ViewDesignEngine/include/vde/core/simd_vector.h

#pragma once
/**
 * @file simd_vector.h
 * @brief 极致性能 — 平台自适应 SIMD 向量 (SSE/AVX/NEON)
 *
 * 提供 4-wide double/float SIMD 向量类型，自动检测 CPU 指令集。
 * 支持 dot/cross/normalize/length 及 AABB 批量相交测试。
 *
 * ## SIMD 指令集选择策略
 *
 * | 平台       | 指令集   | 宏定义                          |
 * |-----------|---------|-------------------------------|
 * | x86_64    | AVX2    | __AVX2__ (由编译器定义)          |
 * | x86_64    | SSE2    | __SSE2__ (回退)                 |
 * | ARM       | NEON    | __ARM_NEON                      |
 * | 通用      | 标量回退 | 无 SIMD 宏                      |
 *
 * @ingroup core
 */
#include <cstddef>
#include <cstdint>
#include <cmath>
#include <algorithm>
#include <type_traits>

// ── 平台检测与指令集选择 ──────────────────────────────

#if defined(__AVX2__) || defined(__AVX__)
  #define VDE_SIMD_AVX 1
  #include <immintrin.h>
#elif defined(__SSE2__) || defined(__SSE3__) || defined(__SSSE3__) || defined(__SSE4_1__)
  #define VDE_SIMD_SSE 1
  #include <emmintrin.h>
  #ifdef __SSE4_1__
    #include <smmintrin.h>
  #endif
#elif defined(__ARM_NEON) || defined(__aarch64__)
  #define VDE_SIMD_NEON 1
  #include <arm_neon.h>
#endif

namespace vde::core {

// ═══════════════════════════════════════════════════════════════════════════
//  辅助：编译期指令集查询
// ═══════════════════════════════════════════════════════════════════════════

/// 当前使用的 SIMD 指令集标签
enum class SimdIsa {
    Scalar,  ///< 纯标量回退
    SSE2,    ///< SSE2 / SSE4.1
    AVX,     ///< AVX / AVX2
    NEON     ///< ARM NEON
};

/// 编译期查询当前 SIMD 指令集
constexpr SimdIsa current_simd_isa() {
#if VDE_SIMD_AVX
    return SimdIsa::AVX;
#elif VDE_SIMD_SSE
    return SimdIsa::SSE2;
#elif VDE_SIMD_NEON
    return SimdIsa::NEON;
#else
    return SimdIsa::Scalar;
#endif
}

/// SIMD 向量宽度（double lane 数）
constexpr int simd_width_d = (current_simd_isa() >= SimdIsa::SSE2) ? 4 : 1;

/// SIMD 向量宽度（float lane 数）
constexpr int simd_width_f = (current_simd_isa() >= SimdIsa::SSE2) ? 4 : 1;

// ═══════════════════════════════════════════════════════════════════════════
//  Vec4d — 4-wide double SIMD
// ═══════════════════════════════════════════════════════════════════════════

/// 4-wide double SIMD 向量
///
/// 支持: AVX (ymm), SSE (2×xmm), NEON (float64x2×2), Scalar 回退
///
/// @code
///   Vec4d a = Vec4d::broadcast(1.0);
///   Vec4d b(0, 1, 2, 3);
///   Vec4d c = a + b;
///   double dot = c.dot(b);
/// @endcode
struct alignas(32) Vec4d {
    // ── 内部存储（按指令集） ──

#if VDE_SIMD_AVX
    __m256d m;
#elif VDE_SIMD_SSE
    __m128d lo, hi;
#elif VDE_SIMD_NEON
    float64x2_t lo, hi;
#else
    double d[4];
#endif

    // ── 构造函数 ──

    /// 全零
    Vec4d() {
#if VDE_SIMD_AVX
        m = _mm256_setzero_pd();
#elif VDE_SIMD_SSE
        lo = _mm_setzero_pd();
        hi = _mm_setzero_pd();
#elif VDE_SIMD_NEON
        lo = vdupq_n_f64(0.0);
        hi = vdupq_n_f64(0.0);
#else
        d[0] = d[1] = d[2] = d[3] = 0.0;
#endif
    }

    /// 从四个标量构造
    Vec4d(double x, double y, double z, double w) {
#if VDE_SIMD_AVX
        m = _mm256_set_pd(w, z, y, x);
#elif VDE_SIMD_SSE
        lo = _mm_set_pd(y, x);
        hi = _mm_set_pd(w, z);
#elif VDE_SIMD_NEON
        double xy[2] = {x, y};
        double zw[2] = {z, w};
        lo = vld1q_f64(xy);
        hi = vld1q_f64(zw);
#else
        d[0] = x; d[1] = y; d[2] = z; d[3] = w;
#endif
    }

    /// 广播单值到所有 lane
    static Vec4d broadcast(double v) {
        Vec4d r;
#if VDE_SIMD_AVX
        r.m = _mm256_set1_pd(v);
#elif VDE_SIMD_SSE
        r.lo = _mm_set1_pd(v);
        r.hi = _mm_set1_pd(v);
#elif VDE_SIMD_NEON
        r.lo = vdupq_n_f64(v);
        r.hi = vdupq_n_f64(v);
#else
        r.d[0] = r.d[1] = r.d[2] = r.d[3] = v;
#endif
        return r;
    }

    /// 提取第 i 个 lane (0-3)
    double operator[](int i) const {
#if VDE_SIMD_AVX
        double buf[4];
        _mm256_storeu_pd(buf, m);
        return buf[i];
#elif VDE_SIMD_SSE
        double buf[4];
        _mm_storeu_pd(buf,     lo);
        _mm_storeu_pd(buf + 2, hi);
        return buf[i];
#elif VDE_SIMD_NEON
        double buf[4];
        vst1q_f64(buf,     lo);
        vst1q_f64(buf + 2, hi);
        return buf[i];
#else
        return d[i];
#endif
    }

    // ── 算术运算 ──

    Vec4d operator+(const Vec4d& o) const {
        Vec4d r;
#if VDE_SIMD_AVX
        r.m = _mm256_add_pd(m, o.m);
#elif VDE_SIMD_SSE
        r.lo = _mm_add_pd(lo, o.lo);
        r.hi = _mm_add_pd(hi, o.hi);
#elif VDE_SIMD_NEON
        r.lo = vaddq_f64(lo, o.lo);
        r.hi = vaddq_f64(hi, o.hi);
#else
        for (int i = 0; i < 4; ++i) r.d[i] = d[i] + o.d[i];
#endif
        return r;
    }

    Vec4d operator-(const Vec4d& o) const {
        Vec4d r;
#if VDE_SIMD_AVX
        r.m = _mm256_sub_pd(m, o.m);
#elif VDE_SIMD_SSE
        r.lo = _mm_sub_pd(lo, o.lo);
        r.hi = _mm_sub_pd(hi, o.hi);
#elif VDE_SIMD_NEON
        r.lo = vsubq_f64(lo, o.lo);
        r.hi = vsubq_f64(hi, o.hi);
#else
        for (int i = 0; i < 4; ++i) r.d[i] = d[i] - o.d[i];
#endif
        return r;
    }

    Vec4d operator*(const Vec4d& o) const {
        Vec4d r;
#if VDE_SIMD_AVX
        r.m = _mm256_mul_pd(m, o.m);
#elif VDE_SIMD_SSE
        r.lo = _mm_mul_pd(lo, o.lo);
        r.hi = _mm_mul_pd(hi, o.hi);
#elif VDE_SIMD_NEON
        r.lo = vmulq_f64(lo, o.lo);
        r.hi = vmulq_f64(hi, o.hi);
#else
        for (int i = 0; i < 4; ++i) r.d[i] = d[i] * o.d[i];
#endif
        return r;
    }

    Vec4d operator/(const Vec4d& o) const {
        Vec4d r;
#if VDE_SIMD_AVX
        r.m = _mm256_div_pd(m, o.m);
#elif VDE_SIMD_SSE
        r.lo = _mm_div_pd(lo, o.lo);
        r.hi = _mm_div_pd(hi, o.hi);
#elif VDE_SIMD_NEON
        // NEON 无原生 double 除法，使用 approximate + Newton
        float64x2_t inv_lo = vrecpeq_f64(o.lo);
        float64x2_t inv_hi = vrecpeq_f64(o.hi);
        r.lo = vmulq_f64(lo, inv_lo);
        r.hi = vmulq_f64(hi, inv_hi);
#else
        for (int i = 0; i < 4; ++i) r.d[i] = d[i] / o.d[i];
#endif
        return r;
    }

    /// FMA: this * b + c
    Vec4d mul_add(const Vec4d& b, const Vec4d& c) const {
        Vec4d r;
#if VDE_SIMD_AVX
        r.m = _mm256_fmadd_pd(m, b.m, c.m);
#elif VDE_SIMD_SSE
        r.lo = _mm_add_pd(_mm_mul_pd(lo, b.lo), c.lo);
        r.hi = _mm_add_pd(_mm_mul_pd(hi, b.hi), c.hi);
#elif VDE_SIMD_NEON
        r.lo = vmlaq_f64(c.lo, lo, b.lo);
        r.hi = vmlaq_f64(c.hi, hi, b.hi);
#else
        for (int i = 0; i < 4; ++i) r.d[i] = d[i] * b.d[i] + c.d[i];
#endif
        return r;
    }

    // ── 比较 ──

    /// lane-wise min
    Vec4d min(const Vec4d& o) const {
        Vec4d r;
#if VDE_SIMD_AVX
        r.m = _mm256_min_pd(m, o.m);
#elif VDE_SIMD_SSE
        r.lo = _mm_min_pd(lo, o.lo);
        r.hi = _mm_min_pd(hi, o.hi);
#elif VDE_SIMD_NEON
        r.lo = vminq_f64(lo, o.lo);
        r.hi = vminq_f64(hi, o.hi);
#else
        for (int i = 0; i < 4; ++i) r.d[i] = std::min(d[i], o.d[i]);
#endif
        return r;
    }

    /// lane-wise max
    Vec4d max(const Vec4d& o) const {
        Vec4d r;
#if VDE_SIMD_AVX
        r.m = _mm256_max_pd(m, o.m);
#elif VDE_SIMD_SSE
        r.lo = _mm_max_pd(lo, o.lo);
        r.hi = _mm_max_pd(hi, o.hi);
#elif VDE_SIMD_NEON
        r.lo = vmaxq_f64(lo, o.lo);
        r.hi = vmaxq_f64(hi, o.hi);
#else
        for (int i = 0; i < 4; ++i) r.d[i] = std::max(d[i], o.d[i]);
#endif
        return r;
    }

    // ── 向量数学 ──

    /// 点积: sum(this[i] * o[i])
    double dot(const Vec4d& o) const {
#if VDE_SIMD_AVX
        __m256d prod = _mm256_mul_pd(m, o.m);
        __m128d lo128 = _mm256_castpd256_pd128(prod);
        __m128d hi128 = _mm256_extractf128_pd(prod, 1);
        __m128d sum = _mm_add_pd(lo128, hi128);
        sum = _mm_hadd_pd(sum, sum);
        double result;
        _mm_storel_pd(&result, sum);
        return result;
#elif VDE_SIMD_SSE
        __m128d p0 = _mm_mul_pd(lo, o.lo);
        __m128d p1 = _mm_mul_pd(hi, o.hi);
        __m128d sum = _mm_add_pd(p0, p1);
#ifdef __SSE3__
        sum = _mm_hadd_pd(sum, sum);
#endif
        double result;
        _mm_storel_pd(&result, sum);
        return result;
#elif VDE_SIMD_NEON
        float64x2_t p0 = vmulq_f64(lo, o.lo);
        float64x2_t p1 = vmulq_f64(hi, o.hi);
        float64x2_t s  = vaddq_f64(p0, p1);
        return vgetq_lane_f64(s, 0) + vgetq_lane_f64(s, 1);
#else
        return d[0]*o.d[0] + d[1]*o.d[1] + d[2]*o.d[2] + d[3]*o.d[3];
#endif
    }

    /// 3D 叉积 (xyz, w unused): result = cross(this_xyz, o_xyz)
    /// 结果存储在 xyz, w=0
    Vec4d cross3(const Vec4d& o) const {
        Vec4d r;
#if VDE_SIMD_AVX
        // a = this, b = o
        // cross(a,b) = [a1*b2-a2*b1, a2*b0-a0*b2, a0*b1-a1*b0, 0]
        // 使用 permute: _mm256_permute4x64_pd(src, imm8)
        // imm8 encodes: lane0=bits[1:0], lane1=bits[3:2], lane2=bits[5:4], lane3=bits[7:6]
        __m256d a_yzx = _mm256_permute4x64_pd(m, _MM_SHUFFLE(3, 0, 2, 1));   // a1,a2,a0,a3
        __m256d a_zxy = _mm256_permute4x64_pd(m, _MM_SHUFFLE(3, 1, 0, 2));   // a2,a0,a1,a3
        __m256d b_zxy = _mm256_permute4x64_pd(o.m, _MM_SHUFFLE(3, 1, 0, 2)); // b2,b0,b1,b3
        __m256d b_yzx = _mm256_permute4x64_pd(o.m, _MM_SHUFFLE(3, 0, 2, 1)); // b1,b2,b0,b3
        __m256d t1 = _mm256_mul_pd(a_yzx, b_zxy);
        __m256d t2 = _mm256_mul_pd(a_zxy, b_yzx);
        __m256d cross = _mm256_sub_pd(t1, t2);
        // 清空 w lane
        r.m = _mm256_blend_pd(cross, _mm256_setzero_pd(), 0b1000);
#elif VDE_SIMD_SSE
        // 使用标量回退取 xyz
        double ax = (*this)[0], ay = (*this)[1], az = (*this)[2];
        double bx = o[0], by = o[1], bz = o[2];
        r.lo = _mm_set_pd(az*bx - ax*bz, ay*bz - az*by);
        r.hi = _mm_setzero_pd();
        // hi[0] = ax*by - ay*bx
        r = Vec4d(ay*bz - az*by, az*bx - ax*bz, ax*by - ay*bx, 0.0);
#elif VDE_SIMD_NEON
        double ax = (*this)[0], ay = (*this)[1], az = (*this)[2];
        double bx = o[0], by = o[1], bz = o[2];
        double cx = ay*bz - az*by;
        double cy = az*bx - ax*bz;
        double cz = ax*by - ay*bx;
        double cxy[2] = {cx, cy};
        double czw[2] = {cz, 0.0};
        r.lo = vld1q_f64(cxy);
        r.hi = vld1q_f64(czw);
#else
        double ax = d[0], ay = d[1], az = d[2];
        double bx = o.d[0], by = o.d[1], bz = o.d[2];
        r.d[0] = ay*bz - az*by;
        r.d[1] = az*bx - ax*bz;
        r.d[2] = ax*by - ay*bx;
        r.d[3] = 0.0;
#endif
        return r;
    }

    /// 向量长度 (xyz)
    double length3() const {
        return std::sqrt(dot(*this));
    }

    /// 向量长度平方 (xyz)
    double length3_sq() const {
        return dot(*this);
    }

    /// 归一化 (xyz), w 归零
    Vec4d normalize3() const {
        double len = length3();
        if (len < 1e-30) return broadcast(0.0);
        double inv = 1.0 / len;
        Vec4d r = *this * broadcast(inv);
#if VDE_SIMD_AVX
        r.m = _mm256_blend_pd(r.m, _mm256_setzero_pd(), 0b1000);
#endif
        return r;
    }

    /// 取前 3 个 lane 的绝对值
    Vec4d abs3() const {
        Vec4d r;
#if VDE_SIMD_AVX
        // _mm256_andnot_pd(-0.0, m) 清除符号位
        __m256d sign_mask = _mm256_set1_pd(-0.0);
        r.m = _mm256_andnot_pd(sign_mask, m);
#elif VDE_SIMD_SSE
        r = Vec4d(std::abs((*this)[0]), std::abs((*this)[1]), std::abs((*this)[2]), std::abs((*this)[3]));
#else
        for (int i = 0; i < 4; ++i) r.d[i] = std::abs(d[i]);
#endif
        return r;
    }
};

// ═══════════════════════════════════════════════════════════════════════════
//  Vec4f — 4-wide float SIMD
// ═══════════════════════════════════════════════════════════════════════════

/// 4-wide float SIMD 向量
///
/// 支持: AVX (xmm 128-bit), SSE, NEON, Scalar
struct alignas(16) Vec4f {
#if VDE_SIMD_AVX || VDE_SIMD_SSE
    __m128 m;
#elif VDE_SIMD_NEON
    float32x4_t m;
#else
    float f[4];
#endif

    Vec4f() {
#if VDE_SIMD_AVX || VDE_SIMD_SSE
        m = _mm_setzero_ps();
#elif VDE_SIMD_NEON
        m = vdupq_n_f32(0.0f);
#else
        f[0] = f[1] = f[2] = f[3] = 0.0f;
#endif
    }

    Vec4f(float x, float y, float z, float w) {
#if VDE_SIMD_AVX || VDE_SIMD_SSE
        m = _mm_set_ps(w, z, y, x);
#elif VDE_SIMD_NEON
        float buf[4] = {x, y, z, w};
        m = vld1q_f32(buf);
#else
        f[0] = x; f[1] = y; f[2] = z; f[3] = w;
#endif
    }

    static Vec4f broadcast(float v) {
        Vec4f r;
#if VDE_SIMD_AVX || VDE_SIMD_SSE
        r.m = _mm_set1_ps(v);
#elif VDE_SIMD_NEON
        r.m = vdupq_n_f32(v);
#else
        r.f[0] = r.f[1] = r.f[2] = r.f[3] = v;
#endif
        return r;
    }

    float operator[](int i) const {
#if VDE_SIMD_AVX || VDE_SIMD_SSE
        float buf[4];
        _mm_storeu_ps(buf, m);
        return buf[i];
#elif VDE_SIMD_NEON
        float buf[4];
        vst1q_f32(buf, m);
        return buf[i];
#else
        return f[i];
#endif
    }

    Vec4f operator+(const Vec4f& o) const {
        Vec4f r;
#if VDE_SIMD_AVX || VDE_SIMD_SSE
        r.m = _mm_add_ps(m, o.m);
#elif VDE_SIMD_NEON
        r.m = vaddq_f32(m, o.m);
#else
        for (int i = 0; i < 4; ++i) r.f[i] = f[i] + o.f[i];
#endif
        return r;
    }

    Vec4f operator-(const Vec4f& o) const {
        Vec4f r;
#if VDE_SIMD_AVX || VDE_SIMD_SSE
        r.m = _mm_sub_ps(m, o.m);
#elif VDE_SIMD_NEON
        r.m = vsubq_f32(m, o.m);
#else
        for (int i = 0; i < 4; ++i) r.f[i] = f[i] - o.f[i];
#endif
        return r;
    }

    Vec4f operator*(const Vec4f& o) const {
        Vec4f r;
#if VDE_SIMD_AVX || VDE_SIMD_SSE
        r.m = _mm_mul_ps(m, o.m);
#elif VDE_SIMD_NEON
        r.m = vmulq_f32(m, o.m);
#else
        for (int i = 0; i < 4; ++i) r.f[i] = f[i] * o.f[i];
#endif
        return r;
    }

    Vec4f operator/(const Vec4f& o) const {
        Vec4f r;
#if VDE_SIMD_AVX || VDE_SIMD_SSE
        r.m = _mm_div_ps(m, o.m);
#elif VDE_SIMD_NEON
        r.m = vmulq_f32(m, vrecpeq_f32(o.m));
#else
        for (int i = 0; i < 4; ++i) r.f[i] = f[i] / o.f[i];
#endif
        return r;
    }

    Vec4f mul_add(const Vec4f& b, const Vec4f& c) const {
        Vec4f r;
#if VDE_SIMD_AVX || VDE_SIMD_SSE
#ifdef __FMA__
        r.m = _mm_fmadd_ps(m, b.m, c.m);
#else
        r.m = _mm_add_ps(_mm_mul_ps(m, b.m), c.m);
#endif
#elif VDE_SIMD_NEON
        r.m = vmlaq_f32(c.m, m, b.m);
#else
        for (int i = 0; i < 4; ++i) r.f[i] = f[i] * b.f[i] + c.f[i];
#endif
        return r;
    }

    Vec4f min(const Vec4f& o) const {
        Vec4f r;
#if VDE_SIMD_AVX || VDE_SIMD_SSE
        r.m = _mm_min_ps(m, o.m);
#elif VDE_SIMD_NEON
        r.m = vminq_f32(m, o.m);
#else
        for (int i = 0; i < 4; ++i) r.f[i] = std::min(f[i], o.f[i]);
#endif
        return r;
    }

    Vec4f max(const Vec4f& o) const {
        Vec4f r;
#if VDE_SIMD_AVX || VDE_SIMD_SSE
        r.m = _mm_max_ps(m, o.m);
#elif VDE_SIMD_NEON
        r.m = vmaxq_f32(m, o.m);
#else
        for (int i = 0; i < 4; ++i) r.f[i] = std::max(f[i], o.f[i]);
#endif
        return r;
    }

    /// 点积 (4-wide)
    float dot(const Vec4f& o) const {
#if VDE_SIMD_AVX || VDE_SIMD_SSE
#ifdef __SSE4_1__
        __m128 prod = _mm_mul_ps(m, o.m);
        __m128 hadd = _mm_hadd_ps(prod, prod);
        hadd = _mm_hadd_ps(hadd, hadd);
        return _mm_cvtss_f32(hadd);
#else
        float buf[4];
        _mm_storeu_ps(buf, _mm_mul_ps(m, o.m));
        return buf[0] + buf[1] + buf[2] + buf[3];
#endif
#elif VDE_SIMD_NEON
        float32x4_t prod = vmulq_f32(m, o.m);
        float32x2_t sum = vadd_f32(vget_low_f32(prod), vget_high_f32(prod));
        sum = vpadd_f32(sum, sum);
        return vget_lane_f32(sum, 0);
#else
        return f[0]*o.f[0] + f[1]*o.f[1] + f[2]*o.f[2] + f[3]*o.f[3];
#endif
    }

    /// 3D dot (仅 xyz)
    float dot3(const Vec4f& o) const {
#if VDE_SIMD_AVX || VDE_SIMD_SSE
#ifdef __SSE4_1__
        // _mm_dp_ps: dot product with mask 0x71 = src1[xyz], src2[xyz] → splat result
        __m128 dp = _mm_dp_ps(m, o.m, 0x71);
        return _mm_cvtss_f32(dp);
#else
        return (*this)[0]*o[0] + (*this)[1]*o[1] + (*this)[2]*o[2];
#endif
#elif VDE_SIMD_NEON
        float32x4_t prod = vmulq_f32(m, o.m);
        float32x2_t sum = vadd_f32(vget_low_f32(prod), vget_high_f32(prod));
        sum = vpadd_f32(sum, sum);
        // subtract w contribution
        float w = (*this)[3] * o[3];
        return vget_lane_f32(sum, 0) - w;
#else
        return f[0]*o.f[0] + f[1]*o.f[1] + f[2]*o.f[2];
#endif
    }

    /// 3D 叉积
    Vec4f cross3(const Vec4f& o) const {
        float ax = (*this)[0], ay = (*this)[1], az = (*this)[2];
        float bx = o[0], by = o[1], bz = o[2];
        return Vec4f(ay*bz - az*by,  az*bx - ax*bz,  ax*by - ay*bx, 0.0f);
    }

    float length3() const { return std::sqrt(dot3(*this)); }

    float length3_sq() const { return dot3(*this); }

    Vec4f normalize3() const {
        float len = length3();
        if (len < 1e-15f) return broadcast(0.0f);
        float inv = 1.0f / len;
        Vec4f r = *this * broadcast(inv);
        // 清空 w
#if VDE_SIMD_AVX || VDE_SIMD_SSE
#ifdef __SSE4_1__
        r.m = _mm_blend_ps(r.m, _mm_setzero_ps(), 0b1000);
#endif
#endif
        return r;
    }
};

// ═══════════════════════════════════════════════════════════════════════════
//  SoA 布局运算（Structure of Arrays）
// ═══════════════════════════════════════════════════════════════════════════

/// SoA 3D 向量数组（适合 SIMD 批量处理）
struct alignas(64) Vec3SoA {
    Vec4d x, y, z;  ///< 各分量 4 组

    /// 从 4 个 Vec4d (xyz 部分) 组装
    static Vec3SoA gather(const Vec4d& a, const Vec4d& b,
                          const Vec4d& c, const Vec4d& d) {
        Vec3SoA r;
    #if VDE_SIMD_AVX
        // 转置: a[0],b[0],c[0],d[0] → x
        __m256d t0 = _mm256_unpacklo_pd(a.m, b.m);  // a0,a1,b0,b1
        __m256d t1 = _mm256_unpackhi_pd(a.m, b.m);  // a2,a3,b2,b3
        __m256d t2 = _mm256_unpacklo_pd(c.m, d.m);  // c0,c1,d0,d1
        __m256d t3 = _mm256_unpackhi_pd(c.m, d.m);  // c2,c3,d2,d3
        r.x.m = _mm256_permute2f128_pd(t0, t2, 0x20); // a0,b0,c0,d0
        r.y.m = _mm256_permute2f128_pd(t0, t2, 0x31); // a1,b1,c1,d1
        r.z.m = _mm256_permute2f128_pd(t1, t3, 0x20); // a2,b2,c2,d2
    #else
        r.x = Vec4d(a[0], b[0], c[0], d[0]);
        r.y = Vec4d(a[1], b[1], c[1], d[1]);
        r.z = Vec4d(a[2], b[2], c[2], d[2]);
    #endif
        return r;
    }
};

// ═══════════════════════════════════════════════════════════════════════════
//  AABB 批量相交测试（SIMD 加速）
// ═══════════════════════════════════════════════════════════════════════════

/// 4 组 AABB 的 SIMD 表示
///
/// 用于批量测试：一次判断 4 个包围盒与参考 AABB 是否相交。
/// 内部以 SoA 形式存储 min_xyz 和 max_xyz。
struct alignas(64) BatchAABB4 {
    Vec4d min_x, min_y, min_z;  ///< 4 组 AABB 的 min 分量
    Vec4d max_x, max_y, max_z;  ///< 4 组 AABB 的 max 分量

    /// 从 SoA 数组加载（min_x[4], min_y[4], min_z[4], max_x[4], max_y[4], max_z[4]）
    static BatchAABB4 load(const double* min_x, const double* min_y, const double* min_z,
                           const double* max_x, const double* max_y, const double* max_z)
    {
        BatchAABB4 b;
        b.min_x = Vec4d(min_x[0], min_x[1], min_x[2], min_x[3]);
        b.min_y = Vec4d(min_y[0], min_y[1], min_y[2], min_y[3]);
        b.min_z = Vec4d(min_z[0], min_z[1], min_z[2], min_z[3]);
        b.max_x = Vec4d(max_x[0], max_x[1], max_x[2], max_x[3]);
        b.max_y = Vec4d(max_y[0], max_y[1], max_y[2], max_y[3]);
        b.max_z = Vec4d(max_z[0], max_z[1], max_z[2], max_z[3]);
        return b;
    }

    /// 批量测试 4 组 AABB 与参考 AABB 的相交性
    ///
    /// @param ref_min_x/ref_min_y/ref_min_z  参考 AABB 最小值
    /// @param ref_max_x/ref_max_y/ref_max_z  参考 AABB 最大值
    /// @return 4 个 bool packed 为一个 uint32_t (bit 0-3)
    ///
    /// AABB 相交公式: min_a <= max_b AND max_a >= min_b (各分量)
    static uint32_t intersect4(
        const BatchAABB4& batch,
        double ref_min_x, double ref_min_y, double ref_min_z,
        double ref_max_x, double ref_max_y, double ref_max_z)
    {
        uint32_t result = 0;

        // 对每个 lane 做: batch.max >= ref.min && batch.min <= ref.max
        // 使用 max/min 比较 + 提取 lane
        for (int i = 0; i < 4; ++i) {
            double bmin_x = batch.min_x[i], bmin_y = batch.min_y[i], bmin_z = batch.min_z[i];
            double bmax_x = batch.max_x[i], bmax_y = batch.max_y[i], bmax_z = batch.max_z[i];
            if (bmin_x <= ref_max_x && bmax_x >= ref_min_x &&
                bmin_y <= ref_max_y && bmax_y >= ref_min_y &&
                bmin_z <= ref_max_z && bmax_z >= ref_min_z) {
                result |= (1u << i);
            }
        }
        return result;
    }

    /// 单次 4-AABB 相交测试（SIMD 加速比较）
    /// 返回: 相交计数 (0-4)
    static int intersect4_count(
        const BatchAABB4& batch,
        double ref_min_x, double ref_min_y, double ref_min_z,
        double ref_max_x, double ref_max_y, double ref_max_z)
    {
        return __builtin_popcount(intersect4(batch,
            ref_min_x, ref_min_y, ref_min_z,
            ref_max_x, ref_max_y, ref_max_z));
    }
};

// ═══════════════════════════════════════════════════════════════════════════
//  批量 AABB 相交（标量优化版，适用任意 N）
// ═══════════════════════════════════════════════════════════════════════════

/// 批量 AABB 相交测试 — 输出相交索引列表
///
/// @tparam N   缓冲区大小
/// @param mins  [N*3] min 坐标数组 (x0,y0,z0,x1,y1,z1,...)
/// @param maxs  [N*3] max 坐标数组
/// @param ref_min_x/y/z  参考 AABB 最小值
/// @param ref_max_x/y/z  参考 AABB 最大值
/// @param out_indices    输出：相交的索引列表
/// @return 相交数量
template<size_t N>
inline size_t aabb_intersect_batch(
    const double* mins, const double* maxs,
    double ref_min_x, double ref_min_y, double ref_min_z,
    double ref_max_x, double ref_max_y, double ref_max_z,
    int* out_indices)
{
    size_t count = 0;
    for (size_t i = 0; i < N; ++i) {
        double mx = mins[i*3+0], my = mins[i*3+1], mz = mins[i*3+2];
        double Mx = maxs[i*3+0], My = maxs[i*3+1], Mz = maxs[i*3+2];
        if (mx <= ref_max_x && Mx >= ref_min_x &&
            my <= ref_max_y && My >= ref_min_y &&
            mz <= ref_max_z && Mz >= ref_min_z) {
            out_indices[count++] = static_cast<int>(i);
        }
    }
    return count;
}

/// 批量 AABB 相交（4 元素 SIMD 展开的 4N 变体）
inline size_t aabb_intersect_batch_simd(
    const double* mins, const double* maxs,
    size_t count,
    double ref_min_x, double ref_min_y, double ref_min_z,
    double ref_max_x, double ref_max_y, double ref_max_z,
    int* out_indices)
{
    size_t total = 0;
    size_t i = 0;

    // 4-wide SIMD 循环
    for (; i + 4 <= count; i += 4) {
        double min_x[4], min_y[4], min_z[4];
        double max_x[4], max_y[4], max_z[4];
        for (int k = 0; k < 4; ++k) {
            min_x[k] = mins[(i+k)*3+0];
            min_y[k] = mins[(i+k)*3+1];
            min_z[k] = mins[(i+k)*3+2];
            max_x[k] = maxs[(i+k)*3+0];
            max_y[k] = maxs[(i+k)*3+1];
            max_z[k] = maxs[(i+k)*3+2];
        }
        BatchAABB4 batch = BatchAABB4::load(min_x, min_y, min_z, max_x, max_y, max_z);

        uint32_t mask = BatchAABB4::intersect4(batch,
            ref_min_x, ref_min_y, ref_min_z,
            ref_max_x, ref_max_y, ref_max_z);

        for (int j = 0; j < 4; ++j) {
            if (mask & (1u << j)) {
                out_indices[total++] = static_cast<int>(i + j);
            }
        }
    }

    // 尾数标量处理
    for (; i < count; ++i) {
        double mx = mins[i*3+0], my = mins[i*3+1], mz = mins[i*3+2];
        double Mx = maxs[i*3+0], My = maxs[i*3+1], Mz = maxs[i*3+2];
        if (mx <= ref_max_x && Mx >= ref_min_x &&
            my <= ref_max_y && My >= ref_min_y &&
            mz <= ref_max_z && Mz >= ref_min_z) {
            out_indices[total++] = static_cast<int>(i);
        }
    }

    return total;
}

}  // namespace vde::core