.fc,
        };
        let i = FramaInput::from_slices(high, low, close, p);
        frama_with_kernel(&i, self.kernel)
    }
    #[inline(always)]
    pub fn into_stream(self) -> Result<FramaStream, FramaError> {
        let p = FramaParams {
            window: self.window,
            sc: self.sc,
            fc: self.fc,
        };
        FramaStream::try_new(p)
    }
}

#[derive(Debug, Error)]
pub enum FramaError {
    #[error("frama: Input data slice is empty.")]
    EmptyInputData,

    #[error("frama: Mismatched slice lengths: high={high}, low={low}, close={close}")]
    MismatchedInputLength {
        high: usize,
        low: usize,
        close: usize,
    },
    #[error("frama: All values are NaN.")]
    AllValuesNaN,
    #[error("frama: Invalid window: window = {window}, data length = {data_len}")]
    InvalidWindow { window: usize, data_len: usize },
    #[error("frama: Not enough valid data: needed = {needed}, valid = {valid}")]
    NotEnoughValidData { needed: usize, valid: usize },

    #[error("frama: Output slice length mismatch: expected = {expected}, got = {got}")]
    OutputLengthMismatch { expected: usize, got: usize },

    #[error("frama: Invalid range: start={start}, end={end}, step={step}")]
    InvalidRange {
        start: usize,
        end: usize,
        step: usize,
    },

    #[error("frama: Invalid kernel for batch API: {0:?}")]
    InvalidKernelForBatch(Kernel),

    #[error("frama: Invalid smoothing constants: sc={sc}, fc={fc}")]
    InvalidSmoothing { sc: usize, fc: usize },

    #[error("frama: arithmetic overflow while computing {context}")]
    ArithmeticOverflow { context: &'static str },
}

#[inline]
pub fn frama(input: &FramaInput) -> Result<FramaOutput, FramaError> {
    frama_with_kernel(input, Kernel::Auto)
}

#[inline(always)]
fn frama_prepare<'a>(
    input: &'a FramaInput,
    kernel: Kernel,
) -> Result<
    (
        (&'a [f64], &'a [f64], &'a [f64]),
        usize,
        usize,
        usize,
        usize,
        usize,
        usize,
        Kernel,
    ),
    FramaError,
> {
    let (high, low, close) = input.slices();
    let len = high.len();
    if len == 0 {
        return Err(FramaError::EmptyInputData);
    }
    if low.len() != len || close.len() != len {
        return Err(FramaError::MismatchedInputLength {
            high: len,
            low: low.len(),
            close: close.len(),
        });
    }
    let window = input.get_window();
    let sc = input.get_sc();
    let fc = input.get_fc();
    if sc == 0 || fc == 0 {
        return Err(FramaError::InvalidSmoothing { sc, fc });
    }
    let first = (0..len)
        .find(|&i| !high[i].is_nan() && !low[i].is_nan() && !close[i].is_nan())
        .ok_or(FramaError::AllValuesNaN)?;
    if window == 0 || window > len {
        return Err(FramaError::InvalidWindow {
            window,
            data_len: len,
        });
    }

    let mut win = window;
    if win & 1 == 1 {
        win += 1;
    }

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

    let chosen = match kernel {
        Kernel::Auto => match detect_best_kernel() {
            Kernel::Avx512 => Kernel::Avx2,
            k => k,
        },
        other => other,
    };

    let warm = first + win - 1;

    Ok(((high, low, close), window, sc, fc, first, len, warm, chosen))
}

#[inline(always)]
fn frama_compute_into(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    window: usize,
    sc: usize,
    fc: usize,
    first: usize,
    len: usize,
    warm: usize,
    chosen: Kernel,
    out: &mut [f64],
) -> Result<(), FramaError> {
    let mut win = window;
    if win & 1 == 1 {
        win += 1;
    }
    let seed = close[first..first + win].iter().sum::<f64>() / win as f64;
    out[first + win - 1] = seed;

    match chosen {
        Kernel::Scalar | Kernel::ScalarBatch => {
            if win <= 32 {
                unsafe {
                    frama_small_scan(high, low, close, win, sc, fc, first, len, out)?;
                }
            } else {
                frama_scalar_deque(high, low, close, win, sc, fc, first, len, out)?;
            }
        }

        #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
        Kernel::Avx2 | Kernel::Avx2Batch => unsafe {
            frama_avx2_into(high, low, close, win, sc, fc, first, len, out)?;
        },

        #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
        Kernel::Avx512 | Kernel::Avx512Batch => unsafe {
            frama_avx512_into(high, low, close, win, sc, fc, first, len, out)?;
        },

        _ => unreachable!("`Auto` must be resolved above"),
    }

    Ok(())
}

pub fn frama_with_kernel(input: &FramaInput, kernel: Kernel) -> Result<FramaOutput, FramaError> {
    let ((high, low, close), window, sc, fc, first, len, warm, chosen) =
        frama_prepare(input, kernel)?;
    let mut out = alloc_with_nan_prefix(len, warm);
    frama_compute_into(
        high, low, close, window, sc, fc, first, len, warm, chosen, &mut out,
    )?;
    Ok(FramaOutput { values: out })
}

#[cfg(not(all(target_arch = "wasm32", feature = "wasm")))]
#[inline]
pub fn frama_into(input: &FramaInput, out: &mut [f64]) -> Result<(), FramaError> {
    frama_into_slice(out, input, Kernel::Auto)
}

#[derive(Copy, Clone)]
struct MonoDeque<const CAP: usize> {
    buf: [usize; CAP],
    head: usize,
    tail: usize,
}
impl<const CAP: usize> MonoDeque<CAP> {
    #[inline(always)]
    const fn new() -> Self {
        Self {
            buf: [0; CAP],
            head: 0,
            tail: 0,
        }
    }
    #[inline(always)]
    fn clear(&mut self) {
        self.head = 0;
        self.tail = 0;
    }
    #[inline(always)]
    fn is_empty(&self) -> bool {
        self.head == self.tail
    }

    #[inline(always)]
    unsafe fn front(&self) -> usize {
        *self.buf.get_unchecked(self.head)
    }

    #[inline(always)]
    fn expire(&mut self, idx_out: usize) {
        if !self.is_empty() && unsafe { self.front() } == idx_out {
            self.head = (self.head + 1) % CAP;
        }
    }

    #[inline(always)]
    unsafe fn push_max(&mut self, idx: usize, data: &[f64]) {
        while !self.is_empty() {
            let last = self.buf[(self.tail + CAP - 1) % CAP];
            if *data.get_unchecked(last) >= *data.get_unchecked(idx) {
                break;
            }
            self.tail = (self.tail + CAP - 1) % CAP;
        }
        self.buf[self.tail] = idx;
        self.tail = (self.tail + 1) % CAP;
    }

    #[inline(always)]
    unsafe fn push_min(&mut self, idx: usize, data: &[f64]) {
        while !self.is_empty() {
            let last = self.buf[(self.tail + CAP - 1) % CAP];
            if *data.get_unchecked(last) <= *data.get_unchecked(idx) {
                break;
            }
            self.tail = (self.tail + CAP - 1) % CAP;
        }
        self.buf[self.tail] = idx;
        self.tail = (self.tail + 1) % CAP;
    }
}

#[inline(always)]
fn frama_scalar_deque(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    mut window: usize,
    sc: usize,
    fc: usize,
    first: usize,
    len: usize,
    out: &mut [f64],
) -> Result<(), FramaError> {
    if window & 1 == 1 {
        window += 1;
    }
    let half = window / 2;
    const MAX_W: usize = 1024;
    assert!(window <= MAX_W, "window bigger than CAP");

    let mut d_full_max: MonoDeque<MAX_W> = MonoDeque::new();
    let mut d_full_min: MonoDeque<MAX_W> = MonoDeque::new();
    let mut d_left_max: MonoDeque<MAX_W> = MonoDeque::new();
    let mut d_left_min: MonoDeque<MAX_W> = MonoDeque::new();
    let mut d_right_max: MonoDeque<MAX_W> = MonoDeque::new();
    let mut d_right_min: MonoDeque<MAX_W> = MonoDeque::new();

    unsafe {
        for idx in first..(first + window) {
            if !high[idx].is_nan() && !low[idx].is_nan() {
                d_full_max.push_max(idx, high);
                d_full_min.push_min(idx, low);
                if idx < first + half {
                    d_left_max.push_max(idx, high);
                    d_left_min.push_min(idx, low);
                } else {
                    d_right_max.push_max(idx, high);
                    d_right_min.push_min(idx, low);
                }
            }
        }
    }

    let w_ln = (2.0 / (sc as f64 + 1.0)).ln();
    let sc_lim = 2.0 / (sc as f64 + 1.0);
    let mut d_prev = 1.0;

    let mut pm1 = f64::NAN;
    let mut pm2 = f64::NAN;
    let mut pm3 = f64::NAN;
    let mut pn1 = f64::NAN;
    let mut pn2 = f64::NAN;
    let mut pn3 = f64::NAN;

    let mut half_progress = 0usize;

    for i in (first + window)..len {
        let idx_out = i - window;
        d_full_max.expire(idx_out);
        d_full_min.expire(idx_out);
        d_left_max.expire(idx_out);
        d_left_min.expire(idx_out);
        d_right_max.expire(idx_out + half);
        d_right_min.expire(idx_out + half);

        let newest = i - 1;
        if !high[newest].is_nan() && !low[newest].is_nan() {
            unsafe {
                d_full_max.push_max(newest, high);
                d_full_min.push_min(newest, low);

                if newest < (idx_out + half) {
                    d_left_max.push_max(newest, high);
                    d_left_min.push_min(newest, low);
                } else {
                    d_right_max.push_max(newest, high);
                    d_right_min.push_min(newest, low);
                }
            }
        }
        fn front_or(
            dq_max: &MonoDeque<MAX_W>,
            dq_min: &MonoDeque<MAX_W>,
            prev_max: &mut f64,
            prev_min: &mut f64,
            high: &[f64],
            low: &[f64],
        ) -> (f64, f64) {
            let maxv = if !dq_max.is_empty() {
                high[unsafe { dq_max.front() }]
            } else {
                *prev_max
            };
            let minv = if !dq_min.is_empty() {
                low[unsafe { dq_min.front() }]
            } else {
                *prev_min
            };
            *prev_max = maxv;
            *prev_min = minv;
            (maxv, minv)
        }
        let (max1, min1) = front_or(&d_right_max, &d_right_min, &mut pm1, &mut pn1, high, low);
        let (max2, min2) = front_or(&d_left_max, &d_left_min, &mut pm2, &mut pn2, high, low);
        let (max3, min3) = front_or(&d_full_max, &d_full_min, &mut pm3, &mut pn3, high, low);

        if !(high[i].is_nan() || low[i].is_nan() || close[i].is_nan()) {
            let n1 = (max1 - min1) / (half as f64);
            let n2 = (max2 - min2) / (half as f64);
            let n3 = (max3 - min3) / (window as f64);

            let d_cur = if n1 > 0.0 && n2 > 0.0 && n3 > 0.0 {
                ((n1 + n2).ln() - n3.ln()) / std::f64::consts::LN_2
            } else {
                d_prev
            };
            d_prev = d_cur;

            let mut alpha0 = (w_ln * (d_cur - 1.0)).exp();
            if alpha0 < 0.1 {
                alpha0 = 0.1;
            }
            if alpha0 > 1.0 {
                alpha0 = 1.0;
            }
            let old_n = (2.0 - alpha0) / alpha0;
            let new_n = (sc - fc) as f64 * ((old_n - 1.0) / (sc as f64 - 1.0)) + fc as f64;
            let mut alpha = 2.0 / (new_n + 1.0);
            if alpha < sc_lim {
                alpha = sc_lim;
            }
            if alpha > 1.0 {
                alpha = 1.0;
            }

            out[i] = alpha * close[i] + (1.0 - alpha) * out[i - 1];
        } else {
            out[i] = out[i - 1];
        }

        half_progress += 1;
        if half_progress == half {
            swap(&mut d_left_max, &mut d_right_max);
            swap(&mut d_left_min, &mut d_right_min);
            d_right_max.clear();
            d_right_min.clear();
            half_progress = 0;
        }
    }

    Ok(())
}

#[inline(always)]
pub fn frama_scalar(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    window: usize,
    sc: usize,
    fc: usize,
    first: usize,
    len: usize,
) -> Result<FramaOutput, FramaError> {
    let mut win = window;
    if win & 1 == 1 {
        win += 1;
    }
    let warm = first + win - 1;

    let mut out = alloc_with_nan_prefix(len, warm);
    frama_compute_into(
        high,
        low,
        close,
        window,
        sc,
        fc,
        first,
        len,
        warm,
        Kernel::Scalar,
        &mut out,
    )?;
    Ok(FramaOutput { values: out })
}

#[inline(always)]
unsafe fn frama_small_scan(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    win: usize,
    sc: usize,
    fc: usize,
    first: usize,
    len: usize,
    out: &mut [f64],
) -> Result<(), FramaError> {
    let half = win >> 1;
    let win_f64 = win as f64;
    let half_f64 = half as f64;
    let w_ln = (2.0 / (sc as f64 + 1.0)).ln();
    let sc_floor = 2.0 / (sc as f64 + 1.0);
    let mut d_prev = 1.0_f64;

    for i in (first + win)..len {
        let seg_start = i - win;
        let mid = seg_start + half;

        let mut max1 = f64::MIN;
        let mut min1 = f64::MAX;
        let mut max2 = f64::MIN;
        let mut min2 = f64::MAX;

        let mut j = seg_start;
        while j + 1 < mid {
            let h0 = *high.get_unchecked(j);
            let h1 = *high.get_unchecked(j + 1);
            let l0 = *low.get_unchecked(j);
            let l1 = *low.get_unchecked(j + 1);
            max2 = f64::max(max2, f64::max(h0, h1));
            min2 = f64::min(min2, f64::min(l0, l1));
            j += 2;
        }
        if j < mid {
            max2 = f64::max(max2, *high.get_unchecked(j));
            min2 = f64::min(min2, *low.get_unchecked(j));
        }

        j = mid;
        while j + 1 < i {
            let h0 = *high.get_unchecked(j);
            let h1 = *high.get_unchecked(j + 1);
            let l0 = *low.get_unchecked(j);
            let l1 = *low.get_unchecked(j + 1);
            max1 = f64::max(max1, f64::max(h0, h1));
            min1 = f64::min(min1, f64::min(l0, l1));
            j += 2;
        }
        if j < i {
            max1 = f64::max(max1, *high.get_unchecked(j));
            min1 = f64::min(min1, *low.get_unchecked(j));
        }

        let max3 = f64::max(max1, max2);
        let min3 = f64::min(min1, min2);

        let n1 = (max1 - min1) / half_f64;
        let n2 = (max2 - min2) / half_f64;
        let n3 = (max3 - min3) / win_f64;

        let d_cur = if n1 > 0.0 && n2 > 0.0 && n3 > 0.0 {
            ((n1 + n2).ln() - n3.ln()) / std::f64::consts::LN_2
        } else {
            d_prev
        };
        d_prev = d_cur;

        let mut alpha0 = (w_ln * (d_cur - 1.0)).exp().clamp(0.1, 1.0);
        let old_n = (2.0 - alpha0) / alpha0;
        let new_n = (sc - fc) as f64 * ((old_n - 1.0) / (sc as f64 - 1.0)) + fc as f64;
        let alpha = (2.0 / (new_n + 1.0)).clamp(sc_floor, 1.0);

        out[i] = (*close.get_unchecked(i)).mul_add(alpha, (1.0 - alpha) * out[i - 1]);
    }
    Ok(())
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline(always)]
unsafe fn hmax_pd256(v: __m256d) -> f64 {
    let hi = _mm256_extractf128_pd::<1>(v);
    let lo = _mm256_castpd256_pd128(v);
    let m = _mm_max_pd(hi, lo);
    let m = _mm_max_pd(m, _mm_permute_pd::<0b01>(m));
    _mm_cvtsd_f64(m)
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline(always)]
unsafe fn hmin_pd256(v: __m256d) -> f64 {
    let hi = _mm256_extractf128_pd::<1>(v);
    let lo = _mm256_castpd256_pd128(v);
    let m = _mm_min_pd(hi, lo);
    let m = _mm_min_pd(m, _mm_permute_pd::<0b01>(m));
    _mm_cvtsd_f64(m)
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[target_feature(enable = "avx2")]
unsafe fn frama_avx2_small<const WIN: usize>(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    sc: usize,
    fc: usize,
    first: usize,
    len: usize,
    out: &mut [f64],
) {
    const LANES: usize = 4;
    const LN2: f64 = std::f64::consts::LN_2;

    let half = WIN / 2;
    let win_f64 = WIN as f64;
    let half_f64 = half as f64;
    let w_ln = (2.0 / (sc as f64 + 1.0)).ln();
    let sc_floor = 2.0 / (sc as f64 + 1.0);
    let mut d_prev = 1.0;

    for i in (first + WIN)..len {
        if i + 1 < len {
            _mm_prefetch(high.as_ptr().add(i + 1 - WIN) as *const i8, _MM_HINT_T0);
            _mm_prefetch(low.as_ptr().add(i + 1 - WIN) as *const i8, _MM_HINT_T0);
            _mm_prefetch(close.as_ptr().add(i + 1) as *const i8, _MM_HINT_T0);
        }

        if unlikely(
            (*high.get_unchecked(i)).is_nan()
                || (*low.get_unchecked(i)).is_nan()
                || (*close.get_unchecked(i)).is_nan(),
        ) {
            *out.get_unchecked_mut(i) = *out.get_unchecked(i - 1);
            continue;
        }

        let mut v_max_l = _mm256_set1_pd(f64::MIN);
        let mut v_min_l = _mm256_set1_pd(f64::MAX);
        let mut idx_l = i - WIN;

        for _ in 0..(half / LANES) {
            let h = _mm256_loadu_pd(high.as_ptr().add(idx_l));
            let l = _mm256_loadu_pd(low.as_ptr().add(idx_l));
            v_max_l = _mm256_max_pd(v_max_l, h);
            v_min_l = _mm256_min_pd(v_min_l, l);
            idx_l += LANES;
        }

        let mut max_l = hmax_pd256(v_max_l);
        let mut min_l = hmin_pd256(v_min_l);

        for j in idx_l..(i - half) {
            let h = *high.get_unchecked(j);
            let l = *low.get_unchecked(j);
            max_l = max_l.max(h);
            min_l = min_l.min(l);
        }

        let mut v_max_r = _mm256_set1_pd(f64::MIN);
        let mut v_min_r = _mm256_set1_pd(f64::MAX);
        let mut idx_r = i - half;

        for _ in 0..(half / LANES) {
            let h = _mm256_loadu_pd(high.as_ptr().add(idx_r));
            let l = _mm256_loadu_pd(low.as_ptr().add(idx_r));
            v_max_r = _mm256_max_pd(v_max_r, h);
            v_min_r = _mm256_min_pd(v_min_r, l);
            idx_r += LANES;
        }

        let mut max_r = hmax_pd256(v_max_r);
        let mut min_r = hmin_pd256(v_min_r);

        for j in idx_r..i {
            let h = *high.get_unchecked(j);
            let l = *low.get_unchecked(j);
            max_r = max_r.max(h);
            min_r = min_r.min(l);
        }

        let max_w = max_l.max(max_r);
        let min_w = min_l.min(min_r);

        let n1 = (max_r - min_r) / half_f64;
        let n2 = (max_l - min_l) / half_f64;
        let n3 = (max_w - min_w) / win_f64;

        let d = if n1 > 0.0 && n2 > 0.0 && n3 > 0.0 {
            ((n1 + n2).ln() - n3.ln()) / LN2
        } else {
            d_prev
        };
        d_prev = d;

        let mut a0 = (w_ln * (d - 1.0)).exp().clamp(0.1, 1.0);
        let old_n = (2.0 - a0) / a0;
        let new_n = (sc - fc) as f64 * ((old_n - 1.0) / (sc as f64 - 1.0)) + fc as f64;
        let alpha = (2.0 / (new_n + 1.0)).clamp(sc_floor, 1.0);

        *out.get_unchecked_mut(i) =
            (*close.get_unchecked(i)).mul_add(alpha, (1.0 - alpha) * *out.get_unchecked(i - 1));
    }
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[target_feature(enable = "avx512f")]
unsafe fn frama_avx512_small<const WIN: usize>(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    sc: usize,
    fc: usize,
    first: usize,
    len: usize,
    out: &mut [f64],
) {
    const LANES: usize = 8;
    const LN2: f64 = std::f64::consts::LN_2;

    let half = WIN / 2;
    let vec_cnt = half / LANES;
    let tail = (half & (LANES - 1)) as i32;
    let mask = (1u8 << tail) - 1;

    let w_ln = (2.0 / (sc as f64 + 1.0)).ln();
    let sc_floor = 2.0 / (sc as f64 + 1.0);
    let win_f64 = WIN as f64;
    let half_f64 = half as f64;

    let v_min_init = _mm512_set1_pd(f64::MIN);
    let v_max_init = _mm512_set1_pd(f64::MAX);

    let mut d_prev = 1.0;

    for i in (first + WIN)..len {
        if i + 1 < len {
            _mm_prefetch(high.as_ptr().add(i + 1 - WIN) as *const i8, _MM_HINT_T0);
            _mm_prefetch(low.as_ptr().add(i + 1 - WIN) as *const i8, _MM_HINT_T0);
            _mm_prefetch(close.as_ptr().add(i + 1) as *const i8, _MM_HINT_T0);
        }

        if unlikely(
            (*high.get_unchecked(i)).is_nan()
                || (*low.get_unchecked(i)).is_nan()
                || (*close.get_unchecked(i)).is_nan(),
        ) {
            *out.get_unchecked_mut(i) = *out.get_unchecked(i - 1);
            continue;
        }

        let mut v_max_l = v_min_init;
        let mut v_min_l = v_max_init;
        let base_l = i - WIN;

        for k in 0..vec_cnt {
            let off = base_l + k * LANES;
            let h = _mm512_loadu_pd(high.as_ptr().add(off));
            let l = _mm512_loadu_pd(low.as_ptr().add(off));
            v_max_l = _mm512_max_pd(v_max_l, h);
            v_min_l = _mm512_min_pd(v_min_l, l);
        }

        if tail != 0 {
            let off = base_l + vec_cnt * LANES;
            let h_tail =
                _mm512_mask_loadu_pd(_mm512_set1_pd(f64::MIN), mask, high.as_ptr().add(off));
            let l_tail =
                _mm512_mask_loadu_pd(_mm512_set1_pd(f64::MAX), mask, low.as_ptr().add(off));
            v_max_l = _mm512_max_pd(v_max_l, h_tail);
            v_min_l = _mm512_min_pd(v_min_l, l_tail);
        }

        let max_l = _mm512_reduce_max_pd(v_max_l);
        let min_l = _mm512_reduce_min_pd(v_min_l);

        let mut v_max_r = v_min_init;
        let mut v_min_r = v_max_init;
        let base_r = i - half;

        for k in 0..vec_cnt {
            let off = base_r + k * LANES;
            let h = _mm512_loadu_pd(high.as_ptr().add(off));
            let l = _mm512_loadu_pd(low.as_ptr().add(off));
            v_max_r = _mm512_max_pd(v_max_r, h);
            v_min_r = _mm512_min_pd(v_min_r, l);
        }

        if tail != 0 {
            let off = base_r + vec_cnt * LANES;
            let h_tail =
                _mm512_mask_loadu_pd(_mm512_set1_pd(f64::MIN), mask, high.as_ptr().add(off));
            let l_tail =
                _mm512_mask_loadu_pd(_mm512_set1_pd(f64::MAX), mask, low.as_ptr().add(off));
            v_max_r = _mm512_max_pd(v_max_r, h_tail);
            v_min_r = _mm512_min_pd(v_min_r, l_tail);
        }

        let max_r = _mm512_reduce_max_pd(v_max_r);
        let min_r = _mm512_reduce_min_pd(v_min_r);

        let max_w = max_l.max(max_r);
        let min_w = min_l.min(min_r);

        let n1 = (max_r - min_r) / half_f64;
        let n2 = (max_l - min_l) / half_f64;
        let n3 = (max_w - min_w) / win_f64;

        let d = if n1 > 0.0 && n2 > 0.0 && n3 > 0.0 {
            ((n1 + n2).ln() - n3.ln()) / LN2
        } else {
            d_prev
        };
        d_prev = d;

        let mut a0 = (w_ln * (d - 1.0)).exp().clamp(0.1, 1.0);
        let old_n = (2.0 - a0) / a0;
        let new_n = (sc - fc) as f64 * ((old_n - 1.0) / (sc as f64 - 1.0)) + fc as f64;
        let alpha = (2.0 / (new_n + 1.0)).clamp(sc_floor, 1.0);

        *out.get_unchecked_mut(i) =
            (*close.get_unchecked(i)).mul_add(alpha, (1.0 - alpha) * *out.get_unchecked(i - 1));
    }
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline(always)]
unsafe fn frama_avx2_into(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    window: usize,
    sc: usize,
    fc: usize,
    first: usize,
    len: usize,
    out: &mut [f64],
) -> Result<(), FramaError> {
    let mut win = window;
    if win & 1 == 1 {
        win += 1;
    }

    if win <= 32 {
        match win {
            10 => unsafe { frama_avx2_small::<10>(high, low, close, sc, fc, first, len, out) },
            14 => unsafe { frama_avx2_small::<14>(high, low, close, sc, fc, first, len, out) },
            20 => unsafe { frama_avx2_small::<20>(high, low, close, sc, fc, first, len, out) },
            32 => unsafe { frama_avx2_small::<32>(high, low, close, sc, fc, first, len, out) },
            _ => unsafe { frama_small_scan(high, low, close, win, sc, fc, first, len, out)? },
        }
    } else {
        frama_scalar_deque(high, low, close, win, sc, fc, first, len, out)?;
    }
    Ok(())
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline(always)]
unsafe fn frama_avx512_into(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    window: usize,
    sc: usize,
    fc: usize,
    first: usize,
    len: usize,
    out: &mut [f64],
) -> Result<(), FramaError> {
    let mut win = window;
    if win & 1 == 1 {
        win += 1;
    }

    if win <= 32 {
        match win {
            10 => unsafe { frama_avx512_small::<10>(high, low, close, sc, fc, first, len, out) },
            14 => unsafe { frama_avx512_small::<14>(high, low, close, sc, fc, first, len, out) },
            20 => unsafe { frama_avx512_small::<20>(high, low, close, sc, fc, first, len, out) },
            32 => unsafe { frama_avx512_small::<32>(high, low, close, sc, fc, first, len, out) },
            _ => unsafe { frama_small_scan(high, low, close, win, sc, fc, first, len, out)? },
        }
    } else {
        frama_scalar_deque(high, low, close, win, sc, fc, first, len, out)?;
    }
    Ok(())
}

#[derive(Clone, Debug)]
pub struct FramaBatchRange {
    pub window: (usize, usize, usize),
    pub sc: (usize, usize, usize),
    pub fc: (usize, usize, usize),
}
impl Default for FramaBatchRange {
    fn default() -> Self {
        Self {
            window: (10, 259, 1),
            sc: (300, 300, 0),
            fc: (1, 1, 0),
        }
    }
}
#[derive(Clone, Debug, Default)]
pub struct FramaBatchBuilder {
    range: FramaBatchRange,
    kernel: Kernel,
}
impl FramaBatchBuilder {
    pub fn new() -> Self {
        Self::default()
    }
    pub fn kernel(mut self, k: Kernel) -> Self {
        self.kernel = k;
        self
    }
    #[inline]
    pub fn window_range(mut self, start: usize, end: usize, step: usize) -> Self {
        self.range.window = (start, end, step);
        self
    }
    #[inline]
    pub fn sc_range(mut self, start: usize, end: usize, step: usize) -> Self {
        self.range.sc = (start, end, step);
        self
    }
    #[inline]
    pub fn fc_range(mut self, start: usize, end: usize, step: usize) -> Self {
        self.range.fc = (start, end, step);
        self
    }
    pub fn apply_slices(
        self,
        high: &[f64],
        low: &[f64],
        close: &[f64],
    ) -> Result<FramaBatchOutput, FramaError> {
        frama_batch_with_kernel(high, low, close, &self.range, self.kernel)
    }
    pub fn apply_slice(self, slice: &[f64]) -> Result<FramaBatchOutput, FramaError> {
        self.apply_slices(slice, slice, slice)
    }
    pub fn with_default_slices(
        high: &[f64],
        low: &[f64],
        close: &[f64],
        k: Kernel,
    ) -> Result<FramaBatchOutput, FramaError> {
        FramaBatchBuilder::new()
            .kernel(k)
            .apply_slices(high, low, close)
    }
    pub fn apply_candles(self, c: &Candles) -> Result<FramaBatchOutput, FramaError> {
        let h = c.select_candle_field("high").unwrap();
        let l = c.select_candle_field("low").unwrap();
        let o = c.select_candle_field("close").unwrap();
        self.apply_slices(h, l, o)
    }
    pub fn with_default_candles(c: &Candles) -> Result<FramaBatchOutput, FramaError> {
        FramaBatchBuilder::new()
            .kernel(Kernel::Auto)
            .apply_candles(c)
    }
}
#[derive(Clone, Debug)]
pub struct FramaBatchOutput {
    pub values: Vec<f64>,
    pub combos: Vec<FramaParams>,
    pub rows: usize,
    pub cols: usize,
}
impl FramaBatchOutput {
    pub fn row_for_params(&self, p: &FramaParams) -> Option<usize> {
        self.combos.iter().position(|c| {
            c.window.unwrap_or(10) == p.window.unwrap_or(10)
                && c.sc.unwrap_or(300) == p.sc.unwrap_or(300)
                && c.fc.unwrap_or(1) == p.fc.unwrap_or(1)
        })
    }
    pub fn values_for(&self, p: &FramaParams) -> 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: &FramaBatchRange) -> Vec<FramaParams> {
    fn axis_usize((start, end, step): (usize, usize, usize)) -> Vec<usize> {
        if step == 0 || start == end {
            return vec![start];
        }

        let (lo, hi) = if start <= end {
            (start, end)
        } else {
            (end, start)
        };
        let mut v = Vec::new();
        let mut x = lo;
        loop {
            v.push(x);
            match x.checked_add(step) {
                Some(nx) if nx <= hi => x = nx,
                _ => break,
            }
        }
        if start > end {
            v.reverse();
        }
        v
    }
    let windows = axis_usize(r.window);
    let scs = axis_usize(r.sc);
    let fcs = axis_usize(r.fc);

    let cap = windows
        .len()
        .checked_mul(scs.len())
        .and_then(|x| x.checked_mul(fcs.len()))
        .unwrap_or(0);
    let mut out = Vec::with_capacity(cap);
    for &w in &windows {
        for &s in &scs {
            for &f in &fcs {
                out.push(FramaParams {
                    window: Some(w),
                    sc: Some(s),
                    fc: Some(f),
                });
            }
        }
    }
    out
}

pub fn frama_batch_with_kernel(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    sweep: &FramaBatchRange,
    k: Kernel,
) -> Result<FramaBatchOutput, FramaError> {
    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(FramaError::InvalidKernelForBatch(other)),
    };
    let simd = match kernel {
        #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
        Kernel::Avx512Batch => Kernel::Avx512,
        #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
        Kernel::Avx2Batch => Kernel::Avx2,
        Kernel::ScalarBatch => Kernel::Scalar,
        _ => unreachable!(),
    };
    frama_batch_inner(high, low, close, sweep, simd, true)
}

#[inline(always)]
pub fn frama_batch_slice(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    sweep: &FramaBatchRange,
    kern: Kernel,
) -> Result<FramaBatchOutput, FramaError> {
    frama_batch_inner(high, low, close, sweep, kern, false)
}
#[inline(always)]
pub fn frama_batch_par_slice(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    sweep: &FramaBatchRange,
    kern: Kernel,
) -> Result<FramaBatchOutput, FramaError> {
    frama_batch_inner(high, low, close, sweep, kern, true)
}

#[inline(always)]
fn frama_batch_inner(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    sweep: &FramaBatchRange,
    kern: Kernel,
    parallel: bool,
) -> Result<FramaBatchOutput, FramaError> {
    if high.is_empty() || low.is_empty() || close.is_empty() {
        return Err(FramaError::EmptyInputData);
    }

    let combos = expand_grid(sweep);
    if combos.is_empty() {
        return Err(FramaError::InvalidRange {
            start: sweep.window.0,
            end: sweep.window.1,
            step: sweep.window.2,
        });
    }

    let rows = combos.len();
    let cols = close.len();

    let _ = rows
        .checked_mul(cols)
        .ok_or(FramaError::ArithmeticOverflow {
            context: "rows*cols",
        })?;

    let mut buf_mu = make_uninit_matrix(rows, cols);

    let first = (0..cols)
        .find(|&i| !high[i].is_nan() && !low[i].is_nan() && !close[i].is_nan())
        .unwrap_or(0);

    let warm: Vec<usize> = combos
        .iter()
        .map(|p| {
            let mut win = p.window.unwrap();

            if win & 1 == 1 {
                win += 1;
            }
            first + win - 1
        })
        .collect();

    init_matrix_prefixes(&mut buf_mu, cols, &warm);

    let mut buf_guard = core::mem::ManuallyDrop::new(buf_mu);
    let out: &mut [f64] = unsafe {
        core::slice::from_raw_parts_mut(buf_guard.as_mut_ptr() as *mut f64, buf_guard.len())
    };

    let combos_ret = frama_batch_inner_into(high, low, close, sweep, kern, parallel, out)?;

    let values = unsafe {
        Vec::from_raw_parts(
            buf_guard.as_mut_ptr() as *mut f64,
            buf_guard.len(),
            buf_guard.capacity(),
        )
    };

    Ok(FramaBatchOutput {
        values,
        combos: combos_ret,
        rows,
        cols,
    })
}

#[inline(always)]
fn frama_batch_inner_into(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    sweep: &FramaBatchRange,
    kern: Kernel,
    parallel: bool,
    out: &mut [f64],
) -> Result<Vec<FramaParams>, FramaError> {
    let combos = expand_grid(sweep);
    if combos.is_empty() {
        return Err(FramaError::InvalidRange {
            start: sweep.window.0,
            end: sweep.window.1,
            step: sweep.window.2,
        });
    }

    if high.is_empty() {
        return Err(FramaError::EmptyInputData);
    }
    if low.len() != high.len() || close.len() != high.len() {
        return Err(FramaError::MismatchedInputLength {
            high: high.len(),
            low: low.len(),
            close: close.len(),
        });
    }

    let len = high.len();
    let first = (0..len)
        .find(|&i| !high[i].is_nan() && !low[i].is_nan() && !close[i].is_nan())
        .ok_or(FramaError::AllValuesNaN)?;

    let max_even_w = combos
        .iter()
        .map(|c| {
            let w = c.window.unwrap();
            if w & 1 == 1 {
                w + 1
            } else {
                w
            }
        })
        .max()
        .unwrap();

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

    let rows = combos.len();
    let cols = len;

    let do_row = |row: usize, dst: &mut [f64]| unsafe {
        let p = &combos[row];
        let window = p.window.unwrap();
        let sc = p.sc.unwrap();
        let fc = p.fc.unwrap();

        match kern {
            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx512 => frama_row_avx512(high, low, close, first, window, dst, sc, fc),
            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx2 => frama_row_avx2(high, low, close, first, window, dst, sc, fc),
            _ => frama_row_scalar(high, low, close, first, window, dst, sc, fc),
        }
    };

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

    Ok(combos)
}

#[derive(Debug, Clone)]
pub struct FramaStream {
    window: usize,
    sc: usize,
    fc: usize,
    n: usize,
    w: f64,
    buffer: Vec<(f64, f64, f64)>,
    head: usize,
    filled: bool,
    last_val: f64,
    d_prev: f64,
    alpha_prev: f64,

    half: usize,
    idx: usize,

    dq_r_max: DqMax,
    dq_r_min: DqMin,
    dq_l_max: DqMax,
    dq_l_min: DqMin,
    dq_w_max: DqMax,
    dq_w_min: DqMin,

    pm_right: f64,
    pn_right: f64,
    pm_left: f64,
    pn_left: f64,
    pm_full: f64,
    pn_full: f64,

    sc_floor: f64,
}
impl FramaStream {
    pub fn try_new(params: FramaParams) -> Result<Self, FramaError> {
        let window = params.window.unwrap_or(10);
        let sc = params.sc.unwrap_or(300);
        let fc = params.fc.unwrap_or(1);
        if window == 0 {
            return Err(FramaError::InvalidWindow {
                window,
                data_len: 0,
            });
        }
        let mut n = window;
        if n % 2 == 1 {
            n += 1;
        }
        Ok(Self {
            window,
            sc,
            fc,
            n,
            w: (2.0 / (sc as f64 + 1.0)).ln(),
            buffer: vec![(f64::NAN, f64::NAN, f64::NAN); n],
            head: 0,
            filled: false,
            last_val: f64::NAN,
            d_prev: 1.0,
            alpha_prev: 2.0 / (sc as f64 + 1.0),

            half: n / 2,
            idx: 0,
            dq_r_max: DqMax::default(),
            dq_r_min: DqMin::default(),
            dq_l_max: DqMax::default(),
            dq_l_min: DqMin::default(),
            dq_w_max: DqMax::default(),
            dq_w_min: DqMin::default(),
            pm_right: f64::NAN,
            pn_right: f64::NAN,
            pm_left: f64::NAN,
            pn_left: f64::NAN,
            pm_full: f64::NAN,
            pn_full: f64::NAN,
            sc_floor: 2.0 / (sc as f64 + 1.0),
        })
    }
    #[inline(always)]
    pub fn update(&mut self, high: f64, low: f64, close: f64) -> Option<f64> {
        if !self.filled {
            self.buffer[self.head] = (high, low, close);
            self.head += 1;

            if self.head == self.n {
                self.head = 0;
                self.filled = true;

                let sum: f64 = self.buffer.iter().map(|&(_, _, c)| c).sum();
                self.last_val = sum / self.n as f64;

                self.dq_r_max.clear();
                self.dq_r_min.clear();
                self.dq_l_max.clear();
                self.dq_l_min.clear();
                self.dq_w_max.clear();
                self.dq_w_min.clear();

                for j in 0..self.n {
                    let (h, l, _) = self.buffer[j];
                    if !(h.is_nan() || l.is_nan()) {
                        self.dq_w_max.push(j, h);
                        self.dq_w_min.push(j, l);
                        if j < self.half {
                            self.dq_l_max.push(j, h);
                            self.dq_l_min.push(j, l);
                        } else {
                            self.dq_r_max.push(j, h);
                            self.dq_r_min.push(j, l);
                        }
                    }
                }

                self.pm_right = self.dq_r_max.front_val().unwrap_or(f64::NAN);
                self.pn_right = self.dq_r_min.front_val().unwrap_or(f64::NAN);
                self.pm_left = self.dq_l_max.front_val().unwrap_or(f64::NAN);
                self.pn_left = self.dq_l_min.front_val().unwrap_or(f64::NAN);
                self.pm_full = self.dq_w_max.front_val().unwrap_or(f64::NAN);
                self.pn_full = self.dq_w_min.front_val().unwrap_or(f64::NAN);

                self.idx = self.n;

                return Some(self.last_val);
            }

            return None;
        }

        let i = self.idx;

        let right_lb = i.saturating_sub(self.half);
        let left_lb = i.saturating_sub(self.n);
        self.dq_r_max.expire_lt(right_lb);
        self.dq_r_min.expire_lt(right_lb);
        self.dq_l_max.expire_lt(left_lb);
        self.dq_l_min.expire_lt(left_lb);
        self.dq_w_max.expire_lt(left_lb);
        self.dq_w_min.expire_lt(left_lb);

        let (max_r, min_r) = {
            let mr = self.dq_r_max.front_val().unwrap_or(self.pm_right);
            let nr = self.dq_r_min.front_val().unwrap_or(self.pn_right);
            (mr, nr)
        };
        let (max_l, min_l) = {
            let ml = self.dq_l_max.front_val().unwrap_or(self.pm_left);
            let nl = self.dq_l_min.front_val().unwrap_or(self.pn_left);
            (ml, nl)
        };
        let (max_w, min_w) = {
            let mw = self.dq_w_max.front_val().unwrap_or(self.pm_full);
            let nw = self.dq_w_min.front_val().unwrap_or(self.pn_full);
            (mw, nw)
        };

        self.pm_right = max_r;
        self.pn_right = min_r;
        self.pm_left = max_l;
        self.pn_left = min_l;
        self.pm_full = max_w;
        self.pn_full = min_w;

        let half_f = self.half as f64;
        let win_f = self.n as f64;

        let output = if !(high.is_nan() || low.is_nan() || close.is_nan()) {
            let n1 = (max_r - min_r) / half_f;
            let n2 = (max_l - min_l) / half_f;
            let n3 = (max_w - min_w) / win_f;

            let d = if n1 > 0.0 && n2 > 0.0 && n3 > 0.0 {
                ((n1 + n2).ln() - n3.ln()) / std::f64::consts::LN_2
            } else {
                self.d_prev
            };
            self.d_prev = d;

            let mut a0 = (self.w * (d - 1.0)).exp();
            if a0 < 0.1 {
                a0 = 0.1;
            }
            if a0 > 1.0 {
                a0 = 1.0;
            }

            let old_n = (2.0 - a0) / a0;
            let new_n = (self.sc - self.fc) as f64 * ((old_n - 1.0) / (self.sc as f64 - 1.0))
                + self.fc as f64;

            let mut alpha = 2.0 / (new_n + 1.0);
            if alpha < self.sc_floor {
                alpha = self.sc_floor;
            }
            if alpha > 1.0 {
                alpha = 1.0;
            }
            self.alpha_prev = alpha;

            close.mul_add(alpha, (1.0 - alpha) * self.last_val)
        } else {
            self.last_val
        };

        if !(high.is_nan() || low.is_nan()) {
            self.dq_r_max.push(i, high);
            self.dq_r_min.push(i, low);
            self.dq_w_max.push(i, high);
            self.dq_w_min.push(i, low);
        }

        if i >= self.half {
            let j = i - self.half;
            let (h_l, l_l, _) = self.buffer[j % self.n];
            if !(h_l.is_nan() || l_l.is_nan()) {
                self.dq_l_max.push(j, h_l);
                self.dq_l_min.push(j, l_l);
            }
        }

        self.buffer[self.head] = (high, low, close);
        self.head = (self.head + 1) % self.n;

        self.idx += 1;
        self.last_val = output;
        Some(output)
    }
}

#[derive(Default, Debug, Clone)]
struct DqMax {
    q: VecDeque<(usize, f64)>,
}
#[derive(Default, Debug, Clone)]
struct DqMin {
    q: VecDeque<(usize, f64)>,
}

impl DqMax {
    #[inline(always)]
    fn clear(&mut self) {
        self.q.clear();
    }
    #[inline(always)]
    fn expire_lt(&mut self, bound: usize) {
        while let Some(&(i, _)) = self.q.front() {
            if i < bound {
                self.q.pop_front();
            } else {
                break;
            }
        }
    }
    #[inline(always)]
    fn push(&mut self, idx: usize, val: f64) {
        while let Some(&(_, v)) = self.q.back() {
            if v >= val {
                break;
            }
            self.q.pop_back();
        }
        self.q.push_back((idx, val));
    }
    #[inline(always)]
    fn front_val(&self) -> Option<f64> {
        self.q.front().map(|&(_, v)| v)
    }
}
impl DqMin {
    #[inline(always)]
    fn clear(&mut self) {
        self.q.clear();
    }
    #[inline(always)]
    fn expire_lt(&mut self, bound: usize) {
        while let Some(&(i, _)) = self.q.front() {
            if i < bound {
                self.q.pop_front();
            } else {
                break;
            }
        }
    }
    #[inline(always)]
    fn push(&mut self, idx: usize, val: f64) {
        while let Some(&(_, v)) = self.q.back() {
            if v <= val {
                break;
            }
            self.q.pop_back();
        }
        self.q.push_back((idx, val));
    }
    #[inline(always)]
    fn front_val(&self) -> Option<f64> {
        self.q.front().map(|&(_, v)| v)
    }
}

#[inline(always)]
pub unsafe fn frama_row_scalar(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    first: usize,
    window: usize,
    out: &mut [f64],
    sc: usize,
    fc: usize,
) {
    let len = high.len();

    let mut win = window;
    if win & 1 == 1 {
        win += 1;
    }
    seed_sma(close, first, win, out);

    if win <= 32 {
        frama_small_scan(high, low, close, win, sc, fc, first, len, out).unwrap();
    } else {
        frama_scalar_deque(high, low, close, win, sc, fc, first, len, out).unwrap();
    }
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline(always)]
pub unsafe fn frama_row_avx2(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    first: usize,
    window: usize,
    out: &mut [f64],
    sc: usize,
    fc: usize,
) {
    let mut win = window;
    if win & 1 == 1 {
        win += 1;
    }

    seed_sma(close, first, win, out);

    if win <= 32 {
        match win {
            10 => frama_avx2_small::<10>(high, low, close, sc, fc, first, high.len(), out),
            14 => frama_avx2_small::<14>(high, low, close, sc, fc, first, high.len(), out),
            20 => frama_avx2_small::<20>(high, low, close, sc, fc, first, high.len(), out),
            32 => frama_avx2_small::<32>(high, low, close, sc, fc, first, high.len(), out),
            _ => frama_small_scan(high, low, close, win, sc, fc, first, high.len(), out).unwrap(),
        }
    } else {
        frama_scalar_deque(high, low, close, win, sc, fc, first, high.len(), out).unwrap();
    }
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline(always)]
pub unsafe fn frama_row_avx512(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    first: usize,
    window: usize,
    out: &mut [f64],
    sc: usize,
    fc: usize,
) {
    let mut win = window;
    if win & 1 == 1 {
        win += 1;
    }

    seed_sma(close, first, win, out);

    if win <= 32 {
        match win {
            10 => frama_avx512_small::<10>(high, low, close, sc, fc, first, high.len(), out),
            14 => frama_avx512_small::<14>(high, low, close, sc, fc, first, high.len(), out),
            20 => frama_avx512_small::<20>(high, low, close, sc, fc, first, high.len(), out),
            32 => frama_avx512_small::<32>(high, low, close, sc, fc, first, high.len(), out),
            _ => frama_small_scan(high, low, close, win, sc, fc, first, high.len(), out).unwrap(),
        }
    } else {
        frama_scalar_deque(high, low, close, win, sc, fc, first, high.len(), out).unwrap();
    }
}

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

    fn check_frama_partial_params(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn 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 = FramaParams {
            window: None,
            sc: None,
            fc: None,
        };
        let input = FramaInput::from_candles(&candles, default_params);
        let output = frama_with_kernel(&input, kernel)?;
        assert_eq!(output.values.len(), candles.close.len());
        Ok(())
    }
    fn check_frama_accuracy(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn 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 = FramaInput::from_candles(&candles, FramaParams::default());
        let result = frama_with_kernel(&input, kernel)?;
        let expected_last_five = [
            59337.23056930512,
            59321.607512374605,
            59286.677929994796,
            59268.00202402624,
            59160.03888720062,
        ];
        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-1,
                "[{}] FRAMA {:?} mismatch at idx {}: got {}, expected {}",
                test_name,
                kernel,
                i,
                val,
                expected_last_five[i]
            );
        }
        Ok(())
    }
    fn check_frama_zero_window(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn Error>> {
        skip_if_unsupported!(kernel, test_name);
        let input_data = [10.0, 20.0, 30.0];
        let params = FramaParams {
            window: Some(0),
            sc: None,
            fc: None,
        };
        let input = FramaInput::from_slices(&input_data, &input_data, &input_data, params);
        let res = frama_with_kernel(&input, kernel);
        assert!(
            res.is_err(),
            "[{}] FRAMA should fail with zero window",
            test_name
        );
        Ok(())
    }
    fn check_frama_window_exceeds_length(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn Error>> {
        skip_if_unsupported!(kernel, test_name);
        let data_small = [10.0, 20.0, 30.0];
        let params = FramaParams {
            window: Some(10),
            sc: None,
            fc: None,
        };
        let input = FramaInput::from_slices(&data_small, &data_small, &data_small, params);
        let res = frama_with_kernel(&input, kernel);
        assert!(
            res.is_err(),
            "[{}] FRAMA should fail with window exceeding length",
            test_name
        );
        Ok(())
    }
    fn check_frama_very_small_dataset(
        test_name: &str,
        kernel: Kernel,
    ) -> Result<(), Box<dyn Error>> {
        skip_if_unsupported!(kernel, test_name);
        let single_point = [42.0];
        let params = FramaParams {
            window: Some(9),
            sc: None,
            fc: None,
        };
        let input = FramaInput::from_slices(&single_point, &single_point, &single_point, params);
        let res = frama_with_kernel(&input, kernel);
        assert!(
            res.is_err(),
            "[{}] FRAMA should fail with insufficient data",
            test_name
        );
        Ok(())
    }
    fn check_frama_all_nan(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn Error>> {
        skip_if_unsupported!(kernel, test_name);
        let nan_data = [f64::NAN, f64::NAN, f64::NAN];
        let params = FramaParams::default();
        let input = FramaInput::from_slices(&nan_data, &nan_data, &nan_data, params);
        let res = frama_with_kernel(&input, kernel);
        assert!(res.is_err());
        Ok(())
    }
    fn check_frama_empty_input(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn Error>> {
        skip_if_unsupported!(kernel, test_name);
        let empty: [f64; 0] = [];
        let params = FramaParams::default();
        let input = FramaInput::from_slices(&empty, &empty, &empty, params);
        let res = frama_with_kernel(&input, kernel);
        assert!(matches!(res, Err(FramaError::EmptyInputData)));
        Ok(())
    }

    fn check_frama_mismatched_len(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn Error>> {
        skip_if_unsupported!(kernel, test_name);
        let h = [1.0, 2.0, 3.0];
        let l = [1.0, 2.0];
        let c = [1.0, 2.0, 3.0];
        let params = FramaParams::default();
        let input = FramaInput::from_slices(&h, &l, &c, params);
        let res = frama_with_kernel(&input, kernel);
        assert!(matches!(res, Err(FramaError::MismatchedInputLength { .. })));
        Ok(())
    }

    fn check_frama_reinput(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn 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 = FramaParams::default();
        let first_input = FramaInput::from_candles(&candles, params.clone());
        let first_res = frama_with_kernel(&first_input, kernel)?;

        let second_input = FramaInput::from_slices(
            &first_res.values,
            &first_res.values,
            &first_res.values,
            params,
        );
        let second_res = frama_with_kernel(&second_input, kernel)?;
        assert_eq!(first_res.values.len(), second_res.values.len());
        Ok(())
    }

    fn check_frama_nan_handling(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn 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 = FramaInput::from_candles(&candles, FramaParams::default());
        let res = frama_with_kernel(&input, kernel)?;
        if res.values.len() > 240 {
            for (i, &v) in res.values[240..].iter().enumerate() {
                assert!(
                    !v.is_nan(),
                    "[{}] Found unexpected NaN at out-index {}",
                    test_name,
                    240 + i
                );
            }
        }
        Ok(())
    }

    fn check_frama_property(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn 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 high = candles.select_candle_field("high").unwrap();
        let low = candles.select_candle_field("low").unwrap();
        let close = candles.select_candle_field("close").unwrap();

        let data_len = high.len();
        let strat = (
            4usize..=64,
            50usize..500,
            1usize..50,
            0usize..data_len.saturating_sub(200),
            100usize..=200,
        );

        proptest::test_runner::TestRunner::default()
            .run(&strat, |(window, sc, fc, start_idx, slice_len)| {
                let end_idx = (start_idx + slice_len).min(data_len);
                let actual_len = end_idx - start_idx;

                if actual_len < window * 2 {
                    return Ok(());
                }

                let high_slice = &high[start_idx..end_idx];
                let low_slice = &low[start_idx..end_idx];
                let close_slice = &close[start_idx..end_idx];

                let params = FramaParams {
                    window: Some(window),
                    sc: Some(sc),
                    fc: Some(fc),
                };

                let input = FramaInput::from_slices(high_slice, low_slice, close_slice, params);
                let result = frama_with_kernel(&input, kernel);

                prop_assert!(result.is_ok(), "FRAMA failed: {:?}", result.err());
                let FramaOutput { values: out } = result.unwrap();

                let FramaOutput { values: ref_out } =
                    frama_with_kernel(&input, Kernel::Scalar).unwrap();

                let actual_window = if window & 1 == 1 { window + 1 } else { window };

                let first_output_idx = actual_window - 1;

                for i in 0..first_output_idx.min(out.len()) {
                    prop_assert!(
                        out[i].is_nan(),
                        "Expected NaN during warmup at index {}, got {}",
                        i,
                        out[i]
                    );
                }

                for i in first_output_idx..out.len() {
                    let y = out[i];
                    let r = ref_out[i];

                    let all_high_max = high_slice
                        .iter()
                        .filter(|x| x.is_finite())
                        .cloned()
                        .fold(f64::NEG_INFINITY, f64::max);
                    let all_low_min = low_slice
                        .iter()
                        .filter(|x| x.is_finite())
                        .cloned()
                        .fold(f64::INFINITY, f64::min);

                    if all_high_max.is_finite() && all_low_min.is_finite() {
                        let tolerance = (all_high_max - all_low_min) * 0.01;
                        prop_assert!(
                            y.is_nan()
                                || (y >= all_low_min - tolerance && y <= all_high_max + tolerance),
                            "idx {}: {} not in overall range [{}, {}] with tolerance {}",
                            i,
                            y,
                            all_low_min,
                            all_high_max,
                            tolerance
                        );
                    }

                    if !y.is_finite() || !r.is_finite() {
                        prop_assert!(
                            y.to_bits() == r.to_bits(),
                            "NaN mismatch at idx {}: {} vs {}",
                            i,
                            y,
                            r
                        );
                    } else {
                        let ulp_diff = y.to_bits().abs_diff(r.to_bits());
                        prop_assert!(
                            (y - r).abs() <= 1e-9 || ulp_diff <= 4,
                            "mismatch at idx {}: {} vs {} (ULP={})",
                            i,
                            y,
                            r,
                            ulp_diff
                        );
                    }

                    if fc >= sc && i > first_output_idx {
                        let change = (y - out[i - 1]).abs();
                        let price_change = (close_slice[i] - close_slice[i - 1]).abs();
                        prop_assert!(
							change <= price_change + 1e-6,
							"Unexpected large change at idx {} with fc >= sc: {} vs price change {}",
							i,
							change,
							price_change
						);
                    }
                }

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

        Ok(())
    }
    fn check_frama_streaming(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn 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 high = candles.select_candle_field("high").unwrap();
        let low = candles.select_candle_field("low").unwrap();
        let close = candles.select_candle_field("close").unwrap();
        let period = 10;
        let sc = 300;
        let fc = 1;
        let input = FramaInput::from_slices(
            high,
            low,
            close,
            FramaParams {
                window: Some(period),
                sc: Some(sc),
                fc: Some(fc),
            },
        );
        let batch_output = frama_with_kernel(&input, kernel)?.values;
        let mut stream = FramaStream::try_new(FramaParams {
            window: Some(period),
            sc: Some(sc),
            fc: Some(fc),
        })?;
        let mut stream_values = Vec::with_capacity(close.len());
        for ((&h, &l), &c) in high.iter().zip(low.iter()).zip(close.iter()) {
            match stream.update(h, l, c) {
                Some(val) => stream_values.push(val),
                None => stream_values.push(f64::NAN),
            }
        }
        assert_eq!(batch_output.len(), stream_values.len());
        for (i, (&b, &s)) in batch_output.iter().zip(stream_values.iter()).enumerate() {
            if b.is_nan() && s.is_nan() {
                continue;
            }
            let diff = (b - s).abs();
            assert!(
                diff < 1e-7,
                "[{}] FRAMA streaming mismatch at idx {}: batch={}, stream={}",
                test_name,
                i,
                b,
                s
            );
        }
        Ok(())
    }
    fn check_frama_default_candles(test: &str, k: Kernel) -> Result<(), Box<dyn Error>> {
        skip_if_unsupported!(k, test);
        let file = "src/data/2018-09-01-2024-Bitfinex_Spot-4h.csv";
        let c = read_candles_from_csv(file)?;
        let input = FramaInput::with_default_candles(&c);
        match input.data {
            FramaData::Candles { .. } => {}
            _ => panic!("Expected FramaData::Candles"),
        }
        let out = frama_with_kernel(&input, k)?;
        assert_eq!(out.values.len(), c.close.len());
        Ok(())
    }

    #[cfg(debug_assertions)]
    fn check_frama_no_poison(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn 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![
            FramaParams::default(),
            FramaParams {
                window: Some(4),
                sc: Some(300),
                fc: Some(1),
            },
            FramaParams {
                window: Some(8),
                sc: Some(150),
                fc: Some(1),
            },
            FramaParams {
                window: Some(10),
                sc: Some(200),
                fc: Some(2),
            },
            FramaParams {
                window: Some(12),
                sc: Some(400),
                fc: Some(1),
            },
            FramaParams {
                window: Some(20),
                sc: Some(300),
                fc: Some(1),
            },
            FramaParams {
                window: Some(30),
                sc: Some(500),
                fc: Some(3),
            },
            FramaParams {
                window: Some(16),
                sc: Some(100),
                fc: Some(1),
            },
            FramaParams {
                window: Some(14),
                sc: Some(600),
                fc: Some(4),
            },
        ];

        for params in test_cases {
            let input = FramaInput::from_candles(&candles, params.clone());
            let output = frama_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 window={:?}, sc={:?}, fc={:?}",
                        test_name, val, bits, i, params.window, params.sc, params.fc
                    );
                }

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

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

        Ok(())
    }

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

    macro_rules! generate_all_frama_tests {
        ($($test_fn:ident),*) => {
            paste! {
                $(
                    #[test]
                    fn [<$test_fn _scalar_f64>]() {
                        let _ = $test_fn(stringify!([<$test_fn _scalar_f64>]), Kernel::Scalar);
                    }
                )*
                $(
                    #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
                    #[test]
                    fn [<$test_fn _avx2_f64>]() {
                        let _ = $test_fn(stringify!([<$test_fn _avx2_f64>]), Kernel::Avx2);
                    }
                )*
                $(
                    #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
                    #[test]
                    fn [<$test_fn _avx512_f64>]() {
                        let _ = $test_fn(stringify!([<$test_fn _avx512_f64>]), Kernel::Avx512);
                    }
                )*
            }
        }
    }
    generate_all_frama_tests!(
        check_frama_partial_params,
        check_frama_accuracy,
        check_frama_zero_window,
        check_frama_window_exceeds_length,
        check_frama_very_small_dataset,
        check_frama_all_nan,
        check_frama_empty_input,
        check_frama_mismatched_len,
        check_frama_reinput,
        check_frama_nan_handling,
        check_frama_property,
        check_frama_streaming,
        check_frama_default_candles,
        check_frama_no_poison
    );
    fn check_batch_default_row(test: &str, kernel: Kernel) -> Result<(), Box<dyn 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 = FramaBatchBuilder::new().kernel(kernel).apply_candles(&c)?;
        let def = FramaParams::default();
        let row = output.values_for(&def).expect("default row missing");
        assert_eq!(row.len(), c.close.len());
        let expected = [
            59337.23056930512,
            59321.607512374605,
            59286.677929994796,
            59268.00202402624,
            59160.03888720062,
        ];
        let start = row.len() - 5;
        for (i, &v) in row[start..].iter().enumerate() {
            assert!(
                (v - expected[i]).abs() < 1e-1,
                "[{test}] default-row mismatch at idx {i}: {v} vs {expected:?}"
            );
        }
        Ok(())
    }

    #[cfg(debug_assertions)]
    fn check_batch_no_poison(test: &str, kernel: Kernel) -> Result<(), Box<dyn 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 test_configs = vec![
            ((4, 8, 2), (100, 300, 100), (1, 2, 1)),
            ((10, 20, 5), (200, 400, 100), (1, 3, 1)),
            ((20, 30, 5), (300, 600, 150), (1, 4, 1)),
            ((6, 12, 2), (150, 450, 50), (1, 2, 1)),
            ((8, 16, 2), (100, 500, 100), (1, 5, 1)),
        ];

        for (window_range, sc_range, fc_range) in test_configs {
            let output = FramaBatchBuilder::new()
                .kernel(kernel)
                .window_range(window_range.0, window_range.1, window_range.2)
                .sc_range(sc_range.0, sc_range.1, sc_range.2)
                .fc_range(fc_range.0, fc_range.1, fc_range.2)
                .apply_candles(&c)?;

            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 params = &output.combos[row];

                if bits == 0x11111111_11111111 {
                    panic!(
                        "[{}] Found alloc_with_nan_prefix poison value {} (0x{:016X}) at row {} col {} (params: window={:?}, sc={:?}, fc={:?})",
                        test, val, bits, row, col, params.window, params.sc, params.fc
                    );
                }

                if bits == 0x22222222_22222222 {
                    panic!(
                        "[{}] Found init_matrix_prefixes poison value {} (0x{:016X}) at row {} col {} (params: window={:?}, sc={:?}, fc={:?})",
                        test, val, bits, row, col, params.window, params.sc, params.fc
                    );
                }

                if bits == 0x33333333_33333333 {
                    panic!(
                        "[{}] Found make_uninit_matrix poison value {} (0x{:016X}) at row {} col {} (params: window={:?}, sc={:?}, fc={:?})",
                        test, val, bits, row, col, params.window, params.sc, params.fc
                    );
                }
            }
        }

        Ok(())
    }

    #[cfg(not(debug_assertions))]
    fn check_batch_no_poison(_test: &str, _kernel: Kernel) -> Result<(), Box<dyn Error>> {
        Ok(())
    }
    macro_rules! gen_batch_tests {
        ($fn_name:ident) => {
            paste! {
                #[test]
                fn [<$fn_name _scalar>]() {
                    let _ = $fn_name(stringify!([<$fn_name _scalar>]), Kernel::ScalarBatch);
                }
                #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
                #[test]
                fn [<$fn_name _avx2>]() {
                    let _ = $fn_name(stringify!([<$fn_name _avx2>]), Kernel::Avx2Batch);
                }
                #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
                #[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);
                }
            }
        };
    }
    gen_batch_tests!(check_batch_default_row);
    gen_batch_tests!(check_batch_no_poison);

    #[test]
    fn test_frama_into_matches_api() -> Result<(), Box<dyn Error>> {
        let file_path = "src/data/2018-09-01-2024-Bitfinex_Spot-4h.csv";
        let candles = read_candles_from_csv(file_path)?;

        let input = FramaInput::with_default_candles(&candles);
        let baseline = frama(&input)?.values;

        let mut out = vec![0.0; candles.close.len()];

        #[cfg(not(all(target_arch = "wasm32", feature = "wasm")))]
        {
            frama_into(&input, &mut out)?;
        }
        #[cfg(all(target_arch = "wasm32", feature = "wasm"))]
        {
            frama_into_slice(&mut out, &input, Kernel::Auto)?;
        }

        assert_eq!(out.len(), baseline.len());
        for i in 0..out.len() {
            let a = out[i];
            let b = baseline[i];
            if a.is_nan() || b.is_nan() {
                assert!(a.is_nan() && b.is_nan(), "NaN mismatch at index {}", i);
            } else {
                assert!(a == b, "Value mismatch at index {}: {} != {}", i, a, b);
            }
        }
        Ok(())
    }
}

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

#[cfg(all(feature = "python", feature = "cuda"))]
use crate::cuda::moving_averages::DeviceArrayF32;
#[cfg(all(feature = "python", feature = "cuda"))]
use cust::context::Context;
#[cfg(all(feature = "python", feature = "cuda"))]
use cust::memory::DeviceBuffer;
#[cfg(all(feature = "python", feature = "cuda"))]
use std::sync::Arc;

#[cfg(all(feature = "python", feature = "cuda"))]
#[pyclass(
    module = "ta_indicators.cuda",
    name = "DeviceArrayF32Frama",
    unsendable
)]
pub struct DeviceArrayF32FramaPy {
    pub(crate) inner: DeviceArrayF32,
    _ctx_guard: Arc<Context>,
    _device_id: u32,
}

#[cfg(all(feature = "python", feature = "cuda"))]
#[pymethods]
impl DeviceArrayF32FramaPy {
    #[new]
    fn py_new() -> PyResult<Self> {
        Err(pyo3::exceptions::PyTypeError::new_err(
            "use CUDA FRAMA factory functions to create this object",
        ))
    }

    #[getter]
    fn __cuda_array_interface__<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyDict>> {
        let d = PyDict::new(py);
        let item = std::mem::size_of::<f32>();
        d.set_item("shape", (self.inner.rows, self.inner.cols))?;
        d.set_item("typestr", "<f4")?;
        d.set_item("strides", (self.inner.cols * item, item))?;
        let size = self.inner.rows.saturating_mul(self.inner.cols);
        let ptr_val: usize = if size == 0 {
            0
        } else {
            self.inner.buf.as_device_ptr().as_raw() as usize
        };
        d.set_item("data", (ptr_val, false))?;

        d.set_item("version", 3)?;
        Ok(d)
    }

    fn __dlpack_device__(&self) -> (i32, i32) {
        (2, self._device_id as i32)
    }

    #[pyo3(signature = (stream=None, max_version=None, dl_device=None, copy=None))]
    fn __dlpack__<'py>(
        &mut self,
        py: Python<'py>,
        stream: Option<PyObject>,
        max_version: Option<PyObject>,
        dl_device: Option<PyObject>,
        copy: Option<PyObject>,
    ) -> PyResult<PyObject> {
        use crate::utilities::dlpack_cuda::export_f32_cuda_dlpack_2d;

        let (kdl, alloc_dev) = self.__dlpack_device__();
        if let Some(dev_obj) = dl_device.as_ref() {
            if let Ok((dev_ty, dev_id)) = dev_obj.extract::<(i32, i32)>(py) {
                if dev_ty != kdl || dev_id != alloc_dev {
                    let wants_copy = copy
                        .as_ref()
                        .and_then(|c| c.extract::<bool>(py).ok())
                        .unwrap_or(false);
                    if wants_copy {
                        return Err(PyValueError::new_err(
                            "device copy not implemented for __dlpack__",
                        ));
                    } else {
                        return Err(PyValueError::new_err("dl_device mismatch for __dlpack__"));
                    }
                }
            }
        }
        let _ = stream;

        let dummy =
            DeviceBuffer::from_slice(&[]).map_err(|e| PyValueError::new_err(e.to_string()))?;
        let inner = std::mem::replace(
            &mut self.inner,
            DeviceArrayF32 {
                buf: dummy,
                rows: 0,
                cols: 0,
            },
        );

        let rows = inner.rows;
        let cols = inner.cols;
        let buf = inner.buf;

        let max_version_bound = max_version.map(|obj| obj.into_bound(py));

        export_f32_cuda_dlpack_2d(py, buf, rows, cols, alloc_dev, max_version_bound)
    }
}

#[cfg(all(feature = "python", feature = "cuda"))]
impl DeviceArrayF32FramaPy {
    pub fn new(inner: DeviceArrayF32, ctx_guard: Arc<Context>, device_id: u32) -> Self {
        Self {
            inner,
            _ctx_guard: ctx_guard,
            _device_id: device_id,
        }
    }
}

#[cfg(feature = "python")]
#[pyfunction(name = "frama")]
#[pyo3(signature = (high, low, close, window, sc=300, fc=1, kernel=None))]
pub fn frama_py<'py>(
    py: Python<'py>,
    high: PyReadonlyArray1<'py, f64>,
    low: PyReadonlyArray1<'py, f64>,
    close: PyReadonlyArray1<'py, f64>,
    window: usize,
    sc: usize,
    fc: usize,
    kernel: Option<&str>,
) -> PyResult<Bound<'py, numpy::PyArray1<f64>>> {
    let h = high.as_slice()?;
    let l = low.as_slice()?;
    let c = close.as_slice()?;

    let params = FramaParams {
        window: Some(window),
        sc: Some(sc),
        fc: Some(fc),
    };
    let input = FramaInput::from_slices(h, l, c, params);
    let kern = validate_kernel(kernel, false)?;

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

    Ok(out.into_pyarray(py))
}

#[cfg(feature = "python")]
#[pyfunction(name = "frama_batch")]
#[pyo3(signature = (high, low, close, window_range, sc_range, fc_range, kernel=None))]
pub fn frama_batch_py<'py>(
    py: Python<'py>,
    high: PyReadonlyArray1<'py, f64>,
    low: PyReadonlyArray1<'py, f64>,
    close: PyReadonlyArray1<'py, f64>,
    window_range: (usize, usize, usize),
    sc_range: (usize, usize, usize),
    fc_range: (usize, usize, usize),
    kernel: Option<&str>,
) -> PyResult<Bound<'py, PyDict>> {
    use numpy::{PyArray1, PyArrayMethods};

    let high_slice = high.as_slice()?;
    let low_slice = low.as_slice()?;
    let close_slice = close.as_slice()?;
    let kern = validate_kernel(kernel, true)?;

    let range = FramaBatchRange {
        window: window_range,
        sc: sc_range,
        fc: fc_range,
    };

    let combos = expand_grid(&range);
    let rows = combos.len();
    let cols = close_slice.len();

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

    let combos_result = py
        .allow_threads(|| -> Result<Vec<FramaParams>, FramaError> {
            let kernel = match kern {
                Kernel::Auto => match detect_best_batch_kernel() {
                    Kernel::Avx512Batch => Kernel::Avx2Batch,
                    other => other,
                },
                k => k,
            };

            let single_kernel = match kernel {
                Kernel::ScalarBatch => Kernel::Scalar,
                #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
                Kernel::Avx2Batch => Kernel::Avx2,
                #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
                Kernel::Avx512Batch => Kernel::Avx512,
                _ => Kernel::Scalar,
            };

            let first = close_slice
                .iter()
                .enumerate()
                .find(|(i, &v)| !v.is_nan() && !high_slice[*i].is_nan() && !low_slice[*i].is_nan())
                .map(|(i, _)| i)
                .unwrap_or(0);

            for (row_idx, combo) in combos.iter().enumerate() {
                let window = combo.window.unwrap_or(10);
                let warmup_period = first + window - 1;
                let row_start = row_idx * cols;
                for col_idx in 0..warmup_period.min(cols) {
                    slice_out[row_start + col_idx] = f64::NAN;
                }
            }

            frama_batch_inner_into(
                high_slice,
                low_slice,
                close_slice,
                &range,
                single_kernel,
                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))?)?;

    let windows: Vec<u64> = combos_result
        .iter()
        .map(|c| c.window.unwrap_or(10) as u64)
        .collect();
    let scs: Vec<u64> = combos_result
        .iter()
        .map(|c| c.sc.unwrap_or(300) as u64)
        .collect();
    let fcs: Vec<u64> = combos_result
        .iter()
        .map(|c| c.fc.unwrap_or(1) as u64)
        .collect();

    dict.set_item("windows", windows.into_pyarray(py))?;
    dict.set_item("scs", scs.into_pyarray(py))?;
    dict.set_item("fcs", fcs.into_pyarray(py))?;
    Ok(dict)
}

#[cfg(all(feature = "python", feature = "cuda"))]
#[pyfunction(name = "frama_cuda_batch_dev")]
#[pyo3(signature = (high_f32, low_f32, close_f32, window_range, sc_range, fc_range, device_id=0))]
pub fn frama_cuda_batch_dev_py<'py>(
    py: Python<'py>,
    high_f32: numpy::PyReadonlyArray1<'py, f32>,
    low_f32: numpy::PyReadonlyArray1<'py, f32>,
    close_f32: numpy::PyReadonlyArray1<'py, f32>,
    window_range: (usize, usize, usize),
    sc_range: (usize, usize, usize),
    fc_range: (usize, usize, usize),
    device_id: usize,
) -> PyResult<(DeviceArrayF32FramaPy, Bound<'py, PyDict>)> {
    use crate::cuda::cuda_available;
    use numpy::IntoPyArray;
    use pyo3::types::PyDict;

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

    let high_slice = high_f32.as_slice()?;
    let low_slice = low_f32.as_slice()?;
    let close_slice = close_f32.as_slice()?;
    if high_slice.len() != low_slice.len() || high_slice.len() != close_slice.len() {
        return Err(PyValueError::new_err("mismatched slice lengths"));
    }

    let sweep = FramaBatchRange {
        window: window_range,
        sc: sc_range,
        fc: fc_range,
    };

    let (inner, combos, ctx, dev_id) = py.allow_threads(|| {
        let cuda = CudaFrama::new(device_id).map_err(|e| PyValueError::new_err(e.to_string()))?;
        let ctx = cuda.ctx();
        let dev_id = cuda.device_id();
        cuda.frama_batch_dev(high_slice, low_slice, close_slice, &sweep)
            .map_err(|e| PyValueError::new_err(e.to_string()))
            .map(|(d, c)| (d, c, ctx, dev_id))
    })?;

    let dict = PyDict::new(py);
    let windows: Vec<u64> = combos.iter().map(|c| c.window.unwrap() as u64).collect();
    let scs: Vec<u64> = combos.iter().map(|c| c.sc.unwrap() as u64).collect();
    let fcs: Vec<u64> = combos.iter().map(|c| c.fc.unwrap() as u64).collect();
    dict.set_item("windows", windows.into_pyarray(py))?;
    dict.set_item("scs", scs.into_pyarray(py))?;
    dict.set_item("fcs", fcs.into_pyarray(py))?;

    Ok((DeviceArrayF32FramaPy::new(inner, ctx, dev_id), dict))
}

#[cfg(all(feature = "python", feature = "cuda"))]
#[pyfunction(name = "frama_cuda_many_series_one_param_dev")]
#[pyo3(signature = (high_tm_f32, low_tm_f32, close_tm_f32, window, sc, fc, device_id=0))]
pub fn frama_cuda_many_series_one_param_dev_py(
    py: Python<'_>,
    high_tm_f32: numpy::PyReadonlyArray2<'_, f32>,
    low_tm_f32: numpy::PyReadonlyArray2<'_, f32>,
    close_tm_f32: numpy::PyReadonlyArray2<'_, f32>,
    window: usize,
    sc: usize,
    fc: usize,
    device_id: usize,
) -> PyResult<DeviceArrayF32FramaPy> {
    use crate::cuda::cuda_available;
    use numpy::PyUntypedArrayMethods;

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

    let high_shape = high_tm_f32.shape();
    let low_shape = low_tm_f32.shape();
    let close_shape = close_tm_f32.shape();
    if low_shape != high_shape || close_shape != high_shape {
        return Err(PyValueError::new_err(
            "high, low, and close arrays must share the same shape",
        ));
    }

    let rows = high_shape[0];
    let cols = high_shape[1];
    let high_slice = high_tm_f32.as_slice()?;
    let low_slice = low_tm_f32.as_slice()?;
    let close_slice = close_tm_f32.as_slice()?;

    let params = FramaParams {
        window: Some(window),
        sc: Some(sc),
        fc: Some(fc),
    };

    let (inner, ctx, dev_id) = py.allow_threads(|| {
        let cuda = CudaFrama::new(device_id).map_err(|e| PyValueError::new_err(e.to_string()))?;
        let ctx = cuda.ctx();
        let dev_id = cuda.device_id();
        cuda.frama_many_series_one_param_time_major_dev(
            high_slice,
            low_slice,
            close_slice,
            cols,
            rows,
            &params,
        )
        .map_err(|e| PyValueError::new_err(e.to_string()))
        .map(|d| (d, ctx, dev_id))
    })?;

    Ok(DeviceArrayF32FramaPy::new(inner, ctx, dev_id))
}

#[cfg(feature = "python")]
#[pyclass(name = "FramaStream")]
pub struct FramaStreamPy {
    inner: FramaStream,
}

#[cfg(feature = "python")]
#[pymethods]
impl FramaStreamPy {
    #[new]
    fn new(window: usize, sc: usize, fc: usize) -> PyResult<Self> {
        Ok(Self {
            inner: FramaStream::try_new(FramaParams {
                window: Some(window),
                sc: Some(sc),
                fc: Some(fc),
            })
            .map_err(|e| PyValueError::new_err(e.to_string()))?,
        })
    }

    fn update(&mut self, high: f64, low: f64, close: f64) -> Option<f64> {
        self.inner.update(high, low, close)
    }
}

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

#[inline]
pub fn frama_into_slice(
    dst: &mut [f64],
    input: &FramaInput,
    kern: Kernel,
) -> Result<(), FramaError> {
    let ((high, low, close), window, sc, fc, first, len, _warm_from_prepare, chosen) =
        frama_prepare(input, kern)?;

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

    let mut win = window;
    if win & 1 == 1 {
        win += 1;
    }
    let warm = first + win - 1;

    for v in &mut dst[..warm] {
        *v = f64::NAN;
    }

    frama_compute_into(
        high, low, close, window, sc, fc, first, len, warm, chosen, dst,
    )?;

    Ok(())
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen]
pub fn frama_js(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    window: usize,
    sc: usize,
    fc: usize,
) -> Result<Vec<f64>, JsValue> {
    let input = FramaInput::from_slices(
        high,
        low,
        close,
        FramaParams {
            window: Some(window),
            sc: Some(sc),
            fc: Some(fc),
        },
    );

    let ((h, l, c), window, sc, fc, first, len, _warm, _chosen) =
        frama_prepare(&input, Kernel::Scalar).map_err(|e| JsValue::from_str(&e.to_string()))?;

    let mut out = vec![f64::NAN; len];

    let mut stream = FramaStream::try_new(FramaParams {
        window: Some(window),
        sc: Some(sc),
        fc: Some(fc),
    })
    .map_err(|e| JsValue::from_str(&e.to_string()))?;

    for i in first..len {
        if let Some(v) = stream.update(h[i], l[i], c[i]) {
            out[i] = v;
        }
    }

    Ok(out)
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen]
pub fn frama_batch_js(
    high: &[f64],
    low: &[f64],
    close: &[f64],
    window_start: usize,
    window_end: usize,
    window_step: usize,
    sc_start: usize,
    sc_end: usize,
    sc_step: usize,
    fc_start: usize,
    fc_end: usize,
    fc_step: usize,
) -> Result<Vec<f64>, JsValue> {
    let range = FramaBatchRange {
        window: (window_start, window_end, window_step),
        sc: (sc_start, sc_end, sc_step),
        fc: (fc_start, fc_end, fc_step),
    };

    let combos = expand_grid(&range);
    if combos.is_empty() {
        return Err(JsValue::from_str(
            &FramaError::InvalidRange {
                start: window_start,
                end: window_end,
                step: window_step,
            }
            .to_string(),
        ));
    }

    let cols = close.len();
    let rows = combos.len();
    let total = rows.checked_mul(cols).ok_or_else(|| {
        JsValue::from_str(
            &FramaError::ArithmeticOverflow {
                context: "rows*cols",
            }
            .to_string(),
        )
    })?;
    let mut out = vec![f64::NAN; total];

    for (row, p) in combos.iter().enumerate() {
        let window = p.window.unwrap_or(10);
        let sc = p.sc.unwrap_or(300);
        let fc = p.fc.unwrap_or(1);

        let row_out = &mut out[row * cols..(row + 1) * cols];
        let input = FramaInput::from_slices(
            high,
            low,
            close,
            FramaParams {
                window: Some(window),
                sc: Some(sc),
                fc: Some(fc),
            },
        );
        let ((h, l, c), window, sc, fc, first, len, _warm, _chosen) =
            frama_prepare(&input, Kernel::Scalar).map_err(|e| JsValue::from_str(&e.to_string()))?;

        if row_out.len() != len {
            return Err(JsValue::from_str(
                &FramaError::OutputLengthMismatch {
                    expected: len,
                    got: row_out.len(),
                }
                .to_string(),
            ));
        }

        let mut stream = FramaStream::try_new(FramaParams {
            window: Some(window),
            sc: Some(sc),
            fc: Some(fc),
        })
        .map_err(|e| JsValue::from_str(&e.to_string()))?;

        for i in first..len {
            if let Some(v) = stream.update(h[i], l[i], c[i]) {
                row_out[i] = v;
            }
        }
    }

    Ok(out)
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen]
pub fn frama_batch_metadata_js(
    window_start: usize,
    window_end: usize,
    window_step: usize,
    sc_start: usize,
    sc_end: usize,
    sc_step: usize,
    fc_start: usize,
    fc_end: usize,
    fc_step: usize,
) -> Vec<usize> {
    let range = FramaBatchRange {
        window: (window_start, window_end, window_step),
        sc: (sc_start, sc_end, sc_step),
        fc: (fc_start, fc_end, fc_step),
    };

    let combos = expand_grid(&range);
    let mut metadata = Vec::with_capacity(combos.len() * 3);

    for combo in combos {
        metadata.push(combo.window.unwrap_or(10));
        metadata.push(combo.sc.unwrap_or(300));
        metadata.push(combo.fc.unwrap_or(1));
    }

    metadata
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen]
pub fn frama_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 frama_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(js_name = frama_into)]
pub fn frama_into_js(
    high_ptr: *const f64,
    low_ptr: *const f64,
    close_ptr: *const f64,
    out_ptr: *mut f64,
    len: usize,
    window: usize,
    sc: usize,
    fc: usize,
) -> Result<(), JsValue> {
    if high_ptr.is_null() || low_ptr.is_null() || close_ptr.is_null() || out_ptr.is_null() {
        return Err(JsValue::from_str("Null pointer provided"));
    }

    unsafe {
        let h = std::slice::from_raw_parts(high_ptr, len);
        let l = std::slice::from_raw_parts(low_ptr, len);
        let c = std::slice::from_raw_parts(close_ptr, len);
        let out = std::slice::from_raw_parts_mut(out_ptr, len);

        let input = FramaInput::from_slices(
            h,
            l,
            c,
            FramaParams {
                window: Some(window),
                sc: Some(sc),
                fc: Some(fc),
            },
        );

        if out_ptr as *const f64 == high_ptr
            || out_ptr as *const f64 == low_ptr
            || out_ptr as *const f64 == close_ptr
        {
            let mut tmp = vec![0.0; len];
            frama_into_slice(&mut tmp, &input, detect_best_kernel())
                .map_err(|e| JsValue::from_str(&e.to_string()))?;
            out.copy_from_slice(&tmp);
        } else {
            frama_into_slice(out, &input, detect_best_kernel())
                .map_err(|e| JsValue::from_str(&e.to_string()))?;
        }
    }
    Ok(())
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[derive(Serialize, Deserialize)]
pub struct FramaBatchConfig {
    pub window_range: (usize, usize, usize),
    pub sc_range: (usize, usize, usize),
    pub fc_range: (usize, usize, usize),
}

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

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

    let sweep = FramaBatchRange {
        window: config.window_range,
        sc: config.sc_range,
        fc: config.fc_range,
    };

    let output = frama_batch_inner(high, low, close, &sweep, Kernel::Auto, false)
        .map_err(|e| JsValue::from_str(&e.to_string()))?;

    let result = FramaBatchJsOutput {
        values: output.values,
        combos: output.combos,
        rows: output.rows,
        cols: output.cols,
    };

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

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen(js_name = frama_batch_into)]
pub fn frama_batch_into_js(
    high_ptr: *const f64,
    low_ptr: *const f64,
    close_ptr: *const f64,
    out_ptr: *mut f64,
    len: usize,
    w0: usize,
    w1: usize,
    ws: usize,
    s0: usize,
    s1: usize,
    ss: usize,
    f0: usize,
    f1: usize,
    fs: usize,
) -> Result<usize, JsValue> {
    if high_ptr.is_null() || low_ptr.is_null() || close_ptr.is_null() || out_ptr.is_null() {
        return Err(JsValue::from_str("null pointer passed to frama_batch_into"));
    }

    unsafe {
        let h = std::slice::from_raw_parts(high_ptr, len);
        let l = std::slice::from_raw_parts(low_ptr, len);
        let c = std::slice::from_raw_parts(close_ptr, len);
        let sweep = FramaBatchRange {
            window: (w0, w1, ws),
            sc: (s0, s1, ss),
            fc: (f0, f1, fs),
        };

        let combos = expand_grid(&sweep);
        let rows = combos.len();
        let cols = len;
        let out = std::slice::from_raw_parts_mut(out_ptr, rows * cols);
        frama_batch_inner_into(h, l, c, &sweep, detect_best_kernel(), false, out)
            .map_err(|e| JsValue::from_str(&e.to_string()))?;
        Ok(rows)
    }
}