usearch 2.21.3 → 2.25.0

package/numkong/include/numkong/mesh/serial.h

@@ -0,0 +1,721 @@
+/**
+ *  @brief SWAR-accelerated Point Cloud Alignment for SIMD-free CPUs.
+ *  @file include/numkong/mesh/serial.h
+ *  @author Ash Vardanian
+ *  @date December 27, 2025
+ *
+ *  @sa include/numkong/mesh.h
+ */
+#ifndef NK_MESH_SERIAL_H
+#define NK_MESH_SERIAL_H
+
+#include "numkong/types.h"
+#include "numkong/scalar/serial.h"
+#include "numkong/spatial/serial.h" // `nk_f32_sqrt_serial`, `nk_f64_sqrt_serial`
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+/*  Constants for the McAdams 3×3 SVD algorithm.
+ *  γ = (√8 + 3)² / 4 = 5.828427124
+ *  cstar = cos(π/8), sstar = sin(π/8)
+ */
+#define NK_F32_SVD_GAMMA_   5.828427124f
+#define NK_F32_SVD_CSTAR_   0.923879532f
+#define NK_F32_SVD_SSTAR_   0.3826834323f
+#define NK_F32_SVD_EPSILON_ 1e-6f
+
+#define NK_F64_SVD_GAMMA_   5.828427124746190
+#define NK_F64_SVD_CSTAR_   0.9238795325112867
+#define NK_F64_SVD_SSTAR_   0.3826834323650898
+#define NK_F64_SVD_EPSILON_ 1e-12
+
+/*  Type-Generic SVD Helper Macros
+ *  These macros generate f32 and f64 versions of SVD helper functions
+ *  used by the Kabsch and Umeyama algorithms.
+ */
+
+#define nk_define_cond_swap_(type)                                                       \
+    NK_INTERNAL void nk_cond_swap_##type##_(int c, nk_##type##_t *x, nk_##type##_t *y) { \
+        nk_##type##_t temp = *x;                                                         \
+        *x = c ? *y : *x;                                                                \
+        *y = c ? temp : *y;                                                              \
+    }
+
+#define nk_define_conditional_negating_swap_(type)                                                       \
+    NK_INTERNAL void nk_conditional_negating_swap_##type##_(int c, nk_##type##_t *x, nk_##type##_t *y) { \
+        nk_##type##_t neg_x = -*x;                                                                       \
+        *x = c ? *y : *x;                                                                                \
+        *y = c ? neg_x : *y;                                                                             \
+    }
+
+#define nk_define_approximate_givens_quaternion_(type, gamma, cstar, sstar, compute_rsqrt)                           \
+    NK_INTERNAL void nk_approximate_givens_quaternion_##type##_(                                                     \
+        nk_##type##_t a11, nk_##type##_t a12, nk_##type##_t a22, nk_##type##_t *cos_half, nk_##type##_t *sin_half) { \
+        *cos_half = 2 * (a11 - a22), *sin_half = a12;                                                                \
+        int use_givens = gamma * (*sin_half) * (*sin_half) < (*cos_half) * (*cos_half);                              \
+        nk_##type##_t w = compute_rsqrt((*cos_half) * (*cos_half) + (*sin_half) * (*sin_half));                      \
+        *cos_half = use_givens ? w * (*cos_half) : cstar;                                                            \
+        *sin_half = use_givens ? w * (*sin_half) : sstar;                                                            \
+    }
+
+#define nk_define_jacobi_conjugation_(type)                                                          \
+    NK_INTERNAL void nk_jacobi_conjugation_##type##_(                                                \
+        int idx_x, int idx_y, int idx_z, nk_##type##_t *s11, nk_##type##_t *s21, nk_##type##_t *s22, \
+        nk_##type##_t *s31, nk_##type##_t *s32, nk_##type##_t *s33, nk_##type##_t *quaternion) {     \
+        nk_##type##_t cos_half, sin_half;                                                            \
+        nk_approximate_givens_quaternion_##type##_(*s11, *s21, *s22, &cos_half, &sin_half);          \
+        nk_##type##_t scale = cos_half * cos_half + sin_half * sin_half;                             \
+        nk_##type##_t cos_theta = (cos_half * cos_half - sin_half * sin_half) / scale;               \
+        nk_##type##_t sin_theta = (2 * sin_half * cos_half) / scale;                                 \
+        nk_##type##_t s11_old = *s11, s21_old = *s21, s22_old = *s22;                                \
+        nk_##type##_t s31_old = *s31, s32_old = *s32, s33_old = *s33;                                \
+        *s11 = cos_theta * (cos_theta * s11_old + sin_theta * s21_old) +                             \
+               sin_theta * (cos_theta * s21_old + sin_theta * s22_old);                              \
+        *s21 = cos_theta * (-sin_theta * s11_old + cos_theta * s21_old) +                            \
+               sin_theta * (-sin_theta * s21_old + cos_theta * s22_old);                             \
+        *s22 = -sin_theta * (-sin_theta * s11_old + cos_theta * s21_old) +                           \
+               cos_theta * (-sin_theta * s21_old + cos_theta * s22_old);                             \
+        *s31 = cos_theta * s31_old + sin_theta * s32_old;                                            \
+        *s32 = -sin_theta * s31_old + cos_theta * s32_old;                                           \
+        *s33 = s33_old;                                                                              \
+        /* Update quaternion accumulator */                                                          \
+        nk_##type##_t quat_temp[3];                                                                  \
+        quat_temp[0] = quaternion[0] * sin_half;                                                     \
+        quat_temp[1] = quaternion[1] * sin_half;                                                     \
+        quat_temp[2] = quaternion[2] * sin_half;                                                     \
+        sin_half *= quaternion[3];                                                                   \
+        quaternion[0] *= cos_half, quaternion[1] *= cos_half;                                        \
+        quaternion[2] *= cos_half, quaternion[3] *= cos_half;                                        \
+        quaternion[idx_z] += sin_half, quaternion[3] -= quat_temp[idx_z];                            \
+        quaternion[idx_x] += quat_temp[idx_y], quaternion[idx_y] -= quat_temp[idx_x];                \
+        /* Cyclic permutation of matrix elements */                                                  \
+        s11_old = *s22, s21_old = *s32, s22_old = *s33;                                              \
+        s31_old = *s21, s32_old = *s31, s33_old = *s11;                                              \
+        *s11 = s11_old, *s21 = s21_old, *s22 = s22_old;                                              \
+        *s31 = s31_old, *s32 = s32_old, *s33 = s33_old;                                              \
+    }
+
+#define nk_define_quaternion_to_mat3x3_(type)                                                              \
+    NK_INTERNAL void nk_quaternion_to_mat3x3_##type##_(nk_##type##_t const *quat, nk_##type##_t *matrix) { \
+        nk_##type##_t w = quat[3], x = quat[0], y = quat[1], z = quat[2];                                  \
+        nk_##type##_t q_xx = x * x, q_yy = y * y, q_zz = z * z;                                            \
+        nk_##type##_t q_xz = x * z, q_xy = x * y, q_yz = y * z;                                            \
+        nk_##type##_t q_wx = w * x, q_wy = w * y, q_wz = w * z;                                            \
+        matrix[0] = 1 - 2 * (q_yy + q_zz), matrix[1] = 2 * (q_xy - q_wz);                                  \
+        matrix[2] = 2 * (q_xz + q_wy);                                                                     \
+        matrix[3] = 2 * (q_xy + q_wz), matrix[4] = 1 - 2 * (q_xx + q_zz);                                  \
+        matrix[5] = 2 * (q_yz - q_wx);                                                                     \
+        matrix[6] = 2 * (q_xz - q_wy), matrix[7] = 2 * (q_yz + q_wx);                                      \
+        matrix[8] = 1 - 2 * (q_xx + q_yy);                                                                 \
+    }
+
+#define nk_define_jacobi_eigenanalysis_(type, compute_rsqrt)                                                       \
+    NK_INTERNAL void nk_jacobi_eigenanalysis_##type##_(nk_##type##_t *s11, nk_##type##_t *s21, nk_##type##_t *s22, \
+                                                       nk_##type##_t *s31, nk_##type##_t *s32, nk_##type##_t *s33, \
+                                                       nk_##type##_t *quaternion) {                                \
+        quaternion[0] = 0, quaternion[1] = 0, quaternion[2] = 0, quaternion[3] = 1;                                \
+        /* 16 iterations for better convergence with repeated eigenvalues and identity-like matrices */            \
+        for (int iter = 0; iter < 16; iter++) {                                                                    \
+            nk_jacobi_conjugation_##type##_(0, 1, 2, s11, s21, s22, s31, s32, s33, quaternion);                    \
+            nk_jacobi_conjugation_##type##_(1, 2, 0, s11, s21, s22, s31, s32, s33, quaternion);                    \
+            nk_jacobi_conjugation_##type##_(2, 0, 1, s11, s21, s22, s31, s32, s33, quaternion);                    \
+        }                                                                                                          \
+        nk_##type##_t norm = compute_rsqrt(quaternion[0] * quaternion[0] + quaternion[1] * quaternion[1] +         \
+                                           quaternion[2] * quaternion[2] + quaternion[3] * quaternion[3]);         \
+        quaternion[0] *= norm, quaternion[1] *= norm;                                                              \
+        quaternion[2] *= norm, quaternion[3] *= norm;                                                              \
+    }
+
+#define nk_define_qr_givens_quaternion_(type, epsilon, compute_rsqrt)                                               \
+    NK_INTERNAL void nk_qr_givens_quaternion_##type##_(nk_##type##_t a1, nk_##type##_t a2, nk_##type##_t *cos_half, \
+                                                       nk_##type##_t *sin_half) {                                   \
+        nk_##type##_t a1_sq_plus_a2_sq = a1 * a1 + a2 * a2;                                                         \
+        nk_##type##_t rho = a1_sq_plus_a2_sq * compute_rsqrt(a1_sq_plus_a2_sq);                                     \
+        rho = a1_sq_plus_a2_sq > epsilon ? rho : 0;                                                                 \
+        *sin_half = rho > epsilon ? a2 : 0;                                                                         \
+        nk_##type##_t abs_a1 = a1 < 0 ? -a1 : a1;                                                                   \
+        nk_##type##_t max_rho = rho > epsilon ? rho : epsilon;                                                      \
+        *cos_half = abs_a1 + max_rho;                                                                               \
+        int should_swap = a1 < 0;                                                                                   \
+        nk_cond_swap_##type##_(should_swap, sin_half, cos_half);                                                    \
+        nk_##type##_t w = compute_rsqrt((*cos_half) * (*cos_half) + (*sin_half) * (*sin_half));                     \
+        *cos_half *= w, *sin_half *= w;                                                                             \
+    }
+
+#define nk_define_sort_singular_values_(type)                                                \
+    NK_INTERNAL void nk_sort_singular_values_##type##_(nk_##type##_t *b, nk_##type##_t *v) { \
+        nk_##type##_t rho1 = b[0] * b[0] + b[3] * b[3] + b[6] * b[6];                        \
+        nk_##type##_t rho2 = b[1] * b[1] + b[4] * b[4] + b[7] * b[7];                        \
+        nk_##type##_t rho3 = b[2] * b[2] + b[5] * b[5] + b[8] * b[8];                        \
+        int should_swap;                                                                     \
+        /* Sort columns by descending singular value magnitude */                            \
+        should_swap = rho1 < rho2;                                                           \
+        nk_conditional_negating_swap_##type##_(should_swap, &b[0], &b[1]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &v[0], &v[1]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &b[3], &b[4]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &v[3], &v[4]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &b[6], &b[7]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &v[6], &v[7]);                   \
+        nk_cond_swap_##type##_(should_swap, &rho1, &rho2);                                   \
+        should_swap = rho1 < rho3;                                                           \
+        nk_conditional_negating_swap_##type##_(should_swap, &b[0], &b[2]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &v[0], &v[2]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &b[3], &b[5]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &v[3], &v[5]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &b[6], &b[8]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &v[6], &v[8]);                   \
+        nk_cond_swap_##type##_(should_swap, &rho1, &rho3);                                   \
+        should_swap = rho2 < rho3;                                                           \
+        nk_conditional_negating_swap_##type##_(should_swap, &b[1], &b[2]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &v[1], &v[2]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &b[4], &b[5]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &v[4], &v[5]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &b[7], &b[8]);                   \
+        nk_conditional_negating_swap_##type##_(should_swap, &v[7], &v[8]);                   \
+    }
+
+#define nk_define_qr_decomposition_(type)                                                                            \
+    NK_INTERNAL void nk_qr_decomposition_##type##_(nk_##type##_t const *input, nk_##type##_t *q, nk_##type##_t *r) { \
+        nk_##type##_t cos_half_1, sin_half_1;                                                                        \
+        nk_##type##_t cos_half_2, sin_half_2;                                                                        \
+        nk_##type##_t cos_half_3, sin_half_3;                                                                        \
+        nk_##type##_t cos_theta, sin_theta;                                                                          \
+        nk_##type##_t rotation_temp[9], matrix_temp[9];                                                              \
+        /* First Givens rotation (zero input[3]) */                                                                  \
+        nk_qr_givens_quaternion_##type##_(input[0], input[3], &cos_half_1, &sin_half_1);                             \
+        cos_theta = 1 - 2 * sin_half_1 * sin_half_1;                                                                 \
+        sin_theta = 2 * cos_half_1 * sin_half_1;                                                                     \
+        rotation_temp[0] = cos_theta * input[0] + sin_theta * input[3];                                              \
+        rotation_temp[1] = cos_theta * input[1] + sin_theta * input[4];                                              \
+        rotation_temp[2] = cos_theta * input[2] + sin_theta * input[5];                                              \
+        rotation_temp[3] = -sin_theta * input[0] + cos_theta * input[3];                                             \
+        rotation_temp[4] = -sin_theta * input[1] + cos_theta * input[4];                                             \
+        rotation_temp[5] = -sin_theta * input[2] + cos_theta * input[5];                                             \
+        rotation_temp[6] = input[6], rotation_temp[7] = input[7];                                                    \
+        rotation_temp[8] = input[8];                                                                                 \
+        /* Second Givens rotation (zero rotation_temp[6]) */                                                         \
+        nk_qr_givens_quaternion_##type##_(rotation_temp[0], rotation_temp[6], &cos_half_2, &sin_half_2);             \
+        cos_theta = 1 - 2 * sin_half_2 * sin_half_2;                                                                 \
+        sin_theta = 2 * cos_half_2 * sin_half_2;                                                                     \
+        matrix_temp[0] = cos_theta * rotation_temp[0] + sin_theta * rotation_temp[6];                                \
+        matrix_temp[1] = cos_theta * rotation_temp[1] + sin_theta * rotation_temp[7];                                \
+        matrix_temp[2] = cos_theta * rotation_temp[2] + sin_theta * rotation_temp[8];                                \
+        matrix_temp[3] = rotation_temp[3], matrix_temp[4] = rotation_temp[4];                                        \
+        matrix_temp[5] = rotation_temp[5];                                                                           \
+        matrix_temp[6] = -sin_theta * rotation_temp[0] + cos_theta * rotation_temp[6];                               \
+        matrix_temp[7] = -sin_theta * rotation_temp[1] + cos_theta * rotation_temp[7];                               \
+        matrix_temp[8] = -sin_theta * rotation_temp[2] + cos_theta * rotation_temp[8];                               \
+        /* Third Givens rotation (zero matrix_temp[7]) */                                                            \
+        nk_qr_givens_quaternion_##type##_(matrix_temp[4], matrix_temp[7], &cos_half_3, &sin_half_3);                 \
+        cos_theta = 1 - 2 * sin_half_3 * sin_half_3;                                                                 \
+        sin_theta = 2 * cos_half_3 * sin_half_3;                                                                     \
+        r[0] = matrix_temp[0], r[1] = matrix_temp[1], r[2] = matrix_temp[2];                                         \
+        r[3] = cos_theta * matrix_temp[3] + sin_theta * matrix_temp[6];                                              \
+        r[4] = cos_theta * matrix_temp[4] + sin_theta * matrix_temp[7];                                              \
+        r[5] = cos_theta * matrix_temp[5] + sin_theta * matrix_temp[8];                                              \
+        r[6] = -sin_theta * matrix_temp[3] + cos_theta * matrix_temp[6];                                             \
+        r[7] = -sin_theta * matrix_temp[4] + cos_theta * matrix_temp[7];                                             \
+        r[8] = -sin_theta * matrix_temp[5] + cos_theta * matrix_temp[8];                                             \
+        /* Construct Q = Q1 * Q2 * Q3 (closed-form expressions) */                                                   \
+        nk_##type##_t sin_half_1_sq = sin_half_1 * sin_half_1;                                                       \
+        nk_##type##_t sin_half_2_sq = sin_half_2 * sin_half_2;                                                       \
+        nk_##type##_t sin_half_3_sq = sin_half_3 * sin_half_3;                                                       \
+        q[0] = (-1 + 2 * sin_half_1_sq) * (-1 + 2 * sin_half_2_sq);                                                  \
+        q[1] = 4 * cos_half_2 * cos_half_3 * (-1 + 2 * sin_half_1_sq) * sin_half_2 * sin_half_3 +                    \
+               2 * cos_half_1 * sin_half_1 * (-1 + 2 * sin_half_3_sq);                                               \
+        q[2] = 4 * cos_half_1 * cos_half_3 * sin_half_1 * sin_half_3 -                                               \
+               2 * cos_half_2 * (-1 + 2 * sin_half_1_sq) * sin_half_2 * (-1 + 2 * sin_half_3_sq);                    \
+        q[3] = 2 * cos_half_1 * sin_half_1 * (1 - 2 * sin_half_2_sq);                                                \
+        q[4] = -8 * cos_half_1 * cos_half_2 * cos_half_3 * sin_half_1 * sin_half_2 * sin_half_3 +                    \
+               (-1 + 2 * sin_half_1_sq) * (-1 + 2 * sin_half_3_sq);                                                  \
+        q[5] = -2 * cos_half_3 * sin_half_3 + 4 * sin_half_1 *                                                       \
+                                                  (cos_half_3 * sin_half_1 * sin_half_3 +                            \
+                                                   cos_half_1 * cos_half_2 * sin_half_2 * (-1 + 2 * sin_half_3_sq)); \
+        q[6] = 2 * cos_half_2 * sin_half_2;                                                                          \
+        q[7] = 2 * cos_half_3 * (1 - 2 * sin_half_2_sq) * sin_half_3;                                                \
+        q[8] = (-1 + 2 * sin_half_2_sq) * (-1 + 2 * sin_half_3_sq);                                                  \
+    }
+
+#define nk_define_svd3x3_(type, compute_sqrt)                                                                \
+    NK_INTERNAL void nk_svd3x3_##type##_(nk_##type##_t const *a, nk_##type##_t *svd_u, nk_##type##_t *svd_s, \
+                                         nk_##type##_t *svd_v) {                                             \
+        /* Compute Aᵀ * A (symmetric) */                                                                     \
+        nk_##type##_t ata[9];                                                                                \
+        ata[0] = nk_sum_three_squares_##type##_(a[0], a[3], a[6]);                                           \
+        ata[1] = nk_sum_three_products_##type##_(a[0], a[1], a[3], a[4], a[6], a[7]);                        \
+        ata[2] = nk_sum_three_products_##type##_(a[0], a[2], a[3], a[5], a[6], a[8]);                        \
+        ata[3] = ata[1];                                                                                     \
+        ata[4] = nk_sum_three_squares_##type##_(a[1], a[4], a[7]);                                           \
+        ata[5] = nk_sum_three_products_##type##_(a[1], a[2], a[4], a[5], a[7], a[8]);                        \
+        ata[6] = ata[2];                                                                                     \
+        ata[7] = ata[5];                                                                                     \
+        ata[8] = nk_sum_three_squares_##type##_(a[2], a[5], a[8]);                                           \
+        /* Jacobi eigenanalysis of Aᵀ * A */                                                                 \
+        nk_##type##_t quaternion[4];                                                                         \
+        nk_jacobi_eigenanalysis_##type##_(&ata[0], &ata[1], &ata[4], &ata[2], &ata[5], &ata[8], quaternion); \
+        nk_quaternion_to_mat3x3_##type##_(quaternion, svd_v);                                                \
+        /* B = A * V */                                                                                      \
+        nk_##type##_t product[9];                                                                            \
+        product[0] = nk_sum_three_products_##type##_(a[0], svd_v[0], a[1], svd_v[3], a[2], svd_v[6]);        \
+        product[1] = nk_sum_three_products_##type##_(a[0], svd_v[1], a[1], svd_v[4], a[2], svd_v[7]);        \
+        product[2] = nk_sum_three_products_##type##_(a[0], svd_v[2], a[1], svd_v[5], a[2], svd_v[8]);        \
+        product[3] = nk_sum_three_products_##type##_(a[3], svd_v[0], a[4], svd_v[3], a[5], svd_v[6]);        \
+        product[4] = nk_sum_three_products_##type##_(a[3], svd_v[1], a[4], svd_v[4], a[5], svd_v[7]);        \
+        product[5] = nk_sum_three_products_##type##_(a[3], svd_v[2], a[4], svd_v[5], a[5], svd_v[8]);        \
+        product[6] = nk_sum_three_products_##type##_(a[6], svd_v[0], a[7], svd_v[3], a[8], svd_v[6]);        \
+        product[7] = nk_sum_three_products_##type##_(a[6], svd_v[1], a[7], svd_v[4], a[8], svd_v[7]);        \
+        product[8] = nk_sum_three_products_##type##_(a[6], svd_v[2], a[7], svd_v[5], a[8], svd_v[8]);        \
+        /* Sort singular values and update V */                                                              \
+        nk_sort_singular_values_##type##_(product, svd_v);                                                   \
+        /* Compute singular values from column norms of sorted B (before QR orthogonalizes them) */          \
+        /* These are the true singular values: √(‖colᵢ‖²) */                                                 \
+        nk_##type##_t s1_sq = nk_sum_three_squares_##type##_(product[0], product[3], product[6]);            \
+        nk_##type##_t s2_sq = nk_sum_three_squares_##type##_(product[1], product[4], product[7]);            \
+        nk_##type##_t s3_sq = nk_sum_three_squares_##type##_(product[2], product[5], product[8]);            \
+        /* QR decomposition: B = U * R (we only need U for the rotation) */                                  \
+        nk_##type##_t qr_r[9];                                                                               \
+        nk_qr_decomposition_##type##_(product, svd_u, qr_r);                                                 \
+        /* Store singular values in diagonal of svd_s (rest is zero for compatibility) */                    \
+        svd_s[0] = compute_sqrt(s1_sq), svd_s[1] = 0, svd_s[2] = 0;                                          \
+        svd_s[3] = 0, svd_s[4] = compute_sqrt(s2_sq), svd_s[5] = 0;                                          \
+        svd_s[6] = 0, svd_s[7] = 0, svd_s[8] = compute_sqrt(s3_sq);                                          \
+    }
+
+#define nk_define_det3x3_(type)                                                          \
+    NK_INTERNAL nk_##type##_t nk_det3x3_##type##_(nk_##type##_t const *m) {              \
+        return m[0] * (m[4] * m[8] - m[5] * m[7]) - m[1] * (m[3] * m[8] - m[5] * m[6]) + \
+               m[2] * (m[3] * m[7] - m[4] * m[6]);                                       \
+    }
+
+/* Optimize serial fallbacks for size — see dots/serial.h for rationale. */
+#if defined(NDEBUG)
+#if defined(_MSC_VER)
+#pragma optimize("s", on)
+#elif defined(__clang__)
+#pragma clang attribute push(__attribute__((minsize)), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC optimize("Os")
+#endif
+#endif
+
+NK_INTERNAL nk_f32_t nk_sum_three_products_f32_(nk_f32_t left_0, nk_f32_t right_0, nk_f32_t left_1, nk_f32_t right_1,
+                                                nk_f32_t left_2, nk_f32_t right_2) {
+    return left_0 * right_0 + left_1 * right_1 + left_2 * right_2;
+}
+NK_INTERNAL nk_f64_t nk_sum_three_products_f64_(nk_f64_t left_0, nk_f64_t right_0, nk_f64_t left_1, nk_f64_t right_1,
+                                                nk_f64_t left_2, nk_f64_t right_2) {
+    nk_f64_t sum = 0.0, compensation = 0.0;
+    nk_f64_dot2_(&sum, &compensation, left_0, right_0);
+    nk_f64_dot2_(&sum, &compensation, left_1, right_1);
+    nk_f64_dot2_(&sum, &compensation, left_2, right_2);
+    return sum + compensation;
+}
+
+NK_INTERNAL nk_f32_t nk_sum_three_squares_f32_(nk_f32_t value_0, nk_f32_t value_1, nk_f32_t value_2) {
+    return value_0 * value_0 + value_1 * value_1 + value_2 * value_2;
+}
+NK_INTERNAL nk_f64_t nk_sum_three_squares_f64_(nk_f64_t value_0, nk_f64_t value_1, nk_f64_t value_2) {
+    nk_f64_t sum = 0.0, compensation = 0.0;
+    nk_f64_dot2_(&sum, &compensation, value_0, value_0);
+    nk_f64_dot2_(&sum, &compensation, value_1, value_1);
+    nk_f64_dot2_(&sum, &compensation, value_2, value_2);
+    return sum + compensation;
+}
+
+NK_INTERNAL void nk_accumulate_sum_f32_(nk_f32_t *sum, nk_f32_t *compensation, nk_f32_t value) {
+    nk_unused_(compensation);
+    *sum += value;
+}
+NK_INTERNAL void nk_accumulate_sum_f64_(nk_f64_t *sum, nk_f64_t *compensation, nk_f64_t value) {
+    nk_f64_t running_sum = *sum + value;
+    *compensation += (nk_f64_abs_(*sum) >= nk_f64_abs_(value)) ? ((*sum - running_sum) + value)
+                                                               : ((value - running_sum) + *sum);
+    *sum = running_sum;
+}
+
+NK_INTERNAL void nk_accumulate_product_f32_(nk_f32_t *sum, nk_f32_t *compensation, nk_f32_t left, nk_f32_t right) {
+    nk_unused_(compensation);
+    *sum += left * right;
+}
+NK_INTERNAL void nk_accumulate_product_f64_(nk_f64_t *sum, nk_f64_t *compensation, nk_f64_t left, nk_f64_t right) {
+    nk_f64_dot2_(sum, compensation, left, right);
+}
+
+NK_INTERNAL void nk_accumulate_square_f32_(nk_f32_t *sum, nk_f32_t *compensation, nk_f32_t value) {
+    nk_unused_(compensation);
+    *sum += value * value;
+}
+NK_INTERNAL void nk_accumulate_square_f64_(nk_f64_t *sum, nk_f64_t *compensation, nk_f64_t value) {
+    nk_f64_dot2_(sum, compensation, value, value);
+}
+
+NK_INTERNAL void nk_rotation_from_svd_f32_serial_(nk_f32_t const *svd_u, nk_f32_t const *svd_v, nk_f32_t *rotation) {
+    rotation[0] = nk_sum_three_products_f32_(svd_v[0], svd_u[0], svd_v[1], svd_u[1], svd_v[2], svd_u[2]);
+    rotation[1] = nk_sum_three_products_f32_(svd_v[0], svd_u[3], svd_v[1], svd_u[4], svd_v[2], svd_u[5]);
+    rotation[2] = nk_sum_three_products_f32_(svd_v[0], svd_u[6], svd_v[1], svd_u[7], svd_v[2], svd_u[8]);
+    rotation[3] = nk_sum_three_products_f32_(svd_v[3], svd_u[0], svd_v[4], svd_u[1], svd_v[5], svd_u[2]);
+    rotation[4] = nk_sum_three_products_f32_(svd_v[3], svd_u[3], svd_v[4], svd_u[4], svd_v[5], svd_u[5]);
+    rotation[5] = nk_sum_three_products_f32_(svd_v[3], svd_u[6], svd_v[4], svd_u[7], svd_v[5], svd_u[8]);
+    rotation[6] = nk_sum_three_products_f32_(svd_v[6], svd_u[0], svd_v[7], svd_u[1], svd_v[8], svd_u[2]);
+    rotation[7] = nk_sum_three_products_f32_(svd_v[6], svd_u[3], svd_v[7], svd_u[4], svd_v[8], svd_u[5]);
+    rotation[8] = nk_sum_three_products_f32_(svd_v[6], svd_u[6], svd_v[7], svd_u[7], svd_v[8], svd_u[8]);
+}
+NK_INTERNAL void nk_rotation_from_svd_f64_serial_(nk_f64_t const *svd_u, nk_f64_t const *svd_v, nk_f64_t *rotation) {
+    rotation[0] = nk_sum_three_products_f64_(svd_v[0], svd_u[0], svd_v[1], svd_u[1], svd_v[2], svd_u[2]);
+    rotation[1] = nk_sum_three_products_f64_(svd_v[0], svd_u[3], svd_v[1], svd_u[4], svd_v[2], svd_u[5]);
+    rotation[2] = nk_sum_three_products_f64_(svd_v[0], svd_u[6], svd_v[1], svd_u[7], svd_v[2], svd_u[8]);
+    rotation[3] = nk_sum_three_products_f64_(svd_v[3], svd_u[0], svd_v[4], svd_u[1], svd_v[5], svd_u[2]);
+    rotation[4] = nk_sum_three_products_f64_(svd_v[3], svd_u[3], svd_v[4], svd_u[4], svd_v[5], svd_u[5]);
+    rotation[5] = nk_sum_three_products_f64_(svd_v[3], svd_u[6], svd_v[4], svd_u[7], svd_v[5], svd_u[8]);
+    rotation[6] = nk_sum_three_products_f64_(svd_v[6], svd_u[0], svd_v[7], svd_u[1], svd_v[8], svd_u[2]);
+    rotation[7] = nk_sum_three_products_f64_(svd_v[6], svd_u[3], svd_v[7], svd_u[4], svd_v[8], svd_u[5]);
+    rotation[8] = nk_sum_three_products_f64_(svd_v[6], svd_u[6], svd_v[7], svd_u[7], svd_v[8], svd_u[8]);
+}
+
+nk_define_cond_swap_(f32)
+nk_define_conditional_negating_swap_(f32)
+nk_define_approximate_givens_quaternion_(f32, NK_F32_SVD_GAMMA_, NK_F32_SVD_CSTAR_, NK_F32_SVD_SSTAR_,
+                                         nk_f32_rsqrt_serial)
+nk_define_jacobi_conjugation_(f32)
+nk_define_quaternion_to_mat3x3_(f32)
+nk_define_jacobi_eigenanalysis_(f32, nk_f32_rsqrt_serial)
+nk_define_qr_givens_quaternion_(f32, NK_F32_SVD_EPSILON_, nk_f32_rsqrt_serial)
+nk_define_sort_singular_values_(f32)
+nk_define_qr_decomposition_(f32)
+nk_define_svd3x3_(f32, nk_f32_sqrt_serial)
+nk_define_det3x3_(f32)
+
+nk_define_cond_swap_(f64)
+nk_define_conditional_negating_swap_(f64)
+nk_define_approximate_givens_quaternion_(f64, NK_F64_SVD_GAMMA_, NK_F64_SVD_CSTAR_, NK_F64_SVD_SSTAR_,
+                                         nk_f64_rsqrt_serial)
+nk_define_jacobi_conjugation_(f64)
+nk_define_quaternion_to_mat3x3_(f64)
+nk_define_jacobi_eigenanalysis_(f64, nk_f64_rsqrt_serial)
+nk_define_qr_givens_quaternion_(f64, NK_F64_SVD_EPSILON_, nk_f64_rsqrt_serial)
+nk_define_sort_singular_values_(f64)
+nk_define_qr_decomposition_(f64)
+nk_define_svd3x3_(f64, nk_f64_sqrt_serial)
+nk_define_det3x3_(f64)
+
+/*  RMSD (Root Mean Square Deviation) without optimal superposition.
+ *  Simply computes the RMS of distances between corresponding points.
+ */
+#define nk_define_rmsd_(input_type, accumulator_type, output_type, result_type, load_and_convert, compute_sqrt)    \
+    NK_PUBLIC void nk_rmsd_##input_type##_serial(nk_##input_type##_t const *a, nk_##input_type##_t const *b,       \
+                                                 nk_size_t n, nk_##output_type##_t *a_centroid,                    \
+                                                 nk_##output_type##_t *b_centroid, nk_##output_type##_t *rotation, \
+                                                 nk_##output_type##_t *scale, nk_##result_type##_t *result) {      \
+        if (a_centroid) a_centroid[0] = 0, a_centroid[1] = 0, a_centroid[2] = 0;                                   \
+        if (b_centroid) b_centroid[0] = 0, b_centroid[1] = 0, b_centroid[2] = 0;                                   \
+        if (rotation)                                                                                              \
+            rotation[0] = 1, rotation[1] = 0, rotation[2] = 0, rotation[3] = 0, rotation[4] = 1, rotation[5] = 0,  \
+            rotation[6] = 0, rotation[7] = 0, rotation[8] = 1;                                                     \
+        if (scale) *scale = (nk_##output_type##_t)1;                                                               \
+        nk_##accumulator_type##_t sum_squared = 0, sum_squared_compensation = 0;                                   \
+        nk_##accumulator_type##_t val_a_x, val_a_y, val_a_z, val_b_x, val_b_y, val_b_z;                            \
+        for (nk_size_t i = 0; i < n; ++i) {                                                                        \
+            load_and_convert(a + i * 3 + 0, &val_a_x), load_and_convert(b + i * 3 + 0, &val_b_x);                  \
+            load_and_convert(a + i * 3 + 1, &val_a_y), load_and_convert(b + i * 3 + 1, &val_b_y);                  \
+            load_and_convert(a + i * 3 + 2, &val_a_z), load_and_convert(b + i * 3 + 2, &val_b_z);                  \
+            nk_##accumulator_type##_t dx = val_a_x - val_b_x;                                                      \
+            nk_##accumulator_type##_t dy = val_a_y - val_b_y;                                                      \
+            nk_##accumulator_type##_t dz = val_a_z - val_b_z;                                                      \
+            nk_accumulate_square_##accumulator_type##_(&sum_squared, &sum_squared_compensation, dx);               \
+            nk_accumulate_square_##accumulator_type##_(&sum_squared, &sum_squared_compensation, dy);               \
+            nk_accumulate_square_##accumulator_type##_(&sum_squared, &sum_squared_compensation, dz);               \
+        }                                                                                                          \
+        nk_##accumulator_type##_t msd = (sum_squared + sum_squared_compensation) / (nk_##accumulator_type##_t)n;   \
+        *result = msd > 0 ? (nk_##result_type##_t)compute_sqrt(msd) : 0;                                           \
+    }
+
+/*  Kabsch algorithm for optimal rigid body superposition.
+ *  Finds the rotation matrix R that minimizes RMSD between the two point sets.
+ */
+#define nk_define_kabsch_(input_type, accumulator_type, output_type, result_type, svd_type, load_and_convert,        \
+                          compute_sqrt)                                                                              \
+    NK_PUBLIC void nk_kabsch_##input_type##_serial(nk_##input_type##_t const *a, nk_##input_type##_t const *b,       \
+                                                   nk_size_t n, nk_##output_type##_t *a_centroid,                    \
+                                                   nk_##output_type##_t *b_centroid, nk_##output_type##_t *rotation, \
+                                                   nk_##output_type##_t *scale, nk_##result_type##_t *result) {      \
+        /* Step 1: Compute centroids */                                                                              \
+        nk_##accumulator_type##_t sum_a_x = 0, sum_a_y = 0, sum_a_z = 0;                                             \
+        nk_##accumulator_type##_t sum_b_x = 0, sum_b_y = 0, sum_b_z = 0;                                             \
+        nk_##accumulator_type##_t sum_a_x_compensation = 0, sum_a_y_compensation = 0, sum_a_z_compensation = 0;      \
+        nk_##accumulator_type##_t sum_b_x_compensation = 0, sum_b_y_compensation = 0, sum_b_z_compensation = 0;      \
+        nk_##accumulator_type##_t val_a_x, val_a_y, val_a_z, val_b_x, val_b_y, val_b_z;                              \
+        for (nk_size_t i = 0; i < n; ++i) {                                                                          \
+            load_and_convert(a + i * 3 + 0, &val_a_x), load_and_convert(a + i * 3 + 1, &val_a_y);                    \
+            load_and_convert(a + i * 3 + 2, &val_a_z), load_and_convert(b + i * 3 + 0, &val_b_x);                    \
+            load_and_convert(b + i * 3 + 1, &val_b_y), load_and_convert(b + i * 3 + 2, &val_b_z);                    \
+            nk_accumulate_sum_##accumulator_type##_(&sum_a_x, &sum_a_x_compensation, val_a_x);                       \
+            nk_accumulate_sum_##accumulator_type##_(&sum_a_y, &sum_a_y_compensation, val_a_y);                       \
+            nk_accumulate_sum_##accumulator_type##_(&sum_a_z, &sum_a_z_compensation, val_a_z);                       \
+            nk_accumulate_sum_##accumulator_type##_(&sum_b_x, &sum_b_x_compensation, val_b_x);                       \
+            nk_accumulate_sum_##accumulator_type##_(&sum_b_y, &sum_b_y_compensation, val_b_y);                       \
+            nk_accumulate_sum_##accumulator_type##_(&sum_b_z, &sum_b_z_compensation, val_b_z);                       \
+        }                                                                                                            \
+        nk_##accumulator_type##_t inv_n = (nk_##accumulator_type##_t)1.0 / (nk_##accumulator_type##_t)n;             \
+        nk_##accumulator_type##_t centroid_a_x = (sum_a_x + sum_a_x_compensation) * inv_n;                           \
+        nk_##accumulator_type##_t centroid_a_y = (sum_a_y + sum_a_y_compensation) * inv_n;                           \
+        nk_##accumulator_type##_t centroid_a_z = (sum_a_z + sum_a_z_compensation) * inv_n;                           \
+        nk_##accumulator_type##_t centroid_b_x = (sum_b_x + sum_b_x_compensation) * inv_n;                           \
+        nk_##accumulator_type##_t centroid_b_y = (sum_b_y + sum_b_y_compensation) * inv_n;                           \
+        nk_##accumulator_type##_t centroid_b_z = (sum_b_z + sum_b_z_compensation) * inv_n;                           \
+        if (a_centroid)                                                                                              \
+            a_centroid[0] = (nk_##output_type##_t)centroid_a_x, a_centroid[1] = (nk_##output_type##_t)centroid_a_y,  \
+            a_centroid[2] = (nk_##output_type##_t)centroid_a_z;                                                      \
+        if (b_centroid)                                                                                              \
+            b_centroid[0] = (nk_##output_type##_t)centroid_b_x, b_centroid[1] = (nk_##output_type##_t)centroid_b_y,  \
+            b_centroid[2] = (nk_##output_type##_t)centroid_b_z;                                                      \
+        /* Step 2: Build 3×3 covariance matrix H = (A - Ā)ᵀ × (B - B̄) */                                             \
+        nk_##accumulator_type##_t h[9] = {0, 0, 0, 0, 0, 0, 0, 0, 0};                                                \
+        nk_##accumulator_type##_t h_compensation[9] = {0, 0, 0, 0, 0, 0, 0, 0, 0};                                   \
+        for (nk_size_t i = 0; i < n; ++i) {                                                                          \
+            load_and_convert(a + i * 3 + 0, &val_a_x), load_and_convert(b + i * 3 + 0, &val_b_x);                    \
+            load_and_convert(a + i * 3 + 1, &val_a_y), load_and_convert(b + i * 3 + 1, &val_b_y);                    \
+            load_and_convert(a + i * 3 + 2, &val_a_z), load_and_convert(b + i * 3 + 2, &val_b_z);                    \
+            val_a_x -= centroid_a_x, val_a_y -= centroid_a_y, val_a_z -= centroid_a_z;                               \
+            val_b_x -= centroid_b_x, val_b_y -= centroid_b_y, val_b_z -= centroid_b_z;                               \
+            nk_accumulate_product_##accumulator_type##_(&h[0], &h_compensation[0], val_a_x, val_b_x);                \
+            nk_accumulate_product_##accumulator_type##_(&h[1], &h_compensation[1], val_a_x, val_b_y);                \
+            nk_accumulate_product_##accumulator_type##_(&h[2], &h_compensation[2], val_a_x, val_b_z);                \
+            nk_accumulate_product_##accumulator_type##_(&h[3], &h_compensation[3], val_a_y, val_b_x);                \
+            nk_accumulate_product_##accumulator_type##_(&h[4], &h_compensation[4], val_a_y, val_b_y);                \
+            nk_accumulate_product_##accumulator_type##_(&h[5], &h_compensation[5], val_a_y, val_b_z);                \
+            nk_accumulate_product_##accumulator_type##_(&h[6], &h_compensation[6], val_a_z, val_b_x);                \
+            nk_accumulate_product_##accumulator_type##_(&h[7], &h_compensation[7], val_a_z, val_b_y);                \
+            nk_accumulate_product_##accumulator_type##_(&h[8], &h_compensation[8], val_a_z, val_b_z);                \
+        }                                                                                                            \
+        /* Convert to svd_type for SVD */                                                                            \
+        nk_##svd_type##_t cross_covariance[9];                                                                       \
+        for (int j = 0; j < 9; ++j) cross_covariance[j] = (nk_##svd_type##_t)(h[j] + h_compensation[j]);             \
+        /* Step 3: SVD of H = U * S * Vᵀ */                                                                          \
+        nk_##svd_type##_t svd_u[9], svd_s[9], svd_v[9];                                                              \
+        nk_svd3x3_##svd_type##_(cross_covariance, svd_u, svd_s, svd_v);                                              \
+        /* Step 4: R = V * Uᵀ */                                                                                     \
+        nk_##svd_type##_t rotation_matrix[9];                                                                        \
+        nk_rotation_from_svd_##svd_type##_serial_(svd_u, svd_v, rotation_matrix);                                    \
+        /* Handle reflection: if det(R) < 0, negate third column of V and recompute R */                             \
+        nk_##svd_type##_t rotation_det = nk_det3x3_##svd_type##_(rotation_matrix);                                   \
+        if (rotation_det < 0) {                                                                                      \
+            svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];                                        \
+            nk_rotation_from_svd_##svd_type##_serial_(svd_u, svd_v, rotation_matrix);                                \
+        }                                                                                                            \
+        /* Output rotation matrix and scale=1.0 */                                                                   \
+        if (rotation)                                                                                                \
+            for (int j = 0; j < 9; ++j) rotation[j] = (nk_##output_type##_t)rotation_matrix[j];                      \
+        if (scale) *scale = (nk_##output_type##_t)1;                                                                 \
+        /* Step 5: Compute RMSD after rotation */                                                                    \
+        nk_##accumulator_type##_t sum_squared = 0, sum_squared_compensation = 0;                                     \
+        for (nk_size_t i = 0; i < n; ++i) {                                                                          \
+            nk_##svd_type##_t point_a[3], point_b[3], rotated_point_a[3];                                            \
+            load_and_convert(a + i * 3 + 0, &val_a_x), load_and_convert(a + i * 3 + 1, &val_a_y);                    \
+            load_and_convert(a + i * 3 + 2, &val_a_z), load_and_convert(b + i * 3 + 0, &val_b_x);                    \
+            load_and_convert(b + i * 3 + 1, &val_b_y), load_and_convert(b + i * 3 + 2, &val_b_z);                    \
+            point_a[0] = (nk_##svd_type##_t)(val_a_x - centroid_a_x),                                                \
+            point_a[1] = (nk_##svd_type##_t)(val_a_y - centroid_a_y),                                                \
+            point_a[2] = (nk_##svd_type##_t)(val_a_z - centroid_a_z);                                                \
+            point_b[0] = (nk_##svd_type##_t)(val_b_x - centroid_b_x),                                                \
+            point_b[1] = (nk_##svd_type##_t)(val_b_y - centroid_b_y),                                                \
+            point_b[2] = (nk_##svd_type##_t)(val_b_z - centroid_b_z);                                                \
+            rotated_point_a[0] = rotation_matrix[0] * point_a[0] + rotation_matrix[1] * point_a[1] +                 \
+                                 rotation_matrix[2] * point_a[2];                                                    \
+            rotated_point_a[1] = rotation_matrix[3] * point_a[0] + rotation_matrix[4] * point_a[1] +                 \
+                                 rotation_matrix[5] * point_a[2];                                                    \
+            rotated_point_a[2] = rotation_matrix[6] * point_a[0] + rotation_matrix[7] * point_a[1] +                 \
+                                 rotation_matrix[8] * point_a[2];                                                    \
+            nk_##svd_type##_t dx = rotated_point_a[0] - point_b[0];                                                  \
+            nk_##svd_type##_t dy = rotated_point_a[1] - point_b[1];                                                  \
+            nk_##svd_type##_t dz = rotated_point_a[2] - point_b[2];                                                  \
+            nk_accumulate_square_##accumulator_type##_(&sum_squared, &sum_squared_compensation,                      \
+                                                       (nk_##accumulator_type##_t)dx);                               \
+            nk_accumulate_square_##accumulator_type##_(&sum_squared, &sum_squared_compensation,                      \
+                                                       (nk_##accumulator_type##_t)dy);                               \
+            nk_accumulate_square_##accumulator_type##_(&sum_squared, &sum_squared_compensation,                      \
+                                                       (nk_##accumulator_type##_t)dz);                               \
+        }                                                                                                            \
+        *result = (nk_##result_type##_t)compute_sqrt((sum_squared + sum_squared_compensation) * inv_n);              \
+    }
+
+/*  Umeyama algorithm for optimal similarity transformation (rotation + uniform scale).
+ *  Finds the rotation matrix R and scale factor c that minimizes ‖c × R × A - B‖.
+ *  Reference: S. Umeyama, "Least-squares estimation of transformation parameters
+ *  between two point patterns", IEEE TPAMI 1991.
+ */
+#define nk_define_umeyama_(input_type, accumulator_type, output_type, result_type, svd_type, load_and_convert,        \
+                           compute_sqrt)                                                                              \
+    NK_PUBLIC void nk_umeyama_##input_type##_serial(nk_##input_type##_t const *a, nk_##input_type##_t const *b,       \
+                                                    nk_size_t n, nk_##output_type##_t *a_centroid,                    \
+                                                    nk_##output_type##_t *b_centroid, nk_##output_type##_t *rotation, \
+                                                    nk_##output_type##_t *scale, nk_##result_type##_t *result) {      \
+        /* Step 1: Compute centroids */                                                                               \
+        nk_##accumulator_type##_t sum_a_x = 0, sum_a_y = 0, sum_a_z = 0;                                              \
+        nk_##accumulator_type##_t sum_b_x = 0, sum_b_y = 0, sum_b_z = 0;                                              \
+        nk_##accumulator_type##_t sum_a_x_compensation = 0, sum_a_y_compensation = 0, sum_a_z_compensation = 0;       \
+        nk_##accumulator_type##_t sum_b_x_compensation = 0, sum_b_y_compensation = 0, sum_b_z_compensation = 0;       \
+        nk_##accumulator_type##_t val_a_x, val_a_y, val_a_z, val_b_x, val_b_y, val_b_z;                               \
+        for (nk_size_t i = 0; i < n; ++i) {                                                                           \
+            load_and_convert(a + i * 3 + 0, &val_a_x), load_and_convert(a + i * 3 + 1, &val_a_y);                     \
+            load_and_convert(a + i * 3 + 2, &val_a_z), load_and_convert(b + i * 3 + 0, &val_b_x);                     \
+            load_and_convert(b + i * 3 + 1, &val_b_y), load_and_convert(b + i * 3 + 2, &val_b_z);                     \
+            nk_accumulate_sum_##accumulator_type##_(&sum_a_x, &sum_a_x_compensation, val_a_x);                        \
+            nk_accumulate_sum_##accumulator_type##_(&sum_a_y, &sum_a_y_compensation, val_a_y);                        \
+            nk_accumulate_sum_##accumulator_type##_(&sum_a_z, &sum_a_z_compensation, val_a_z);                        \
+            nk_accumulate_sum_##accumulator_type##_(&sum_b_x, &sum_b_x_compensation, val_b_x);                        \
+            nk_accumulate_sum_##accumulator_type##_(&sum_b_y, &sum_b_y_compensation, val_b_y);                        \
+            nk_accumulate_sum_##accumulator_type##_(&sum_b_z, &sum_b_z_compensation, val_b_z);                        \
+        }                                                                                                             \
+        nk_##accumulator_type##_t inv_n = (nk_##accumulator_type##_t)1.0 / (nk_##accumulator_type##_t)n;              \
+        nk_##accumulator_type##_t centroid_a_x = (sum_a_x + sum_a_x_compensation) * inv_n;                            \
+        nk_##accumulator_type##_t centroid_a_y = (sum_a_y + sum_a_y_compensation) * inv_n;                            \
+        nk_##accumulator_type##_t centroid_a_z = (sum_a_z + sum_a_z_compensation) * inv_n;                            \
+        nk_##accumulator_type##_t centroid_b_x = (sum_b_x + sum_b_x_compensation) * inv_n;                            \
+        nk_##accumulator_type##_t centroid_b_y = (sum_b_y + sum_b_y_compensation) * inv_n;                            \
+        nk_##accumulator_type##_t centroid_b_z = (sum_b_z + sum_b_z_compensation) * inv_n;                            \
+        if (a_centroid)                                                                                               \
+            a_centroid[0] = (nk_##output_type##_t)centroid_a_x, a_centroid[1] = (nk_##output_type##_t)centroid_a_y,   \
+            a_centroid[2] = (nk_##output_type##_t)centroid_a_z;                                                       \
+        if (b_centroid)                                                                                               \
+            b_centroid[0] = (nk_##output_type##_t)centroid_b_x, b_centroid[1] = (nk_##output_type##_t)centroid_b_y,   \
+            b_centroid[2] = (nk_##output_type##_t)centroid_b_z;                                                       \
+        /* Step 2: Build covariance matrix H and compute variance of A */                                             \
+        nk_##accumulator_type##_t h[9] = {0, 0, 0, 0, 0, 0, 0, 0, 0};                                                 \
+        nk_##accumulator_type##_t h_compensation[9] = {0, 0, 0, 0, 0, 0, 0, 0, 0};                                    \
+        nk_##accumulator_type##_t variance_a = 0, variance_a_compensation = 0;                                        \
+        for (nk_size_t i = 0; i < n; ++i) {                                                                           \
+            load_and_convert(a + i * 3 + 0, &val_a_x), load_and_convert(b + i * 3 + 0, &val_b_x);                     \
+            load_and_convert(a + i * 3 + 1, &val_a_y), load_and_convert(b + i * 3 + 1, &val_b_y);                     \
+            load_and_convert(a + i * 3 + 2, &val_a_z), load_and_convert(b + i * 3 + 2, &val_b_z);                     \
+            val_a_x -= centroid_a_x, val_a_y -= centroid_a_y, val_a_z -= centroid_a_z;                                \
+            val_b_x -= centroid_b_x, val_b_y -= centroid_b_y, val_b_z -= centroid_b_z;                                \
+            nk_accumulate_square_##accumulator_type##_(&variance_a, &variance_a_compensation, val_a_x);               \
+            nk_accumulate_square_##accumulator_type##_(&variance_a, &variance_a_compensation, val_a_y);               \
+            nk_accumulate_square_##accumulator_type##_(&variance_a, &variance_a_compensation, val_a_z);               \
+            nk_accumulate_product_##accumulator_type##_(&h[0], &h_compensation[0], val_a_x, val_b_x);                 \
+            nk_accumulate_product_##accumulator_type##_(&h[1], &h_compensation[1], val_a_x, val_b_y);                 \
+            nk_accumulate_product_##accumulator_type##_(&h[2], &h_compensation[2], val_a_x, val_b_z);                 \
+            nk_accumulate_product_##accumulator_type##_(&h[3], &h_compensation[3], val_a_y, val_b_x);                 \
+            nk_accumulate_product_##accumulator_type##_(&h[4], &h_compensation[4], val_a_y, val_b_y);                 \
+            nk_accumulate_product_##accumulator_type##_(&h[5], &h_compensation[5], val_a_y, val_b_z);                 \
+            nk_accumulate_product_##accumulator_type##_(&h[6], &h_compensation[6], val_a_z, val_b_x);                 \
+            nk_accumulate_product_##accumulator_type##_(&h[7], &h_compensation[7], val_a_z, val_b_y);                 \
+            nk_accumulate_product_##accumulator_type##_(&h[8], &h_compensation[8], val_a_z, val_b_z);                 \
+        }                                                                                                             \
+        variance_a = (variance_a + variance_a_compensation) * inv_n;                                                  \
+        /* Convert to svd_type for SVD */                                                                             \
+        nk_##svd_type##_t cross_covariance[9];                                                                        \
+        for (int j = 0; j < 9; ++j) cross_covariance[j] = (nk_##svd_type##_t)(h[j] + h_compensation[j]);              \
+        /* Step 3: SVD of H = U * S * Vᵀ */                                                                           \
+        nk_##svd_type##_t svd_u[9], svd_s[9], svd_v[9];                                                               \
+        nk_svd3x3_##svd_type##_(cross_covariance, svd_u, svd_s, svd_v);                                               \
+        /* Step 4: R = V * Uᵀ */                                                                                      \
+        nk_##svd_type##_t rotation_matrix[9];                                                                         \
+        nk_rotation_from_svd_##svd_type##_serial_(svd_u, svd_v, rotation_matrix);                                     \
+        /* Handle reflection and compute scale: c = trace(D × S) / variance(a) */                                     \
+        /* D = diag(1, 1, det(R)), svd_s contains proper positive singular values on diagonal */                      \
+        nk_##svd_type##_t rotation_det = nk_det3x3_##svd_type##_(rotation_matrix);                                    \
+        nk_##svd_type##_t sign_det = rotation_det < 0 ? (nk_##svd_type##_t) - 1.0 : (nk_##svd_type##_t)1.0;           \
+        nk_##svd_type##_t trace_scaled_s = svd_s[0] + svd_s[4] + sign_det * svd_s[8];                                 \
+        nk_##accumulator_type##_t scale_factor = (nk_##accumulator_type##_t)trace_scaled_s /                          \
+                                                 ((nk_##accumulator_type##_t)n * variance_a);                         \
+        if (scale) *scale = (nk_##output_type##_t)scale_factor;                                                       \
+        if (rotation_det < 0) {                                                                                       \
+            svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];                                         \
+            nk_rotation_from_svd_##svd_type##_serial_(svd_u, svd_v, rotation_matrix);                                 \
+        }                                                                                                             \
+        /* Output rotation matrix */                                                                                  \
+        if (rotation)                                                                                                 \
+            for (int j = 0; j < 9; ++j) rotation[j] = (nk_##output_type##_t)rotation_matrix[j];                       \
+        /* Step 5: Compute RMSD after similarity transform: ‖c × R × a - b‖ */                                        \
+        nk_##accumulator_type##_t sum_squared = 0, sum_squared_compensation = 0;                                      \
+        for (nk_size_t i = 0; i < n; ++i) {                                                                           \
+            nk_##svd_type##_t point_a[3], point_b[3], rotated_point_a[3];                                             \
+            load_and_convert(a + i * 3 + 0, &val_a_x), load_and_convert(a + i * 3 + 1, &val_a_y);                     \
+            load_and_convert(a + i * 3 + 2, &val_a_z), load_and_convert(b + i * 3 + 0, &val_b_x);                     \
+            load_and_convert(b + i * 3 + 1, &val_b_y), load_and_convert(b + i * 3 + 2, &val_b_z);                     \
+            point_a[0] = (nk_##svd_type##_t)(val_a_x - centroid_a_x),                                                 \
+            point_a[1] = (nk_##svd_type##_t)(val_a_y - centroid_a_y),                                                 \
+            point_a[2] = (nk_##svd_type##_t)(val_a_z - centroid_a_z);                                                 \
+            point_b[0] = (nk_##svd_type##_t)(val_b_x - centroid_b_x),                                                 \
+            point_b[1] = (nk_##svd_type##_t)(val_b_y - centroid_b_y),                                                 \
+            point_b[2] = (nk_##svd_type##_t)(val_b_z - centroid_b_z);                                                 \
+            rotated_point_a[0] = (nk_##svd_type##_t)scale_factor *                                                    \
+                                 (rotation_matrix[0] * point_a[0] + rotation_matrix[1] * point_a[1] +                 \
+                                  rotation_matrix[2] * point_a[2]);                                                   \
+            rotated_point_a[1] = (nk_##svd_type##_t)scale_factor *                                                    \
+                                 (rotation_matrix[3] * point_a[0] + rotation_matrix[4] * point_a[1] +                 \
+                                  rotation_matrix[5] * point_a[2]);                                                   \
+            rotated_point_a[2] = (nk_##svd_type##_t)scale_factor *                                                    \
+                                 (rotation_matrix[6] * point_a[0] + rotation_matrix[7] * point_a[1] +                 \
+                                  rotation_matrix[8] * point_a[2]);                                                   \
+            nk_##svd_type##_t dx = rotated_point_a[0] - point_b[0];                                                   \
+            nk_##svd_type##_t dy = rotated_point_a[1] - point_b[1];                                                   \
+            nk_##svd_type##_t dz = rotated_point_a[2] - point_b[2];                                                   \
+            nk_accumulate_square_##accumulator_type##_(&sum_squared, &sum_squared_compensation,                       \
+                                                       (nk_##accumulator_type##_t)dx);                                \
+            nk_accumulate_square_##accumulator_type##_(&sum_squared, &sum_squared_compensation,                       \
+                                                       (nk_##accumulator_type##_t)dy);                                \
+            nk_accumulate_square_##accumulator_type##_(&sum_squared, &sum_squared_compensation,                       \
+                                                       (nk_##accumulator_type##_t)dz);                                \
+        }                                                                                                             \
+        *result = (nk_##result_type##_t)compute_sqrt((sum_squared + sum_squared_compensation) * inv_n);               \
+    }
+
+nk_define_rmsd_(f64, f64, f64, f64, nk_assign_from_to_, nk_f64_sqrt_serial)         // nk_rmsd_f64_serial
+nk_define_kabsch_(f64, f64, f64, f64, f64, nk_assign_from_to_, nk_f64_sqrt_serial)  // nk_kabsch_f64_serial
+nk_define_umeyama_(f64, f64, f64, f64, f64, nk_assign_from_to_, nk_f64_sqrt_serial) // nk_umeyama_f64_serial
+
+nk_define_rmsd_(f32, f64, f32, f64, nk_assign_from_to_, nk_f64_sqrt_serial)         // nk_rmsd_f32_serial
+nk_define_kabsch_(f32, f64, f32, f64, f64, nk_assign_from_to_, nk_f64_sqrt_serial)  // nk_kabsch_f32_serial
+nk_define_umeyama_(f32, f64, f32, f64, f64, nk_assign_from_to_, nk_f64_sqrt_serial) // nk_umeyama_f32_serial
+
+nk_define_rmsd_(f16, f32, f32, f32, nk_f16_to_f32_serial, nk_f32_sqrt_serial)         // nk_rmsd_f16_serial
+nk_define_kabsch_(f16, f32, f32, f32, f32, nk_f16_to_f32_serial, nk_f32_sqrt_serial)  // nk_kabsch_f16_serial
+nk_define_umeyama_(f16, f32, f32, f32, f32, nk_f16_to_f32_serial, nk_f32_sqrt_serial) // nk_umeyama_f16_serial
+
+nk_define_rmsd_(bf16, f32, f32, f32, nk_bf16_to_f32_serial, nk_f32_sqrt_serial)         // nk_rmsd_bf16_serial
+nk_define_kabsch_(bf16, f32, f32, f32, f32, nk_bf16_to_f32_serial, nk_f32_sqrt_serial)  // nk_kabsch_bf16_serial
+nk_define_umeyama_(bf16, f32, f32, f32, f32, nk_bf16_to_f32_serial, nk_f32_sqrt_serial) // nk_umeyama_bf16_serial
+
+#undef NK_F32_SVD_GAMMA_
+#undef NK_F32_SVD_CSTAR_
+#undef NK_F32_SVD_SSTAR_
+#undef NK_F32_SVD_EPSILON_
+#undef NK_F64_SVD_GAMMA_
+#undef NK_F64_SVD_CSTAR_
+#undef NK_F64_SVD_SSTAR_
+#undef NK_F64_SVD_EPSILON_
+#undef nk_define_cond_swap_
+#undef nk_define_conditional_negating_swap_
+#undef nk_define_approximate_givens_quaternion_
+#undef nk_define_jacobi_conjugation_
+#undef nk_define_quaternion_to_mat3x3_
+#undef nk_define_jacobi_eigenanalysis_
+#undef nk_define_qr_givens_quaternion_
+#undef nk_define_sort_singular_values_
+#undef nk_define_qr_decomposition_
+#undef nk_define_svd3x3_
+#undef nk_define_det3x3_
+#undef nk_define_rmsd_
+#undef nk_define_kabsch_
+#undef nk_define_umeyama_
+
+#if defined(NDEBUG)
+#if defined(_MSC_VER)
+#pragma optimize("", on)
+#elif defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_MESH_SERIAL_H