vector_ta/indicators/moving_averages/

nma.rs

#[cfg(all(feature = "python", feature = "cuda"))]
use crate::cuda::moving_averages::{CudaNma, DeviceArrayF32};
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,
};
use aligned_vec::{AVec, CACHELINE_ALIGN};
#[cfg(all(target_arch = "wasm32", target_feature = "simd128"))]
use core::arch::wasm32::*;
#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
use core::arch::x86_64::*;
#[cfg(all(feature = "python", feature = "cuda"))]
use cust::context::Context;
#[cfg(all(feature = "python", feature = "cuda"))]
use cust::memory::DeviceBuffer;
#[cfg(not(target_arch = "wasm32"))]
use rayon::prelude::*;
use std::convert::AsRef;
use std::error::Error;
use std::mem::MaybeUninit;
#[cfg(all(feature = "python", feature = "cuda"))]
use std::sync::Arc;
use thiserror::Error;

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

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

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

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

impl Default for NmaParams {
    fn default() -> Self {
        Self { period: Some(40) }
    }
}

#[derive(Debug, Clone)]
pub struct NmaInput<'a> {
    pub data: NmaData<'a>,
    pub params: NmaParams,
}

impl<'a> NmaInput<'a> {
    #[inline]
    pub fn from_candles(c: &'a Candles, s: &'a str, p: NmaParams) -> Self {
        Self {
            data: NmaData::Candles {
                candles: c,
                source: s,
            },
            params: p,
        }
    }
    #[inline]
    pub fn from_slice(sl: &'a [f64], p: NmaParams) -> Self {
        Self {
            data: NmaData::Slice(sl),
            params: p,
        }
    }
    #[inline]
    pub fn with_default_candles(c: &'a Candles) -> Self {
        Self::from_candles(c, "close", NmaParams::default())
    }
    #[inline]
    pub fn get_period(&self) -> usize {
        self.params.period.unwrap_or(40)
    }
}

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

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

impl NmaBuilder {
    #[inline(always)]
    pub fn new() -> Self {
        Self::default()
    }
    #[inline(always)]
    pub fn period(mut self, n: usize) -> Self {
        self.period = Some(n);
        self
    }
    #[inline(always)]
    pub fn kernel(mut self, k: Kernel) -> Self {
        self.kernel = k;
        self
    }
    #[inline(always)]
    pub fn apply(self, c: &Candles) -> Result<NmaOutput, NmaError> {
        let p = NmaParams {
            period: self.period,
        };
        let i = NmaInput::from_candles(c, "close", p);
        nma_with_kernel(&i, self.kernel)
    }
    #[inline(always)]
    pub fn apply_slice(self, d: &[f64]) -> Result<NmaOutput, NmaError> {
        let p = NmaParams {
            period: self.period,
        };
        let i = NmaInput::from_slice(d, p);
        nma_with_kernel(&i, self.kernel)
    }
    #[inline(always)]
    pub fn into_stream(self) -> Result<NmaStream, NmaError> {
        let p = NmaParams {
            period: self.period,
        };
        NmaStream::try_new(p)
    }
}

#[derive(Debug, Error)]
pub enum NmaError {
    #[error("nma: Input data slice is empty.")]
    EmptyInputData,
    #[error("nma: All values are NaN.")]
    AllValuesNaN,
    #[error("nma: Invalid period: period = {period}, data length = {data_len}")]
    InvalidPeriod { period: usize, data_len: usize },
    #[error("nma: Not enough valid data: needed = {needed}, valid = {valid}")]
    NotEnoughValidData { needed: usize, valid: usize },
    #[error("nma: Output length mismatch: expected = {expected}, got = {got}")]
    OutputLengthMismatch { expected: usize, got: usize },
    #[error("nma: Invalid range: start = {start}, end = {end}, step = {step}")]
    InvalidRange {
        start: usize,
        end: usize,
        step: usize,
    },
    #[error("nma: Invalid kernel for batch path: {0:?}")]
    InvalidKernelForBatch(Kernel),
    #[error("nma: invalid input: {0}")]
    InvalidInput(String),
}

#[inline]
pub fn nma(input: &NmaInput) -> Result<NmaOutput, NmaError> {
    nma_with_kernel(input, Kernel::Auto)
}

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

    if len == 0 {
        return Err(NmaError::EmptyInputData);
    }

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

    let period = input.get_period();

    if period == 0 || period > len {
        return Err(NmaError::InvalidPeriod {
            period,
            data_len: len,
        });
    }
    if (len - first) < (period + 1) {
        return Err(NmaError::NotEnoughValidData {
            needed: period + 1,
            valid: len - first,
        });
    }

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

    let mut ln_values = alloc_with_nan_prefix(len, 0);
    if matches!(chosen, Kernel::Scalar | Kernel::ScalarBatch) {
        for i in 0..len {
            ln_values[i] = data[i].max(1e-10).ln();
        }
    }

    let mut sqrt_diffs = vec![0.0; period];
    for i in 0..period {
        let s0 = (i as f64).sqrt();
        let s1 = ((i + 1) as f64).sqrt();
        sqrt_diffs[i] = s1 - s0;
    }

    Ok((data, period, first, ln_values, sqrt_diffs, chosen))
}

fn nma_compute_into(
    data: &[f64],
    period: usize,
    first: usize,
    ln_values: &mut [f64],
    sqrt_diffs: &mut [f64],
    kernel: Kernel,
    out: &mut [f64],
) {
    unsafe {
        #[cfg(all(target_arch = "wasm32", target_feature = "simd128"))]
        {
            if matches!(kernel, Kernel::Scalar | Kernel::ScalarBatch) {
                nma_simd128(data, period, first, ln_values, sqrt_diffs, out);
                return;
            }
        }

        match kernel {
            Kernel::Scalar | Kernel::ScalarBatch => {
                nma_scalar_with_precomputed(data, period, first, ln_values, sqrt_diffs, out)
            }

            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx2 | Kernel::Avx2Batch => {
                nma_avx2(data, period, first, ln_values, sqrt_diffs, out)
            }

            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx512 | Kernel::Avx512Batch => {
                nma_avx512_v2(data, period, first, ln_values, sqrt_diffs, out)
            }

            #[cfg(not(all(feature = "nightly-avx", target_arch = "x86_64")))]
            Kernel::Avx2 | Kernel::Avx2Batch | Kernel::Avx512 | Kernel::Avx512Batch => {
                nma_scalar_with_precomputed(data, period, first, ln_values, sqrt_diffs, out)
            }
            _ => unreachable!(),
        }
    }
}

pub fn nma_with_kernel(input: &NmaInput, kernel: Kernel) -> Result<NmaOutput, NmaError> {
    let (data, period, first, mut ln_values, mut sqrt_diffs, chosen) = nma_prepare(input, kernel)?;

    let warm = first + period;
    let mut out = alloc_with_nan_prefix(data.len(), warm);

    nma_compute_into(
        data,
        period,
        first,
        &mut ln_values,
        &mut sqrt_diffs,
        chosen,
        &mut out,
    );

    Ok(NmaOutput { values: out })
}
#[inline]
pub fn nma_scalar(data: &[f64], period: usize, first: usize, out: &mut [f64]) {
    let len = data.len();

    let mut ln_values = alloc_with_nan_prefix(len, 0);
    for i in 0..len {
        ln_values[i] = data[i].max(1e-10).ln();
    }

    let mut sqrt_diffs = vec![0.0; period];
    for i in 0..period {
        let s0 = (i as f64).sqrt();
        let s1 = ((i + 1) as f64).sqrt();
        sqrt_diffs[i] = s1 - s0;
    }

    for j in (first + period)..len {
        let mut num = 0.0;
        let mut denom = 0.0;

        for i in 0..period {
            let oi = (ln_values[j - i] - ln_values[j - i - 1]).abs();
            num += oi * sqrt_diffs[i];
            denom += oi;
        }

        let ratio = if denom == 0.0 { 0.0 } else { num / denom };

        let i = period - 1;
        out[j] = data[j - i] * ratio + data[j - i - 1] * (1.0 - ratio);
    }
}

#[inline]
pub fn nma_scalar_with_precomputed(
    data: &[f64],
    period: usize,
    first: usize,
    ln_values: &[f64],
    sqrt_diffs: &[f64],
    out: &mut [f64],
) {
    let len = data.len();

    for j in (first + period)..len {
        let base = j - period;

        let mut num = 0.0f64;
        let mut denom = 0.0f64;

        let mut prev = ln_values[base];
        for t in 0..period {
            let cur = ln_values[base + t + 1];
            let diff = (cur - prev).abs();
            prev = cur;

            num += diff * sqrt_diffs[period - 1 - t];
            denom += diff;
        }

        let ratio = if denom == 0.0 { 0.0 } else { num / denom };

        let x0 = data[j - period];
        let x1 = data[j - period + 1];
        out[j] = (x1 - x0).mul_add(ratio, x0);
    }
}

#[cfg(all(target_arch = "wasm32", target_feature = "simd128"))]
#[inline]
unsafe fn nma_simd128(
    data: &[f64],
    period: usize,
    first: usize,
    ln_values: &[f64],
    sqrt_diffs: &[f64],
    out: &mut [f64],
) {
    use core::arch::wasm32::*;

    const STEP: usize = 2;
    let len = data.len();

    for j in (first + period)..len {
        let chunks = period / STEP;
        let tail = period % STEP;

        let mut num_acc = f64x2_splat(0.0);
        let mut denom_acc = f64x2_splat(0.0);

        for blk in 0..chunks {
            let i = blk * STEP;

            let ln_curr_0 = f64x2(ln_values[j - i], ln_values[j - i - 1]);
            let ln_prev_0 = f64x2(ln_values[j - i - 1], ln_values[j - i - 2]);

            let diff = f64x2_sub(ln_curr_0, ln_prev_0);
            let abs_diff = f64x2_abs(diff);

            let sqrt_d = v128_load(sqrt_diffs.as_ptr().add(i) as *const v128);

            num_acc = f64x2_add(num_acc, f64x2_mul(abs_diff, sqrt_d));
            denom_acc = f64x2_add(denom_acc, abs_diff);
        }

        let mut num = f64x2_extract_lane::<0>(num_acc) + f64x2_extract_lane::<1>(num_acc);
        let mut denom = f64x2_extract_lane::<0>(denom_acc) + f64x2_extract_lane::<1>(denom_acc);

        for i in (chunks * STEP)..period {
            let oi = (ln_values[j - i] - ln_values[j - i - 1]).abs();
            num += oi * sqrt_diffs[i];
            denom += oi;
        }

        let ratio = if denom == 0.0 { 0.0 } else { num / denom };
        let i = period - 1;
        out[j] = data[j - i] * ratio + data[j - i - 1] * (1.0 - ratio);
    }
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline]
#[target_feature(enable = "avx512f,avx512dq,avx512vl,avx512bw,fma")]
unsafe fn fast_ln_avx512_hi(x: __m512d) -> __m512d {
    let one = _mm512_set1_pd(1.0);
    let two = _mm512_set1_pd(2.0);
    let half = _mm512_set1_pd(0.5);
    let ln2 = _mm512_set1_pd(std::f64::consts::LN_2);
    let sqrt_half = _mm512_set1_pd(0.7071067811865475244);

    let threshold = _mm512_set1_pd(0.2);
    let x_minus_1 = _mm512_sub_pd(x, one);
    let abs_x_minus_1 = _mm512_abs_pd(x_minus_1);
    let near_one_mask = _mm512_cmp_pd_mask(abs_x_minus_1, threshold, _CMP_LT_OQ);

    let c2 = _mm512_set1_pd(-0.5);
    let c3 = _mm512_set1_pd(1.0 / 3.0);
    let c4 = _mm512_set1_pd(-0.25);
    let c5 = _mm512_set1_pd(0.2);
    let c6 = _mm512_set1_pd(-1.0 / 6.0);
    let c7 = _mm512_set1_pd(1.0 / 7.0);
    let c8 = _mm512_set1_pd(-0.125);

    let y = x_minus_1;
    let y2 = _mm512_mul_pd(y, y);
    let y3 = _mm512_mul_pd(y2, y);
    let y4 = _mm512_mul_pd(y2, y2);

    let mut taylor = y;
    taylor = _mm512_fmadd_pd(y2, c2, taylor);
    taylor = _mm512_fmadd_pd(y3, c3, taylor);
    taylor = _mm512_fmadd_pd(y4, c4, taylor);
    let y5 = _mm512_mul_pd(y4, y);
    let y6 = _mm512_mul_pd(y4, y2);
    let y7 = _mm512_mul_pd(y4, y3);
    let y8 = _mm512_mul_pd(y4, y4);
    taylor = _mm512_fmadd_pd(y5, c5, taylor);
    taylor = _mm512_fmadd_pd(y6, c6, taylor);
    taylor = _mm512_fmadd_pd(y7, c7, taylor);
    taylor = _mm512_fmadd_pd(y8, c8, taylor);

    let ix = _mm512_castpd_si512(x);
    let exp_mask = _mm512_set1_epi64(0x7FF0000000000000u64 as i64);
    let mantissa_mask = _mm512_set1_epi64(0x000FFFFFFFFFFFFFu64 as i64);
    let bias = _mm512_set1_epi64(1023);

    let exp_bits = _mm512_and_si512(ix, exp_mask);
    let exp_shifted = _mm512_srli_epi64::<52>(exp_bits);
    let e = _mm512_sub_epi64(exp_shifted, bias);
    let e_f64 = _mm512_cvtepi64_pd(e);

    let mantissa_bits = _mm512_and_si512(ix, mantissa_mask);
    let one_bits = _mm512_set1_epi64(0x3FF0000000000000u64 as i64);
    let m_bits = _mm512_or_si512(mantissa_bits, one_bits);
    let mut m = _mm512_castsi512_pd(m_bits);

    let needs_fold = _mm512_cmp_pd_mask(m, sqrt_half, _CMP_LT_OQ);
    m = _mm512_mask_mul_pd(m, needs_fold, m, two);
    let e_adjust = _mm512_mask_sub_pd(e_f64, needs_fold, e_f64, one);

    let f = _mm512_sub_pd(m, one);

    let two_plus_f = _mm512_add_pd(two, f);
    let s = _mm512_div_pd(f, two_plus_f);
    let z = _mm512_mul_pd(s, s);
    let w = _mm512_mul_pd(z, z);

    let lg1 = _mm512_set1_pd(6.666666666666735130e-01);
    let lg2 = _mm512_set1_pd(3.999999999940941908e-01);
    let lg3 = _mm512_set1_pd(2.857142874366239149e-01);
    let lg4 = _mm512_set1_pd(2.222219843214978396e-01);
    let lg5 = _mm512_set1_pd(1.818357216161805012e-01);
    let lg6 = _mm512_set1_pd(1.531383769920937332e-01);
    let lg7 = _mm512_set1_pd(1.479819860511658591e-01);

    let lg8 = _mm512_set1_pd(1.333355814642869980e-01);
    let lg9 = _mm512_set1_pd(1.253141636393179328e-01);

    let mut r1 = lg9;
    r1 = _mm512_fmadd_pd(r1, z, lg7);
    r1 = _mm512_fmadd_pd(r1, z, lg5);
    r1 = _mm512_fmadd_pd(r1, z, lg3);
    r1 = _mm512_fmadd_pd(r1, z, lg1);
    r1 = _mm512_mul_pd(r1, z);

    let mut r2 = lg8;
    r2 = _mm512_fmadd_pd(r2, z, lg6);
    r2 = _mm512_fmadd_pd(r2, z, lg4);
    r2 = _mm512_fmadd_pd(r2, z, lg2);
    r2 = _mm512_mul_pd(r2, w);

    let r = _mm512_add_pd(r1, r2);

    let hfsq = _mm512_mul_pd(_mm512_mul_pd(half, f), f);

    let ln1pf = _mm512_sub_pd(f, hfsq);
    let s_squared_times_f = _mm512_mul_pd(_mm512_mul_pd(s, s), f);
    let ln1pf = _mm512_fmadd_pd(s_squared_times_f, r, ln1pf);

    let general_result = _mm512_fmadd_pd(e_adjust, ln2, ln1pf);

    _mm512_mask_blend_pd(near_one_mask, general_result, taylor)
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline]
#[target_feature(enable = "avx2,fma")]
pub unsafe fn nma_avx2(
    data: &[f64],
    period: usize,
    first: usize,
    ln_values: &mut [f64],
    sqrt_diffs: &mut [f64],
    out: &mut [f64],
) {
    let len = data.len();

    let epsilon = _mm256_set1_pd(1e-10);

    let one = _mm256_set1_pd(1.0);
    let zero = _mm256_setzero_pd();

    let mut i = 0;
    while i + 4 <= len {
        let vals = _mm256_loadu_pd(data.as_ptr().add(i));
        let clamped = _mm256_max_pd(vals, epsilon);

        let mut ln_vals = [0.0f64; 4];
        _mm256_storeu_pd(ln_vals.as_mut_ptr(), clamped);
        for j in 0..4 {
            ln_vals[j] = ln_vals[j].ln();
        }
        let ln_result = _mm256_loadu_pd(ln_vals.as_ptr());

        _mm256_storeu_pd(ln_values.as_mut_ptr().add(i), ln_result);

        i += 4;
    }

    for j in i..len {
        ln_values[j] = data[j].max(1e-10).ln();
    }

    for j in (first + period)..len {
        let mut num_accum = zero;
        let mut denom_accum = zero;

        let mut idx = 0;
        while idx + 4 <= period {
            let mut diffs = [0.0f64; 4];
            for k in 0..4 {
                let i = idx + k;
                let diff = (ln_values[j - i] - ln_values[j - i - 1]).abs();
                diffs[k] = diff;
            }
            let oi_vec = _mm256_loadu_pd(diffs.as_ptr());

            let weights = _mm256_loadu_pd(sqrt_diffs.as_ptr().add(idx));

            num_accum = _mm256_fmadd_pd(oi_vec, weights, num_accum);
            denom_accum = _mm256_add_pd(denom_accum, oi_vec);

            idx += 4;
        }

        let num_scalar = horizontal_sum_avx2(num_accum);
        let denom_scalar = horizontal_sum_avx2(denom_accum);

        let mut num_final = num_scalar;
        let mut denom_final = denom_scalar;

        for i in idx..period {
            let oi = (ln_values[j - i] - ln_values[j - i - 1]).abs();
            num_final += oi * sqrt_diffs[i];
            denom_final += oi;
        }

        let ratio = if denom_final == 0.0 {
            0.0
        } else {
            num_final / denom_final
        };
        let i = period - 1;
        out[j] = data[j - i] * ratio + data[j - i - 1] * (1.0 - ratio);
    }
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline]
#[target_feature(enable = "avx2")]
unsafe fn horizontal_sum_avx2(v: __m256d) -> f64 {
    let vlow = _mm256_castpd256_pd128(v);
    let vhigh = _mm256_extractf128_pd(v, 1);

    let sum128 = _mm_add_pd(vlow, vhigh);

    let high64 = _mm_unpackhi_pd(sum128, sum128);

    _mm_cvtsd_f64(_mm_add_sd(sum128, high64))
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline]
#[target_feature(enable = "avx2")]
unsafe fn fast_ln_avx2_hi(x: __m256d) -> __m256d {
    let one = _mm256_set1_pd(1.0);
    let two = _mm256_set1_pd(2.0);
    let half = _mm256_set1_pd(0.5);
    let ln2 = _mm256_set1_pd(std::f64::consts::LN_2);
    let sqrt_half = _mm256_set1_pd(0.7071067811865475244);

    let mut mantissa = [0.0f64; 4];
    let mut exponent = [0i32; 4];
    _mm256_storeu_pd(mantissa.as_mut_ptr(), x);

    for j in 0..4 {
        let bits = mantissa[j].to_bits();
        let exp_bits = ((bits >> 52) & 0x7FF) as i32;
        exponent[j] = exp_bits - 1023;

        let mantissa_bits = (bits & !0x7FF0000000000000) | 0x3FF0000000000000;
        mantissa[j] = f64::from_bits(mantissa_bits);
    }

    let mut m = _mm256_loadu_pd(mantissa.as_ptr());
    let e_vals = [
        exponent[0] as f64,
        exponent[1] as f64,
        exponent[2] as f64,
        exponent[3] as f64,
    ];
    let mut e_f64 = _mm256_loadu_pd(e_vals.as_ptr());

    let mask = _mm256_cmp_pd(m, sqrt_half, _CMP_LT_OQ);
    m = _mm256_blendv_pd(m, _mm256_mul_pd(m, two), mask);
    e_f64 = _mm256_blendv_pd(e_f64, _mm256_sub_pd(e_f64, one), mask);

    let f = _mm256_sub_pd(m, one);

    let two_plus_f = _mm256_add_pd(two, f);
    let s = _mm256_div_pd(f, two_plus_f);
    let z = _mm256_mul_pd(s, s);
    let w = _mm256_mul_pd(z, z);

    let lg1 = _mm256_set1_pd(6.666666666666735130e-01);
    let lg2 = _mm256_set1_pd(3.999999999940941908e-01);
    let lg3 = _mm256_set1_pd(2.857142874366239149e-01);
    let lg4 = _mm256_set1_pd(2.222219843214978396e-01);
    let lg5 = _mm256_set1_pd(1.818357216161805012e-01);
    let lg6 = _mm256_set1_pd(1.531383769920937332e-01);
    let lg7 = _mm256_set1_pd(1.479819860511658591e-01);

    let mut r1 = lg7;
    r1 = _mm256_fmadd_pd(r1, z, lg5);
    r1 = _mm256_fmadd_pd(r1, z, lg3);
    r1 = _mm256_fmadd_pd(r1, z, lg1);
    r1 = _mm256_mul_pd(r1, z);

    let mut r2 = lg6;
    r2 = _mm256_fmadd_pd(r2, z, lg4);
    r2 = _mm256_fmadd_pd(r2, z, lg2);
    r2 = _mm256_mul_pd(r2, w);

    let r = _mm256_add_pd(r1, r2);

    let hfsq = _mm256_mul_pd(_mm256_mul_pd(half, f), f);
    let f_times_hfsq = _mm256_mul_pd(f, hfsq);
    let ln1pf = _mm256_sub_pd(f, hfsq);
    let ln1pf = _mm256_fmadd_pd(f_times_hfsq, r, ln1pf);

    _mm256_fmadd_pd(e_f64, ln2, ln1pf)
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline]
#[target_feature(enable = "avx512f")]
unsafe fn _mm512_abs_pd(a: __m512d) -> __m512d {
    let sign_mask = _mm512_set1_pd(-0.0);
    _mm512_andnot_pd(sign_mask, a)
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline]
#[target_feature(enable = "avx512f,avx512dq,avx512vl,avx512bw,fma")]
pub unsafe fn nma_avx512(
    data: &[f64],
    period: usize,
    first: usize,
    ln_values: &mut [f64],
    sqrt_diffs: &mut [f64],
    out: &mut [f64],
) {
    let len = data.len();

    let one = _mm512_set1_pd(1.0);
    let zero = _mm512_setzero_pd();

    for i in 0..len {
        ln_values[i] = data[i].max(1e-10).ln();
    }

    for j in (first + period)..len {
        let mut num_accum = zero;
        let mut denom_accum = zero;

        let mut idx = 0;
        while idx + 8 <= period {
            if j >= idx + 8 {
                let base_ptr = ln_values.as_ptr().add(j - idx - 8);

                let prev = _mm512_loadu_pd(base_ptr);

                let curr = _mm512_loadu_pd(base_ptr.add(1));

                let diff = _mm512_sub_pd(curr, prev);
                let abs_diff = _mm512_abs_pd(diff);

                let perm_indices = _mm512_set_epi64(7, 6, 5, 4, 3, 2, 1, 0);
                let oi_vec = _mm512_permutexvar_pd(perm_indices, abs_diff);

                let weights = _mm512_loadu_pd(sqrt_diffs.as_ptr().add(idx));

                num_accum = _mm512_fmadd_pd(oi_vec, weights, num_accum);
                denom_accum = _mm512_add_pd(denom_accum, oi_vec);
            } else {
                for k in 0..8 {
                    let i = idx + k;
                    let oi = (ln_values[j - i] - ln_values[j - i - 1]).abs();
                    let weight = sqrt_diffs[i];
                    num_accum = _mm512_mask_add_pd(
                        num_accum,
                        1 << k,
                        num_accum,
                        _mm512_set1_pd(oi * weight),
                    );
                    denom_accum =
                        _mm512_mask_add_pd(denom_accum, 1 << k, denom_accum, _mm512_set1_pd(oi));
                }
            }

            idx += 8;
        }

        let mut num_scalar = _mm512_reduce_add_pd(num_accum);
        let mut denom_scalar = _mm512_reduce_add_pd(denom_accum);

        for i in idx..period {
            let oi = (ln_values[j - i] - ln_values[j - i - 1]).abs();
            num_scalar += oi * sqrt_diffs[i];
            denom_scalar += oi;
        }

        let ratio = if denom_scalar == 0.0 {
            0.0
        } else {
            num_scalar / denom_scalar
        };
        let i = period - 1;
        out[j] = data[j - i] * ratio + data[j - i - 1] * (1.0 - ratio);
    }
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline]
#[target_feature(enable = "avx512f,avx512dq,avx512vl,fma")]
pub unsafe fn nma_avx512_v2(
    data: &[f64],
    period: usize,
    first: usize,
    ln_values: &mut [f64],
    sqrt_diffs: &mut [f64],
    out: &mut [f64],
) {
    use aligned_vec::AVec;
    use core::arch::x86_64::*;

    let len = data.len();
    debug_assert!(len == ln_values.len());
    debug_assert!(period >= 1 && period <= len);

    for i in 0..len {
        ln_values[i] = data[i].max(1e-10).ln();
    }

    for i in 0..len - 1 {
        ln_values[i] = (ln_values[i + 1] - ln_values[i]).abs();
    }
    ln_values[len - 1] = 0.0;
    let d = ln_values;

    let mut s = alloc_with_nan_prefix(len + 1, 0);
    s[0] = 0.0;
    for k in 0..len {
        s[k + 1] = s[k] + d[k];
    }

    let wlen_padded = (period + 7) & !7;
    let mut w_rev = AVec::<f64>::with_capacity(64, wlen_padded);
    w_rev.resize(wlen_padded, 0.0);
    for i in 0..period {
        w_rev[i] = sqrt_diffs[period - 1 - i];
    }

    let warm = first + period;
    let zero = _mm512_setzero_pd();

    for j in warm..len {
        let base = j - period;

        let denom = s[j] - s[j - period];

        let mut num_acc = zero;
        let mut t = 0usize;

        while t + 16 <= period {
            let d0 = _mm512_loadu_pd(d.as_ptr().add(base + t));
            let w0 = _mm512_loadu_pd(w_rev.as_ptr().add(t));
            let d1 = _mm512_loadu_pd(d.as_ptr().add(base + t + 8));
            let w1 = _mm512_loadu_pd(w_rev.as_ptr().add(t + 8));
            num_acc = _mm512_fmadd_pd(d0, w0, num_acc);
            num_acc = _mm512_fmadd_pd(d1, w1, num_acc);
            t += 16;
        }
        while t + 8 <= period {
            let d0 = _mm512_loadu_pd(d.as_ptr().add(base + t));
            let w0 = _mm512_loadu_pd(w_rev.as_ptr().add(t));
            num_acc = _mm512_fmadd_pd(d0, w0, num_acc);
            t += 8;
        }
        if t < period {
            let tail = (period - t) as u32;
            let mask: __mmask8 = ((1u32 << tail) - 1) as u8;
            let d0 = _mm512_maskz_loadu_pd(mask, d.as_ptr().add(base + t));
            let w0 = _mm512_maskz_loadu_pd(mask, w_rev.as_ptr().add(t));
            num_acc = _mm512_fmadd_pd(d0, w0, num_acc);
        }

        let num = _mm512_reduce_add_pd(num_acc);
        let ratio = if denom == 0.0 { 0.0 } else { num / denom };

        let i0 = period - 1;
        let x2 = data[j - i0 - 1];
        let dx = data[j - i0] - x2;
        out[j] = ratio.mul_add(dx, x2);
    }
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[target_feature(enable = "avx512f,avx512dq,avx512vl,fma")]
unsafe fn nma_batch_avx512_optimized(
    data: &[f64],
    sweep: &NmaBatchRange,
    first: usize,
    parallel: bool,
) -> Result<NmaBatchOutput, NmaError> {
    use aligned_vec::AVec;
    use core::arch::x86_64::*;

    let combos = expand_grid(sweep)?;
    if combos.is_empty() {
        return Err(NmaError::InvalidPeriod {
            period: 0,
            data_len: 0,
        });
    }

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

    let mut ln_values = alloc_with_nan_prefix(len, 0);
    for i in 0..len {
        ln_values[i] = data[i].max(1e-10).ln();
    }

    for i in 0..len - 1 {
        ln_values[i] = (ln_values[i + 1] - ln_values[i]).abs();
    }
    ln_values[len - 1] = 0.0;
    let d = &mut ln_values;

    let mut s = alloc_with_nan_prefix(len + 1, 0);
    s[0] = 0.0;
    for k in 0..len {
        s[k + 1] = s[k] + d[k];
    }

    let warm: Vec<usize> = combos.iter().map(|c| first + c.period.unwrap()).collect();
    let mut raw = make_uninit_matrix(rows, cols);
    unsafe { init_matrix_prefixes(&mut raw, cols, &warm) };

    let do_row = |row: usize, dst_mu: &mut [MaybeUninit<f64>]| unsafe {
        let period = combos[row].period.unwrap();
        let warm = first + period;

        let out_row =
            core::slice::from_raw_parts_mut(dst_mu.as_mut_ptr() as *mut f64, dst_mu.len());

        let wlen_padded = (period + 7) & !7;
        let mut w_rev = AVec::<f64>::with_capacity(64, wlen_padded);
        w_rev.resize(wlen_padded, 0.0);

        for i in 0..period {
            let s0 = ((period - 1 - i) as f64).sqrt();
            let s1 = ((period - i) as f64).sqrt();
            w_rev[i] = s1 - s0;
        }

        let zero = _mm512_setzero_pd();

        for j in warm..len {
            let base = j - period;

            let denom = s[j] - s[j - period];

            let mut num_acc = zero;
            let mut t = 0usize;

            while t + 16 <= period {
                let d0 = _mm512_loadu_pd(d.as_ptr().add(base + t));
                let w0 = _mm512_loadu_pd(w_rev.as_ptr().add(t));
                let d1 = _mm512_loadu_pd(d.as_ptr().add(base + t + 8));
                let w1 = _mm512_loadu_pd(w_rev.as_ptr().add(t + 8));
                num_acc = _mm512_fmadd_pd(d0, w0, num_acc);
                num_acc = _mm512_fmadd_pd(d1, w1, num_acc);
                t += 16;
            }
            while t + 8 <= period {
                let d0 = _mm512_loadu_pd(d.as_ptr().add(base + t));
                let w0 = _mm512_loadu_pd(w_rev.as_ptr().add(t));
                num_acc = _mm512_fmadd_pd(d0, w0, num_acc);
                t += 8;
            }
            if t < period {
                let tail = (period - t) as u32;
                let mask: __mmask8 = ((1u32 << tail) - 1) as u8;
                let d0 = _mm512_maskz_loadu_pd(mask, d.as_ptr().add(base + t));
                let w0 = _mm512_maskz_loadu_pd(mask, w_rev.as_ptr().add(t));
                num_acc = _mm512_fmadd_pd(d0, w0, num_acc);
            }

            let num = _mm512_reduce_add_pd(num_acc);
            let ratio = if denom == 0.0 { 0.0 } else { num / denom };

            let i0 = period - 1;
            let x2 = data[j - i0 - 1];
            let dx = data[j - i0] - x2;
            out_row[j] = ratio.mul_add(dx, x2);
        }
    };

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

    let values: Vec<f64> = unsafe { std::mem::transmute(raw) };

    Ok(NmaBatchOutput {
        values,
        combos,
        rows,
        cols,
    })
}

#[inline(always)]
pub fn nma_batch_with_kernel(
    data: &[f64],
    sweep: &NmaBatchRange,
    k: Kernel,
) -> Result<NmaBatchOutput, NmaError> {
    let kernel = match k {
        Kernel::Auto => detect_best_batch_kernel(),
        other if other.is_batch() => other,
        _ => return Err(NmaError::InvalidKernelForBatch(k)),
    };

    let simd = match kernel {
        Kernel::Avx512Batch => Kernel::Avx512,
        Kernel::Avx2Batch => Kernel::Avx2,
        Kernel::ScalarBatch => Kernel::Scalar,
        _ => Kernel::Scalar,
    };
    nma_batch_par_slice(data, sweep, simd)
}

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

impl Default for NmaBatchRange {
    fn default() -> Self {
        Self {
            period: (40, 289, 1),
        }
    }
}

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

impl NmaBatchBuilder {
    pub fn new() -> Self {
        Self::default()
    }
    pub fn kernel(mut self, k: Kernel) -> Self {
        self.kernel = k;
        self
    }

    #[inline]
    pub fn period_range(mut self, start: usize, end: usize, step: usize) -> Self {
        self.range.period = (start, end, step);
        self
    }
    #[inline]
    pub fn period_static(mut self, p: usize) -> Self {
        self.range.period = (p, p, 0);
        self
    }

    pub fn apply_slice(self, data: &[f64]) -> Result<NmaBatchOutput, NmaError> {
        nma_batch_with_kernel(data, &self.range, self.kernel)
    }

    pub fn with_default_slice(data: &[f64], k: Kernel) -> Result<NmaBatchOutput, NmaError> {
        NmaBatchBuilder::new().kernel(k).apply_slice(data)
    }

    pub fn apply_candles(self, c: &Candles, src: &str) -> Result<NmaBatchOutput, NmaError> {
        let slice = source_type(c, src);
        self.apply_slice(slice)
    }

    pub fn with_default_candles(c: &Candles) -> Result<NmaBatchOutput, NmaError> {
        NmaBatchBuilder::new()
            .kernel(Kernel::Auto)
            .apply_candles(c, "close")
    }
}

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

impl NmaBatchOutput {
    pub fn row_for_params(&self, p: &NmaParams) -> Option<usize> {
        self.combos
            .iter()
            .position(|c| c.period.unwrap_or(40) == p.period.unwrap_or(40))
    }

    pub fn values_for(&self, p: &NmaParams) -> 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: &NmaBatchRange) -> Result<Vec<NmaParams>, NmaError> {
    fn axis_usize((start, end, step): (usize, usize, usize)) -> Result<Vec<usize>, NmaError> {
        if step == 0 || start == end {
            return Ok(vec![start]);
        }
        if start < end {
            let mut v = Vec::new();
            let mut cur = start;
            while cur <= end {
                v.push(cur);
                cur = cur
                    .checked_add(step)
                    .ok_or_else(|| NmaError::InvalidRange { start, end, step })?;
            }
            if v.is_empty() {
                return Err(NmaError::InvalidRange { start, end, step });
            }
            Ok(v)
        } else {
            Err(NmaError::InvalidRange { start, end, step })
        }
    }
    let periods = axis_usize(r.period)?;

    let mut out = Vec::with_capacity(periods.len());
    for &p in &periods {
        out.push(NmaParams { period: Some(p) });
    }
    Ok(out)
}

#[inline]
fn round_up8(x: usize) -> usize {
    (x + 7) & !7
}

#[inline(always)]
fn nma_batch_inner_into_scalar_reuse(
    data: &[f64],
    sweep: &NmaBatchRange,
    parallel: bool,
    out: &mut [f64],
) -> Result<Vec<NmaParams>, NmaError> {
    let combos = expand_grid(sweep)?;
    if combos.is_empty() {
        return Err(NmaError::InvalidInput("no parameter combinations".into()));
    }

    let len = data.len();
    let first = data
        .iter()
        .position(|x| !x.is_nan())
        .ok_or(NmaError::AllValuesNaN)?;
    let max_p = combos.iter().map(|c| c.period.unwrap()).max().unwrap();
    if len - first < max_p + 1 {
        return Err(NmaError::NotEnoughValidData {
            needed: max_p + 1,
            valid: len - first,
        });
    }

    let rows = combos.len();
    let cols = len;
    let warm: Vec<usize> = combos.iter().map(|c| first + c.period.unwrap()).collect();
    let out_mu = unsafe {
        std::slice::from_raw_parts_mut(out.as_mut_ptr() as *mut MaybeUninit<f64>, out.len())
    };
    unsafe { init_matrix_prefixes(out_mu, cols, &warm) };

    let mut ln = alloc_with_nan_prefix(len, 0);
    for i in 0..len {
        ln[i] = data[i].max(1e-10).ln();
    }
    for i in 0..len.saturating_sub(1) {
        ln[i] = (ln[i + 1] - ln[i]).abs();
    }
    ln[len.saturating_sub(1)] = 0.0;
    let d = &ln;

    let mut s = alloc_with_nan_prefix(len + 1, 0);
    s[0] = 0.0;
    for i in 0..len {
        s[i + 1] = s[i] + d[i];
    }

    let do_row = |row: usize, dst_mu: &mut [MaybeUninit<f64>]| {
        let p = combos[row].period.unwrap();
        let warm = first + p;

        let mut w_rev = Vec::with_capacity(p);
        for i in 0..p {
            let s0 = ((p - 1 - i) as f64).sqrt();
            let s1 = ((p - i) as f64).sqrt();
            w_rev.push(s1 - s0);
        }
        let dst = unsafe {
            std::slice::from_raw_parts_mut(dst_mu.as_mut_ptr() as *mut f64, dst_mu.len())
        };

        for j in warm..len {
            let base = j - p;
            let denom = s[j] - s[j - p];

            let mut num = 0.0;

            for t in 0..p {
                num += d[base + t] * w_rev[t];
            }

            let ratio = if denom == 0.0 { 0.0 } else { num / denom };
            let x2 = data[j - p];
            let x1 = data[j - p + 1];
            dst[j] = ratio.mul_add(x1 - x2, x2);
        }
    };

    if parallel {
        #[cfg(not(target_arch = "wasm32"))]
        {
            use rayon::prelude::*;
            out_mu
                .par_chunks_mut(cols)
                .enumerate()
                .for_each(|(r, row)| do_row(r, row));
        }
        #[cfg(target_arch = "wasm32")]
        for (r, row) in out_mu.chunks_mut(cols).enumerate() {
            do_row(r, row);
        }
    } else {
        for (r, row) in out_mu.chunks_mut(cols).enumerate() {
            do_row(r, row);
        }
    }

    Ok(combos)
}

#[inline(always)]
pub fn nma_batch_slice(
    data: &[f64],
    sweep: &NmaBatchRange,
    kern: Kernel,
) -> Result<NmaBatchOutput, NmaError> {
    nma_batch_inner(data, sweep, kern, false)
}

#[inline(always)]
pub fn nma_batch_par_slice(
    data: &[f64],
    sweep: &NmaBatchRange,
    kern: Kernel,
) -> Result<NmaBatchOutput, NmaError> {
    nma_batch_inner(data, sweep, kern, true)
}

#[inline(always)]
fn nma_batch_inner(
    data: &[f64],
    sweep: &NmaBatchRange,
    kern: Kernel,
    parallel: bool,
) -> Result<NmaBatchOutput, NmaError> {
    #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
    if kern == Kernel::Avx512 {
        let first = data
            .iter()
            .position(|x| !x.is_nan())
            .ok_or(NmaError::AllValuesNaN)?;
        return unsafe { nma_batch_avx512_optimized(data, sweep, first, parallel) };
    }

    let combos = expand_grid(sweep)?;
    if combos.is_empty() {
        return Err(NmaError::InvalidInput("no parameter combinations".into()));
    }
    let rows = combos.len();
    let cols = data.len();
    let _ = rows
        .checked_mul(cols)
        .ok_or_else(|| NmaError::InvalidInput("rows*cols overflow".into()))?;

    if kern == Kernel::Scalar {
        let first = data
            .iter()
            .position(|x| !x.is_nan())
            .ok_or(NmaError::AllValuesNaN)?;
        let warm: Vec<usize> = combos.iter().map(|c| first + c.period.unwrap()).collect();
        let mut raw = make_uninit_matrix(rows, cols);
        unsafe { init_matrix_prefixes(&mut raw, cols, &warm) };

        let out: &mut [f64] =
            unsafe { std::slice::from_raw_parts_mut(raw.as_mut_ptr() as *mut f64, raw.len()) };
        let combos = nma_batch_inner_into_scalar_reuse(data, sweep, parallel, out)?;

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

    let first = data
        .iter()
        .position(|x| !x.is_nan())
        .ok_or(NmaError::AllValuesNaN)?;
    let max_p = combos
        .iter()
        .map(|c| round_up8(c.period.unwrap()))
        .max()
        .unwrap();
    if data.len() - first < max_p + 1 {
        return Err(NmaError::NotEnoughValidData {
            needed: max_p + 1,
            valid: data.len() - first,
        });
    }

    let warm: Vec<usize> = combos.iter().map(|c| first + c.period.unwrap()).collect();

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

    let do_row = |row: usize, dst_mu: &mut [MaybeUninit<f64>]| {
        let period = combos[row].period.unwrap();

        let out_row = unsafe {
            core::slice::from_raw_parts_mut(dst_mu.as_mut_ptr() as *mut f64, dst_mu.len())
        };

        match kern {
            Kernel::Scalar => nma_row_scalar(data, first, period, out_row),
            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx2 => unsafe { nma_row_avx2(data, first, period, out_row) },
            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx512 => unsafe { nma_row_avx512(data, first, period, out_row) },
            #[cfg(not(all(feature = "nightly-avx", target_arch = "x86_64")))]
            Kernel::Avx2 | Kernel::Avx512 => nma_row_scalar(data, first, period, out_row),
            _ => nma_row_scalar(data, first, period, out_row),
        }
    };

    if parallel {
        #[cfg(not(target_arch = "wasm32"))]
        {
            use rayon::prelude::*;
            raw.par_chunks_mut(cols)
                .enumerate()
                .for_each(|(row, slice)| do_row(row, slice));
        }

        #[cfg(target_arch = "wasm32")]
        {
            for (row, slice) in raw.chunks_mut(cols).enumerate() {
                do_row(row, slice);
            }
        }
    } else {
        for (row, slice) in raw.chunks_mut(cols).enumerate() {
            do_row(row, slice);
        }
    }

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

    Ok(NmaBatchOutput {
        values,
        combos,
        rows,
        cols,
    })
}

#[inline(always)]
fn nma_batch_inner_into(
    data: &[f64],
    sweep: &NmaBatchRange,
    kern: Kernel,
    parallel: bool,
    out: &mut [f64],
) -> Result<Vec<NmaParams>, NmaError> {
    let combos = expand_grid(sweep)?;
    if combos.is_empty() {
        return Err(NmaError::InvalidInput("no parameter combinations".into()));
    }

    let first = data
        .iter()
        .position(|x| !x.is_nan())
        .ok_or(NmaError::AllValuesNaN)?;
    let max_p = combos
        .iter()
        .map(|c| round_up8(c.period.unwrap()))
        .max()
        .unwrap();
    if data.len() - first < max_p + 1 {
        return Err(NmaError::NotEnoughValidData {
            needed: max_p + 1,
            valid: data.len() - first,
        });
    }

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

    let warm: Vec<usize> = combos.iter().map(|c| first + c.period.unwrap()).collect();

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

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

    let do_row = |row: usize, dst_mu: &mut [MaybeUninit<f64>]| {
        let period = combos[row].period.unwrap();

        let out_row = unsafe {
            core::slice::from_raw_parts_mut(dst_mu.as_mut_ptr() as *mut f64, dst_mu.len())
        };

        match kern {
            Kernel::Scalar => nma_row_scalar(data, first, period, out_row),
            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx2 => unsafe { nma_row_avx2(data, first, period, out_row) },
            #[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
            Kernel::Avx512 => unsafe { nma_row_avx512(data, first, period, out_row) },
            #[cfg(not(all(feature = "nightly-avx", target_arch = "x86_64")))]
            Kernel::Avx2 | Kernel::Avx512 => nma_row_scalar(data, first, period, out_row),
            _ => nma_row_scalar(data, first, period, out_row),
        }
    };

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

    Ok(combos)
}

#[inline(always)]
fn nma_row_scalar(data: &[f64], first: usize, period: usize, out: &mut [f64]) {
    nma_scalar(data, period, first, out)
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline(always)]
unsafe fn nma_row_avx2(data: &[f64], first: usize, period: usize, out: &mut [f64]) {
    let len = data.len();
    let mut ln_values = alloc_with_nan_prefix(len, 0);

    let mut sqrt_diffs = vec![0.0; period];

    for i in 0..len {
        ln_values[i] = data[i].max(1e-10).ln();
    }

    for k in 0..period {
        let s0 = (k as f64).sqrt();
        let s1 = ((k + 1) as f64).sqrt();
        sqrt_diffs[k] = s1 - s0;
    }

    nma_avx2(data, period, first, &mut ln_values, &mut sqrt_diffs, out);
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline(always)]
pub unsafe fn nma_row_avx512(data: &[f64], first: usize, period: usize, out: &mut [f64]) {
    let len = data.len();
    let mut ln_values = alloc_with_nan_prefix(len, 0);

    let mut sqrt_diffs = vec![0.0; period];

    for i in 0..len {
        ln_values[i] = data[i].max(1e-10).ln();
    }

    for k in 0..period {
        let s0 = (k as f64).sqrt();
        let s1 = ((k + 1) as f64).sqrt();
        sqrt_diffs[k] = s1 - s0;
    }

    nma_avx512_v2(data, period, first, &mut ln_values, &mut sqrt_diffs, out);
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline(always)]
pub unsafe fn nma_row_avx512_short(data: &[f64], first: usize, period: usize, out: &mut [f64]) {
    nma_row_avx512(data, first, period, out)
}

#[cfg(all(feature = "nightly-avx", target_arch = "x86_64"))]
#[inline(always)]
pub unsafe fn nma_row_avx512_long(data: &[f64], first: usize, period: usize, out: &mut [f64]) {
    nma_row_avx512(data, first, period, out)
}

#[derive(Debug, Clone)]
pub struct NmaStream {
    period: usize,

    m: usize,

    alpha: Vec<f64>,
    beta: Vec<f64>,
    beta_pow_p: Vec<f64>,

    d_ring: Vec<f64>,
    d_head: usize,
    d_count: usize,
    denom: f64,
    x_acc: Vec<f64>,

    buffer: Vec<f64>,
    ln_buffer: Vec<f64>,
    head: usize,
    filled: bool,

    sqrt_diffs: Vec<f64>,
}

#[inline(always)]
fn ln_pos(x: f64) -> f64 {
    debug_assert!(x > 0.0);
    x.ln()
}

impl NmaStream {
    pub fn try_new(params: NmaParams) -> Result<Self, NmaError> {
        let period = params.period.unwrap_or(40);
        if period == 0 {
            return Err(NmaError::InvalidPeriod {
                period,
                data_len: 0,
            });
        }

        let mut sqrt_diffs = Vec::with_capacity(period);
        for i in 0..period {
            let s0 = (i as f64).sqrt();
            let s1 = ((i + 1) as f64).sqrt();
            sqrt_diffs.push(s1 - s0);
        }

        const GAMMAS: [f64; 4] = [0.25, 1.2, 3.0, 8.0];
        let m = if period <= 64 { 3 } else { 4 };
        let mut beta = Vec::with_capacity(m);
        for g in GAMMAS.iter().take(m) {
            beta.push((-g / (period as f64)).exp());
        }

        let alpha = fit_exp_weights_least_squares(&sqrt_diffs, &beta);

        let mut beta_pow_p = Vec::with_capacity(m);
        for &b in &beta {
            beta_pow_p.push(b.powi(period as i32));
        }

        Ok(Self {
            period,
            m,
            alpha,
            beta,
            beta_pow_p,
            d_ring: vec![0.0; period],
            d_head: 0,
            d_count: 0,
            denom: 0.0,
            x_acc: vec![0.0; m],

            buffer: vec![f64::NAN; period + 1],
            ln_buffer: vec![f64::NAN; period + 1],
            head: 0,
            filled: false,

            sqrt_diffs,
        })
    }

    #[inline(always)]
    pub fn update(&mut self, value: f64) -> Option<f64> {
        if !value.is_finite() {
            self.reset_state();
            return None;
        }

        let ln_val = ln_pos(value.max(1e-10));

        let prev_idx = (self.head + self.period) % (self.period + 1);
        let prev_ln = self.ln_buffer[prev_idx];

        self.buffer[self.head] = value;
        self.ln_buffer[self.head] = ln_val;

        self.head = (self.head + 1) % (self.period + 1);
        if !self.filled && self.head == 0 {
            self.filled = true;
        }

        if prev_ln.is_nan() {
            return None;
        }

        let d_new = (ln_val - prev_ln).abs();

        if self.d_count < self.period {
            self.d_ring[self.d_head] = d_new;
            self.d_head = (self.d_head + 1) % self.period;
            self.d_count += 1;
            self.denom += d_new;

            for m in 0..self.m {
                self.x_acc[m] = self.beta[m] * self.x_acc[m] + d_new;
            }
        } else {
            let d_old = self.d_ring[self.d_head];
            self.d_ring[self.d_head] = d_new;
            self.d_head = (self.d_head + 1) % self.period;

            self.denom += d_new - d_old;

            for m in 0..self.m {
                self.x_acc[m] = self.beta[m] * self.x_acc[m] + d_new - self.beta_pow_p[m] * d_old;
            }
        }

        if !self.filled {
            return None;
        }

        let mut num = 0.0f64;
        for m in 0..self.m {
            num = (self.alpha[m] * self.x_acc[m]).mul_add(1.0, num);
        }
        let ratio = if self.denom == 0.0 {
            0.0
        } else {
            num / self.denom
        };

        let x0 = self.buffer[self.head];
        let x1 = self.buffer[(self.head + 1) % (self.period + 1)];

        Some((x1 - x0).mul_add(ratio, x0))
    }

    #[inline(always)]
    fn reset_state(&mut self) {
        self.d_head = 0;
        self.d_count = 0;
        self.denom = 0.0;
        for v in &mut self.d_ring {
            *v = 0.0;
        }
        for v in &mut self.x_acc {
            *v = 0.0;
        }
        for v in &mut self.buffer {
            *v = f64::NAN;
        }
        for v in &mut self.ln_buffer {
            *v = f64::NAN;
        }
        self.head = 0;
        self.filled = false;
    }
}

fn fit_exp_weights_least_squares(w: &[f64], beta: &[f64]) -> Vec<f64> {
    let p = w.len();
    let m = beta.len();

    let mut ata = vec![0.0f64; m * m];
    for u in 0..m {
        for v in u..m {
            let r = beta[u] * beta[v];
            let s = if (1.0 - r).abs() < 1e-15 {
                p as f64
            } else {
                (1.0 - r.powi(p as i32)) / (1.0 - r)
            };
            ata[u * m + v] = s;
            ata[v * m + u] = s;
        }
    }

    let mut atw = vec![0.0f64; m];
    for u in 0..m {
        let mut pow = 1.0f64;
        let bu = beta[u];
        let mut sum = 0.0f64;
        for i in 0..p {
            sum += w[i] * pow;
            pow *= bu;
        }
        atw[u] = sum;
    }

    let lambda = 1e-12;
    for i in 0..m {
        ata[i * m + i] += lambda;
    }

    solve_linear_system(&mut ata, &mut atw, m)
}

fn solve_linear_system(a: &mut [f64], b: &mut [f64], n: usize) -> Vec<f64> {
    for k in 0..n {
        let mut piv = k;
        let mut maxv = a[k * n + k].abs();
        for i in (k + 1)..n {
            let v = a[i * n + k].abs();
            if v > maxv {
                maxv = v;
                piv = i;
            }
        }
        if piv != k {
            for j in k..n {
                a.swap(k * n + j, piv * n + j);
            }
            b.swap(k, piv);
        }
        let akk = a[k * n + k];
        if akk.abs() < 1e-18 {
            a[k * n + k] = 1e-18;
        }

        for i in (k + 1)..n {
            let f = a[i * n + k] / a[k * n + k];
            if f != 0.0 {
                for j in k..n {
                    a[i * n + j] -= f * a[k * n + j];
                }
                b[i] -= f * b[k];
            }
        }
    }

    let mut x = vec![0.0f64; n];
    for i in (0..n).rev() {
        let mut s = b[i];
        for j in (i + 1)..n {
            s -= a[i * n + j] * x[j];
        }
        x[i] = s / a[i * n + i];
    }
    x
}

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

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

    let slice_in = data.as_slice()?;
    let kern = validate_kernel(kernel, false)?;
    let params = NmaParams {
        period: Some(period),
    };
    let nma_in = NmaInput::from_slice(slice_in, params);

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

    Ok(result_vec.into_pyarray(py))
}

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

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

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

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

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

    let sweep = NmaBatchRange {
        period: period_range,
    };

    let combos = expand_grid(&sweep).map_err(|e| PyValueError::new_err(e.to_string()))?;
    let rows = combos.len();
    let cols = slice_in.len();
    let expected = rows
        .checked_mul(cols)
        .ok_or_else(|| PyValueError::new_err("rows*cols overflow"))?;

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

    let combos = py
        .allow_threads(|| {
            let kernel = match kern {
                Kernel::Auto => detect_best_batch_kernel(),
                k => k,
            };
            let simd = match kernel {
                Kernel::Avx512Batch => Kernel::Avx512,
                Kernel::Avx2Batch => Kernel::Avx2,
                Kernel::ScalarBatch => Kernel::Scalar,
                _ => kernel,
            };

            nma_batch_inner_into(slice_in, &sweep, simd, true, slice_out)
        })
        .map_err(|e| PyValueError::new_err(e.to_string()))?;

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

    Ok(dict)
}

#[cfg(all(feature = "python", feature = "cuda"))]
#[pyfunction(name = "nma_cuda_batch_dev")]
#[pyo3(signature = (data_f32, period_range, device_id=0))]
pub fn nma_cuda_batch_dev_py<'py>(
    py: Python<'py>,
    data_f32: numpy::PyReadonlyArray1<'py, f32>,
    period_range: (usize, usize, usize),
    device_id: usize,
) -> PyResult<(NmaDeviceArrayF32Py, 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 slice_in = data_f32.as_slice()?;
    let sweep = NmaBatchRange {
        period: period_range,
    };

    let (inner, combos, ctx_arc, dev_id) = py.allow_threads(|| {
        let cuda = CudaNma::new(device_id).map_err(|e| PyValueError::new_err(e.to_string()))?;
        let (dev, combos) = cuda
            .nma_batch_dev(slice_in, &sweep)
            .map_err(|e| PyValueError::new_err(e.to_string()))?;
        cuda.synchronize()
            .map_err(|e| PyValueError::new_err(e.to_string()))?;
        Ok::<_, PyErr>((dev, combos, cuda.context_arc_clone(), cuda.device_id()))
    })?;

    let dict = PyDict::new(py);
    let periods: Vec<u64> = combos.iter().map(|c| c.period.unwrap() as u64).collect();
    dict.set_item("periods", periods.into_pyarray(py))?;

    Ok((
        NmaDeviceArrayF32Py {
            inner,
            _ctx: ctx_arc,
            device_id: dev_id,
        },
        dict,
    ))
}

#[cfg(all(feature = "python", feature = "cuda"))]
#[pyfunction(name = "nma_cuda_many_series_one_param_dev")]
#[pyo3(signature = (data_tm_f32, period, device_id=0))]
pub fn nma_cuda_many_series_one_param_dev_py(
    py: Python<'_>,
    data_tm_f32: numpy::PyReadonlyArray2<'_, f32>,
    period: usize,
    device_id: usize,
) -> PyResult<NmaDeviceArrayF32Py> {
    use crate::cuda::cuda_available;
    use numpy::PyUntypedArrayMethods;

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

    let flat_in = data_tm_f32.as_slice()?;
    let rows = data_tm_f32.shape()[0];
    let cols = data_tm_f32.shape()[1];
    let params = NmaParams {
        period: Some(period),
    };

    let (inner, ctx_arc, dev_id) = py.allow_threads(|| {
        let cuda = CudaNma::new(device_id).map_err(|e| PyValueError::new_err(e.to_string()))?;
        let dev = cuda
            .nma_multi_series_one_param_time_major_dev(flat_in, cols, rows, &params)
            .map_err(|e| PyValueError::new_err(e.to_string()))?;
        cuda.synchronize()
            .map_err(|e| PyValueError::new_err(e.to_string()))?;
        Ok::<_, PyErr>((dev, cuda.context_arc_clone(), cuda.device_id()))
    })?;

    Ok(NmaDeviceArrayF32Py {
        inner,
        _ctx: ctx_arc,
        device_id: dev_id,
    })
}

#[cfg(all(feature = "python", feature = "cuda"))]
#[pyclass(module = "ta_indicators.cuda", name = "NmaDeviceArrayF32", unsendable)]
pub struct NmaDeviceArrayF32Py {
    pub(crate) inner: DeviceArrayF32,
    pub(crate) _ctx: Arc<Context>,
    pub(crate) device_id: u32,
}

#[cfg(all(feature = "python", feature = "cuda"))]
#[pymethods]
impl NmaDeviceArrayF32Py {
    #[getter]
    fn __cuda_array_interface__<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyDict>> {
        let d = PyDict::new(py);
        d.set_item("shape", (self.inner.rows, self.inner.cols))?;
        d.set_item("typestr", "<f4")?;
        d.set_item(
            "strides",
            (
                self.inner.cols * std::mem::size_of::<f32>(),
                std::mem::size_of::<f32>(),
            ),
        )?;
        d.set_item("data", (self.inner.device_ptr() as usize, 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<pyo3::PyObject>,
        max_version: Option<pyo3::PyObject>,
        dl_device: Option<pyo3::PyObject>,
        copy: Option<pyo3::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)
    }
}

pub fn nma_into_slice(dst: &mut [f64], input: &NmaInput, kern: Kernel) -> Result<(), NmaError> {
    let (data, period, first, mut ln_values, mut sqrt_diffs, chosen) = nma_prepare(input, kern)?;

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

    nma_compute_into(
        data,
        period,
        first,
        &mut ln_values,
        &mut sqrt_diffs,
        chosen,
        dst,
    );

    let warmup_end = first + period;
    let qnan = f64::from_bits(0x7ff8_0000_0000_0000);
    for v in &mut dst[..warmup_end] {
        *v = qnan;
    }

    Ok(())
}

#[cfg(not(all(target_arch = "wasm32", feature = "wasm")))]
pub fn nma_into(input: &NmaInput, out: &mut [f64]) -> Result<(), NmaError> {
    nma_into_slice(out, input, Kernel::Auto)
}

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

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

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

    nma_into_slice(&mut output, &input, detect_best_kernel())
        .map_err(|e| JsValue::from_str(&e.to_string()))?;

    Ok(output)
}

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

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

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

    let sweep = NmaBatchRange {
        period: config.period_range,
    };

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

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

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

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

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

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

    let combos = expand_grid(&sweep).map_err(|e| JsValue::from_str(&e.to_string()))?;
    let metadata: Vec<f64> = combos
        .iter()
        .map(|combo| combo.period.unwrap() as f64)
        .collect();

    Ok(metadata)
}

#[cfg(all(target_arch = "wasm32", feature = "wasm"))]
#[wasm_bindgen]
pub fn nma_batch_rows_cols_js(
    period_start: usize,
    period_end: usize,
    period_step: usize,
    data_len: usize,
) -> Vec<usize> {
    let sweep = NmaBatchRange {
        period: (period_start, period_end, period_step),
    };
    let combos = expand_grid(&sweep).unwrap_or_else(|_| Vec::new());
    vec![combos.len(), data_len]
}

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

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

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

        if period == 0 || period > len {
            return Err(JsValue::from_str("Invalid period"));
        }

        let params = NmaParams {
            period: Some(period),
        };
        let input = NmaInput::from_slice(data, params);

        if in_ptr == out_ptr {
            let mut temp = alloc_with_nan_prefix(len, 0);
            nma_into_slice(&mut temp, &input, Kernel::Scalar)
                .map_err(|e| JsValue::from_str(&e.to_string()))?;

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

        Ok(())
    }
}

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

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

        let sweep = NmaBatchRange {
            period: (period_start, period_end, period_step),
        };

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

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

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

        Ok(rows)
    }
}

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

    fn check_nma_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 = NmaParams { period: None };
        let input = NmaInput::from_candles(&candles, "close", default_params);
        let output = nma_with_kernel(&input, kernel)?;
        assert_eq!(output.values.len(), candles.close.len());

        Ok(())
    }

    #[test]
    fn test_nma_into_matches_api() -> Result<(), Box<dyn Error>> {
        let n = 256usize;
        let mut data = vec![0.0f64; n];
        for i in 0..n {
            let t = i as f64;
            data[i] = 100.0 + 0.1 * t + (t * 0.07).sin();
        }

        let params = NmaParams::default();
        let input = NmaInput::from_slice(&data, params);

        let baseline = nma(&input)?.values;

        let mut out = vec![0.0f64; n];
        #[cfg(not(all(target_arch = "wasm32", feature = "wasm")))]
        {
            nma_into(&input, &mut out)?;
        }
        #[cfg(all(target_arch = "wasm32", feature = "wasm"))]
        {
            nma_into_slice(&mut out, &input, detect_best_kernel())?;
        }

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

        fn eq_or_both_nan(a: f64, b: f64) -> bool {
            (a.is_nan() && b.is_nan()) || (a - b).abs() <= 1e-12
        }

        for (i, (&a, &b)) in baseline.iter().zip(out.iter()).enumerate() {
            assert!(
                eq_or_both_nan(a, b),
                "Mismatch at index {}: baseline={} out={}",
                i,
                a,
                b
            );
        }

        Ok(())
    }

    fn check_nma_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 = NmaInput::from_candles(&candles, "close", NmaParams::default());
        let nma_result = nma_with_kernel(&input, kernel)?;

        let expected_last_five_nma = [
            64320.486018271724,
            64227.95719984426,
            64180.9249333126,
            63966.35530620797,
            64039.04719192334,
        ];
        let start_index = nma_result.values.len() - 5;
        let result_last_five_nma = &nma_result.values[start_index..];
        for (i, &value) in result_last_five_nma.iter().enumerate() {
            let expected_value = expected_last_five_nma[i];

            let tolerance = if test_name.contains("avx512") {
                1.0
            } else {
                1e-3
            };
            assert!(
                (value - expected_value).abs() < tolerance,
                "[{}] NMA value mismatch at last-5 index {}: expected {}, got {}",
                test_name,
                i,
                expected_value,
                value
            );
        }
        Ok(())
    }

    fn check_nma_default_candles(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 = NmaInput::with_default_candles(&candles);
        match input.data {
            NmaData::Candles { source, .. } => assert_eq!(source, "close"),
            _ => panic!("Expected NmaData::Candles"),
        }
        let output = nma_with_kernel(&input, kernel)?;
        assert_eq!(output.values.len(), candles.close.len());

        Ok(())
    }

    fn check_nma_zero_period(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 = NmaParams { period: Some(0) };
        let input = NmaInput::from_slice(&input_data, params);
        let res = nma_with_kernel(&input, kernel);
        assert!(
            res.is_err(),
            "[{}] NMA should fail with zero period",
            test_name
        );
        Ok(())
    }

    fn check_nma_period_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 = NmaParams { period: Some(10) };
        let input = NmaInput::from_slice(&data_small, params);
        let res = nma_with_kernel(&input, kernel);
        assert!(
            res.is_err(),
            "[{}] NMA should fail with period exceeding length",
            test_name
        );
        Ok(())
    }

    fn check_nma_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 = NmaParams { period: Some(40) };
        let input = NmaInput::from_slice(&single_point, params);
        let res = nma_with_kernel(&input, kernel);
        assert!(
            res.is_err(),
            "[{}] NMA should fail with insufficient data",
            test_name
        );
        Ok(())
    }

    fn check_nma_empty_input(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn Error>> {
        skip_if_unsupported!(kernel, test_name);
        let empty: [f64; 0] = [];
        let input = NmaInput::from_slice(&empty, NmaParams::default());
        let res = nma_with_kernel(&input, kernel);
        assert!(
            matches!(res, Err(NmaError::EmptyInputData)),
            "[{}] NMA should fail with empty input error",
            test_name
        );
        Ok(())
    }

    fn check_nma_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 first_params = NmaParams { period: Some(40) };
        let first_input = NmaInput::from_candles(&candles, "close", first_params);
        let first_result = nma_with_kernel(&first_input, kernel)?;
        let second_params = NmaParams { period: Some(20) };
        let second_input = NmaInput::from_slice(&first_result.values, second_params);
        let second_result = nma_with_kernel(&second_input, kernel)?;
        assert_eq!(second_result.values.len(), first_result.values.len());
        if second_result.values.len() > 240 {
            for i in 240..second_result.values.len() {
                assert!(second_result.values[i].is_finite());
            }
        }
        Ok(())
    }

    fn check_nma_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 = NmaInput::from_candles(&candles, "close", NmaParams { period: Some(40) });
        let res = nma_with_kernel(&input, kernel)?;
        assert_eq!(res.values.len(), candles.close.len());
        if res.values.len() > 240 {
            for (i, &val) in res.values[240..].iter().enumerate() {
                assert!(
                    !val.is_nan(),
                    "[{}] Found unexpected NaN at out-index {}",
                    test_name,
                    240 + i
                );
            }
        }
        Ok(())
    }

    fn check_nma_property(test_name: &str, kernel: Kernel) -> Result<(), Box<dyn Error>> {
        use proptest::prelude::*;
        skip_if_unsupported!(kernel, test_name);

        let strat = (2usize..=100).prop_flat_map(|period| {
            (
                prop::collection::vec(
                    (-1e6f64..1e6f64).prop_filter("finite", |x| x.is_finite()),
                    (period + 1)..400,
                ),
                Just(period),
            )
        });

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

                let result = nma_with_kernel(&input, kernel);
                prop_assert!(result.is_ok(), "NMA computation failed: {:?}", result.err());
                let out = result.unwrap().values;

                let ref_result = nma_with_kernel(&input, Kernel::Scalar);
                prop_assert!(ref_result.is_ok(), "Reference NMA failed");
                let ref_out = ref_result.unwrap().values;

                prop_assert_eq!(out.len(), data.len(), "Output length mismatch");

                let first_valid = data.iter().position(|x| !x.is_nan()).unwrap_or(0);
                let warmup_end = first_valid + period;

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

                for i in warmup_end..out.len() {
                    prop_assert!(
                        out[i].is_finite(),
                        "Expected finite value at index {} (after warmup), got {}",
                        i,
                        out[i]
                    );
                }

                for i in warmup_end..out.len() {
                    let point1 = data[i - period + 1];
                    let point2 = data[i - period];
                    let min_bound = point1.min(point2);
                    let max_bound = point1.max(point2);

                    let tolerance = if test_name.contains("avx512") {
                        1e-7
                    } else {
                        1e-9
                    };
                    prop_assert!(
                        out[i] >= min_bound - tolerance && out[i] <= max_bound + tolerance,
                        "NMA at index {} = {} not in bounds [{}, {}]",
                        i,
                        out[i],
                        min_bound,
                        max_bound
                    );
                }

                if data.windows(2).all(|w| (w[0] - w[1]).abs() < 1e-12) && !data.is_empty() {
                    for i in warmup_end..out.len() {
                        prop_assert!(
                            (out[i] - data[0]).abs() < 1e-9,
                            "Constant data: NMA[{}] = {} should equal {}",
                            i,
                            out[i],
                            data[0]
                        );
                    }
                }

                if period == 1 {
                    for i in (first_valid + 1)..out.len() {
                        prop_assert!(
                            (out[i] - data[i]).abs() < 1e-6,
                            "Period=1: NMA[{}] = {} should be close to data[{}] = {}",
                            i,
                            out[i],
                            i,
                            data[i]
                        );
                    }
                }

                for i in warmup_end..out.len() {
                    let point1 = data[i - period + 1];
                    let point2 = data[i - period];

                    if (point1 - point2).abs() > 1e-10 {
                        let implied_ratio = (out[i] - point2) / (point1 - point2);
                        prop_assert!(
                            implied_ratio >= -1e-9 && implied_ratio <= 1.0 + 1e-9,
                            "Invalid interpolation ratio {} at index {} (output={}, p1={}, p2={})",
                            implied_ratio,
                            i,
                            out[i],
                            point1,
                            point2
                        );
                    }
                }

                for i in 0..out.len() {
                    if !out[i].is_finite() || !ref_out[i].is_finite() {
                        prop_assert_eq!(
                            out[i].is_nan(),
                            ref_out[i].is_nan(),
                            "NaN mismatch at index {}",
                            i
                        );
                        continue;
                    }

                    let out_bits = out[i].to_bits();
                    let ref_bits = ref_out[i].to_bits();
                    let ulp_diff = out_bits.abs_diff(ref_bits);

                    if test_name.contains("avx512") {
                        let rel_error = if ref_out[i].abs() > 1e-10 {
                            ((out[i] - ref_out[i]) / ref_out[i]).abs()
                        } else {
                            (out[i] - ref_out[i]).abs()
                        };
                        prop_assert!(
                            rel_error < 1e-7 || ulp_diff <= 75,
                            "Kernel mismatch at index {}: {} vs {} (rel_error: {}, ULP diff: {})",
                            i,
                            out[i],
                            ref_out[i],
                            rel_error,
                            ulp_diff
                        );
                    } else {
                        prop_assert!(
                            (out[i] - ref_out[i]).abs() <= 1e-9 || ulp_diff <= 25,
                            "Kernel mismatch at index {}: {} vs {} (ULP diff: {})",
                            i,
                            out[i],
                            ref_out[i],
                            ulp_diff
                        );
                    }
                }

                let has_small_values = data.iter().any(|&x| x > 0.0 && x < 1e-8);
                if has_small_values {
                    for i in warmup_end..out.len() {
                        prop_assert!(
                            out[i].is_finite(),
                            "NMA failed to handle small values at index {}: {}",
                            i,
                            out[i]
                        );
                    }
                }

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

        Ok(())
    }

    macro_rules! generate_all_nma_tests {
        ($($test_fn:ident),*) => {
            paste::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);
                    }
                    #[test]
                    fn [<$test_fn _avx512_f64>]() {
                        let _ = $test_fn(stringify!([<$test_fn _avx512_f64>]), Kernel::Avx512);
                    }
                )*
                #[cfg(all(target_arch = "wasm32", target_feature = "simd128"))]
                $(
                    #[test]
                    fn [<$test_fn _simd128_f64>]() {
                        let _ = $test_fn(stringify!([<$test_fn _simd128_f64>]), Kernel::Scalar);
                    }
                )*
            }
        }
    }

    #[cfg(debug_assertions)]
    fn check_nma_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![
            NmaParams { period: Some(40) },
            NmaParams { period: Some(10) },
            NmaParams { period: Some(5) },
            NmaParams { period: Some(20) },
            NmaParams { period: Some(60) },
            NmaParams { period: Some(100) },
            NmaParams { period: Some(3) },
            NmaParams { period: Some(80) },
            NmaParams { period: None },
        ];

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

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

                let bits = val.to_bits();

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

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

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

        Ok(())
    }

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

    generate_all_nma_tests!(
        check_nma_partial_params,
        check_nma_accuracy,
        check_nma_default_candles,
        check_nma_zero_period,
        check_nma_period_exceeds_length,
        check_nma_very_small_dataset,
        check_nma_empty_input,
        check_nma_reinput,
        check_nma_nan_handling,
        check_nma_no_poison,
        check_nma_property
    );

    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 = NmaBatchBuilder::new()
            .kernel(kernel)
            .apply_candles(&c, "close")?;

        let def = NmaParams::default();
        let row = output.values_for(&def).expect("default row missing");

        assert_eq!(row.len(), c.close.len());

        let expected = [
            64320.486018271724,
            64227.95719984426,
            64180.924933312606,
            63966.35530620797,
            64039.04719192333,
        ];
        let start = row.len() - 5;
        for (i, &v) in row[start..].iter().enumerate() {
            let tolerance = 1e-3;
            assert!(
                (v - expected[i]).abs() < tolerance,
                "[{test}] default-row mismatch at idx {i}: {v} vs {expected:?}"
            );
        }
        Ok(())
    }

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

    #[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 batch_configs = vec![
            (10, 30, 10),
            (40, 40, 0),
            (3, 15, 3),
            (50, 100, 25),
            (5, 25, 5),
            (20, 80, 20),
            (8, 24, 8),
            (60, 120, 30),
        ];

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

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

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

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

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

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

        Ok(())
    }

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

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