src/tensor.rs

// SPF Smart Gateway - Tensor Engine
// Copyright 2026 Joseph Stone - All Rights Reserved
//
// N-dimensional tensor operations for SPF Transformer.
// Pure Rust. No external ML dependencies.
// ARM NEON SIMD acceleration with scalar fallback.
//
// This is the mathematical foundation — all transformer layers build on this.
// Depends on: nothing (Layer 0)

use std::fmt;

// ============================================================================
// TENSOR STRUCT
// ============================================================================

/// N-dimensional tensor with f32 data and shape metadata.
/// Storage is row-major (C-order): last dimension varies fastest.
#[derive(Clone)]
pub struct Tensor {
    /// Flat data storage
    pub data: Vec<f32>,
    /// Shape dimensions (e.g., [batch, seq_len, d_model])
    pub shape: Vec<usize>,
}

impl Tensor {
    /// Create a tensor with given shape, initialized to zeros
    pub fn zeros(shape: &[usize]) -> Self {
        let size = shape.iter().product::<usize>();
        Self {
            data: vec![0.0; size],
            shape: shape.to_vec(),
        }
    }

    /// Create a tensor with given shape, initialized to ones
    pub fn ones(shape: &[usize]) -> Self {
        let size = shape.iter().product::<usize>();
        Self {
            data: vec![1.0; size],
            shape: shape.to_vec(),
        }
    }

    /// Create a tensor from raw data and shape
    pub fn from_data(data: Vec<f32>, shape: Vec<usize>) -> Result<Self, String> {
        let expected = shape.iter().product::<usize>();
        if data.len() != expected {
            return Err(format!(
                "Data length {} doesn't match shape {:?} (expected {})",
                data.len(), shape, expected
            ));
        }
        Ok(Self { data, shape })
    }

    /// Create a tensor with random uniform values in [0, 1)
    /// Uses simple xorshift for reproducibility without external deps
    pub fn rand(shape: &[usize], seed: u64) -> Self {
        let size = shape.iter().product::<usize>();
        let mut data = Vec::with_capacity(size);
        let mut state = seed;
        for _ in 0..size {
            state = xorshift64(state);
            data.push((state as f32) / (u64::MAX as f32));
        }
        Self {
            data,
            shape: shape.to_vec(),
        }
    }

    /// Create a tensor with random normal values (mean=0, std=1)
    /// Box-Muller transform from uniform random pairs
    pub fn randn(shape: &[usize], seed: u64) -> Self {
        let size = shape.iter().product::<usize>();
        let mut data = Vec::with_capacity(size);
        let mut state = seed;
        for i in 0..size {
            state = xorshift64(state);
            let u1 = (state as f64) / (u64::MAX as f64);
            state = xorshift64(state);
            let u2 = (state as f64) / (u64::MAX as f64);
            // Box-Muller: avoid log(0) by clamping
            let u1 = u1.max(1e-10);
            let val = if i % 2 == 0 {
                (-2.0 * u1.ln()).sqrt() * (2.0 * std::f64::consts::PI * u2).cos()
            } else {
                (-2.0 * u1.ln()).sqrt() * (2.0 * std::f64::consts::PI * u2).sin()
            };
            data.push(val as f32);
        }
        Self {
            data,
            shape: shape.to_vec(),
        }
    }

    /// Total number of elements
    pub fn numel(&self) -> usize {
        self.data.len()
    }

    /// Number of dimensions
    pub fn ndim(&self) -> usize {
        self.shape.len()
    }

    // ========================================================================
    // SHAPE OPERATIONS
    // ========================================================================

    /// Reshape tensor (total elements must match)
    pub fn reshape(&self, new_shape: &[usize]) -> Result<Self, String> {
        let new_size: usize = new_shape.iter().product();
        if new_size != self.numel() {
            return Err(format!(
                "Cannot reshape {:?} ({}) to {:?} ({})",
                self.shape, self.numel(), new_shape, new_size
            ));
        }
        Ok(Self {
            data: self.data.clone(),
            shape: new_shape.to_vec(),
        })
    }

    /// Transpose a 2D tensor
    pub fn transpose_2d(&self) -> Result<Self, String> {
        if self.ndim() != 2 {
            return Err(format!("transpose_2d requires 2D tensor, got {}D", self.ndim()));
        }
        let (rows, cols) = (self.shape[0], self.shape[1]);
        let mut out = vec![0.0f32; rows * cols];
        for r in 0..rows {
            for c in 0..cols {
                out[c * rows + r] = self.data[r * cols + c];
            }
        }
        Tensor::from_data(out, vec![cols, rows])
    }

    /// Slice along first dimension: tensor[start..end]
    pub fn slice(&self, start: usize, end: usize) -> Result<Self, String> {
        if self.ndim() == 0 {
            return Err("Cannot slice scalar tensor".to_string());
        }
        if start >= end || end > self.shape[0] {
            return Err(format!(
                "Invalid slice [{}..{}) for dim 0 size {}",
                start, end, self.shape[0]
            ));
        }
        let inner_size: usize = self.shape[1..].iter().product();
        let data = self.data[start * inner_size..end * inner_size].to_vec();
        let mut new_shape = self.shape.clone();
        new_shape[0] = end - start;
        Tensor::from_data(data, new_shape)
    }

    /// Concatenate tensors along first dimension
    pub fn concat(tensors: &[&Tensor]) -> Result<Self, String> {
        if tensors.is_empty() {
            return Err("Cannot concat empty list".to_string());
        }
        let base_shape = &tensors[0].shape[1..];
        for (i, t) in tensors.iter().enumerate().skip(1) {
            if t.shape[1..] != *base_shape {
                return Err(format!(
                    "Shape mismatch at index {}: {:?} vs {:?}",
                    i, &t.shape[1..], base_shape
                ));
            }
        }
        let total_first: usize = tensors.iter().map(|t| t.shape[0]).sum();
        let mut data = Vec::with_capacity(total_first * base_shape.iter().product::<usize>());
        for t in tensors {
            data.extend_from_slice(&t.data);
        }
        let mut new_shape = vec![total_first];
        new_shape.extend_from_slice(base_shape);
        Tensor::from_data(data, new_shape)
    }

    // ========================================================================
    // MATRIX OPERATIONS
    // ========================================================================

    /// Matrix multiply: [M, K] × [K, N] → [M, N]
    /// Uses SIMD-accelerated inner loop on aarch64, scalar fallback otherwise.
    pub fn matmul(&self, other: &Tensor) -> Result<Self, String> {
        if self.ndim() != 2 || other.ndim() != 2 {
            return Err(format!(
                "matmul requires 2D tensors, got {}D × {}D",
                self.ndim(), other.ndim()
            ));
        }
        let (m, k1) = (self.shape[0], self.shape[1]);
        let (k2, n) = (other.shape[0], other.shape[1]);
        if k1 != k2 {
            return Err(format!(
                "matmul inner dims mismatch: [{}, {}] × [{}, {}]",
                m, k1, k2, n
            ));
        }
        let mut out = vec![0.0f32; m * n];
        matmul_inner(&self.data, &other.data, &mut out, m, k1, n);
        Tensor::from_data(out, vec![m, n])
    }

    /// Batched matrix multiply: [B, M, K] × [B, K, N] → [B, M, N]
    /// Supports broadcasting: [B, M, K] × [K, N] broadcasts the right operand.
    pub fn bmm(&self, other: &Tensor) -> Result<Self, String> {
        if self.ndim() != 3 {
            return Err(format!("bmm requires 3D left tensor, got {}D", self.ndim()));
        }
        let b = self.shape[0];
        let m = self.shape[1];
        let k = self.shape[2];

        let (other_k, n, broadcast) = if other.ndim() == 2 {
            (other.shape[0], other.shape[1], true)
        } else if other.ndim() == 3 {
            if other.shape[0] != b {
                return Err(format!("bmm batch mismatch: {} vs {}", b, other.shape[0]));
            }
            (other.shape[1], other.shape[2], false)
        } else {
            return Err(format!("bmm requires 2D or 3D right tensor, got {}D", other.ndim()));
        };

        if k != other_k {
            return Err(format!("bmm inner dims mismatch: {} vs {}", k, other_k));
        }

        let mut out = vec![0.0f32; b * m * n];
        let batch_a = m * k;
        let batch_b = if broadcast { 0 } else { other_k * n };
        let batch_out = m * n;

        for bi in 0..b {
            let a_off = bi * batch_a;
            let b_off = bi * batch_b;
            let o_off = bi * batch_out;
            matmul_inner(
                &self.data[a_off..a_off + batch_a],
                &other.data[b_off..b_off + other_k * n],
                &mut out[o_off..o_off + batch_out],
                m, k, n,
            );
        }
        Tensor::from_data(out, vec![b, m, n])
    }

    // ========================================================================
    // ELEMENT-WISE OPERATIONS
    // ========================================================================

    /// Element-wise addition (shapes must match or broadcast along last dim)
    pub fn add(&self, other: &Tensor) -> Result<Self, String> {
        if self.shape == other.shape {
            let data: Vec<f32> = self.data.iter().zip(&other.data).map(|(a, b)| a + b).collect();
            return Tensor::from_data(data, self.shape.clone());
        }
        // Broadcast: other has shape matching last dims of self
        if other.numel() == *self.shape.last().unwrap_or(&0) {
            let last = *self.shape.last().unwrap();
            let data: Vec<f32> = self.data.iter().enumerate()
                .map(|(i, &v)| v + other.data[i % last])
                .collect();
            return Tensor::from_data(data, self.shape.clone());
        }
        Err(format!("Cannot add shapes {:?} and {:?}", self.shape, other.shape))
    }

    /// Element-wise subtraction
    pub fn sub(&self, other: &Tensor) -> Result<Self, String> {
        if self.shape != other.shape {
            return Err(format!("Cannot sub shapes {:?} and {:?}", self.shape, other.shape));
        }
        let data: Vec<f32> = self.data.iter().zip(&other.data).map(|(a, b)| a - b).collect();
        Tensor::from_data(data, self.shape.clone())
    }

    /// Element-wise multiplication (Hadamard product)
    pub fn mul(&self, other: &Tensor) -> Result<Self, String> {
        if self.shape == other.shape {
            let data: Vec<f32> = self.data.iter().zip(&other.data).map(|(a, b)| a * b).collect();
            return Tensor::from_data(data, self.shape.clone());
        }
        // Broadcast along last dim
        if other.numel() == *self.shape.last().unwrap_or(&0) {
            let last = *self.shape.last().unwrap();
            let data: Vec<f32> = self.data.iter().enumerate()
                .map(|(i, &v)| v * other.data[i % last])
                .collect();
            return Tensor::from_data(data, self.shape.clone());
        }
        Err(format!("Cannot mul shapes {:?} and {:?}", self.shape, other.shape))
    }

    /// Scalar multiply
    pub fn scale(&self, s: f32) -> Self {
        Self {
            data: self.data.iter().map(|&v| v * s).collect(),
            shape: self.shape.clone(),
        }
    }

    /// Scalar add
    pub fn add_scalar(&self, s: f32) -> Self {
        Self {
            data: self.data.iter().map(|&v| v + s).collect(),
            shape: self.shape.clone(),
        }
    }

    // ========================================================================
    // ACTIVATION FUNCTIONS
    // ========================================================================

    /// ReLU activation: max(0, x)
    pub fn relu(&self) -> Self {
        Self {
            data: self.data.iter().map(|&v| v.max(0.0)).collect(),
            shape: self.shape.clone(),
        }
    }

    /// GELU activation: x × Φ(x) ≈ 0.5x(1 + tanh(√(2/π)(x + 0.044715x³)))
    pub fn gelu(&self) -> Self {
        let sqrt_2_over_pi = (2.0_f32 / std::f32::consts::PI).sqrt();
        Self {
            data: self.data.iter().map(|&x| {
                let inner = sqrt_2_over_pi * (x + 0.044715 * x * x * x);
                0.5 * x * (1.0 + inner.tanh())
            }).collect(),
            shape: self.shape.clone(),
        }
    }

    /// Softmax along last dimension
    /// Numerically stable: subtract max before exp
    pub fn softmax(&self) -> Result<Self, String> {
        if self.shape.is_empty() {
            return Err("Cannot softmax scalar".to_string());
        }
        let last_dim = *self.shape.last().unwrap();
        let num_rows = self.numel() / last_dim;
        let mut out = vec![0.0f32; self.numel()];

        for row in 0..num_rows {
            let offset = row * last_dim;
            let slice = &self.data[offset..offset + last_dim];

            // Find max for numerical stability
            let max_val = slice.iter().cloned().fold(f32::NEG_INFINITY, f32::max);

            // Exp and sum
            let mut sum = 0.0f32;
            for i in 0..last_dim {
                let exp_val = (slice[i] - max_val).exp();
                out[offset + i] = exp_val;
                sum += exp_val;
            }

            // Normalize
            if sum > 0.0 {
                for i in 0..last_dim {
                    out[offset + i] /= sum;
                }
            }
        }
        Tensor::from_data(out, self.shape.clone())
    }

    // ========================================================================
    // NORMALIZATION
    // ========================================================================

    /// Layer normalization along last dimension
    /// output = (x - mean) / sqrt(var + eps) * gamma + beta
    pub fn layer_norm(&self, gamma: &Tensor, beta: &Tensor, eps: f32) -> Result<Self, String> {
        let last_dim = *self.shape.last().ok_or("Cannot layer_norm scalar")?;
        if gamma.numel() != last_dim || beta.numel() != last_dim {
            return Err(format!(
                "layer_norm: gamma/beta size {} doesn't match last dim {}",
                gamma.numel(), last_dim
            ));
        }
        let num_rows = self.numel() / last_dim;
        let mut out = vec![0.0f32; self.numel()];

        for row in 0..num_rows {
            let offset = row * last_dim;
            let slice = &self.data[offset..offset + last_dim];

            // Compute mean
            let mean: f32 = slice.iter().sum::<f32>() / last_dim as f32;

            // Compute variance
            let var: f32 = slice.iter()
                .map(|&x| (x - mean) * (x - mean))
                .sum::<f32>() / last_dim as f32;

            let inv_std = 1.0 / (var + eps).sqrt();

            // Normalize and apply affine
            for i in 0..last_dim {
                out[offset + i] = (slice[i] - mean) * inv_std * gamma.data[i] + beta.data[i];
            }
        }
        Tensor::from_data(out, self.shape.clone())
    }

    // ========================================================================
    // REDUCTION OPERATIONS
    // ========================================================================

    /// Sum all elements
    pub fn sum(&self) -> f32 {
        self.data.iter().sum()
    }

    /// Mean of all elements
    pub fn mean(&self) -> f32 {
        if self.data.is_empty() { return 0.0; }
        self.sum() / self.numel() as f32
    }

    /// Max of all elements
    pub fn max(&self) -> f32 {
        self.data.iter().cloned().fold(f32::NEG_INFINITY, f32::max)
    }

    /// Argmax along last dimension
    pub fn argmax(&self) -> Vec<usize> {
        let last_dim = *self.shape.last().unwrap_or(&1);
        let num_rows = self.numel() / last_dim;
        let mut indices = Vec::with_capacity(num_rows);
        for row in 0..num_rows {
            let offset = row * last_dim;
            let slice = &self.data[offset..offset + last_dim];
            let (idx, _) = slice.iter().enumerate()
                .fold((0, f32::NEG_INFINITY), |(best_i, best_v), (i, &v)| {
                    if v > best_v { (i, v) } else { (best_i, best_v) }
                });
            indices.push(idx);
        }
        indices
    }
}

// ============================================================================
// DISPLAY
// ============================================================================

impl fmt::Debug for Tensor {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        write!(f, "Tensor(shape={:?}, numel={})", self.shape, self.numel())
    }
}

impl fmt::Display for Tensor {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        if self.numel() <= 20 {
            write!(f, "Tensor({:?}, {:?})", self.shape, self.data)
        } else {
            write!(f, "Tensor({:?}, [{:.4}, {:.4}, ... {:.4}, {:.4}])",
                self.shape,
                self.data[0], self.data[1],
                self.data[self.numel() - 2], self.data[self.numel() - 1])
        }
    }
}

// ============================================================================
// MATRIX MULTIPLY INNER — SIMD-accelerated on aarch64, scalar fallback
// ============================================================================

/// Core matmul kernel: C[m×n] = A[m×k] × B[k×n]
/// Dispatches to SIMD on aarch64, scalar otherwise.
/// Small matrices (m×n < 16) use scalar even on aarch64 — NEON setup overhead
/// exceeds SIMD benefit for tiny dimensions.
fn matmul_inner(a: &[f32], b: &[f32], c: &mut [f32], m: usize, k: usize, n: usize) {
    #[cfg(target_arch = "aarch64")]
    {
        if m * n < 16 {
            matmul_scalar(a, b, c, m, k, n);
        } else {
            matmul_neon(a, b, c, m, k, n);
        }
    }
    #[cfg(not(target_arch = "aarch64"))]
    {
        matmul_scalar(a, b, c, m, k, n);
    }
}

/// Scalar matmul — portable fallback
fn matmul_scalar(a: &[f32], b: &[f32], c: &mut [f32], m: usize, k: usize, n: usize) {
    for i in 0..m {
        for j in 0..n {
            let mut sum = 0.0f32;
            for p in 0..k {
                sum += a[i * k + p] * b[p * n + j];
            }
            c[i * n + j] = sum;
        }
    }
}

/// ARM NEON SIMD matmul — processes 4 floats at a time
#[cfg(target_arch = "aarch64")]
fn matmul_neon(a: &[f32], b: &[f32], c: &mut [f32], m: usize, k: usize, n: usize) {
    use std::arch::aarch64::*;

    // Process 4 columns at a time using NEON
    for i in 0..m {
        let mut j = 0;
        // SIMD path: 4 columns at a time
        while j + 4 <= n {
            let mut acc = unsafe { vdupq_n_f32(0.0) };
            for p in 0..k {
                let a_val = a[i * k + p];
                let a_vec = unsafe { vdupq_n_f32(a_val) };
                let b_vec = unsafe { vld1q_f32(b.as_ptr().add(p * n + j)) };
                acc = unsafe { vfmaq_f32(acc, a_vec, b_vec) };
            }
            unsafe {
                vst1q_f32(c.as_mut_ptr().add(i * n + j), acc);
            }
            j += 4;
        }
        // Scalar remainder for columns not divisible by 4
        while j < n {
            let mut sum = 0.0f32;
            for p in 0..k {
                sum += a[i * k + p] * b[p * n + j];
            }
            c[i * n + j] = sum;
            j += 1;
        }
    }
}

// ============================================================================
// UTILITIES
// ============================================================================

/// xorshift64 PRNG — fast, deterministic, no dependencies
fn xorshift64(mut state: u64) -> u64 {
    state ^= state << 13;
    state ^= state >> 7;
    state ^= state << 17;
    state
}

// ============================================================================
// TESTS
// ============================================================================

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

    #[test]
    fn test_zeros_and_ones() {
        let z = Tensor::zeros(&[2, 3]);
        assert_eq!(z.numel(), 6);
        assert_eq!(z.shape, vec![2, 3]);
        assert!(z.data.iter().all(|&v| v == 0.0));

        let o = Tensor::ones(&[3, 4]);
        assert_eq!(o.numel(), 12);
        assert!(o.data.iter().all(|&v| v == 1.0));
    }

    #[test]
    fn test_from_data_validation() {
        assert!(Tensor::from_data(vec![1.0, 2.0, 3.0], vec![3]).is_ok());
        assert!(Tensor::from_data(vec![1.0, 2.0], vec![3]).is_err());
    }

    #[test]
    fn test_reshape() {
        let t = Tensor::from_data(vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0], vec![2, 3]).unwrap();
        let r = t.reshape(&[3, 2]).unwrap();
        assert_eq!(r.shape, vec![3, 2]);
        assert_eq!(r.data, t.data);
        assert!(t.reshape(&[4, 2]).is_err());
    }

    #[test]
    fn test_transpose_2d() {
        let t = Tensor::from_data(vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0], vec![2, 3]).unwrap();
        let tr = t.transpose_2d().unwrap();
        assert_eq!(tr.shape, vec![3, 2]);
        assert_eq!(tr.data, vec![1.0, 4.0, 2.0, 5.0, 3.0, 6.0]);
    }

    #[test]
    fn test_slice() {
        let t = Tensor::from_data(
            vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0], vec![3, 2]
        ).unwrap();
        let s = t.slice(1, 3).unwrap();
        assert_eq!(s.shape, vec![2, 2]);
        assert_eq!(s.data, vec![3.0, 4.0, 5.0, 6.0]);
    }

    #[test]
    fn test_concat() {
        let a = Tensor::from_data(vec![1.0, 2.0], vec![1, 2]).unwrap();
        let b = Tensor::from_data(vec![3.0, 4.0, 5.0, 6.0], vec![2, 2]).unwrap();
        let c = Tensor::concat(&[&a, &b]).unwrap();
        assert_eq!(c.shape, vec![3, 2]);
        assert_eq!(c.data, vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]);
    }

    #[test]
    fn test_matmul_identity() {
        // [2,2] × [2,2] identity = same matrix
        let a = Tensor::from_data(vec![1.0, 2.0, 3.0, 4.0], vec![2, 2]).unwrap();
        let eye = Tensor::from_data(vec![1.0, 0.0, 0.0, 1.0], vec![2, 2]).unwrap();
        let c = a.matmul(&eye).unwrap();
        assert_eq!(c.data, vec![1.0, 2.0, 3.0, 4.0]);
    }

    #[test]
    fn test_matmul_known() {
        // [1,3] × [3,1] = [1,1] = dot product
        let a = Tensor::from_data(vec![1.0, 2.0, 3.0], vec![1, 3]).unwrap();
        let b = Tensor::from_data(vec![4.0, 5.0, 6.0], vec![3, 1]).unwrap();
        let c = a.matmul(&b).unwrap();
        assert_eq!(c.shape, vec![1, 1]);
        assert!((c.data[0] - 32.0).abs() < 1e-5); // 1×4 + 2×5 + 3×6 = 32
    }

    #[test]
    fn test_matmul_dimension_mismatch() {
        let a = Tensor::zeros(&[2, 3]);
        let b = Tensor::zeros(&[4, 5]);
        assert!(a.matmul(&b).is_err());
    }

    #[test]
    fn test_bmm() {
        // Batch of 2: [2, 1, 2] × [2, 2, 1] = [2, 1, 1]
        let a = Tensor::from_data(vec![1.0, 2.0, 3.0, 4.0], vec![2, 1, 2]).unwrap();
        let b = Tensor::from_data(vec![5.0, 6.0, 7.0, 8.0], vec![2, 2, 1]).unwrap();
        let c = a.bmm(&b).unwrap();
        assert_eq!(c.shape, vec![2, 1, 1]);
        assert!((c.data[0] - 17.0).abs() < 1e-5); // 1×5 + 2×6 = 17
        assert!((c.data[1] - 53.0).abs() < 1e-5); // 3×7 + 4×8 = 53
    }

    #[test]
    fn test_add_same_shape() {
        let a = Tensor::from_data(vec![1.0, 2.0, 3.0], vec![3]).unwrap();
        let b = Tensor::from_data(vec![10.0, 20.0, 30.0], vec![3]).unwrap();
        let c = a.add(&b).unwrap();
        assert_eq!(c.data, vec![11.0, 22.0, 33.0]);
    }

    #[test]
    fn test_add_broadcast() {
        let a = Tensor::from_data(vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0], vec![2, 3]).unwrap();
        let bias = Tensor::from_data(vec![10.0, 20.0, 30.0], vec![3]).unwrap();
        let c = a.add(&bias).unwrap();
        assert_eq!(c.data, vec![11.0, 22.0, 33.0, 14.0, 25.0, 36.0]);
    }

    #[test]
    fn test_scale() {
        let a = Tensor::from_data(vec![1.0, 2.0, 3.0], vec![3]).unwrap();
        let b = a.scale(2.0);
        assert_eq!(b.data, vec![2.0, 4.0, 6.0]);
    }

    #[test]
    fn test_relu() {
        let a = Tensor::from_data(vec![-2.0, -1.0, 0.0, 1.0, 2.0], vec![5]).unwrap();
        let b = a.relu();
        assert_eq!(b.data, vec![0.0, 0.0, 0.0, 1.0, 2.0]);
    }

    #[test]
    fn test_gelu() {
        let a = Tensor::from_data(vec![0.0, 1.0, -1.0], vec![3]).unwrap();
        let b = a.gelu();
        assert!((b.data[0] - 0.0).abs() < 1e-4); // GELU(0) ≈ 0
        assert!((b.data[1] - 0.8412).abs() < 1e-3); // GELU(1) ≈ 0.8412
        assert!((b.data[2] - (-0.1588)).abs() < 1e-3); // GELU(-1) ≈ -0.1588
    }

    #[test]
    fn test_softmax_sum_to_one() {
        let a = Tensor::from_data(vec![1.0, 2.0, 3.0, 4.0], vec![2, 2]).unwrap();
        let s = a.softmax().unwrap();
        // Each row should sum to 1.0
        assert!((s.data[0] + s.data[1] - 1.0).abs() < 1e-5);
        assert!((s.data[2] + s.data[3] - 1.0).abs() < 1e-5);
        // All values non-negative
        assert!(s.data.iter().all(|&v| v >= 0.0));
    }

    #[test]
    fn test_softmax_large_values() {
        // Numerical stability: large values shouldn't overflow
        let a = Tensor::from_data(vec![1000.0, 1001.0, 1002.0], vec![3]).unwrap();
        let s = a.softmax().unwrap();
        assert!((s.sum() - 1.0).abs() < 1e-5);
        assert!(s.data.iter().all(|v| v.is_finite()));
    }

    #[test]
    fn test_layer_norm() {
        let x = Tensor::from_data(vec![1.0, 2.0, 3.0, 4.0], vec![2, 2]).unwrap();
        let gamma = Tensor::ones(&[2]);
        let beta = Tensor::zeros(&[2]);
        let out = x.layer_norm(&gamma, &beta, 1e-5).unwrap();
        // Each row should be normalized (mean ≈ 0, std ≈ 1)
        let row1_mean = (out.data[0] + out.data[1]) / 2.0;
        assert!(row1_mean.abs() < 1e-5);
    }

    #[test]
    fn test_argmax() {
        let a = Tensor::from_data(vec![0.1, 0.9, 0.5, 0.8, 0.2, 0.7], vec![2, 3]).unwrap();
        let idx = a.argmax();
        assert_eq!(idx, vec![1, 0]); // max at index 1 in row 0, index 0 in row 1
    }

    #[test]
    fn test_rand_deterministic() {
        let a = Tensor::rand(&[100], 42);
        let b = Tensor::rand(&[100], 42);
        assert_eq!(a.data, b.data); // Same seed = same values
    }

    #[test]
    fn test_randn_distribution() {
        let t = Tensor::randn(&[10000], 42);
        let mean = t.mean();
        let var: f32 = t.data.iter().map(|&x| (x - mean) * (x - mean)).sum::<f32>() / t.numel() as f32;
        assert!(mean.abs() < 0.1); // Mean ≈ 0
        assert!((var - 1.0).abs() < 0.15); // Variance ≈ 1
    }

    #[cfg(target_arch = "aarch64")]
    #[test]
    fn test_simd_matches_scalar() {
        // Verify NEON produces same results as scalar
        let a = Tensor::rand(&[8, 16], 42);
        let b = Tensor::rand(&[16, 12], 43);

        let mut c_simd = vec![0.0f32; 8 * 12];
        let mut c_scalar = vec![0.0f32; 8 * 12];

        matmul_neon(&a.data, &b.data, &mut c_simd, 8, 16, 12);
        matmul_scalar(&a.data, &b.data, &mut c_scalar, 8, 16, 12);

        for i in 0..c_simd.len() {
            assert!(
                (c_simd[i] - c_scalar[i]).abs() < 1e-3,
                "Mismatch at {}: SIMD={}, scalar={}", i, c_simd[i], c_scalar[i]
            );
        }
    }
}