usearch 2.21.3 → 2.25.0

package/numkong/include/numkong/mesh/neon.h

@@ -0,0 +1,1821 @@
+/**
+ *  @brief SIMD-accelerated Point Cloud Alignment for NEON.
+ *  @file include/numkong/mesh/neon.h
+ *  @author Ash Vardanian
+ *  @date December 27, 2025
+ *
+ *  @sa include/numkong/mesh.h
+ *
+ *  @section neon_mesh_instructions Key NEON Mesh Instructions
+ *
+ *  Point cloud operations use these ARM NEON instructions:
+ *
+ *      Intrinsic    Instruction                    A76       M5
+ *      vfmaq_f32    FMLA (V.4S, V.4S, V.4S)        4cy @ 2p  3cy @ 4p
+ *      vmulq_n_f32  FMUL (V.4S, V.4S, V.S[0])      3cy @ 2p  3cy @ 4p
+ *      vsubq_f32    FSUB (V.4S, V.4S, V.4S)        2cy @ 2p  2cy @ 4p
+ *      vaddvq_f32   FADDP+FADDP (reduce)           5cy @ 1p  8cy @ 1p
+ *      vld3q_f32    LD3 ({Vt.4S, Vt2.4S, Vt3.4S})  4cy @ 1p  4cy @ 1p
+ *
+ *  LD3 provides hardware stride-3 deinterleaving for XYZ point data. The 6cy latency and
+ *  1/cy throughput make it the memory bottleneck regardless of core microarchitecture.
+ *
+ *  FMA throughput doubles on 4-pipe cores (Apple M4+, Graviton3+, Oryon). Using 2x loop
+ *  unrolling with independent accumulators hides FMA latency and saturates 2 FP pipes on
+ *  A76-class cores; 4x unrolling may further benefit 4-pipe cores.
+ */
+#ifndef NK_MESH_NEON_H
+#define NK_MESH_NEON_H
+
+#if NK_TARGET_ARM64_
+#if NK_TARGET_NEON
+
+#include "numkong/types.h"
+#include "numkong/dot/neon.h"
+#include "numkong/mesh/serial.h"
+#include "numkong/spatial/neon.h" // `nk_f32_sqrt_neon`, `nk_f64_sqrt_neon`
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((target("arch=armv8-a+simd"))), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("arch=armv8-a+simd")
+#endif
+
+NK_INTERNAL void nk_deinterleave_f32x4_neon_(nk_f32_t const *ptr, float32x4_t *x_out, float32x4_t *y_out,
+                                             float32x4_t *z_out) {
+    // Deinterleave 12 floats (4 xyz triplets) into separate x, y, z vectors.
+    // Uses NEON vld3q for efficient stride-3 deinterleaving.
+    //
+    // Input: 12 contiguous floats [x0,y0,z0, x1,y1,z1, x2,y2,z2, x3,y3,z3]
+    // Output: x[4], y[4], z[4] vectors
+    float32x4x3_t xyz_f32x4x3 = vld3q_f32(ptr);
+    *x_out = xyz_f32x4x3.val[0];
+    *y_out = xyz_f32x4x3.val[1];
+    *z_out = xyz_f32x4x3.val[2];
+}
+
+NK_INTERNAL void nk_deinterleave_f64x2_neon_(nk_f64_t const *ptr, float64x2_t *x_out, float64x2_t *y_out,
+                                             float64x2_t *z_out) {
+    // Deinterleave 6 f64 values (2 xyz triplets) into separate x, y, z vectors.
+    //
+    // Input: 6 contiguous f64 [x0,y0,z0, x1,y1,z1]
+    // Output: x[2], y[2], z[2] vectors
+    // NEON doesn't have vld3q_f64, so we use vcombine to avoid stack round-trips
+    // Load 2 xyz triplets: [x0,y0,z0, x1,y1,z1]
+    *x_out = vcombine_f64(vld1_f64(&ptr[0]), vld1_f64(&ptr[3]));
+    *y_out = vcombine_f64(vld1_f64(&ptr[1]), vld1_f64(&ptr[4]));
+    *z_out = vcombine_f64(vld1_f64(&ptr[2]), vld1_f64(&ptr[5]));
+}
+
+NK_INTERNAL nk_f64_t nk_reduce_stable_f64x2_neon_(float64x2_t values_f64x2) {
+    nk_b128_vec_t values;
+    values.f64x2 = values_f64x2;
+    nk_f64_t sum = 0.0, compensation = 0.0;
+    nk_accumulate_sum_f64_(&sum, &compensation, values.f64s[0]);
+    nk_accumulate_sum_f64_(&sum, &compensation, values.f64s[1]);
+    return sum + compensation;
+}
+
+NK_INTERNAL void nk_rotation_from_svd_f64_neon_(nk_f64_t const *svd_u, nk_f64_t const *svd_v, nk_f64_t *rotation) {
+    nk_rotation_from_svd_f64_serial_(svd_u, svd_v, rotation);
+}
+
+NK_INTERNAL void nk_accumulate_square_f64x2_neon_(float64x2_t *sum_f64x2, float64x2_t *compensation_f64x2,
+                                                  float64x2_t values_f64x2) {
+    float64x2_t product_f64x2 = vmulq_f64(values_f64x2, values_f64x2);
+    float64x2_t product_error_f64x2 = vfmaq_f64(vnegq_f64(product_f64x2), values_f64x2, values_f64x2);
+    float64x2_t tentative_sum_f64x2 = vaddq_f64(*sum_f64x2, product_f64x2);
+    float64x2_t virtual_addend_f64x2 = vsubq_f64(tentative_sum_f64x2, *sum_f64x2);
+    float64x2_t sum_error_f64x2 = vaddq_f64(vsubq_f64(*sum_f64x2, vsubq_f64(tentative_sum_f64x2, virtual_addend_f64x2)),
+                                            vsubq_f64(product_f64x2, virtual_addend_f64x2));
+    *sum_f64x2 = tentative_sum_f64x2;
+    *compensation_f64x2 = vaddq_f64(*compensation_f64x2, vaddq_f64(sum_error_f64x2, product_error_f64x2));
+}
+
+NK_INTERNAL nk_f64_t nk_transformed_ssd_f32_neon_( //
+    nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t const *r, nk_f64_t scale, nk_f64_t centroid_a_x,
+    nk_f64_t centroid_a_y, nk_f64_t centroid_a_z, nk_f64_t centroid_b_x, nk_f64_t centroid_b_y, nk_f64_t centroid_b_z) {
+    float64x2_t scaled_rotation_x_x_f64x2 = vdupq_n_f64(scale * r[0]);
+    float64x2_t scaled_rotation_x_y_f64x2 = vdupq_n_f64(scale * r[1]);
+    float64x2_t scaled_rotation_x_z_f64x2 = vdupq_n_f64(scale * r[2]);
+    float64x2_t scaled_rotation_y_x_f64x2 = vdupq_n_f64(scale * r[3]);
+    float64x2_t scaled_rotation_y_y_f64x2 = vdupq_n_f64(scale * r[4]);
+    float64x2_t scaled_rotation_y_z_f64x2 = vdupq_n_f64(scale * r[5]);
+    float64x2_t scaled_rotation_z_x_f64x2 = vdupq_n_f64(scale * r[6]);
+    float64x2_t scaled_rotation_z_y_f64x2 = vdupq_n_f64(scale * r[7]);
+    float64x2_t scaled_rotation_z_z_f64x2 = vdupq_n_f64(scale * r[8]);
+    float64x2_t centroid_a_x_f64x2 = vdupq_n_f64(centroid_a_x), centroid_a_y_f64x2 = vdupq_n_f64(centroid_a_y);
+    float64x2_t centroid_a_z_f64x2 = vdupq_n_f64(centroid_a_z), centroid_b_x_f64x2 = vdupq_n_f64(centroid_b_x);
+    float64x2_t centroid_b_y_f64x2 = vdupq_n_f64(centroid_b_y), centroid_b_z_f64x2 = vdupq_n_f64(centroid_b_z);
+    float64x2_t sum_squared_low_f64x2 = vdupq_n_f64(0.0), sum_squared_high_f64x2 = vdupq_n_f64(0.0);
+    nk_size_t index = 0;
+
+    for (; index + 4 <= n; index += 4) {
+        float32x4_t a_x_f32x4, a_y_f32x4, a_z_f32x4, b_x_f32x4, b_y_f32x4, b_z_f32x4;
+        nk_deinterleave_f32x4_neon_(a + index * 3, &a_x_f32x4, &a_y_f32x4, &a_z_f32x4),
+            nk_deinterleave_f32x4_neon_(b + index * 3, &b_x_f32x4, &b_y_f32x4, &b_z_f32x4);
+
+        float64x2_t centered_a_x_low_f64x2 = vsubq_f64(vcvt_f64_f32(vget_low_f32(a_x_f32x4)), centroid_a_x_f64x2);
+        float64x2_t centered_a_x_high_f64x2 = vsubq_f64(vcvt_high_f64_f32(a_x_f32x4), centroid_a_x_f64x2);
+        float64x2_t centered_a_y_low_f64x2 = vsubq_f64(vcvt_f64_f32(vget_low_f32(a_y_f32x4)), centroid_a_y_f64x2);
+        float64x2_t centered_a_y_high_f64x2 = vsubq_f64(vcvt_high_f64_f32(a_y_f32x4), centroid_a_y_f64x2);
+        float64x2_t centered_a_z_low_f64x2 = vsubq_f64(vcvt_f64_f32(vget_low_f32(a_z_f32x4)), centroid_a_z_f64x2);
+        float64x2_t centered_a_z_high_f64x2 = vsubq_f64(vcvt_high_f64_f32(a_z_f32x4), centroid_a_z_f64x2);
+        float64x2_t centered_b_x_low_f64x2 = vsubq_f64(vcvt_f64_f32(vget_low_f32(b_x_f32x4)), centroid_b_x_f64x2);
+        float64x2_t centered_b_x_high_f64x2 = vsubq_f64(vcvt_high_f64_f32(b_x_f32x4), centroid_b_x_f64x2);
+        float64x2_t centered_b_y_low_f64x2 = vsubq_f64(vcvt_f64_f32(vget_low_f32(b_y_f32x4)), centroid_b_y_f64x2);
+        float64x2_t centered_b_y_high_f64x2 = vsubq_f64(vcvt_high_f64_f32(b_y_f32x4), centroid_b_y_f64x2);
+        float64x2_t centered_b_z_low_f64x2 = vsubq_f64(vcvt_f64_f32(vget_low_f32(b_z_f32x4)), centroid_b_z_f64x2);
+        float64x2_t centered_b_z_high_f64x2 = vsubq_f64(vcvt_high_f64_f32(b_z_f32x4), centroid_b_z_f64x2);
+
+        float64x2_t rotated_a_x_low_f64x2 = vfmaq_f64(
+            vfmaq_f64(vmulq_f64(scaled_rotation_x_x_f64x2, centered_a_x_low_f64x2), scaled_rotation_x_y_f64x2,
+                      centered_a_y_low_f64x2),
+            scaled_rotation_x_z_f64x2, centered_a_z_low_f64x2);
+        float64x2_t rotated_a_x_high_f64x2 = vfmaq_f64(
+            vfmaq_f64(vmulq_f64(scaled_rotation_x_x_f64x2, centered_a_x_high_f64x2), scaled_rotation_x_y_f64x2,
+                      centered_a_y_high_f64x2),
+            scaled_rotation_x_z_f64x2, centered_a_z_high_f64x2);
+        float64x2_t rotated_a_y_low_f64x2 = vfmaq_f64(
+            vfmaq_f64(vmulq_f64(scaled_rotation_y_x_f64x2, centered_a_x_low_f64x2), scaled_rotation_y_y_f64x2,
+                      centered_a_y_low_f64x2),
+            scaled_rotation_y_z_f64x2, centered_a_z_low_f64x2);
+        float64x2_t rotated_a_y_high_f64x2 = vfmaq_f64(
+            vfmaq_f64(vmulq_f64(scaled_rotation_y_x_f64x2, centered_a_x_high_f64x2), scaled_rotation_y_y_f64x2,
+                      centered_a_y_high_f64x2),
+            scaled_rotation_y_z_f64x2, centered_a_z_high_f64x2);
+        float64x2_t rotated_a_z_low_f64x2 = vfmaq_f64(
+            vfmaq_f64(vmulq_f64(scaled_rotation_z_x_f64x2, centered_a_x_low_f64x2), scaled_rotation_z_y_f64x2,
+                      centered_a_y_low_f64x2),
+            scaled_rotation_z_z_f64x2, centered_a_z_low_f64x2);
+        float64x2_t rotated_a_z_high_f64x2 = vfmaq_f64(
+            vfmaq_f64(vmulq_f64(scaled_rotation_z_x_f64x2, centered_a_x_high_f64x2), scaled_rotation_z_y_f64x2,
+                      centered_a_y_high_f64x2),
+            scaled_rotation_z_z_f64x2, centered_a_z_high_f64x2);
+
+        float64x2_t delta_x_low_f64x2 = vsubq_f64(rotated_a_x_low_f64x2, centered_b_x_low_f64x2);
+        float64x2_t delta_x_high_f64x2 = vsubq_f64(rotated_a_x_high_f64x2, centered_b_x_high_f64x2);
+        float64x2_t delta_y_low_f64x2 = vsubq_f64(rotated_a_y_low_f64x2, centered_b_y_low_f64x2);
+        float64x2_t delta_y_high_f64x2 = vsubq_f64(rotated_a_y_high_f64x2, centered_b_y_high_f64x2);
+        float64x2_t delta_z_low_f64x2 = vsubq_f64(rotated_a_z_low_f64x2, centered_b_z_low_f64x2);
+        float64x2_t delta_z_high_f64x2 = vsubq_f64(rotated_a_z_high_f64x2, centered_b_z_high_f64x2);
+
+        sum_squared_low_f64x2 = vfmaq_f64(sum_squared_low_f64x2, delta_x_low_f64x2, delta_x_low_f64x2),
+        sum_squared_high_f64x2 = vfmaq_f64(sum_squared_high_f64x2, delta_x_high_f64x2, delta_x_high_f64x2);
+        sum_squared_low_f64x2 = vfmaq_f64(sum_squared_low_f64x2, delta_y_low_f64x2, delta_y_low_f64x2),
+        sum_squared_high_f64x2 = vfmaq_f64(sum_squared_high_f64x2, delta_y_high_f64x2, delta_y_high_f64x2);
+        sum_squared_low_f64x2 = vfmaq_f64(sum_squared_low_f64x2, delta_z_low_f64x2, delta_z_low_f64x2),
+        sum_squared_high_f64x2 = vfmaq_f64(sum_squared_high_f64x2, delta_z_high_f64x2, delta_z_high_f64x2);
+    }
+
+    nk_f64_t sum_squared = vaddvq_f64(vaddq_f64(sum_squared_low_f64x2, sum_squared_high_f64x2));
+    for (; index < n; ++index) {
+        nk_f64_t centered_a_x = (nk_f64_t)a[index * 3 + 0] - centroid_a_x,
+                 centered_a_y = (nk_f64_t)a[index * 3 + 1] - centroid_a_y,
+                 centered_a_z = (nk_f64_t)a[index * 3 + 2] - centroid_a_z;
+        nk_f64_t centered_b_x = (nk_f64_t)b[index * 3 + 0] - centroid_b_x,
+                 centered_b_y = (nk_f64_t)b[index * 3 + 1] - centroid_b_y,
+                 centered_b_z = (nk_f64_t)b[index * 3 + 2] - centroid_b_z;
+        nk_f64_t rotated_a_x = scale * (r[0] * centered_a_x + r[1] * centered_a_y + r[2] * centered_a_z),
+                 rotated_a_y = scale * (r[3] * centered_a_x + r[4] * centered_a_y + r[5] * centered_a_z),
+                 rotated_a_z = scale * (r[6] * centered_a_x + r[7] * centered_a_y + r[8] * centered_a_z);
+        nk_f64_t delta_x = rotated_a_x - centered_b_x, delta_y = rotated_a_y - centered_b_y,
+                 delta_z = rotated_a_z - centered_b_z;
+        sum_squared += delta_x * delta_x + delta_y * delta_y + delta_z * delta_z;
+    }
+
+    return sum_squared;
+}
+
+/*  Compute sum of squared distances for f64 after applying rotation (and optional scale).
+ *
+ *  Optimization: 2x loop unrolling with multiple accumulators hides FMA latency (3-7 cycles).
+ */
+NK_INTERNAL nk_f64_t nk_transformed_ssd_f64_neon_(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t const *r,
+                                                  nk_f64_t scale, nk_f64_t centroid_a_x, nk_f64_t centroid_a_y,
+                                                  nk_f64_t centroid_a_z, nk_f64_t centroid_b_x, nk_f64_t centroid_b_y,
+                                                  nk_f64_t centroid_b_z) {
+    // Broadcast scaled rotation matrix elements
+    float64x2_t scaled_rotation_x_x_f64x2 = vdupq_n_f64(scale * r[0]);
+    float64x2_t scaled_rotation_x_y_f64x2 = vdupq_n_f64(scale * r[1]);
+    float64x2_t scaled_rotation_x_z_f64x2 = vdupq_n_f64(scale * r[2]);
+    float64x2_t scaled_rotation_y_x_f64x2 = vdupq_n_f64(scale * r[3]);
+    float64x2_t scaled_rotation_y_y_f64x2 = vdupq_n_f64(scale * r[4]);
+    float64x2_t scaled_rotation_y_z_f64x2 = vdupq_n_f64(scale * r[5]);
+    float64x2_t scaled_rotation_z_x_f64x2 = vdupq_n_f64(scale * r[6]);
+    float64x2_t scaled_rotation_z_y_f64x2 = vdupq_n_f64(scale * r[7]);
+    float64x2_t scaled_rotation_z_z_f64x2 = vdupq_n_f64(scale * r[8]);
+
+    // Broadcast centroids
+    float64x2_t centroid_a_x_f64x2 = vdupq_n_f64(centroid_a_x);
+    float64x2_t centroid_a_y_f64x2 = vdupq_n_f64(centroid_a_y);
+    float64x2_t centroid_a_z_f64x2 = vdupq_n_f64(centroid_a_z);
+    float64x2_t centroid_b_x_f64x2 = vdupq_n_f64(centroid_b_x);
+    float64x2_t centroid_b_y_f64x2 = vdupq_n_f64(centroid_b_y);
+    float64x2_t centroid_b_z_f64x2 = vdupq_n_f64(centroid_b_z);
+
+    // Two independent accumulators to hide FMA latency
+    float64x2_t sum_squared_a_f64x2 = vdupq_n_f64(0), sum_squared_a_compensation_f64x2 = vdupq_n_f64(0);
+    float64x2_t sum_squared_b_f64x2 = vdupq_n_f64(0), sum_squared_b_compensation_f64x2 = vdupq_n_f64(0);
+    nk_size_t j = 0;
+
+    // Main loop: process 4 points per iteration (2x unrolled, 2 points per batch)
+    for (; j + 4 <= n; j += 4) {
+        // First batch of 2 points
+        float64x2_t a1_x_f64x2, a1_y_f64x2, a1_z_f64x2, b1_x_f64x2, b1_y_f64x2, b1_z_f64x2;
+        nk_deinterleave_f64x2_neon_(a + j * 3, &a1_x_f64x2, &a1_y_f64x2, &a1_z_f64x2);
+        nk_deinterleave_f64x2_neon_(b + j * 3, &b1_x_f64x2, &b1_y_f64x2, &b1_z_f64x2);
+
+        // Second batch of 2 points
+        float64x2_t a2_x_f64x2, a2_y_f64x2, a2_z_f64x2, b2_x_f64x2, b2_y_f64x2, b2_z_f64x2;
+        nk_deinterleave_f64x2_neon_(a + (j + 2) * 3, &a2_x_f64x2, &a2_y_f64x2, &a2_z_f64x2);
+        nk_deinterleave_f64x2_neon_(b + (j + 2) * 3, &b2_x_f64x2, &b2_y_f64x2, &b2_z_f64x2);
+
+        // Center first batch
+        float64x2_t centered_a1_x_f64x2 = vsubq_f64(a1_x_f64x2, centroid_a_x_f64x2);
+        float64x2_t centered_a1_y_f64x2 = vsubq_f64(a1_y_f64x2, centroid_a_y_f64x2);
+        float64x2_t centered_a1_z_f64x2 = vsubq_f64(a1_z_f64x2, centroid_a_z_f64x2);
+        float64x2_t centered_b1_x_f64x2 = vsubq_f64(b1_x_f64x2, centroid_b_x_f64x2);
+        float64x2_t centered_b1_y_f64x2 = vsubq_f64(b1_y_f64x2, centroid_b_y_f64x2);
+        float64x2_t centered_b1_z_f64x2 = vsubq_f64(b1_z_f64x2, centroid_b_z_f64x2);
+
+        // Center second batch
+        float64x2_t centered_a2_x_f64x2 = vsubq_f64(a2_x_f64x2, centroid_a_x_f64x2);
+        float64x2_t centered_a2_y_f64x2 = vsubq_f64(a2_y_f64x2, centroid_a_y_f64x2);
+        float64x2_t centered_a2_z_f64x2 = vsubq_f64(a2_z_f64x2, centroid_a_z_f64x2);
+        float64x2_t centered_b2_x_f64x2 = vsubq_f64(b2_x_f64x2, centroid_b_x_f64x2);
+        float64x2_t centered_b2_y_f64x2 = vsubq_f64(b2_y_f64x2, centroid_b_y_f64x2);
+        float64x2_t centered_b2_z_f64x2 = vsubq_f64(b2_z_f64x2, centroid_b_z_f64x2);
+
+        // Rotate and scale first batch
+        float64x2_t rotated_a1_x_f64x2 = vfmaq_f64(vfmaq_f64(vmulq_f64(scaled_rotation_x_x_f64x2, centered_a1_x_f64x2),
+                                                             scaled_rotation_x_y_f64x2, centered_a1_y_f64x2),
+                                                   scaled_rotation_x_z_f64x2, centered_a1_z_f64x2);
+        float64x2_t rotated_a1_y_f64x2 = vfmaq_f64(vfmaq_f64(vmulq_f64(scaled_rotation_y_x_f64x2, centered_a1_x_f64x2),
+                                                             scaled_rotation_y_y_f64x2, centered_a1_y_f64x2),
+                                                   scaled_rotation_y_z_f64x2, centered_a1_z_f64x2);
+        float64x2_t rotated_a1_z_f64x2 = vfmaq_f64(vfmaq_f64(vmulq_f64(scaled_rotation_z_x_f64x2, centered_a1_x_f64x2),
+                                                             scaled_rotation_z_y_f64x2, centered_a1_y_f64x2),
+                                                   scaled_rotation_z_z_f64x2, centered_a1_z_f64x2);
+
+        // Rotate and scale second batch
+        float64x2_t rotated_a2_x_f64x2 = vfmaq_f64(vfmaq_f64(vmulq_f64(scaled_rotation_x_x_f64x2, centered_a2_x_f64x2),
+                                                             scaled_rotation_x_y_f64x2, centered_a2_y_f64x2),
+                                                   scaled_rotation_x_z_f64x2, centered_a2_z_f64x2);
+        float64x2_t rotated_a2_y_f64x2 = vfmaq_f64(vfmaq_f64(vmulq_f64(scaled_rotation_y_x_f64x2, centered_a2_x_f64x2),
+                                                             scaled_rotation_y_y_f64x2, centered_a2_y_f64x2),
+                                                   scaled_rotation_y_z_f64x2, centered_a2_z_f64x2);
+        float64x2_t rotated_a2_z_f64x2 = vfmaq_f64(vfmaq_f64(vmulq_f64(scaled_rotation_z_x_f64x2, centered_a2_x_f64x2),
+                                                             scaled_rotation_z_y_f64x2, centered_a2_y_f64x2),
+                                                   scaled_rotation_z_z_f64x2, centered_a2_z_f64x2);
+
+        // Deltas
+        float64x2_t delta1_x_f64x2 = vsubq_f64(rotated_a1_x_f64x2, centered_b1_x_f64x2);
+        float64x2_t delta1_y_f64x2 = vsubq_f64(rotated_a1_y_f64x2, centered_b1_y_f64x2);
+        float64x2_t delta1_z_f64x2 = vsubq_f64(rotated_a1_z_f64x2, centered_b1_z_f64x2);
+        float64x2_t delta2_x_f64x2 = vsubq_f64(rotated_a2_x_f64x2, centered_b2_x_f64x2);
+        float64x2_t delta2_y_f64x2 = vsubq_f64(rotated_a2_y_f64x2, centered_b2_y_f64x2);
+        float64x2_t delta2_z_f64x2 = vsubq_f64(rotated_a2_z_f64x2, centered_b2_z_f64x2);
+
+        // Accumulate to independent accumulators (interleaved for latency hiding)
+        nk_accumulate_square_f64x2_neon_(&sum_squared_a_f64x2, &sum_squared_a_compensation_f64x2, delta1_x_f64x2);
+        nk_accumulate_square_f64x2_neon_(&sum_squared_b_f64x2, &sum_squared_b_compensation_f64x2, delta2_x_f64x2);
+        nk_accumulate_square_f64x2_neon_(&sum_squared_a_f64x2, &sum_squared_a_compensation_f64x2, delta1_y_f64x2);
+        nk_accumulate_square_f64x2_neon_(&sum_squared_b_f64x2, &sum_squared_b_compensation_f64x2, delta2_y_f64x2);
+        nk_accumulate_square_f64x2_neon_(&sum_squared_a_f64x2, &sum_squared_a_compensation_f64x2, delta1_z_f64x2);
+        nk_accumulate_square_f64x2_neon_(&sum_squared_b_f64x2, &sum_squared_b_compensation_f64x2, delta2_z_f64x2);
+    }
+
+    // Handle remaining 2 points
+    if (j + 2 <= n) {
+        float64x2_t a_x_f64x2, a_y_f64x2, a_z_f64x2, b_x_f64x2, b_y_f64x2, b_z_f64x2;
+        nk_deinterleave_f64x2_neon_(a + j * 3, &a_x_f64x2, &a_y_f64x2, &a_z_f64x2);
+        nk_deinterleave_f64x2_neon_(b + j * 3, &b_x_f64x2, &b_y_f64x2, &b_z_f64x2);
+
+        float64x2_t centered_a_x_f64x2 = vsubq_f64(a_x_f64x2, centroid_a_x_f64x2);
+        float64x2_t centered_a_y_f64x2 = vsubq_f64(a_y_f64x2, centroid_a_y_f64x2);
+        float64x2_t centered_a_z_f64x2 = vsubq_f64(a_z_f64x2, centroid_a_z_f64x2);
+        float64x2_t centered_b_x_f64x2 = vsubq_f64(b_x_f64x2, centroid_b_x_f64x2);
+        float64x2_t centered_b_y_f64x2 = vsubq_f64(b_y_f64x2, centroid_b_y_f64x2);
+        float64x2_t centered_b_z_f64x2 = vsubq_f64(b_z_f64x2, centroid_b_z_f64x2);
+
+        float64x2_t rotated_a_x_f64x2 = vfmaq_f64(vfmaq_f64(vmulq_f64(scaled_rotation_x_x_f64x2, centered_a_x_f64x2),
+                                                            scaled_rotation_x_y_f64x2, centered_a_y_f64x2),
+                                                  scaled_rotation_x_z_f64x2, centered_a_z_f64x2);
+        float64x2_t rotated_a_y_f64x2 = vfmaq_f64(vfmaq_f64(vmulq_f64(scaled_rotation_y_x_f64x2, centered_a_x_f64x2),
+                                                            scaled_rotation_y_y_f64x2, centered_a_y_f64x2),
+                                                  scaled_rotation_y_z_f64x2, centered_a_z_f64x2);
+        float64x2_t rotated_a_z_f64x2 = vfmaq_f64(vfmaq_f64(vmulq_f64(scaled_rotation_z_x_f64x2, centered_a_x_f64x2),
+                                                            scaled_rotation_z_y_f64x2, centered_a_y_f64x2),
+                                                  scaled_rotation_z_z_f64x2, centered_a_z_f64x2);
+
+        float64x2_t delta_x_f64x2 = vsubq_f64(rotated_a_x_f64x2, centered_b_x_f64x2);
+        float64x2_t delta_y_f64x2 = vsubq_f64(rotated_a_y_f64x2, centered_b_y_f64x2);
+        float64x2_t delta_z_f64x2 = vsubq_f64(rotated_a_z_f64x2, centered_b_z_f64x2);
+
+        nk_accumulate_square_f64x2_neon_(&sum_squared_a_f64x2, &sum_squared_a_compensation_f64x2, delta_x_f64x2);
+        nk_accumulate_square_f64x2_neon_(&sum_squared_a_f64x2, &sum_squared_a_compensation_f64x2, delta_y_f64x2);
+        nk_accumulate_square_f64x2_neon_(&sum_squared_a_f64x2, &sum_squared_a_compensation_f64x2, delta_z_f64x2);
+        j += 2;
+    }
+
+    // Combine accumulators and reduce
+    float64x2_t sum_squared_f64x2 = vaddq_f64(sum_squared_a_f64x2, sum_squared_b_f64x2);
+    float64x2_t sum_squared_compensation_f64x2 = vaddq_f64(sum_squared_a_compensation_f64x2,
+                                                           sum_squared_b_compensation_f64x2);
+    nk_f64_t sum_squared = nk_dot_stable_sum_f64x2_neon_(sum_squared_f64x2, sum_squared_compensation_f64x2);
+    nk_f64_t sum_squared_compensation = 0.0;
+
+    // Scalar tail
+    for (; j < n; ++j) {
+        nk_f64_t pa_x = a[j * 3 + 0] - centroid_a_x, pa_y = a[j * 3 + 1] - centroid_a_y,
+                 pa_z = a[j * 3 + 2] - centroid_a_z;
+        nk_f64_t pb_x = b[j * 3 + 0] - centroid_b_x, pb_y = b[j * 3 + 1] - centroid_b_y,
+                 pb_z = b[j * 3 + 2] - centroid_b_z;
+
+        nk_f64_t ra_x = scale * (r[0] * pa_x + r[1] * pa_y + r[2] * pa_z),
+                 ra_y = scale * (r[3] * pa_x + r[4] * pa_y + r[5] * pa_z),
+                 ra_z = scale * (r[6] * pa_x + r[7] * pa_y + r[8] * pa_z);
+
+        nk_f64_t delta_x = ra_x - pb_x, delta_y = ra_y - pb_y, delta_z = ra_z - pb_z;
+        nk_accumulate_square_f64_(&sum_squared, &sum_squared_compensation, delta_x);
+        nk_accumulate_square_f64_(&sum_squared, &sum_squared_compensation, delta_y);
+        nk_accumulate_square_f64_(&sum_squared, &sum_squared_compensation, delta_z);
+    }
+
+    return sum_squared + sum_squared_compensation;
+}
+
+NK_PUBLIC void nk_rmsd_f32_neon(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f64_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 = 1.0f;
+
+    float64x2_t zero_f64x2 = vdupq_n_f64(0.0);
+    float64x2_t sum_squared_x_low_f64x2 = zero_f64x2, sum_squared_x_high_f64x2 = zero_f64x2;
+    float64x2_t sum_squared_y_low_f64x2 = zero_f64x2, sum_squared_y_high_f64x2 = zero_f64x2;
+    float64x2_t sum_squared_z_low_f64x2 = zero_f64x2, sum_squared_z_high_f64x2 = zero_f64x2;
+    nk_size_t index = 0;
+
+    for (; index + 4 <= n; index += 4) {
+        float32x4_t a_x_f32x4, a_y_f32x4, a_z_f32x4, b_x_f32x4, b_y_f32x4, b_z_f32x4;
+        nk_deinterleave_f32x4_neon_(a + index * 3, &a_x_f32x4, &a_y_f32x4, &a_z_f32x4),
+            nk_deinterleave_f32x4_neon_(b + index * 3, &b_x_f32x4, &b_y_f32x4, &b_z_f32x4);
+
+        float64x2_t delta_x_low_f64x2 = vsubq_f64(vcvt_f64_f32(vget_low_f32(a_x_f32x4)),
+                                                  vcvt_f64_f32(vget_low_f32(b_x_f32x4)));
+        float64x2_t delta_x_high_f64x2 = vsubq_f64(vcvt_high_f64_f32(a_x_f32x4), vcvt_high_f64_f32(b_x_f32x4));
+        float64x2_t delta_y_low_f64x2 = vsubq_f64(vcvt_f64_f32(vget_low_f32(a_y_f32x4)),
+                                                  vcvt_f64_f32(vget_low_f32(b_y_f32x4)));
+        float64x2_t delta_y_high_f64x2 = vsubq_f64(vcvt_high_f64_f32(a_y_f32x4), vcvt_high_f64_f32(b_y_f32x4));
+        float64x2_t delta_z_low_f64x2 = vsubq_f64(vcvt_f64_f32(vget_low_f32(a_z_f32x4)),
+                                                  vcvt_f64_f32(vget_low_f32(b_z_f32x4)));
+        float64x2_t delta_z_high_f64x2 = vsubq_f64(vcvt_high_f64_f32(a_z_f32x4), vcvt_high_f64_f32(b_z_f32x4));
+
+        sum_squared_x_low_f64x2 = vfmaq_f64(sum_squared_x_low_f64x2, delta_x_low_f64x2, delta_x_low_f64x2),
+        sum_squared_x_high_f64x2 = vfmaq_f64(sum_squared_x_high_f64x2, delta_x_high_f64x2, delta_x_high_f64x2);
+        sum_squared_y_low_f64x2 = vfmaq_f64(sum_squared_y_low_f64x2, delta_y_low_f64x2, delta_y_low_f64x2),
+        sum_squared_y_high_f64x2 = vfmaq_f64(sum_squared_y_high_f64x2, delta_y_high_f64x2, delta_y_high_f64x2);
+        sum_squared_z_low_f64x2 = vfmaq_f64(sum_squared_z_low_f64x2, delta_z_low_f64x2, delta_z_low_f64x2),
+        sum_squared_z_high_f64x2 = vfmaq_f64(sum_squared_z_high_f64x2, delta_z_high_f64x2, delta_z_high_f64x2);
+    }
+
+    nk_f64_t sum_squared_x = vaddvq_f64(vaddq_f64(sum_squared_x_low_f64x2, sum_squared_x_high_f64x2));
+    nk_f64_t sum_squared_y = vaddvq_f64(vaddq_f64(sum_squared_y_low_f64x2, sum_squared_y_high_f64x2));
+    nk_f64_t sum_squared_z = vaddvq_f64(vaddq_f64(sum_squared_z_low_f64x2, sum_squared_z_high_f64x2));
+
+    for (; index < n; ++index) {
+        nk_f64_t delta_x = (nk_f64_t)a[index * 3 + 0] - (nk_f64_t)b[index * 3 + 0];
+        nk_f64_t delta_y = (nk_f64_t)a[index * 3 + 1] - (nk_f64_t)b[index * 3 + 1];
+        nk_f64_t delta_z = (nk_f64_t)a[index * 3 + 2] - (nk_f64_t)b[index * 3 + 2];
+        sum_squared_x += delta_x * delta_x, sum_squared_y += delta_y * delta_y, sum_squared_z += delta_z * delta_z;
+    }
+
+    *result = nk_f64_sqrt_neon((sum_squared_x + sum_squared_y + sum_squared_z) / (nk_f64_t)n);
+}
+
+NK_PUBLIC void nk_rmsd_f64_neon(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *a_centroid,
+                                nk_f64_t *b_centroid, nk_f64_t *rotation, nk_f64_t *scale, nk_f64_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 = 1.0;
+
+    float64x2_t const zeros_f64x2 = vdupq_n_f64(0);
+    float64x2_t sum_squared_x_f64x2 = zeros_f64x2, sum_squared_y_f64x2 = zeros_f64x2, sum_squared_z_f64x2 = zeros_f64x2;
+
+    float64x2_t a_x_f64x2, a_y_f64x2, a_z_f64x2, b_x_f64x2, b_y_f64x2, b_z_f64x2;
+    nk_size_t i = 0;
+
+    for (; i + 2 <= n; i += 2) {
+        nk_deinterleave_f64x2_neon_(a + i * 3, &a_x_f64x2, &a_y_f64x2, &a_z_f64x2);
+        nk_deinterleave_f64x2_neon_(b + i * 3, &b_x_f64x2, &b_y_f64x2, &b_z_f64x2);
+
+        float64x2_t delta_x_f64x2 = vsubq_f64(a_x_f64x2, b_x_f64x2);
+        float64x2_t delta_y_f64x2 = vsubq_f64(a_y_f64x2, b_y_f64x2);
+        float64x2_t delta_z_f64x2 = vsubq_f64(a_z_f64x2, b_z_f64x2);
+
+        sum_squared_x_f64x2 = vfmaq_f64(sum_squared_x_f64x2, delta_x_f64x2, delta_x_f64x2);
+        sum_squared_y_f64x2 = vfmaq_f64(sum_squared_y_f64x2, delta_y_f64x2, delta_y_f64x2);
+        sum_squared_z_f64x2 = vfmaq_f64(sum_squared_z_f64x2, delta_z_f64x2, delta_z_f64x2);
+    }
+
+    nk_f64_t total_squared_x = nk_reduce_stable_f64x2_neon_(sum_squared_x_f64x2), total_squared_x_compensation = 0.0;
+    nk_f64_t total_squared_y = nk_reduce_stable_f64x2_neon_(sum_squared_y_f64x2), total_squared_y_compensation = 0.0;
+    nk_f64_t total_squared_z = nk_reduce_stable_f64x2_neon_(sum_squared_z_f64x2), total_squared_z_compensation = 0.0;
+
+    for (; i < n; ++i) {
+        nk_f64_t delta_x = a[i * 3 + 0] - b[i * 3 + 0];
+        nk_f64_t delta_y = a[i * 3 + 1] - b[i * 3 + 1];
+        nk_f64_t delta_z = a[i * 3 + 2] - b[i * 3 + 2];
+        nk_accumulate_square_f64_(&total_squared_x, &total_squared_x_compensation, delta_x);
+        nk_accumulate_square_f64_(&total_squared_y, &total_squared_y_compensation, delta_y);
+        nk_accumulate_square_f64_(&total_squared_z, &total_squared_z_compensation, delta_z);
+    }
+
+    total_squared_x += total_squared_x_compensation, total_squared_y += total_squared_y_compensation,
+        total_squared_z += total_squared_z_compensation;
+
+    *result = nk_f64_sqrt_neon((total_squared_x + total_squared_y + total_squared_z) / (nk_f64_t)n);
+}
+
+NK_PUBLIC void nk_kabsch_f32_neon(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                  nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f64_t *result) {
+    float64x2_t zero_f64x2 = vdupq_n_f64(0.0);
+
+    // Centroid accumulators (f64, lower/upper halves of f32x4)
+    float64x2_t sum_a_x_low_f64x2 = zero_f64x2, sum_a_x_high_f64x2 = zero_f64x2;
+    float64x2_t sum_a_y_low_f64x2 = zero_f64x2, sum_a_y_high_f64x2 = zero_f64x2;
+    float64x2_t sum_a_z_low_f64x2 = zero_f64x2, sum_a_z_high_f64x2 = zero_f64x2;
+    float64x2_t sum_b_x_low_f64x2 = zero_f64x2, sum_b_x_high_f64x2 = zero_f64x2;
+    float64x2_t sum_b_y_low_f64x2 = zero_f64x2, sum_b_y_high_f64x2 = zero_f64x2;
+    float64x2_t sum_b_z_low_f64x2 = zero_f64x2, sum_b_z_high_f64x2 = zero_f64x2;
+
+    // Covariance accumulators (f64, lower/upper halves)
+    float64x2_t cov_xx_low_f64x2 = zero_f64x2, cov_xx_high_f64x2 = zero_f64x2;
+    float64x2_t cov_xy_low_f64x2 = zero_f64x2, cov_xy_high_f64x2 = zero_f64x2;
+    float64x2_t cov_xz_low_f64x2 = zero_f64x2, cov_xz_high_f64x2 = zero_f64x2;
+    float64x2_t cov_yx_low_f64x2 = zero_f64x2, cov_yx_high_f64x2 = zero_f64x2;
+    float64x2_t cov_yy_low_f64x2 = zero_f64x2, cov_yy_high_f64x2 = zero_f64x2;
+    float64x2_t cov_yz_low_f64x2 = zero_f64x2, cov_yz_high_f64x2 = zero_f64x2;
+    float64x2_t cov_zx_low_f64x2 = zero_f64x2, cov_zx_high_f64x2 = zero_f64x2;
+    float64x2_t cov_zy_low_f64x2 = zero_f64x2, cov_zy_high_f64x2 = zero_f64x2;
+    float64x2_t cov_zz_low_f64x2 = zero_f64x2, cov_zz_high_f64x2 = zero_f64x2;
+
+    nk_size_t index = 0;
+    for (; index + 4 <= n; index += 4) {
+        float32x4_t a_x_f32x4, a_y_f32x4, a_z_f32x4, b_x_f32x4, b_y_f32x4, b_z_f32x4;
+        nk_deinterleave_f32x4_neon_(a + index * 3, &a_x_f32x4, &a_y_f32x4, &a_z_f32x4),
+            nk_deinterleave_f32x4_neon_(b + index * 3, &b_x_f32x4, &b_y_f32x4, &b_z_f32x4);
+
+        float64x2_t a_x_low_f64x2 = vcvt_f64_f32(vget_low_f32(a_x_f32x4));
+        float64x2_t a_x_high_f64x2 = vcvt_high_f64_f32(a_x_f32x4);
+        float64x2_t a_y_low_f64x2 = vcvt_f64_f32(vget_low_f32(a_y_f32x4));
+        float64x2_t a_y_high_f64x2 = vcvt_high_f64_f32(a_y_f32x4);
+        float64x2_t a_z_low_f64x2 = vcvt_f64_f32(vget_low_f32(a_z_f32x4));
+        float64x2_t a_z_high_f64x2 = vcvt_high_f64_f32(a_z_f32x4);
+        float64x2_t b_x_low_f64x2 = vcvt_f64_f32(vget_low_f32(b_x_f32x4));
+        float64x2_t b_x_high_f64x2 = vcvt_high_f64_f32(b_x_f32x4);
+        float64x2_t b_y_low_f64x2 = vcvt_f64_f32(vget_low_f32(b_y_f32x4));
+        float64x2_t b_y_high_f64x2 = vcvt_high_f64_f32(b_y_f32x4);
+        float64x2_t b_z_low_f64x2 = vcvt_f64_f32(vget_low_f32(b_z_f32x4));
+        float64x2_t b_z_high_f64x2 = vcvt_high_f64_f32(b_z_f32x4);
+
+        // Accumulate centroids
+        sum_a_x_low_f64x2 = vaddq_f64(sum_a_x_low_f64x2, a_x_low_f64x2),
+        sum_a_x_high_f64x2 = vaddq_f64(sum_a_x_high_f64x2, a_x_high_f64x2);
+        sum_a_y_low_f64x2 = vaddq_f64(sum_a_y_low_f64x2, a_y_low_f64x2),
+        sum_a_y_high_f64x2 = vaddq_f64(sum_a_y_high_f64x2, a_y_high_f64x2);
+        sum_a_z_low_f64x2 = vaddq_f64(sum_a_z_low_f64x2, a_z_low_f64x2),
+        sum_a_z_high_f64x2 = vaddq_f64(sum_a_z_high_f64x2, a_z_high_f64x2);
+        sum_b_x_low_f64x2 = vaddq_f64(sum_b_x_low_f64x2, b_x_low_f64x2),
+        sum_b_x_high_f64x2 = vaddq_f64(sum_b_x_high_f64x2, b_x_high_f64x2);
+        sum_b_y_low_f64x2 = vaddq_f64(sum_b_y_low_f64x2, b_y_low_f64x2),
+        sum_b_y_high_f64x2 = vaddq_f64(sum_b_y_high_f64x2, b_y_high_f64x2);
+        sum_b_z_low_f64x2 = vaddq_f64(sum_b_z_low_f64x2, b_z_low_f64x2),
+        sum_b_z_high_f64x2 = vaddq_f64(sum_b_z_high_f64x2, b_z_high_f64x2);
+
+        // Accumulate raw outer products (uncentered)
+        cov_xx_low_f64x2 = vfmaq_f64(cov_xx_low_f64x2, a_x_low_f64x2, b_x_low_f64x2),
+        cov_xx_high_f64x2 = vfmaq_f64(cov_xx_high_f64x2, a_x_high_f64x2, b_x_high_f64x2);
+        cov_xy_low_f64x2 = vfmaq_f64(cov_xy_low_f64x2, a_x_low_f64x2, b_y_low_f64x2),
+        cov_xy_high_f64x2 = vfmaq_f64(cov_xy_high_f64x2, a_x_high_f64x2, b_y_high_f64x2);
+        cov_xz_low_f64x2 = vfmaq_f64(cov_xz_low_f64x2, a_x_low_f64x2, b_z_low_f64x2),
+        cov_xz_high_f64x2 = vfmaq_f64(cov_xz_high_f64x2, a_x_high_f64x2, b_z_high_f64x2);
+        cov_yx_low_f64x2 = vfmaq_f64(cov_yx_low_f64x2, a_y_low_f64x2, b_x_low_f64x2),
+        cov_yx_high_f64x2 = vfmaq_f64(cov_yx_high_f64x2, a_y_high_f64x2, b_x_high_f64x2);
+        cov_yy_low_f64x2 = vfmaq_f64(cov_yy_low_f64x2, a_y_low_f64x2, b_y_low_f64x2),
+        cov_yy_high_f64x2 = vfmaq_f64(cov_yy_high_f64x2, a_y_high_f64x2, b_y_high_f64x2);
+        cov_yz_low_f64x2 = vfmaq_f64(cov_yz_low_f64x2, a_y_low_f64x2, b_z_low_f64x2),
+        cov_yz_high_f64x2 = vfmaq_f64(cov_yz_high_f64x2, a_y_high_f64x2, b_z_high_f64x2);
+        cov_zx_low_f64x2 = vfmaq_f64(cov_zx_low_f64x2, a_z_low_f64x2, b_x_low_f64x2),
+        cov_zx_high_f64x2 = vfmaq_f64(cov_zx_high_f64x2, a_z_high_f64x2, b_x_high_f64x2);
+        cov_zy_low_f64x2 = vfmaq_f64(cov_zy_low_f64x2, a_z_low_f64x2, b_y_low_f64x2),
+        cov_zy_high_f64x2 = vfmaq_f64(cov_zy_high_f64x2, a_z_high_f64x2, b_y_high_f64x2);
+        cov_zz_low_f64x2 = vfmaq_f64(cov_zz_low_f64x2, a_z_low_f64x2, b_z_low_f64x2),
+        cov_zz_high_f64x2 = vfmaq_f64(cov_zz_high_f64x2, a_z_high_f64x2, b_z_high_f64x2);
+    }
+
+    // Reduce centroid accumulators
+    nk_f64_t sum_a_x = vaddvq_f64(vaddq_f64(sum_a_x_low_f64x2, sum_a_x_high_f64x2));
+    nk_f64_t sum_a_y = vaddvq_f64(vaddq_f64(sum_a_y_low_f64x2, sum_a_y_high_f64x2));
+    nk_f64_t sum_a_z = vaddvq_f64(vaddq_f64(sum_a_z_low_f64x2, sum_a_z_high_f64x2));
+    nk_f64_t sum_b_x = vaddvq_f64(vaddq_f64(sum_b_x_low_f64x2, sum_b_x_high_f64x2));
+    nk_f64_t sum_b_y = vaddvq_f64(vaddq_f64(sum_b_y_low_f64x2, sum_b_y_high_f64x2));
+    nk_f64_t sum_b_z = vaddvq_f64(vaddq_f64(sum_b_z_low_f64x2, sum_b_z_high_f64x2));
+
+    // Reduce covariance accumulators
+    nk_f64_t covariance_x_x = vaddvq_f64(vaddq_f64(cov_xx_low_f64x2, cov_xx_high_f64x2));
+    nk_f64_t covariance_x_y = vaddvq_f64(vaddq_f64(cov_xy_low_f64x2, cov_xy_high_f64x2));
+    nk_f64_t covariance_x_z = vaddvq_f64(vaddq_f64(cov_xz_low_f64x2, cov_xz_high_f64x2));
+    nk_f64_t covariance_y_x = vaddvq_f64(vaddq_f64(cov_yx_low_f64x2, cov_yx_high_f64x2));
+    nk_f64_t covariance_y_y = vaddvq_f64(vaddq_f64(cov_yy_low_f64x2, cov_yy_high_f64x2));
+    nk_f64_t covariance_y_z = vaddvq_f64(vaddq_f64(cov_yz_low_f64x2, cov_yz_high_f64x2));
+    nk_f64_t covariance_z_x = vaddvq_f64(vaddq_f64(cov_zx_low_f64x2, cov_zx_high_f64x2));
+    nk_f64_t covariance_z_y = vaddvq_f64(vaddq_f64(cov_zy_low_f64x2, cov_zy_high_f64x2));
+    nk_f64_t covariance_z_z = vaddvq_f64(vaddq_f64(cov_zz_low_f64x2, cov_zz_high_f64x2));
+
+    // Scalar tail
+    for (; index < n; ++index) {
+        nk_f64_t ax = (nk_f64_t)a[index * 3 + 0], ay = (nk_f64_t)a[index * 3 + 1], az = (nk_f64_t)a[index * 3 + 2];
+        nk_f64_t bx = (nk_f64_t)b[index * 3 + 0], by = (nk_f64_t)b[index * 3 + 1], bz = (nk_f64_t)b[index * 3 + 2];
+        sum_a_x += ax, sum_a_y += ay, sum_a_z += az;
+        sum_b_x += bx, sum_b_y += by, sum_b_z += bz;
+        covariance_x_x += ax * bx, covariance_x_y += ax * by, covariance_x_z += ax * bz;
+        covariance_y_x += ay * bx, covariance_y_y += ay * by, covariance_y_z += ay * bz;
+        covariance_z_x += az * bx, covariance_z_y += az * by, covariance_z_z += az * bz;
+    }
+
+    // Compute centroids
+    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
+    nk_f64_t centroid_a_x = sum_a_x * inv_n, centroid_a_y = sum_a_y * inv_n, centroid_a_z = sum_a_z * inv_n;
+    nk_f64_t centroid_b_x = sum_b_x * inv_n, centroid_b_y = sum_b_y * inv_n, centroid_b_z = sum_b_z * inv_n;
+    if (a_centroid)
+        a_centroid[0] = (nk_f32_t)centroid_a_x, a_centroid[1] = (nk_f32_t)centroid_a_y,
+        a_centroid[2] = (nk_f32_t)centroid_a_z;
+    if (b_centroid)
+        b_centroid[0] = (nk_f32_t)centroid_b_x, b_centroid[1] = (nk_f32_t)centroid_b_y,
+        b_centroid[2] = (nk_f32_t)centroid_b_z;
+
+    // Apply centering correction: H_centered = sum(a * bᵀ) - n * centroid_a * centroid_bᵀ
+    nk_f64_t h[9];
+    h[0] = covariance_x_x - (nk_f64_t)n * centroid_a_x * centroid_b_x;
+    h[1] = covariance_x_y - (nk_f64_t)n * centroid_a_x * centroid_b_y;
+    h[2] = covariance_x_z - (nk_f64_t)n * centroid_a_x * centroid_b_z;
+    h[3] = covariance_y_x - (nk_f64_t)n * centroid_a_y * centroid_b_x;
+    h[4] = covariance_y_y - (nk_f64_t)n * centroid_a_y * centroid_b_y;
+    h[5] = covariance_y_z - (nk_f64_t)n * centroid_a_y * centroid_b_z;
+    h[6] = covariance_z_x - (nk_f64_t)n * centroid_a_z * centroid_b_x;
+    h[7] = covariance_z_y - (nk_f64_t)n * centroid_a_z * centroid_b_y;
+    h[8] = covariance_z_z - (nk_f64_t)n * centroid_a_z * centroid_b_z;
+
+    nk_f64_t svd_u[9], svd_s[9], svd_v[9];
+    nk_svd3x3_f64_(h, svd_u, svd_s, svd_v);
+
+    nk_f64_t r[9];
+    r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+    r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+    r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+    r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+    r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+    r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+    r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+    r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+    r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+
+    if (nk_det3x3_f64_(r) < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+        r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+        r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+        r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+        r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+        r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+        r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+        r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+        r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+    }
+
+    if (rotation)
+        for (int j = 0; j != 9; ++j) rotation[j] = (nk_f32_t)r[j];
+    if (scale) *scale = 1.0f;
+    *result = nk_f64_sqrt_neon(nk_transformed_ssd_f32_neon_(a, b, n, r, 1.0, centroid_a_x, centroid_a_y, centroid_a_z,
+                                                            centroid_b_x, centroid_b_y, centroid_b_z) /
+                               (nk_f64_t)n);
+}
+
+NK_PUBLIC void nk_kabsch_f64_neon(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *a_centroid,
+                                  nk_f64_t *b_centroid, nk_f64_t *rotation, nk_f64_t *scale, nk_f64_t *result) {
+    float64x2_t const zeros_f64x2 = vdupq_n_f64(0);
+
+    // 2x unrolling with dual accumulators to hide FMA latency.
+    float64x2_t sum_a_x_a_f64x2 = zeros_f64x2, sum_a_y_a_f64x2 = zeros_f64x2, sum_a_z_a_f64x2 = zeros_f64x2;
+    float64x2_t sum_b_x_a_f64x2 = zeros_f64x2, sum_b_y_a_f64x2 = zeros_f64x2, sum_b_z_a_f64x2 = zeros_f64x2;
+    float64x2_t sum_a_x_b_f64x2 = zeros_f64x2, sum_a_y_b_f64x2 = zeros_f64x2, sum_a_z_b_f64x2 = zeros_f64x2;
+    float64x2_t sum_b_x_b_f64x2 = zeros_f64x2, sum_b_y_b_f64x2 = zeros_f64x2, sum_b_z_b_f64x2 = zeros_f64x2;
+
+    float64x2_t cov_xx_a_f64x2 = zeros_f64x2, cov_xy_a_f64x2 = zeros_f64x2, cov_xz_a_f64x2 = zeros_f64x2;
+    float64x2_t cov_yx_a_f64x2 = zeros_f64x2, cov_yy_a_f64x2 = zeros_f64x2, cov_yz_a_f64x2 = zeros_f64x2;
+    float64x2_t cov_zx_a_f64x2 = zeros_f64x2, cov_zy_a_f64x2 = zeros_f64x2, cov_zz_a_f64x2 = zeros_f64x2;
+    float64x2_t cov_xx_b_f64x2 = zeros_f64x2, cov_xy_b_f64x2 = zeros_f64x2, cov_xz_b_f64x2 = zeros_f64x2;
+    float64x2_t cov_yx_b_f64x2 = zeros_f64x2, cov_yy_b_f64x2 = zeros_f64x2, cov_yz_b_f64x2 = zeros_f64x2;
+    float64x2_t cov_zx_b_f64x2 = zeros_f64x2, cov_zy_b_f64x2 = zeros_f64x2, cov_zz_b_f64x2 = zeros_f64x2;
+
+    nk_size_t i = 0;
+    float64x2_t a1_x_f64x2, a1_y_f64x2, a1_z_f64x2, b1_x_f64x2, b1_y_f64x2, b1_z_f64x2;
+    float64x2_t a2_x_f64x2, a2_y_f64x2, a2_z_f64x2, b2_x_f64x2, b2_y_f64x2, b2_z_f64x2;
+
+    // Main loop: 4 points per iteration (2x unrolled)
+    for (; i + 4 <= n; i += 4) {
+        nk_deinterleave_f64x2_neon_(a + i * 3, &a1_x_f64x2, &a1_y_f64x2, &a1_z_f64x2);
+        nk_deinterleave_f64x2_neon_(b + i * 3, &b1_x_f64x2, &b1_y_f64x2, &b1_z_f64x2);
+        nk_deinterleave_f64x2_neon_(a + (i + 2) * 3, &a2_x_f64x2, &a2_y_f64x2, &a2_z_f64x2);
+        nk_deinterleave_f64x2_neon_(b + (i + 2) * 3, &b2_x_f64x2, &b2_y_f64x2, &b2_z_f64x2);
+
+        // Interleaved accumulation
+        sum_a_x_a_f64x2 = vaddq_f64(sum_a_x_a_f64x2, a1_x_f64x2);
+        sum_a_x_b_f64x2 = vaddq_f64(sum_a_x_b_f64x2, a2_x_f64x2);
+        sum_a_y_a_f64x2 = vaddq_f64(sum_a_y_a_f64x2, a1_y_f64x2);
+        sum_a_y_b_f64x2 = vaddq_f64(sum_a_y_b_f64x2, a2_y_f64x2);
+        sum_a_z_a_f64x2 = vaddq_f64(sum_a_z_a_f64x2, a1_z_f64x2);
+        sum_a_z_b_f64x2 = vaddq_f64(sum_a_z_b_f64x2, a2_z_f64x2);
+        sum_b_x_a_f64x2 = vaddq_f64(sum_b_x_a_f64x2, b1_x_f64x2);
+        sum_b_x_b_f64x2 = vaddq_f64(sum_b_x_b_f64x2, b2_x_f64x2);
+        sum_b_y_a_f64x2 = vaddq_f64(sum_b_y_a_f64x2, b1_y_f64x2);
+        sum_b_y_b_f64x2 = vaddq_f64(sum_b_y_b_f64x2, b2_y_f64x2);
+        sum_b_z_a_f64x2 = vaddq_f64(sum_b_z_a_f64x2, b1_z_f64x2);
+        sum_b_z_b_f64x2 = vaddq_f64(sum_b_z_b_f64x2, b2_z_f64x2);
+
+        cov_xx_a_f64x2 = vfmaq_f64(cov_xx_a_f64x2, a1_x_f64x2, b1_x_f64x2);
+        cov_xx_b_f64x2 = vfmaq_f64(cov_xx_b_f64x2, a2_x_f64x2, b2_x_f64x2);
+        cov_xy_a_f64x2 = vfmaq_f64(cov_xy_a_f64x2, a1_x_f64x2, b1_y_f64x2);
+        cov_xy_b_f64x2 = vfmaq_f64(cov_xy_b_f64x2, a2_x_f64x2, b2_y_f64x2);
+        cov_xz_a_f64x2 = vfmaq_f64(cov_xz_a_f64x2, a1_x_f64x2, b1_z_f64x2);
+        cov_xz_b_f64x2 = vfmaq_f64(cov_xz_b_f64x2, a2_x_f64x2, b2_z_f64x2);
+        cov_yx_a_f64x2 = vfmaq_f64(cov_yx_a_f64x2, a1_y_f64x2, b1_x_f64x2);
+        cov_yx_b_f64x2 = vfmaq_f64(cov_yx_b_f64x2, a2_y_f64x2, b2_x_f64x2);
+        cov_yy_a_f64x2 = vfmaq_f64(cov_yy_a_f64x2, a1_y_f64x2, b1_y_f64x2);
+        cov_yy_b_f64x2 = vfmaq_f64(cov_yy_b_f64x2, a2_y_f64x2, b2_y_f64x2);
+        cov_yz_a_f64x2 = vfmaq_f64(cov_yz_a_f64x2, a1_y_f64x2, b1_z_f64x2);
+        cov_yz_b_f64x2 = vfmaq_f64(cov_yz_b_f64x2, a2_y_f64x2, b2_z_f64x2);
+        cov_zx_a_f64x2 = vfmaq_f64(cov_zx_a_f64x2, a1_z_f64x2, b1_x_f64x2);
+        cov_zx_b_f64x2 = vfmaq_f64(cov_zx_b_f64x2, a2_z_f64x2, b2_x_f64x2);
+        cov_zy_a_f64x2 = vfmaq_f64(cov_zy_a_f64x2, a1_z_f64x2, b1_y_f64x2);
+        cov_zy_b_f64x2 = vfmaq_f64(cov_zy_b_f64x2, a2_z_f64x2, b2_y_f64x2);
+        cov_zz_a_f64x2 = vfmaq_f64(cov_zz_a_f64x2, a1_z_f64x2, b1_z_f64x2);
+        cov_zz_b_f64x2 = vfmaq_f64(cov_zz_b_f64x2, a2_z_f64x2, b2_z_f64x2);
+    }
+
+    // 2-point tail
+    for (; i + 2 <= n; i += 2) {
+        nk_deinterleave_f64x2_neon_(a + i * 3, &a1_x_f64x2, &a1_y_f64x2, &a1_z_f64x2);
+        nk_deinterleave_f64x2_neon_(b + i * 3, &b1_x_f64x2, &b1_y_f64x2, &b1_z_f64x2);
+        sum_a_x_a_f64x2 = vaddq_f64(sum_a_x_a_f64x2, a1_x_f64x2);
+        sum_a_y_a_f64x2 = vaddq_f64(sum_a_y_a_f64x2, a1_y_f64x2);
+        sum_a_z_a_f64x2 = vaddq_f64(sum_a_z_a_f64x2, a1_z_f64x2);
+        sum_b_x_a_f64x2 = vaddq_f64(sum_b_x_a_f64x2, b1_x_f64x2);
+        sum_b_y_a_f64x2 = vaddq_f64(sum_b_y_a_f64x2, b1_y_f64x2);
+        sum_b_z_a_f64x2 = vaddq_f64(sum_b_z_a_f64x2, b1_z_f64x2);
+        cov_xx_a_f64x2 = vfmaq_f64(cov_xx_a_f64x2, a1_x_f64x2, b1_x_f64x2);
+        cov_xy_a_f64x2 = vfmaq_f64(cov_xy_a_f64x2, a1_x_f64x2, b1_y_f64x2);
+        cov_xz_a_f64x2 = vfmaq_f64(cov_xz_a_f64x2, a1_x_f64x2, b1_z_f64x2);
+        cov_yx_a_f64x2 = vfmaq_f64(cov_yx_a_f64x2, a1_y_f64x2, b1_x_f64x2);
+        cov_yy_a_f64x2 = vfmaq_f64(cov_yy_a_f64x2, a1_y_f64x2, b1_y_f64x2);
+        cov_yz_a_f64x2 = vfmaq_f64(cov_yz_a_f64x2, a1_y_f64x2, b1_z_f64x2);
+        cov_zx_a_f64x2 = vfmaq_f64(cov_zx_a_f64x2, a1_z_f64x2, b1_x_f64x2);
+        cov_zy_a_f64x2 = vfmaq_f64(cov_zy_a_f64x2, a1_z_f64x2, b1_y_f64x2);
+        cov_zz_a_f64x2 = vfmaq_f64(cov_zz_a_f64x2, a1_z_f64x2, b1_z_f64x2);
+    }
+
+    // Combine dual accumulators
+    float64x2_t sum_a_x_f64x2 = vaddq_f64(sum_a_x_a_f64x2, sum_a_x_b_f64x2);
+    float64x2_t sum_a_y_f64x2 = vaddq_f64(sum_a_y_a_f64x2, sum_a_y_b_f64x2);
+    float64x2_t sum_a_z_f64x2 = vaddq_f64(sum_a_z_a_f64x2, sum_a_z_b_f64x2);
+    float64x2_t sum_b_x_f64x2 = vaddq_f64(sum_b_x_a_f64x2, sum_b_x_b_f64x2);
+    float64x2_t sum_b_y_f64x2 = vaddq_f64(sum_b_y_a_f64x2, sum_b_y_b_f64x2);
+    float64x2_t sum_b_z_f64x2 = vaddq_f64(sum_b_z_a_f64x2, sum_b_z_b_f64x2);
+    float64x2_t cov_xx_f64x2 = vaddq_f64(cov_xx_a_f64x2, cov_xx_b_f64x2);
+    float64x2_t cov_xy_f64x2 = vaddq_f64(cov_xy_a_f64x2, cov_xy_b_f64x2);
+    float64x2_t cov_xz_f64x2 = vaddq_f64(cov_xz_a_f64x2, cov_xz_b_f64x2);
+    float64x2_t cov_yx_f64x2 = vaddq_f64(cov_yx_a_f64x2, cov_yx_b_f64x2);
+    float64x2_t cov_yy_f64x2 = vaddq_f64(cov_yy_a_f64x2, cov_yy_b_f64x2);
+    float64x2_t cov_yz_f64x2 = vaddq_f64(cov_yz_a_f64x2, cov_yz_b_f64x2);
+    float64x2_t cov_zx_f64x2 = vaddq_f64(cov_zx_a_f64x2, cov_zx_b_f64x2);
+    float64x2_t cov_zy_f64x2 = vaddq_f64(cov_zy_a_f64x2, cov_zy_b_f64x2);
+    float64x2_t cov_zz_f64x2 = vaddq_f64(cov_zz_a_f64x2, cov_zz_b_f64x2);
+
+    // Reduce vector accumulators.
+    nk_f64_t sum_a_x = nk_reduce_stable_f64x2_neon_(sum_a_x_f64x2), sum_a_x_compensation = 0.0;
+    nk_f64_t sum_a_y = nk_reduce_stable_f64x2_neon_(sum_a_y_f64x2), sum_a_y_compensation = 0.0;
+    nk_f64_t sum_a_z = nk_reduce_stable_f64x2_neon_(sum_a_z_f64x2), sum_a_z_compensation = 0.0;
+    nk_f64_t sum_b_x = nk_reduce_stable_f64x2_neon_(sum_b_x_f64x2), sum_b_x_compensation = 0.0;
+    nk_f64_t sum_b_y = nk_reduce_stable_f64x2_neon_(sum_b_y_f64x2), sum_b_y_compensation = 0.0;
+    nk_f64_t sum_b_z = nk_reduce_stable_f64x2_neon_(sum_b_z_f64x2), sum_b_z_compensation = 0.0;
+
+    nk_f64_t covariance_x_x = nk_reduce_stable_f64x2_neon_(cov_xx_f64x2), covariance_x_x_compensation = 0.0;
+    nk_f64_t covariance_x_y = nk_reduce_stable_f64x2_neon_(cov_xy_f64x2), covariance_x_y_compensation = 0.0;
+    nk_f64_t covariance_x_z = nk_reduce_stable_f64x2_neon_(cov_xz_f64x2), covariance_x_z_compensation = 0.0;
+    nk_f64_t covariance_y_x = nk_reduce_stable_f64x2_neon_(cov_yx_f64x2), covariance_y_x_compensation = 0.0;
+    nk_f64_t covariance_y_y = nk_reduce_stable_f64x2_neon_(cov_yy_f64x2), covariance_y_y_compensation = 0.0;
+    nk_f64_t covariance_y_z = nk_reduce_stable_f64x2_neon_(cov_yz_f64x2), covariance_y_z_compensation = 0.0;
+    nk_f64_t covariance_z_x = nk_reduce_stable_f64x2_neon_(cov_zx_f64x2), covariance_z_x_compensation = 0.0;
+    nk_f64_t covariance_z_y = nk_reduce_stable_f64x2_neon_(cov_zy_f64x2), covariance_z_y_compensation = 0.0;
+    nk_f64_t covariance_z_z = nk_reduce_stable_f64x2_neon_(cov_zz_f64x2), covariance_z_z_compensation = 0.0;
+
+    // Scalar tail
+    for (; i < n; ++i) {
+        nk_f64_t ax = a[i * 3 + 0], ay = a[i * 3 + 1], az = a[i * 3 + 2];
+        nk_f64_t bx = b[i * 3 + 0], by = b[i * 3 + 1], bz = b[i * 3 + 2];
+        nk_accumulate_sum_f64_(&sum_a_x, &sum_a_x_compensation, ax),
+            nk_accumulate_sum_f64_(&sum_a_y, &sum_a_y_compensation, ay),
+            nk_accumulate_sum_f64_(&sum_a_z, &sum_a_z_compensation, az);
+        nk_accumulate_sum_f64_(&sum_b_x, &sum_b_x_compensation, bx),
+            nk_accumulate_sum_f64_(&sum_b_y, &sum_b_y_compensation, by),
+            nk_accumulate_sum_f64_(&sum_b_z, &sum_b_z_compensation, bz);
+        nk_accumulate_product_f64_(&covariance_x_x, &covariance_x_x_compensation, ax, bx),
+            nk_accumulate_product_f64_(&covariance_x_y, &covariance_x_y_compensation, ax, by),
+            nk_accumulate_product_f64_(&covariance_x_z, &covariance_x_z_compensation, ax, bz);
+        nk_accumulate_product_f64_(&covariance_y_x, &covariance_y_x_compensation, ay, bx),
+            nk_accumulate_product_f64_(&covariance_y_y, &covariance_y_y_compensation, ay, by),
+            nk_accumulate_product_f64_(&covariance_y_z, &covariance_y_z_compensation, ay, bz);
+        nk_accumulate_product_f64_(&covariance_z_x, &covariance_z_x_compensation, az, bx),
+            nk_accumulate_product_f64_(&covariance_z_y, &covariance_z_y_compensation, az, by),
+            nk_accumulate_product_f64_(&covariance_z_z, &covariance_z_z_compensation, az, bz);
+    }
+
+    sum_a_x += sum_a_x_compensation, sum_a_y += sum_a_y_compensation, sum_a_z += sum_a_z_compensation;
+    sum_b_x += sum_b_x_compensation, sum_b_y += sum_b_y_compensation, sum_b_z += sum_b_z_compensation;
+    covariance_x_x += covariance_x_x_compensation, covariance_x_y += covariance_x_y_compensation,
+        covariance_x_z += covariance_x_z_compensation;
+    covariance_y_x += covariance_y_x_compensation, covariance_y_y += covariance_y_y_compensation,
+        covariance_y_z += covariance_y_z_compensation;
+    covariance_z_x += covariance_z_x_compensation, covariance_z_y += covariance_z_y_compensation,
+        covariance_z_z += covariance_z_z_compensation;
+
+    // Compute centroids
+    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
+    nk_f64_t centroid_a_x = sum_a_x * inv_n, centroid_a_y = sum_a_y * inv_n, centroid_a_z = sum_a_z * inv_n;
+    nk_f64_t centroid_b_x = sum_b_x * inv_n, centroid_b_y = sum_b_y * inv_n, centroid_b_z = sum_b_z * inv_n;
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Apply centering correction: H_centered = H - n * centroid_a * centroid_bᵀ
+    covariance_x_x -= (nk_f64_t)n * centroid_a_x * centroid_b_x;
+    covariance_x_y -= (nk_f64_t)n * centroid_a_x * centroid_b_y;
+    covariance_x_z -= (nk_f64_t)n * centroid_a_x * centroid_b_z;
+    covariance_y_x -= (nk_f64_t)n * centroid_a_y * centroid_b_x;
+    covariance_y_y -= (nk_f64_t)n * centroid_a_y * centroid_b_y;
+    covariance_y_z -= (nk_f64_t)n * centroid_a_y * centroid_b_z;
+    covariance_z_x -= (nk_f64_t)n * centroid_a_z * centroid_b_x;
+    covariance_z_y -= (nk_f64_t)n * centroid_a_z * centroid_b_y;
+    covariance_z_z -= (nk_f64_t)n * centroid_a_z * centroid_b_z;
+
+    // Compute SVD and optimal rotation
+    nk_f64_t cross_covariance[9] = {covariance_x_x, covariance_x_y, covariance_x_z, covariance_y_x, covariance_y_y,
+                                    covariance_y_z, covariance_z_x, covariance_z_y, covariance_z_z};
+    nk_f64_t svd_u[9], svd_s[9], svd_v[9];
+    nk_svd3x3_f64_(cross_covariance, svd_u, svd_s, svd_v);
+
+    nk_f64_t r[9];
+    nk_rotation_from_svd_f64_neon_(svd_u, svd_v, r);
+
+    // Handle reflection: if det(R) < 0, negate third column of V and recompute R
+    if (nk_det3x3_f64_(r) < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        nk_rotation_from_svd_f64_neon_(svd_u, svd_v, r);
+    }
+
+    // Output rotation matrix and scale=1.0.
+    if (rotation)
+        for (int j = 0; j < 9; ++j) rotation[j] = r[j];
+
+    if (scale) *scale = 1.0;
+
+    // Compute RMSD after optimal rotation
+    nk_f64_t sum_squared = nk_transformed_ssd_f64_neon_(a, b, n, r, 1.0, centroid_a_x, centroid_a_y, centroid_a_z,
+                                                        centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f64_sqrt_neon(sum_squared * inv_n);
+}
+
+NK_PUBLIC void nk_umeyama_f32_neon(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                   nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f64_t *result) {
+    float64x2_t zero_f64x2 = vdupq_n_f64(0.0);
+
+    // Centroid accumulators (f64, lower/upper halves of f32x4)
+    float64x2_t sum_a_x_low_f64x2 = zero_f64x2, sum_a_x_high_f64x2 = zero_f64x2;
+    float64x2_t sum_a_y_low_f64x2 = zero_f64x2, sum_a_y_high_f64x2 = zero_f64x2;
+    float64x2_t sum_a_z_low_f64x2 = zero_f64x2, sum_a_z_high_f64x2 = zero_f64x2;
+    float64x2_t sum_b_x_low_f64x2 = zero_f64x2, sum_b_x_high_f64x2 = zero_f64x2;
+    float64x2_t sum_b_y_low_f64x2 = zero_f64x2, sum_b_y_high_f64x2 = zero_f64x2;
+    float64x2_t sum_b_z_low_f64x2 = zero_f64x2, sum_b_z_high_f64x2 = zero_f64x2;
+
+    // Covariance accumulators (f64, lower/upper halves)
+    float64x2_t cov_xx_low_f64x2 = zero_f64x2, cov_xx_high_f64x2 = zero_f64x2;
+    float64x2_t cov_xy_low_f64x2 = zero_f64x2, cov_xy_high_f64x2 = zero_f64x2;
+    float64x2_t cov_xz_low_f64x2 = zero_f64x2, cov_xz_high_f64x2 = zero_f64x2;
+    float64x2_t cov_yx_low_f64x2 = zero_f64x2, cov_yx_high_f64x2 = zero_f64x2;
+    float64x2_t cov_yy_low_f64x2 = zero_f64x2, cov_yy_high_f64x2 = zero_f64x2;
+    float64x2_t cov_yz_low_f64x2 = zero_f64x2, cov_yz_high_f64x2 = zero_f64x2;
+    float64x2_t cov_zx_low_f64x2 = zero_f64x2, cov_zx_high_f64x2 = zero_f64x2;
+    float64x2_t cov_zy_low_f64x2 = zero_f64x2, cov_zy_high_f64x2 = zero_f64x2;
+    float64x2_t cov_zz_low_f64x2 = zero_f64x2, cov_zz_high_f64x2 = zero_f64x2;
+
+    // Variance of A accumulator
+    float64x2_t variance_low_f64x2 = zero_f64x2, variance_high_f64x2 = zero_f64x2;
+
+    nk_size_t index = 0;
+    for (; index + 4 <= n; index += 4) {
+        float32x4_t a_x_f32x4, a_y_f32x4, a_z_f32x4, b_x_f32x4, b_y_f32x4, b_z_f32x4;
+        nk_deinterleave_f32x4_neon_(a + index * 3, &a_x_f32x4, &a_y_f32x4, &a_z_f32x4),
+            nk_deinterleave_f32x4_neon_(b + index * 3, &b_x_f32x4, &b_y_f32x4, &b_z_f32x4);
+
+        float64x2_t a_x_low_f64x2 = vcvt_f64_f32(vget_low_f32(a_x_f32x4));
+        float64x2_t a_x_high_f64x2 = vcvt_high_f64_f32(a_x_f32x4);
+        float64x2_t a_y_low_f64x2 = vcvt_f64_f32(vget_low_f32(a_y_f32x4));
+        float64x2_t a_y_high_f64x2 = vcvt_high_f64_f32(a_y_f32x4);
+        float64x2_t a_z_low_f64x2 = vcvt_f64_f32(vget_low_f32(a_z_f32x4));
+        float64x2_t a_z_high_f64x2 = vcvt_high_f64_f32(a_z_f32x4);
+        float64x2_t b_x_low_f64x2 = vcvt_f64_f32(vget_low_f32(b_x_f32x4));
+        float64x2_t b_x_high_f64x2 = vcvt_high_f64_f32(b_x_f32x4);
+        float64x2_t b_y_low_f64x2 = vcvt_f64_f32(vget_low_f32(b_y_f32x4));
+        float64x2_t b_y_high_f64x2 = vcvt_high_f64_f32(b_y_f32x4);
+        float64x2_t b_z_low_f64x2 = vcvt_f64_f32(vget_low_f32(b_z_f32x4));
+        float64x2_t b_z_high_f64x2 = vcvt_high_f64_f32(b_z_f32x4);
+
+        // Accumulate centroids
+        sum_a_x_low_f64x2 = vaddq_f64(sum_a_x_low_f64x2, a_x_low_f64x2),
+        sum_a_x_high_f64x2 = vaddq_f64(sum_a_x_high_f64x2, a_x_high_f64x2);
+        sum_a_y_low_f64x2 = vaddq_f64(sum_a_y_low_f64x2, a_y_low_f64x2),
+        sum_a_y_high_f64x2 = vaddq_f64(sum_a_y_high_f64x2, a_y_high_f64x2);
+        sum_a_z_low_f64x2 = vaddq_f64(sum_a_z_low_f64x2, a_z_low_f64x2),
+        sum_a_z_high_f64x2 = vaddq_f64(sum_a_z_high_f64x2, a_z_high_f64x2);
+        sum_b_x_low_f64x2 = vaddq_f64(sum_b_x_low_f64x2, b_x_low_f64x2),
+        sum_b_x_high_f64x2 = vaddq_f64(sum_b_x_high_f64x2, b_x_high_f64x2);
+        sum_b_y_low_f64x2 = vaddq_f64(sum_b_y_low_f64x2, b_y_low_f64x2),
+        sum_b_y_high_f64x2 = vaddq_f64(sum_b_y_high_f64x2, b_y_high_f64x2);
+        sum_b_z_low_f64x2 = vaddq_f64(sum_b_z_low_f64x2, b_z_low_f64x2),
+        sum_b_z_high_f64x2 = vaddq_f64(sum_b_z_high_f64x2, b_z_high_f64x2);
+
+        // Accumulate raw outer products (uncentered)
+        cov_xx_low_f64x2 = vfmaq_f64(cov_xx_low_f64x2, a_x_low_f64x2, b_x_low_f64x2),
+        cov_xx_high_f64x2 = vfmaq_f64(cov_xx_high_f64x2, a_x_high_f64x2, b_x_high_f64x2);
+        cov_xy_low_f64x2 = vfmaq_f64(cov_xy_low_f64x2, a_x_low_f64x2, b_y_low_f64x2),
+        cov_xy_high_f64x2 = vfmaq_f64(cov_xy_high_f64x2, a_x_high_f64x2, b_y_high_f64x2);
+        cov_xz_low_f64x2 = vfmaq_f64(cov_xz_low_f64x2, a_x_low_f64x2, b_z_low_f64x2),
+        cov_xz_high_f64x2 = vfmaq_f64(cov_xz_high_f64x2, a_x_high_f64x2, b_z_high_f64x2);
+        cov_yx_low_f64x2 = vfmaq_f64(cov_yx_low_f64x2, a_y_low_f64x2, b_x_low_f64x2),
+        cov_yx_high_f64x2 = vfmaq_f64(cov_yx_high_f64x2, a_y_high_f64x2, b_x_high_f64x2);
+        cov_yy_low_f64x2 = vfmaq_f64(cov_yy_low_f64x2, a_y_low_f64x2, b_y_low_f64x2),
+        cov_yy_high_f64x2 = vfmaq_f64(cov_yy_high_f64x2, a_y_high_f64x2, b_y_high_f64x2);
+        cov_yz_low_f64x2 = vfmaq_f64(cov_yz_low_f64x2, a_y_low_f64x2, b_z_low_f64x2),
+        cov_yz_high_f64x2 = vfmaq_f64(cov_yz_high_f64x2, a_y_high_f64x2, b_z_high_f64x2);
+        cov_zx_low_f64x2 = vfmaq_f64(cov_zx_low_f64x2, a_z_low_f64x2, b_x_low_f64x2),
+        cov_zx_high_f64x2 = vfmaq_f64(cov_zx_high_f64x2, a_z_high_f64x2, b_x_high_f64x2);
+        cov_zy_low_f64x2 = vfmaq_f64(cov_zy_low_f64x2, a_z_low_f64x2, b_y_low_f64x2),
+        cov_zy_high_f64x2 = vfmaq_f64(cov_zy_high_f64x2, a_z_high_f64x2, b_y_high_f64x2);
+        cov_zz_low_f64x2 = vfmaq_f64(cov_zz_low_f64x2, a_z_low_f64x2, b_z_low_f64x2),
+        cov_zz_high_f64x2 = vfmaq_f64(cov_zz_high_f64x2, a_z_high_f64x2, b_z_high_f64x2);
+
+        // Accumulate variance of A (sum of squared coordinates)
+        variance_low_f64x2 = vfmaq_f64(variance_low_f64x2, a_x_low_f64x2, a_x_low_f64x2),
+        variance_high_f64x2 = vfmaq_f64(variance_high_f64x2, a_x_high_f64x2, a_x_high_f64x2);
+        variance_low_f64x2 = vfmaq_f64(variance_low_f64x2, a_y_low_f64x2, a_y_low_f64x2),
+        variance_high_f64x2 = vfmaq_f64(variance_high_f64x2, a_y_high_f64x2, a_y_high_f64x2);
+        variance_low_f64x2 = vfmaq_f64(variance_low_f64x2, a_z_low_f64x2, a_z_low_f64x2),
+        variance_high_f64x2 = vfmaq_f64(variance_high_f64x2, a_z_high_f64x2, a_z_high_f64x2);
+    }
+
+    // Reduce centroid accumulators
+    nk_f64_t sum_a_x = vaddvq_f64(vaddq_f64(sum_a_x_low_f64x2, sum_a_x_high_f64x2));
+    nk_f64_t sum_a_y = vaddvq_f64(vaddq_f64(sum_a_y_low_f64x2, sum_a_y_high_f64x2));
+    nk_f64_t sum_a_z = vaddvq_f64(vaddq_f64(sum_a_z_low_f64x2, sum_a_z_high_f64x2));
+    nk_f64_t sum_b_x = vaddvq_f64(vaddq_f64(sum_b_x_low_f64x2, sum_b_x_high_f64x2));
+    nk_f64_t sum_b_y = vaddvq_f64(vaddq_f64(sum_b_y_low_f64x2, sum_b_y_high_f64x2));
+    nk_f64_t sum_b_z = vaddvq_f64(vaddq_f64(sum_b_z_low_f64x2, sum_b_z_high_f64x2));
+
+    // Reduce covariance accumulators
+    nk_f64_t covariance_x_x = vaddvq_f64(vaddq_f64(cov_xx_low_f64x2, cov_xx_high_f64x2));
+    nk_f64_t covariance_x_y = vaddvq_f64(vaddq_f64(cov_xy_low_f64x2, cov_xy_high_f64x2));
+    nk_f64_t covariance_x_z = vaddvq_f64(vaddq_f64(cov_xz_low_f64x2, cov_xz_high_f64x2));
+    nk_f64_t covariance_y_x = vaddvq_f64(vaddq_f64(cov_yx_low_f64x2, cov_yx_high_f64x2));
+    nk_f64_t covariance_y_y = vaddvq_f64(vaddq_f64(cov_yy_low_f64x2, cov_yy_high_f64x2));
+    nk_f64_t covariance_y_z = vaddvq_f64(vaddq_f64(cov_yz_low_f64x2, cov_yz_high_f64x2));
+    nk_f64_t covariance_z_x = vaddvq_f64(vaddq_f64(cov_zx_low_f64x2, cov_zx_high_f64x2));
+    nk_f64_t covariance_z_y = vaddvq_f64(vaddq_f64(cov_zy_low_f64x2, cov_zy_high_f64x2));
+    nk_f64_t covariance_z_z = vaddvq_f64(vaddq_f64(cov_zz_low_f64x2, cov_zz_high_f64x2));
+    nk_f64_t sum_sq_a = vaddvq_f64(vaddq_f64(variance_low_f64x2, variance_high_f64x2));
+
+    // Scalar tail
+    for (; index < n; ++index) {
+        nk_f64_t ax = (nk_f64_t)a[index * 3 + 0], ay = (nk_f64_t)a[index * 3 + 1], az = (nk_f64_t)a[index * 3 + 2];
+        nk_f64_t bx = (nk_f64_t)b[index * 3 + 0], by = (nk_f64_t)b[index * 3 + 1], bz = (nk_f64_t)b[index * 3 + 2];
+        sum_a_x += ax, sum_a_y += ay, sum_a_z += az;
+        sum_b_x += bx, sum_b_y += by, sum_b_z += bz;
+        covariance_x_x += ax * bx, covariance_x_y += ax * by, covariance_x_z += ax * bz;
+        covariance_y_x += ay * bx, covariance_y_y += ay * by, covariance_y_z += ay * bz;
+        covariance_z_x += az * bx, covariance_z_y += az * by, covariance_z_z += az * bz;
+        sum_sq_a += ax * ax + ay * ay + az * az;
+    }
+
+    // Compute centroids
+    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
+    nk_f64_t centroid_a_x = sum_a_x * inv_n, centroid_a_y = sum_a_y * inv_n, centroid_a_z = sum_a_z * inv_n;
+    nk_f64_t centroid_b_x = sum_b_x * inv_n, centroid_b_y = sum_b_y * inv_n, centroid_b_z = sum_b_z * inv_n;
+    if (a_centroid)
+        a_centroid[0] = (nk_f32_t)centroid_a_x, a_centroid[1] = (nk_f32_t)centroid_a_y,
+        a_centroid[2] = (nk_f32_t)centroid_a_z;
+    if (b_centroid)
+        b_centroid[0] = (nk_f32_t)centroid_b_x, b_centroid[1] = (nk_f32_t)centroid_b_y,
+        b_centroid[2] = (nk_f32_t)centroid_b_z;
+
+    // Compute variance of A (centered): var = sum(a^2)/n - centroid^2
+    nk_f64_t variance_a = sum_sq_a * inv_n -
+                          (centroid_a_x * centroid_a_x + centroid_a_y * centroid_a_y + centroid_a_z * centroid_a_z);
+
+    // Apply centering correction: H_centered = sum(a * bᵀ) - n * centroid_a * centroid_bᵀ
+    nk_f64_t h[9];
+    h[0] = covariance_x_x - (nk_f64_t)n * centroid_a_x * centroid_b_x;
+    h[1] = covariance_x_y - (nk_f64_t)n * centroid_a_x * centroid_b_y;
+    h[2] = covariance_x_z - (nk_f64_t)n * centroid_a_x * centroid_b_z;
+    h[3] = covariance_y_x - (nk_f64_t)n * centroid_a_y * centroid_b_x;
+    h[4] = covariance_y_y - (nk_f64_t)n * centroid_a_y * centroid_b_y;
+    h[5] = covariance_y_z - (nk_f64_t)n * centroid_a_y * centroid_b_z;
+    h[6] = covariance_z_x - (nk_f64_t)n * centroid_a_z * centroid_b_x;
+    h[7] = covariance_z_y - (nk_f64_t)n * centroid_a_z * centroid_b_y;
+    h[8] = covariance_z_z - (nk_f64_t)n * centroid_a_z * centroid_b_z;
+
+    nk_f64_t svd_u[9], svd_s[9], svd_v[9];
+    nk_svd3x3_f64_(h, svd_u, svd_s, svd_v);
+
+    nk_f64_t r[9];
+    r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+    r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+    r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+    r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+    r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+    r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+    r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+    r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+    r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+
+    nk_f64_t det = nk_det3x3_f64_(r), sign_correction = det < 0 ? -1.0 : 1.0;
+    if (det < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+        r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+        r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+        r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+        r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+        r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+        r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+        r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+        r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+    }
+
+    nk_f64_t applied_scale = (svd_s[0] + svd_s[4] + sign_correction * svd_s[8]) / ((nk_f64_t)n * variance_a);
+    if (rotation)
+        for (int j = 0; j != 9; ++j) rotation[j] = (nk_f32_t)r[j];
+    if (scale) *scale = (nk_f32_t)applied_scale;
+    *result = nk_f64_sqrt_neon(nk_transformed_ssd_f32_neon_(a, b, n, r, applied_scale, centroid_a_x, centroid_a_y,
+                                                            centroid_a_z, centroid_b_x, centroid_b_y, centroid_b_z) /
+                               (nk_f64_t)n);
+}
+
+NK_PUBLIC void nk_umeyama_f64_neon(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *a_centroid,
+                                   nk_f64_t *b_centroid, nk_f64_t *rotation, nk_f64_t *scale, nk_f64_t *result) {
+    float64x2_t const zeros_f64x2 = vdupq_n_f64(0);
+
+    // 2x unrolling with dual accumulators to hide FMA latency.
+    float64x2_t sum_a_x_a_f64x2 = zeros_f64x2, sum_a_y_a_f64x2 = zeros_f64x2, sum_a_z_a_f64x2 = zeros_f64x2;
+    float64x2_t sum_b_x_a_f64x2 = zeros_f64x2, sum_b_y_a_f64x2 = zeros_f64x2, sum_b_z_a_f64x2 = zeros_f64x2;
+    float64x2_t sum_a_x_b_f64x2 = zeros_f64x2, sum_a_y_b_f64x2 = zeros_f64x2, sum_a_z_b_f64x2 = zeros_f64x2;
+    float64x2_t sum_b_x_b_f64x2 = zeros_f64x2, sum_b_y_b_f64x2 = zeros_f64x2, sum_b_z_b_f64x2 = zeros_f64x2;
+
+    float64x2_t cov_xx_a_f64x2 = zeros_f64x2, cov_xy_a_f64x2 = zeros_f64x2, cov_xz_a_f64x2 = zeros_f64x2;
+    float64x2_t cov_yx_a_f64x2 = zeros_f64x2, cov_yy_a_f64x2 = zeros_f64x2, cov_yz_a_f64x2 = zeros_f64x2;
+    float64x2_t cov_zx_a_f64x2 = zeros_f64x2, cov_zy_a_f64x2 = zeros_f64x2, cov_zz_a_f64x2 = zeros_f64x2;
+    float64x2_t cov_xx_b_f64x2 = zeros_f64x2, cov_xy_b_f64x2 = zeros_f64x2, cov_xz_b_f64x2 = zeros_f64x2;
+    float64x2_t cov_yx_b_f64x2 = zeros_f64x2, cov_yy_b_f64x2 = zeros_f64x2, cov_yz_b_f64x2 = zeros_f64x2;
+    float64x2_t cov_zx_b_f64x2 = zeros_f64x2, cov_zy_b_f64x2 = zeros_f64x2, cov_zz_b_f64x2 = zeros_f64x2;
+    float64x2_t variance_a_a_f64x2 = zeros_f64x2, variance_a_b_f64x2 = zeros_f64x2;
+
+    nk_size_t i = 0;
+    float64x2_t a1_x_f64x2, a1_y_f64x2, a1_z_f64x2, b1_x_f64x2, b1_y_f64x2, b1_z_f64x2;
+    float64x2_t a2_x_f64x2, a2_y_f64x2, a2_z_f64x2, b2_x_f64x2, b2_y_f64x2, b2_z_f64x2;
+
+    // Main loop: 4 points per iteration (2x unrolled)
+    for (; i + 4 <= n; i += 4) {
+        nk_deinterleave_f64x2_neon_(a + i * 3, &a1_x_f64x2, &a1_y_f64x2, &a1_z_f64x2);
+        nk_deinterleave_f64x2_neon_(b + i * 3, &b1_x_f64x2, &b1_y_f64x2, &b1_z_f64x2);
+        nk_deinterleave_f64x2_neon_(a + (i + 2) * 3, &a2_x_f64x2, &a2_y_f64x2, &a2_z_f64x2);
+        nk_deinterleave_f64x2_neon_(b + (i + 2) * 3, &b2_x_f64x2, &b2_y_f64x2, &b2_z_f64x2);
+
+        // Interleaved accumulation
+        sum_a_x_a_f64x2 = vaddq_f64(sum_a_x_a_f64x2, a1_x_f64x2);
+        sum_a_x_b_f64x2 = vaddq_f64(sum_a_x_b_f64x2, a2_x_f64x2);
+        sum_a_y_a_f64x2 = vaddq_f64(sum_a_y_a_f64x2, a1_y_f64x2);
+        sum_a_y_b_f64x2 = vaddq_f64(sum_a_y_b_f64x2, a2_y_f64x2);
+        sum_a_z_a_f64x2 = vaddq_f64(sum_a_z_a_f64x2, a1_z_f64x2);
+        sum_a_z_b_f64x2 = vaddq_f64(sum_a_z_b_f64x2, a2_z_f64x2);
+        sum_b_x_a_f64x2 = vaddq_f64(sum_b_x_a_f64x2, b1_x_f64x2);
+        sum_b_x_b_f64x2 = vaddq_f64(sum_b_x_b_f64x2, b2_x_f64x2);
+        sum_b_y_a_f64x2 = vaddq_f64(sum_b_y_a_f64x2, b1_y_f64x2);
+        sum_b_y_b_f64x2 = vaddq_f64(sum_b_y_b_f64x2, b2_y_f64x2);
+        sum_b_z_a_f64x2 = vaddq_f64(sum_b_z_a_f64x2, b1_z_f64x2);
+        sum_b_z_b_f64x2 = vaddq_f64(sum_b_z_b_f64x2, b2_z_f64x2);
+
+        cov_xx_a_f64x2 = vfmaq_f64(cov_xx_a_f64x2, a1_x_f64x2, b1_x_f64x2);
+        cov_xx_b_f64x2 = vfmaq_f64(cov_xx_b_f64x2, a2_x_f64x2, b2_x_f64x2);
+        cov_xy_a_f64x2 = vfmaq_f64(cov_xy_a_f64x2, a1_x_f64x2, b1_y_f64x2);
+        cov_xy_b_f64x2 = vfmaq_f64(cov_xy_b_f64x2, a2_x_f64x2, b2_y_f64x2);
+        cov_xz_a_f64x2 = vfmaq_f64(cov_xz_a_f64x2, a1_x_f64x2, b1_z_f64x2);
+        cov_xz_b_f64x2 = vfmaq_f64(cov_xz_b_f64x2, a2_x_f64x2, b2_z_f64x2);
+        cov_yx_a_f64x2 = vfmaq_f64(cov_yx_a_f64x2, a1_y_f64x2, b1_x_f64x2);
+        cov_yx_b_f64x2 = vfmaq_f64(cov_yx_b_f64x2, a2_y_f64x2, b2_x_f64x2);
+        cov_yy_a_f64x2 = vfmaq_f64(cov_yy_a_f64x2, a1_y_f64x2, b1_y_f64x2);
+        cov_yy_b_f64x2 = vfmaq_f64(cov_yy_b_f64x2, a2_y_f64x2, b2_y_f64x2);
+        cov_yz_a_f64x2 = vfmaq_f64(cov_yz_a_f64x2, a1_y_f64x2, b1_z_f64x2);
+        cov_yz_b_f64x2 = vfmaq_f64(cov_yz_b_f64x2, a2_y_f64x2, b2_z_f64x2);
+        cov_zx_a_f64x2 = vfmaq_f64(cov_zx_a_f64x2, a1_z_f64x2, b1_x_f64x2);
+        cov_zx_b_f64x2 = vfmaq_f64(cov_zx_b_f64x2, a2_z_f64x2, b2_x_f64x2);
+        cov_zy_a_f64x2 = vfmaq_f64(cov_zy_a_f64x2, a1_z_f64x2, b1_y_f64x2);
+        cov_zy_b_f64x2 = vfmaq_f64(cov_zy_b_f64x2, a2_z_f64x2, b2_y_f64x2);
+        cov_zz_a_f64x2 = vfmaq_f64(cov_zz_a_f64x2, a1_z_f64x2, b1_z_f64x2);
+        cov_zz_b_f64x2 = vfmaq_f64(cov_zz_b_f64x2, a2_z_f64x2, b2_z_f64x2);
+
+        variance_a_a_f64x2 = vfmaq_f64(variance_a_a_f64x2, a1_x_f64x2, a1_x_f64x2);
+        variance_a_b_f64x2 = vfmaq_f64(variance_a_b_f64x2, a2_x_f64x2, a2_x_f64x2);
+        variance_a_a_f64x2 = vfmaq_f64(variance_a_a_f64x2, a1_y_f64x2, a1_y_f64x2);
+        variance_a_b_f64x2 = vfmaq_f64(variance_a_b_f64x2, a2_y_f64x2, a2_y_f64x2);
+        variance_a_a_f64x2 = vfmaq_f64(variance_a_a_f64x2, a1_z_f64x2, a1_z_f64x2);
+        variance_a_b_f64x2 = vfmaq_f64(variance_a_b_f64x2, a2_z_f64x2, a2_z_f64x2);
+    }
+
+    // 2-point tail
+    for (; i + 2 <= n; i += 2) {
+        nk_deinterleave_f64x2_neon_(a + i * 3, &a1_x_f64x2, &a1_y_f64x2, &a1_z_f64x2);
+        nk_deinterleave_f64x2_neon_(b + i * 3, &b1_x_f64x2, &b1_y_f64x2, &b1_z_f64x2);
+        sum_a_x_a_f64x2 = vaddq_f64(sum_a_x_a_f64x2, a1_x_f64x2);
+        sum_a_y_a_f64x2 = vaddq_f64(sum_a_y_a_f64x2, a1_y_f64x2);
+        sum_a_z_a_f64x2 = vaddq_f64(sum_a_z_a_f64x2, a1_z_f64x2);
+        sum_b_x_a_f64x2 = vaddq_f64(sum_b_x_a_f64x2, b1_x_f64x2);
+        sum_b_y_a_f64x2 = vaddq_f64(sum_b_y_a_f64x2, b1_y_f64x2);
+        sum_b_z_a_f64x2 = vaddq_f64(sum_b_z_a_f64x2, b1_z_f64x2);
+        cov_xx_a_f64x2 = vfmaq_f64(cov_xx_a_f64x2, a1_x_f64x2, b1_x_f64x2);
+        cov_xy_a_f64x2 = vfmaq_f64(cov_xy_a_f64x2, a1_x_f64x2, b1_y_f64x2);
+        cov_xz_a_f64x2 = vfmaq_f64(cov_xz_a_f64x2, a1_x_f64x2, b1_z_f64x2);
+        cov_yx_a_f64x2 = vfmaq_f64(cov_yx_a_f64x2, a1_y_f64x2, b1_x_f64x2);
+        cov_yy_a_f64x2 = vfmaq_f64(cov_yy_a_f64x2, a1_y_f64x2, b1_y_f64x2);
+        cov_yz_a_f64x2 = vfmaq_f64(cov_yz_a_f64x2, a1_y_f64x2, b1_z_f64x2);
+        cov_zx_a_f64x2 = vfmaq_f64(cov_zx_a_f64x2, a1_z_f64x2, b1_x_f64x2);
+        cov_zy_a_f64x2 = vfmaq_f64(cov_zy_a_f64x2, a1_z_f64x2, b1_y_f64x2);
+        cov_zz_a_f64x2 = vfmaq_f64(cov_zz_a_f64x2, a1_z_f64x2, b1_z_f64x2);
+        variance_a_a_f64x2 = vfmaq_f64(variance_a_a_f64x2, a1_x_f64x2, a1_x_f64x2);
+        variance_a_a_f64x2 = vfmaq_f64(variance_a_a_f64x2, a1_y_f64x2, a1_y_f64x2);
+        variance_a_a_f64x2 = vfmaq_f64(variance_a_a_f64x2, a1_z_f64x2, a1_z_f64x2);
+    }
+
+    // Combine dual accumulators
+    float64x2_t sum_a_x_f64x2 = vaddq_f64(sum_a_x_a_f64x2, sum_a_x_b_f64x2);
+    float64x2_t sum_a_y_f64x2 = vaddq_f64(sum_a_y_a_f64x2, sum_a_y_b_f64x2);
+    float64x2_t sum_a_z_f64x2 = vaddq_f64(sum_a_z_a_f64x2, sum_a_z_b_f64x2);
+    float64x2_t sum_b_x_f64x2 = vaddq_f64(sum_b_x_a_f64x2, sum_b_x_b_f64x2);
+    float64x2_t sum_b_y_f64x2 = vaddq_f64(sum_b_y_a_f64x2, sum_b_y_b_f64x2);
+    float64x2_t sum_b_z_f64x2 = vaddq_f64(sum_b_z_a_f64x2, sum_b_z_b_f64x2);
+    float64x2_t cov_xx_f64x2 = vaddq_f64(cov_xx_a_f64x2, cov_xx_b_f64x2);
+    float64x2_t cov_xy_f64x2 = vaddq_f64(cov_xy_a_f64x2, cov_xy_b_f64x2);
+    float64x2_t cov_xz_f64x2 = vaddq_f64(cov_xz_a_f64x2, cov_xz_b_f64x2);
+    float64x2_t cov_yx_f64x2 = vaddq_f64(cov_yx_a_f64x2, cov_yx_b_f64x2);
+    float64x2_t cov_yy_f64x2 = vaddq_f64(cov_yy_a_f64x2, cov_yy_b_f64x2);
+    float64x2_t cov_yz_f64x2 = vaddq_f64(cov_yz_a_f64x2, cov_yz_b_f64x2);
+    float64x2_t cov_zx_f64x2 = vaddq_f64(cov_zx_a_f64x2, cov_zx_b_f64x2);
+    float64x2_t cov_zy_f64x2 = vaddq_f64(cov_zy_a_f64x2, cov_zy_b_f64x2);
+    float64x2_t cov_zz_f64x2 = vaddq_f64(cov_zz_a_f64x2, cov_zz_b_f64x2);
+    float64x2_t variance_a_f64x2 = vaddq_f64(variance_a_a_f64x2, variance_a_b_f64x2);
+
+    // Reduce vector accumulators.
+    nk_f64_t sum_a_x = nk_reduce_stable_f64x2_neon_(sum_a_x_f64x2), sum_a_x_compensation = 0.0;
+    nk_f64_t sum_a_y = nk_reduce_stable_f64x2_neon_(sum_a_y_f64x2), sum_a_y_compensation = 0.0;
+    nk_f64_t sum_a_z = nk_reduce_stable_f64x2_neon_(sum_a_z_f64x2), sum_a_z_compensation = 0.0;
+    nk_f64_t sum_b_x = nk_reduce_stable_f64x2_neon_(sum_b_x_f64x2), sum_b_x_compensation = 0.0;
+    nk_f64_t sum_b_y = nk_reduce_stable_f64x2_neon_(sum_b_y_f64x2), sum_b_y_compensation = 0.0;
+    nk_f64_t sum_b_z = nk_reduce_stable_f64x2_neon_(sum_b_z_f64x2), sum_b_z_compensation = 0.0;
+    nk_f64_t covariance_x_x = nk_reduce_stable_f64x2_neon_(cov_xx_f64x2), covariance_x_x_compensation = 0.0;
+    nk_f64_t covariance_x_y = nk_reduce_stable_f64x2_neon_(cov_xy_f64x2), covariance_x_y_compensation = 0.0;
+    nk_f64_t covariance_x_z = nk_reduce_stable_f64x2_neon_(cov_xz_f64x2), covariance_x_z_compensation = 0.0;
+    nk_f64_t covariance_y_x = nk_reduce_stable_f64x2_neon_(cov_yx_f64x2), covariance_y_x_compensation = 0.0;
+    nk_f64_t covariance_y_y = nk_reduce_stable_f64x2_neon_(cov_yy_f64x2), covariance_y_y_compensation = 0.0;
+    nk_f64_t covariance_y_z = nk_reduce_stable_f64x2_neon_(cov_yz_f64x2), covariance_y_z_compensation = 0.0;
+    nk_f64_t covariance_z_x = nk_reduce_stable_f64x2_neon_(cov_zx_f64x2), covariance_z_x_compensation = 0.0;
+    nk_f64_t covariance_z_y = nk_reduce_stable_f64x2_neon_(cov_zy_f64x2), covariance_z_y_compensation = 0.0;
+    nk_f64_t covariance_z_z = nk_reduce_stable_f64x2_neon_(cov_zz_f64x2), covariance_z_z_compensation = 0.0;
+    nk_f64_t sum_sq_a = nk_reduce_stable_f64x2_neon_(variance_a_f64x2), sum_sq_a_compensation = 0.0;
+
+    // Scalar tail
+    for (; i < n; ++i) {
+        nk_f64_t ax = a[i * 3 + 0], ay = a[i * 3 + 1], az = a[i * 3 + 2];
+        nk_f64_t bx = b[i * 3 + 0], by = b[i * 3 + 1], bz = b[i * 3 + 2];
+        nk_accumulate_sum_f64_(&sum_a_x, &sum_a_x_compensation, ax),
+            nk_accumulate_sum_f64_(&sum_a_y, &sum_a_y_compensation, ay),
+            nk_accumulate_sum_f64_(&sum_a_z, &sum_a_z_compensation, az);
+        nk_accumulate_sum_f64_(&sum_b_x, &sum_b_x_compensation, bx),
+            nk_accumulate_sum_f64_(&sum_b_y, &sum_b_y_compensation, by),
+            nk_accumulate_sum_f64_(&sum_b_z, &sum_b_z_compensation, bz);
+        nk_accumulate_product_f64_(&covariance_x_x, &covariance_x_x_compensation, ax, bx),
+            nk_accumulate_product_f64_(&covariance_x_y, &covariance_x_y_compensation, ax, by),
+            nk_accumulate_product_f64_(&covariance_x_z, &covariance_x_z_compensation, ax, bz);
+        nk_accumulate_product_f64_(&covariance_y_x, &covariance_y_x_compensation, ay, bx),
+            nk_accumulate_product_f64_(&covariance_y_y, &covariance_y_y_compensation, ay, by),
+            nk_accumulate_product_f64_(&covariance_y_z, &covariance_y_z_compensation, ay, bz);
+        nk_accumulate_product_f64_(&covariance_z_x, &covariance_z_x_compensation, az, bx),
+            nk_accumulate_product_f64_(&covariance_z_y, &covariance_z_y_compensation, az, by),
+            nk_accumulate_product_f64_(&covariance_z_z, &covariance_z_z_compensation, az, bz);
+        nk_accumulate_square_f64_(&sum_sq_a, &sum_sq_a_compensation, ax),
+            nk_accumulate_square_f64_(&sum_sq_a, &sum_sq_a_compensation, ay),
+            nk_accumulate_square_f64_(&sum_sq_a, &sum_sq_a_compensation, az);
+    }
+
+    sum_a_x += sum_a_x_compensation, sum_a_y += sum_a_y_compensation, sum_a_z += sum_a_z_compensation;
+    sum_b_x += sum_b_x_compensation, sum_b_y += sum_b_y_compensation, sum_b_z += sum_b_z_compensation;
+    covariance_x_x += covariance_x_x_compensation, covariance_x_y += covariance_x_y_compensation,
+        covariance_x_z += covariance_x_z_compensation;
+    covariance_y_x += covariance_y_x_compensation, covariance_y_y += covariance_y_y_compensation,
+        covariance_y_z += covariance_y_z_compensation;
+    covariance_z_x += covariance_z_x_compensation, covariance_z_y += covariance_z_y_compensation,
+        covariance_z_z += covariance_z_z_compensation;
+    sum_sq_a += sum_sq_a_compensation;
+
+    // Compute centroids
+    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
+    nk_f64_t centroid_a_x = sum_a_x * inv_n, centroid_a_y = sum_a_y * inv_n, centroid_a_z = sum_a_z * inv_n;
+    nk_f64_t centroid_b_x = sum_b_x * inv_n, centroid_b_y = sum_b_y * inv_n, centroid_b_z = sum_b_z * inv_n;
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Compute variance of A (centered)
+    nk_f64_t centroid_sq = centroid_a_x * centroid_a_x + centroid_a_y * centroid_a_y + centroid_a_z * centroid_a_z;
+    nk_f64_t var_a = sum_sq_a * inv_n - centroid_sq;
+
+    // Apply centering correction: H_centered = H - n * centroid_a * centroid_bᵀ
+    covariance_x_x -= (nk_f64_t)n * centroid_a_x * centroid_b_x;
+    covariance_x_y -= (nk_f64_t)n * centroid_a_x * centroid_b_y;
+    covariance_x_z -= (nk_f64_t)n * centroid_a_x * centroid_b_z;
+    covariance_y_x -= (nk_f64_t)n * centroid_a_y * centroid_b_x;
+    covariance_y_y -= (nk_f64_t)n * centroid_a_y * centroid_b_y;
+    covariance_y_z -= (nk_f64_t)n * centroid_a_y * centroid_b_z;
+    covariance_z_x -= (nk_f64_t)n * centroid_a_z * centroid_b_x;
+    covariance_z_y -= (nk_f64_t)n * centroid_a_z * centroid_b_y;
+    covariance_z_z -= (nk_f64_t)n * centroid_a_z * centroid_b_z;
+
+    // Compute SVD
+    nk_f64_t cross_covariance[9] = {covariance_x_x, covariance_x_y, covariance_x_z, covariance_y_x, covariance_y_y,
+                                    covariance_y_z, covariance_z_x, covariance_z_y, covariance_z_z};
+    nk_f64_t svd_u[9], svd_s[9], svd_v[9];
+    nk_svd3x3_f64_(cross_covariance, svd_u, svd_s, svd_v);
+
+    nk_f64_t r[9];
+    nk_rotation_from_svd_f64_neon_(svd_u, svd_v, r);
+
+    // Handle reflection and compute scale
+    nk_f64_t det = nk_det3x3_f64_(r);
+    nk_f64_t trace_d_s = svd_s[0] + svd_s[4] + (det < 0 ? -svd_s[8] : svd_s[8]);
+    nk_f64_t computed_scale = trace_d_s / ((nk_f64_t)n * var_a);
+
+    if (det < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        nk_rotation_from_svd_f64_neon_(svd_u, svd_v, r);
+    }
+
+    if (rotation)
+        for (int j = 0; j < 9; ++j) rotation[j] = r[j];
+    if (scale) *scale = computed_scale;
+
+    // Compute RMSD after transformation
+    nk_f64_t sum_squared = nk_transformed_ssd_f64_neon_(a, b, n, r, computed_scale, centroid_a_x, centroid_a_y,
+                                                        centroid_a_z, centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f64_sqrt_neon(sum_squared * inv_n);
+}
+
+NK_INTERNAL void nk_deinterleave_f16x8_to_f32x4x2_neon_(nk_f16_t const *ptr,                             //
+                                                        float32x4_t *x_low_out, float32x4_t *x_high_out, //
+                                                        float32x4_t *y_low_out, float32x4_t *y_high_out, //
+                                                        float32x4_t *z_low_out, float32x4_t *z_high_out) {
+    // Deinterleave 24 f16 values (8 xyz triplets) into separate x, y, z vectors.
+    // Uses NEON vld3q_u16 for efficient stride-3 deinterleaving, then converts to f32.
+    // Avoids vld3q_f16 which is unavailable on MSVC for ARM.
+    //
+    // Input: 24 contiguous f16 values [x0,y0,z0, ..., x7,y7,z7]
+    // Output: x_low[4]+x_high[4], y_low[4]+y_high[4], z_low[4]+z_high[4] vectors in f32
+    uint16x8x3_t xyz_u16x8x3 = vld3q_u16((nk_u16_t const *)ptr);
+    float16x8_t x_f16x8 = vreinterpretq_f16_u16(xyz_u16x8x3.val[0]);
+    float16x8_t y_f16x8 = vreinterpretq_f16_u16(xyz_u16x8x3.val[1]);
+    float16x8_t z_f16x8 = vreinterpretq_f16_u16(xyz_u16x8x3.val[2]);
+    *x_low_out = vcvt_f32_f16(vget_low_f16(x_f16x8));
+    *x_high_out = vcvt_high_f32_f16(x_f16x8);
+    *y_low_out = vcvt_f32_f16(vget_low_f16(y_f16x8));
+    *y_high_out = vcvt_high_f32_f16(y_f16x8);
+    *z_low_out = vcvt_f32_f16(vget_low_f16(z_f16x8));
+    *z_high_out = vcvt_high_f32_f16(z_f16x8);
+}
+
+NK_INTERNAL void nk_partial_deinterleave_f16_to_f32x4x2_neon_(nk_f16_t const *ptr, nk_size_t n_points,         //
+                                                              float32x4_t *x_low_out, float32x4_t *x_high_out, //
+                                                              float32x4_t *y_low_out, float32x4_t *y_high_out, //
+                                                              float32x4_t *z_low_out, float32x4_t *z_high_out) {
+    nk_u16_t buf[24] = {0};
+    nk_u16_t const *src = (nk_u16_t const *)ptr;
+    for (nk_size_t k = 0; k < n_points * 3; ++k) buf[k] = src[k];
+    nk_deinterleave_f16x8_to_f32x4x2_neon_((nk_f16_t const *)buf, x_low_out, x_high_out, y_low_out, y_high_out,
+                                           z_low_out, z_high_out);
+}
+
+NK_INTERNAL nk_f32_t nk_transformed_ssd_f16_neon_(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t const *r,
+                                                  nk_f32_t scale, nk_f32_t centroid_a_x, nk_f32_t centroid_a_y,
+                                                  nk_f32_t centroid_a_z, nk_f32_t centroid_b_x, nk_f32_t centroid_b_y,
+                                                  nk_f32_t centroid_b_z) {
+    // Compute sum of squared differences after rigid transformation.
+    // Used by Kabsch algorithm for RMSD computation after rotation is applied.
+    float32x4_t const centroid_a_x_f32x4 = vdupq_n_f32(centroid_a_x);
+    float32x4_t const centroid_a_y_f32x4 = vdupq_n_f32(centroid_a_y);
+    float32x4_t const centroid_a_z_f32x4 = vdupq_n_f32(centroid_a_z);
+    float32x4_t const centroid_b_x_f32x4 = vdupq_n_f32(centroid_b_x);
+    float32x4_t const centroid_b_y_f32x4 = vdupq_n_f32(centroid_b_y);
+    float32x4_t const centroid_b_z_f32x4 = vdupq_n_f32(centroid_b_z);
+    float32x4_t const scale_f32x4 = vdupq_n_f32(scale);
+
+    // Load rotation matrix elements
+    float32x4_t const r00_f32x4 = vdupq_n_f32(r[0]), r01_f32x4 = vdupq_n_f32(r[1]), r02_f32x4 = vdupq_n_f32(r[2]);
+    float32x4_t const r10_f32x4 = vdupq_n_f32(r[3]), r11_f32x4 = vdupq_n_f32(r[4]), r12_f32x4 = vdupq_n_f32(r[5]);
+    float32x4_t const r20_f32x4 = vdupq_n_f32(r[6]), r21_f32x4 = vdupq_n_f32(r[7]), r22_f32x4 = vdupq_n_f32(r[8]);
+
+    float32x4_t sum_squared_f32x4 = vdupq_n_f32(0);
+    float32x4_t a_x_low_f32x4, a_x_high_f32x4, a_y_low_f32x4, a_y_high_f32x4, a_z_low_f32x4, a_z_high_f32x4;
+    float32x4_t b_x_low_f32x4, b_x_high_f32x4, b_y_low_f32x4, b_y_high_f32x4, b_z_low_f32x4, b_z_high_f32x4;
+
+    nk_size_t j = 0;
+    for (; j + 8 <= n; j += 8) {
+        nk_deinterleave_f16x8_to_f32x4x2_neon_(a + j * 3, &a_x_low_f32x4, &a_x_high_f32x4, &a_y_low_f32x4,
+                                               &a_y_high_f32x4, &a_z_low_f32x4, &a_z_high_f32x4);
+        nk_deinterleave_f16x8_to_f32x4x2_neon_(b + j * 3, &b_x_low_f32x4, &b_x_high_f32x4, &b_y_low_f32x4,
+                                               &b_y_high_f32x4, &b_z_low_f32x4, &b_z_high_f32x4);
+
+        // Center points → low half
+        float32x4_t pa_x_f32x4 = vsubq_f32(a_x_low_f32x4, centroid_a_x_f32x4);
+        float32x4_t pa_y_f32x4 = vsubq_f32(a_y_low_f32x4, centroid_a_y_f32x4);
+        float32x4_t pa_z_f32x4 = vsubq_f32(a_z_low_f32x4, centroid_a_z_f32x4);
+        float32x4_t pb_x_f32x4 = vsubq_f32(b_x_low_f32x4, centroid_b_x_f32x4);
+        float32x4_t pb_y_f32x4 = vsubq_f32(b_y_low_f32x4, centroid_b_y_f32x4);
+        float32x4_t pb_z_f32x4 = vsubq_f32(b_z_low_f32x4, centroid_b_z_f32x4);
+        float32x4_t ra_x_f32x4 = vmulq_f32(
+            scale_f32x4,
+            vfmaq_f32(vfmaq_f32(vmulq_f32(r00_f32x4, pa_x_f32x4), r01_f32x4, pa_y_f32x4), r02_f32x4, pa_z_f32x4));
+        float32x4_t ra_y_f32x4 = vmulq_f32(
+            scale_f32x4,
+            vfmaq_f32(vfmaq_f32(vmulq_f32(r10_f32x4, pa_x_f32x4), r11_f32x4, pa_y_f32x4), r12_f32x4, pa_z_f32x4));
+        float32x4_t ra_z_f32x4 = vmulq_f32(
+            scale_f32x4,
+            vfmaq_f32(vfmaq_f32(vmulq_f32(r20_f32x4, pa_x_f32x4), r21_f32x4, pa_y_f32x4), r22_f32x4, pa_z_f32x4));
+        float32x4_t delta_x_f32x4 = vsubq_f32(ra_x_f32x4, pb_x_f32x4);
+        float32x4_t delta_y_f32x4 = vsubq_f32(ra_y_f32x4, pb_y_f32x4);
+        float32x4_t delta_z_f32x4 = vsubq_f32(ra_z_f32x4, pb_z_f32x4);
+        sum_squared_f32x4 = vfmaq_f32(sum_squared_f32x4, delta_x_f32x4, delta_x_f32x4);
+        sum_squared_f32x4 = vfmaq_f32(sum_squared_f32x4, delta_y_f32x4, delta_y_f32x4);
+        sum_squared_f32x4 = vfmaq_f32(sum_squared_f32x4, delta_z_f32x4, delta_z_f32x4);
+
+        // Center points → high half
+        pa_x_f32x4 = vsubq_f32(a_x_high_f32x4, centroid_a_x_f32x4);
+        pa_y_f32x4 = vsubq_f32(a_y_high_f32x4, centroid_a_y_f32x4);
+        pa_z_f32x4 = vsubq_f32(a_z_high_f32x4, centroid_a_z_f32x4);
+        pb_x_f32x4 = vsubq_f32(b_x_high_f32x4, centroid_b_x_f32x4);
+        pb_y_f32x4 = vsubq_f32(b_y_high_f32x4, centroid_b_y_f32x4);
+        pb_z_f32x4 = vsubq_f32(b_z_high_f32x4, centroid_b_z_f32x4);
+        ra_x_f32x4 = vmulq_f32(
+            scale_f32x4,
+            vfmaq_f32(vfmaq_f32(vmulq_f32(r00_f32x4, pa_x_f32x4), r01_f32x4, pa_y_f32x4), r02_f32x4, pa_z_f32x4));
+        ra_y_f32x4 = vmulq_f32(
+            scale_f32x4,
+            vfmaq_f32(vfmaq_f32(vmulq_f32(r10_f32x4, pa_x_f32x4), r11_f32x4, pa_y_f32x4), r12_f32x4, pa_z_f32x4));
+        ra_z_f32x4 = vmulq_f32(
+            scale_f32x4,
+            vfmaq_f32(vfmaq_f32(vmulq_f32(r20_f32x4, pa_x_f32x4), r21_f32x4, pa_y_f32x4), r22_f32x4, pa_z_f32x4));
+        delta_x_f32x4 = vsubq_f32(ra_x_f32x4, pb_x_f32x4);
+        delta_y_f32x4 = vsubq_f32(ra_y_f32x4, pb_y_f32x4);
+        delta_z_f32x4 = vsubq_f32(ra_z_f32x4, pb_z_f32x4);
+        sum_squared_f32x4 = vfmaq_f32(sum_squared_f32x4, delta_x_f32x4, delta_x_f32x4);
+        sum_squared_f32x4 = vfmaq_f32(sum_squared_f32x4, delta_y_f32x4, delta_y_f32x4);
+        sum_squared_f32x4 = vfmaq_f32(sum_squared_f32x4, delta_z_f32x4, delta_z_f32x4);
+    }
+
+    // Reduce to scalar
+    nk_f32_t sum_squared = vaddvq_f32(sum_squared_f32x4);
+
+    if (j < n) {
+        nk_partial_deinterleave_f16_to_f32x4x2_neon_(a + j * 3, n - j, &a_x_low_f32x4, &a_x_high_f32x4, &a_y_low_f32x4,
+                                                     &a_y_high_f32x4, &a_z_low_f32x4, &a_z_high_f32x4);
+        nk_partial_deinterleave_f16_to_f32x4x2_neon_(b + j * 3, n - j, &b_x_low_f32x4, &b_x_high_f32x4, &b_y_low_f32x4,
+                                                     &b_y_high_f32x4, &b_z_low_f32x4, &b_z_high_f32x4);
+
+        // Low half
+        float32x4_t pa_x_f32x4 = vsubq_f32(a_x_low_f32x4, centroid_a_x_f32x4);
+        float32x4_t pa_y_f32x4 = vsubq_f32(a_y_low_f32x4, centroid_a_y_f32x4);
+        float32x4_t pa_z_f32x4 = vsubq_f32(a_z_low_f32x4, centroid_a_z_f32x4);
+        float32x4_t pb_x_f32x4 = vsubq_f32(b_x_low_f32x4, centroid_b_x_f32x4);
+        float32x4_t pb_y_f32x4 = vsubq_f32(b_y_low_f32x4, centroid_b_y_f32x4);
+        float32x4_t pb_z_f32x4 = vsubq_f32(b_z_low_f32x4, centroid_b_z_f32x4);
+        float32x4_t ra_x_f32x4 = vmulq_f32(
+            scale_f32x4,
+            vfmaq_f32(vfmaq_f32(vmulq_f32(r00_f32x4, pa_x_f32x4), r01_f32x4, pa_y_f32x4), r02_f32x4, pa_z_f32x4));
+        float32x4_t ra_y_f32x4 = vmulq_f32(
+            scale_f32x4,
+            vfmaq_f32(vfmaq_f32(vmulq_f32(r10_f32x4, pa_x_f32x4), r11_f32x4, pa_y_f32x4), r12_f32x4, pa_z_f32x4));
+        float32x4_t ra_z_f32x4 = vmulq_f32(
+            scale_f32x4,
+            vfmaq_f32(vfmaq_f32(vmulq_f32(r20_f32x4, pa_x_f32x4), r21_f32x4, pa_y_f32x4), r22_f32x4, pa_z_f32x4));
+        float32x4_t delta_x_f32x4 = vsubq_f32(ra_x_f32x4, pb_x_f32x4);
+        float32x4_t delta_y_f32x4 = vsubq_f32(ra_y_f32x4, pb_y_f32x4);
+        float32x4_t delta_z_f32x4 = vsubq_f32(ra_z_f32x4, pb_z_f32x4);
+        float32x4_t tail_sum_f32x4 = vmulq_f32(delta_x_f32x4, delta_x_f32x4);
+        tail_sum_f32x4 = vfmaq_f32(tail_sum_f32x4, delta_y_f32x4, delta_y_f32x4);
+        tail_sum_f32x4 = vfmaq_f32(tail_sum_f32x4, delta_z_f32x4, delta_z_f32x4);
+
+        // High half
+        pa_x_f32x4 = vsubq_f32(a_x_high_f32x4, centroid_a_x_f32x4);
+        pa_y_f32x4 = vsubq_f32(a_y_high_f32x4, centroid_a_y_f32x4);
+        pa_z_f32x4 = vsubq_f32(a_z_high_f32x4, centroid_a_z_f32x4);
+        pb_x_f32x4 = vsubq_f32(b_x_high_f32x4, centroid_b_x_f32x4);
+        pb_y_f32x4 = vsubq_f32(b_y_high_f32x4, centroid_b_y_f32x4);
+        pb_z_f32x4 = vsubq_f32(b_z_high_f32x4, centroid_b_z_f32x4);
+        ra_x_f32x4 = vmulq_f32(
+            scale_f32x4,
+            vfmaq_f32(vfmaq_f32(vmulq_f32(r00_f32x4, pa_x_f32x4), r01_f32x4, pa_y_f32x4), r02_f32x4, pa_z_f32x4));
+        ra_y_f32x4 = vmulq_f32(
+            scale_f32x4,
+            vfmaq_f32(vfmaq_f32(vmulq_f32(r10_f32x4, pa_x_f32x4), r11_f32x4, pa_y_f32x4), r12_f32x4, pa_z_f32x4));
+        ra_z_f32x4 = vmulq_f32(
+            scale_f32x4,
+            vfmaq_f32(vfmaq_f32(vmulq_f32(r20_f32x4, pa_x_f32x4), r21_f32x4, pa_y_f32x4), r22_f32x4, pa_z_f32x4));
+        delta_x_f32x4 = vsubq_f32(ra_x_f32x4, pb_x_f32x4);
+        delta_y_f32x4 = vsubq_f32(ra_y_f32x4, pb_y_f32x4);
+        delta_z_f32x4 = vsubq_f32(ra_z_f32x4, pb_z_f32x4);
+        tail_sum_f32x4 = vfmaq_f32(tail_sum_f32x4, delta_x_f32x4, delta_x_f32x4);
+        tail_sum_f32x4 = vfmaq_f32(tail_sum_f32x4, delta_y_f32x4, delta_y_f32x4);
+        tail_sum_f32x4 = vfmaq_f32(tail_sum_f32x4, delta_z_f32x4, delta_z_f32x4);
+        sum_squared += vaddvq_f32(tail_sum_f32x4);
+    }
+
+    return sum_squared;
+}
+
+/**
+ *  @brief RMSD (Root Mean Square Deviation) computation using NEON FP16 with widening to FP32.
+ *  Computes the RMS of distances between corresponding points after centroid alignment.
+ */
+NK_PUBLIC void nk_rmsd_f16_neon(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
+    // RMSD uses identity rotation and scale=1.0
+    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 = 1.0f;
+
+    float32x4_t const zeros_f32x4 = vdupq_n_f32(0);
+
+    // Accumulators for centroids and squared differences (all in f32)
+    float32x4_t sum_a_x_f32x4 = zeros_f32x4, sum_a_y_f32x4 = zeros_f32x4, sum_a_z_f32x4 = zeros_f32x4;
+    float32x4_t sum_b_x_f32x4 = zeros_f32x4, sum_b_y_f32x4 = zeros_f32x4, sum_b_z_f32x4 = zeros_f32x4;
+    float32x4_t sum_squared_x_f32x4 = zeros_f32x4, sum_squared_y_f32x4 = zeros_f32x4, sum_squared_z_f32x4 = zeros_f32x4;
+
+    float32x4_t a_x_low_f32x4, a_x_high_f32x4, a_y_low_f32x4, a_y_high_f32x4, a_z_low_f32x4, a_z_high_f32x4;
+    float32x4_t b_x_low_f32x4, b_x_high_f32x4, b_y_low_f32x4, b_y_high_f32x4, b_z_low_f32x4, b_z_high_f32x4;
+    nk_size_t i = 0;
+
+    for (; i + 8 <= n; i += 8) {
+        nk_deinterleave_f16x8_to_f32x4x2_neon_(a + i * 3, &a_x_low_f32x4, &a_x_high_f32x4, &a_y_low_f32x4,
+                                               &a_y_high_f32x4, &a_z_low_f32x4, &a_z_high_f32x4);
+        nk_deinterleave_f16x8_to_f32x4x2_neon_(b + i * 3, &b_x_low_f32x4, &b_x_high_f32x4, &b_y_low_f32x4,
+                                               &b_y_high_f32x4, &b_z_low_f32x4, &b_z_high_f32x4);
+
+        sum_a_x_f32x4 = vaddq_f32(vaddq_f32(sum_a_x_f32x4, a_x_low_f32x4), a_x_high_f32x4);
+        sum_a_y_f32x4 = vaddq_f32(vaddq_f32(sum_a_y_f32x4, a_y_low_f32x4), a_y_high_f32x4);
+        sum_a_z_f32x4 = vaddq_f32(vaddq_f32(sum_a_z_f32x4, a_z_low_f32x4), a_z_high_f32x4);
+        sum_b_x_f32x4 = vaddq_f32(vaddq_f32(sum_b_x_f32x4, b_x_low_f32x4), b_x_high_f32x4);
+        sum_b_y_f32x4 = vaddq_f32(vaddq_f32(sum_b_y_f32x4, b_y_low_f32x4), b_y_high_f32x4);
+        sum_b_z_f32x4 = vaddq_f32(vaddq_f32(sum_b_z_f32x4, b_z_low_f32x4), b_z_high_f32x4);
+
+        float32x4_t delta_x_f32x4 = vsubq_f32(a_x_low_f32x4, b_x_low_f32x4);
+        float32x4_t delta_y_f32x4 = vsubq_f32(a_y_low_f32x4, b_y_low_f32x4);
+        float32x4_t delta_z_f32x4 = vsubq_f32(a_z_low_f32x4, b_z_low_f32x4);
+        sum_squared_x_f32x4 = vfmaq_f32(sum_squared_x_f32x4, delta_x_f32x4, delta_x_f32x4);
+        sum_squared_y_f32x4 = vfmaq_f32(sum_squared_y_f32x4, delta_y_f32x4, delta_y_f32x4);
+        sum_squared_z_f32x4 = vfmaq_f32(sum_squared_z_f32x4, delta_z_f32x4, delta_z_f32x4);
+
+        delta_x_f32x4 = vsubq_f32(a_x_high_f32x4, b_x_high_f32x4);
+        delta_y_f32x4 = vsubq_f32(a_y_high_f32x4, b_y_high_f32x4);
+        delta_z_f32x4 = vsubq_f32(a_z_high_f32x4, b_z_high_f32x4);
+        sum_squared_x_f32x4 = vfmaq_f32(sum_squared_x_f32x4, delta_x_f32x4, delta_x_f32x4);
+        sum_squared_y_f32x4 = vfmaq_f32(sum_squared_y_f32x4, delta_y_f32x4, delta_y_f32x4);
+        sum_squared_z_f32x4 = vfmaq_f32(sum_squared_z_f32x4, delta_z_f32x4, delta_z_f32x4);
+    }
+
+    if (i < n) {
+        nk_partial_deinterleave_f16_to_f32x4x2_neon_(a + i * 3, n - i, &a_x_low_f32x4, &a_x_high_f32x4, &a_y_low_f32x4,
+                                                     &a_y_high_f32x4, &a_z_low_f32x4, &a_z_high_f32x4);
+        nk_partial_deinterleave_f16_to_f32x4x2_neon_(b + i * 3, n - i, &b_x_low_f32x4, &b_x_high_f32x4, &b_y_low_f32x4,
+                                                     &b_y_high_f32x4, &b_z_low_f32x4, &b_z_high_f32x4);
+
+        sum_a_x_f32x4 = vaddq_f32(vaddq_f32(sum_a_x_f32x4, a_x_low_f32x4), a_x_high_f32x4);
+        sum_a_y_f32x4 = vaddq_f32(vaddq_f32(sum_a_y_f32x4, a_y_low_f32x4), a_y_high_f32x4);
+        sum_a_z_f32x4 = vaddq_f32(vaddq_f32(sum_a_z_f32x4, a_z_low_f32x4), a_z_high_f32x4);
+        sum_b_x_f32x4 = vaddq_f32(vaddq_f32(sum_b_x_f32x4, b_x_low_f32x4), b_x_high_f32x4);
+        sum_b_y_f32x4 = vaddq_f32(vaddq_f32(sum_b_y_f32x4, b_y_low_f32x4), b_y_high_f32x4);
+        sum_b_z_f32x4 = vaddq_f32(vaddq_f32(sum_b_z_f32x4, b_z_low_f32x4), b_z_high_f32x4);
+
+        float32x4_t delta_x_f32x4 = vsubq_f32(a_x_low_f32x4, b_x_low_f32x4);
+        float32x4_t delta_y_f32x4 = vsubq_f32(a_y_low_f32x4, b_y_low_f32x4);
+        float32x4_t delta_z_f32x4 = vsubq_f32(a_z_low_f32x4, b_z_low_f32x4);
+        sum_squared_x_f32x4 = vfmaq_f32(sum_squared_x_f32x4, delta_x_f32x4, delta_x_f32x4);
+        sum_squared_y_f32x4 = vfmaq_f32(sum_squared_y_f32x4, delta_y_f32x4, delta_y_f32x4);
+        sum_squared_z_f32x4 = vfmaq_f32(sum_squared_z_f32x4, delta_z_f32x4, delta_z_f32x4);
+
+        delta_x_f32x4 = vsubq_f32(a_x_high_f32x4, b_x_high_f32x4);
+        delta_y_f32x4 = vsubq_f32(a_y_high_f32x4, b_y_high_f32x4);
+        delta_z_f32x4 = vsubq_f32(a_z_high_f32x4, b_z_high_f32x4);
+        sum_squared_x_f32x4 = vfmaq_f32(sum_squared_x_f32x4, delta_x_f32x4, delta_x_f32x4);
+        sum_squared_y_f32x4 = vfmaq_f32(sum_squared_y_f32x4, delta_y_f32x4, delta_y_f32x4);
+        sum_squared_z_f32x4 = vfmaq_f32(sum_squared_z_f32x4, delta_z_f32x4, delta_z_f32x4);
+    }
+
+    // Reduce vectors to scalars
+    nk_f32_t total_ax = vaddvq_f32(sum_a_x_f32x4);
+    nk_f32_t total_ay = vaddvq_f32(sum_a_y_f32x4);
+    nk_f32_t total_az = vaddvq_f32(sum_a_z_f32x4);
+    nk_f32_t total_bx = vaddvq_f32(sum_b_x_f32x4);
+    nk_f32_t total_by = vaddvq_f32(sum_b_y_f32x4);
+    nk_f32_t total_bz = vaddvq_f32(sum_b_z_f32x4);
+    nk_f32_t total_sq_x = vaddvq_f32(sum_squared_x_f32x4);
+    nk_f32_t total_sq_y = vaddvq_f32(sum_squared_y_f32x4);
+    nk_f32_t total_sq_z = vaddvq_f32(sum_squared_z_f32x4);
+
+    // Compute centroids
+    nk_f32_t inv_n = 1.0f / (nk_f32_t)n;
+    nk_f32_t centroid_a_x = total_ax * inv_n;
+    nk_f32_t centroid_a_y = total_ay * inv_n;
+    nk_f32_t centroid_a_z = total_az * inv_n;
+    nk_f32_t centroid_b_x = total_bx * inv_n;
+    nk_f32_t centroid_b_y = total_by * inv_n;
+    nk_f32_t centroid_b_z = total_bz * inv_n;
+
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Compute RMSD
+    nk_f32_t mean_diff_x = centroid_a_x - centroid_b_x;
+    nk_f32_t mean_diff_y = centroid_a_y - centroid_b_y;
+    nk_f32_t mean_diff_z = centroid_a_z - centroid_b_z;
+    nk_f32_t sum_squared = total_sq_x + total_sq_y + total_sq_z;
+    nk_f32_t mean_diff_sq = mean_diff_x * mean_diff_x + mean_diff_y * mean_diff_y + mean_diff_z * mean_diff_z;
+
+    *result = nk_f32_sqrt_neon(sum_squared * inv_n - mean_diff_sq);
+}
+
+/**
+ *  @brief Kabsch algorithm for optimal rigid body superposition using NEON FP16 with widening to FP32.
+ *  Finds the rotation matrix R that minimizes RMSD between two point sets.
+ */
+NK_PUBLIC void nk_kabsch_f16_neon(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                  nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
+    // Fused single-pass: load f16, convert to f32, compute centroids and covariance
+    float32x4_t const zeros_f32x4 = vdupq_n_f32(0);
+
+    // Accumulators for centroids (f32)
+    float32x4_t sum_a_x_f32x4 = zeros_f32x4, sum_a_y_f32x4 = zeros_f32x4, sum_a_z_f32x4 = zeros_f32x4;
+    float32x4_t sum_b_x_f32x4 = zeros_f32x4, sum_b_y_f32x4 = zeros_f32x4, sum_b_z_f32x4 = zeros_f32x4;
+
+    // Accumulators for covariance matrix (sum of outer products)
+    float32x4_t cov_xx_f32x4 = zeros_f32x4, cov_xy_f32x4 = zeros_f32x4, cov_xz_f32x4 = zeros_f32x4;
+    float32x4_t cov_yx_f32x4 = zeros_f32x4, cov_yy_f32x4 = zeros_f32x4, cov_yz_f32x4 = zeros_f32x4;
+    float32x4_t cov_zx_f32x4 = zeros_f32x4, cov_zy_f32x4 = zeros_f32x4, cov_zz_f32x4 = zeros_f32x4;
+
+    nk_size_t i = 0;
+    float32x4_t a_x_low_f32x4, a_x_high_f32x4, a_y_low_f32x4, a_y_high_f32x4, a_z_low_f32x4, a_z_high_f32x4;
+    float32x4_t b_x_low_f32x4, b_x_high_f32x4, b_y_low_f32x4, b_y_high_f32x4, b_z_low_f32x4, b_z_high_f32x4;
+
+    for (; i + 8 <= n; i += 8) {
+        nk_deinterleave_f16x8_to_f32x4x2_neon_(a + i * 3, &a_x_low_f32x4, &a_x_high_f32x4, &a_y_low_f32x4,
+                                               &a_y_high_f32x4, &a_z_low_f32x4, &a_z_high_f32x4);
+        nk_deinterleave_f16x8_to_f32x4x2_neon_(b + i * 3, &b_x_low_f32x4, &b_x_high_f32x4, &b_y_low_f32x4,
+                                               &b_y_high_f32x4, &b_z_low_f32x4, &b_z_high_f32x4);
+
+        sum_a_x_f32x4 = vaddq_f32(vaddq_f32(sum_a_x_f32x4, a_x_low_f32x4), a_x_high_f32x4);
+        sum_a_y_f32x4 = vaddq_f32(vaddq_f32(sum_a_y_f32x4, a_y_low_f32x4), a_y_high_f32x4);
+        sum_a_z_f32x4 = vaddq_f32(vaddq_f32(sum_a_z_f32x4, a_z_low_f32x4), a_z_high_f32x4);
+        sum_b_x_f32x4 = vaddq_f32(vaddq_f32(sum_b_x_f32x4, b_x_low_f32x4), b_x_high_f32x4);
+        sum_b_y_f32x4 = vaddq_f32(vaddq_f32(sum_b_y_f32x4, b_y_low_f32x4), b_y_high_f32x4);
+        sum_b_z_f32x4 = vaddq_f32(vaddq_f32(sum_b_z_f32x4, b_z_low_f32x4), b_z_high_f32x4);
+
+        cov_xx_f32x4 = vfmaq_f32(vfmaq_f32(cov_xx_f32x4, a_x_low_f32x4, b_x_low_f32x4), a_x_high_f32x4, b_x_high_f32x4);
+        cov_xy_f32x4 = vfmaq_f32(vfmaq_f32(cov_xy_f32x4, a_x_low_f32x4, b_y_low_f32x4), a_x_high_f32x4, b_y_high_f32x4);
+        cov_xz_f32x4 = vfmaq_f32(vfmaq_f32(cov_xz_f32x4, a_x_low_f32x4, b_z_low_f32x4), a_x_high_f32x4, b_z_high_f32x4);
+        cov_yx_f32x4 = vfmaq_f32(vfmaq_f32(cov_yx_f32x4, a_y_low_f32x4, b_x_low_f32x4), a_y_high_f32x4, b_x_high_f32x4);
+        cov_yy_f32x4 = vfmaq_f32(vfmaq_f32(cov_yy_f32x4, a_y_low_f32x4, b_y_low_f32x4), a_y_high_f32x4, b_y_high_f32x4);
+        cov_yz_f32x4 = vfmaq_f32(vfmaq_f32(cov_yz_f32x4, a_y_low_f32x4, b_z_low_f32x4), a_y_high_f32x4, b_z_high_f32x4);
+        cov_zx_f32x4 = vfmaq_f32(vfmaq_f32(cov_zx_f32x4, a_z_low_f32x4, b_x_low_f32x4), a_z_high_f32x4, b_x_high_f32x4);
+        cov_zy_f32x4 = vfmaq_f32(vfmaq_f32(cov_zy_f32x4, a_z_low_f32x4, b_y_low_f32x4), a_z_high_f32x4, b_y_high_f32x4);
+        cov_zz_f32x4 = vfmaq_f32(vfmaq_f32(cov_zz_f32x4, a_z_low_f32x4, b_z_low_f32x4), a_z_high_f32x4, b_z_high_f32x4);
+    }
+
+    if (i < n) {
+        nk_partial_deinterleave_f16_to_f32x4x2_neon_(a + i * 3, n - i, &a_x_low_f32x4, &a_x_high_f32x4, &a_y_low_f32x4,
+                                                     &a_y_high_f32x4, &a_z_low_f32x4, &a_z_high_f32x4);
+        nk_partial_deinterleave_f16_to_f32x4x2_neon_(b + i * 3, n - i, &b_x_low_f32x4, &b_x_high_f32x4, &b_y_low_f32x4,
+                                                     &b_y_high_f32x4, &b_z_low_f32x4, &b_z_high_f32x4);
+
+        sum_a_x_f32x4 = vaddq_f32(vaddq_f32(sum_a_x_f32x4, a_x_low_f32x4), a_x_high_f32x4);
+        sum_a_y_f32x4 = vaddq_f32(vaddq_f32(sum_a_y_f32x4, a_y_low_f32x4), a_y_high_f32x4);
+        sum_a_z_f32x4 = vaddq_f32(vaddq_f32(sum_a_z_f32x4, a_z_low_f32x4), a_z_high_f32x4);
+        sum_b_x_f32x4 = vaddq_f32(vaddq_f32(sum_b_x_f32x4, b_x_low_f32x4), b_x_high_f32x4);
+        sum_b_y_f32x4 = vaddq_f32(vaddq_f32(sum_b_y_f32x4, b_y_low_f32x4), b_y_high_f32x4);
+        sum_b_z_f32x4 = vaddq_f32(vaddq_f32(sum_b_z_f32x4, b_z_low_f32x4), b_z_high_f32x4);
+
+        cov_xx_f32x4 = vfmaq_f32(vfmaq_f32(cov_xx_f32x4, a_x_low_f32x4, b_x_low_f32x4), a_x_high_f32x4, b_x_high_f32x4);
+        cov_xy_f32x4 = vfmaq_f32(vfmaq_f32(cov_xy_f32x4, a_x_low_f32x4, b_y_low_f32x4), a_x_high_f32x4, b_y_high_f32x4);
+        cov_xz_f32x4 = vfmaq_f32(vfmaq_f32(cov_xz_f32x4, a_x_low_f32x4, b_z_low_f32x4), a_x_high_f32x4, b_z_high_f32x4);
+        cov_yx_f32x4 = vfmaq_f32(vfmaq_f32(cov_yx_f32x4, a_y_low_f32x4, b_x_low_f32x4), a_y_high_f32x4, b_x_high_f32x4);
+        cov_yy_f32x4 = vfmaq_f32(vfmaq_f32(cov_yy_f32x4, a_y_low_f32x4, b_y_low_f32x4), a_y_high_f32x4, b_y_high_f32x4);
+        cov_yz_f32x4 = vfmaq_f32(vfmaq_f32(cov_yz_f32x4, a_y_low_f32x4, b_z_low_f32x4), a_y_high_f32x4, b_z_high_f32x4);
+        cov_zx_f32x4 = vfmaq_f32(vfmaq_f32(cov_zx_f32x4, a_z_low_f32x4, b_x_low_f32x4), a_z_high_f32x4, b_x_high_f32x4);
+        cov_zy_f32x4 = vfmaq_f32(vfmaq_f32(cov_zy_f32x4, a_z_low_f32x4, b_y_low_f32x4), a_z_high_f32x4, b_y_high_f32x4);
+        cov_zz_f32x4 = vfmaq_f32(vfmaq_f32(cov_zz_f32x4, a_z_low_f32x4, b_z_low_f32x4), a_z_high_f32x4, b_z_high_f32x4);
+    }
+
+    // Reduce vector accumulators
+    nk_f32_t sum_a_x = vaddvq_f32(sum_a_x_f32x4);
+    nk_f32_t sum_a_y = vaddvq_f32(sum_a_y_f32x4);
+    nk_f32_t sum_a_z = vaddvq_f32(sum_a_z_f32x4);
+    nk_f32_t sum_b_x = vaddvq_f32(sum_b_x_f32x4);
+    nk_f32_t sum_b_y = vaddvq_f32(sum_b_y_f32x4);
+    nk_f32_t sum_b_z = vaddvq_f32(sum_b_z_f32x4);
+
+    nk_f32_t covariance_x_x = vaddvq_f32(cov_xx_f32x4);
+    nk_f32_t covariance_x_y = vaddvq_f32(cov_xy_f32x4);
+    nk_f32_t covariance_x_z = vaddvq_f32(cov_xz_f32x4);
+    nk_f32_t covariance_y_x = vaddvq_f32(cov_yx_f32x4);
+    nk_f32_t covariance_y_y = vaddvq_f32(cov_yy_f32x4);
+    nk_f32_t covariance_y_z = vaddvq_f32(cov_yz_f32x4);
+    nk_f32_t covariance_z_x = vaddvq_f32(cov_zx_f32x4);
+    nk_f32_t covariance_z_y = vaddvq_f32(cov_zy_f32x4);
+    nk_f32_t covariance_z_z = vaddvq_f32(cov_zz_f32x4);
+
+    // Compute centroids
+    nk_f32_t inv_n = 1.0f / (nk_f32_t)n;
+    nk_f32_t centroid_a_x = sum_a_x * inv_n;
+    nk_f32_t centroid_a_y = sum_a_y * inv_n;
+    nk_f32_t centroid_a_z = sum_a_z * inv_n;
+    nk_f32_t centroid_b_x = sum_b_x * inv_n;
+    nk_f32_t centroid_b_y = sum_b_y * inv_n;
+    nk_f32_t centroid_b_z = sum_b_z * inv_n;
+
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Compute centered covariance: H = (A - centroid_A)ᵀ * (B - centroid_B)
+    // H = sum(a * bᵀ) - n * centroid_a * centroid_bᵀ
+    nk_f32_t h[9];
+    h[0] = covariance_x_x - (nk_f32_t)n * centroid_a_x * centroid_b_x;
+    h[1] = covariance_x_y - (nk_f32_t)n * centroid_a_x * centroid_b_y;
+    h[2] = covariance_x_z - (nk_f32_t)n * centroid_a_x * centroid_b_z;
+    h[3] = covariance_y_x - (nk_f32_t)n * centroid_a_y * centroid_b_x;
+    h[4] = covariance_y_y - (nk_f32_t)n * centroid_a_y * centroid_b_y;
+    h[5] = covariance_y_z - (nk_f32_t)n * centroid_a_y * centroid_b_z;
+    h[6] = covariance_z_x - (nk_f32_t)n * centroid_a_z * centroid_b_x;
+    h[7] = covariance_z_y - (nk_f32_t)n * centroid_a_z * centroid_b_y;
+    h[8] = covariance_z_z - (nk_f32_t)n * centroid_a_z * centroid_b_z;
+
+    // SVD of H = U * S * Vᵀ
+    nk_f32_t svd_u[9], svd_s[9], svd_v[9];
+    nk_svd3x3_f32_(h, svd_u, svd_s, svd_v);
+
+    // R = V * Uᵀ
+    nk_f32_t r[9];
+    r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+    r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+    r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+    r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+    r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+    r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+    r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+    r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+    r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+
+    // Handle reflection: if det(R) < 0, negate third column of V and recompute
+    nk_f32_t det_r = nk_det3x3_f32_(r);
+    if (det_r < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+        r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+        r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+        r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+        r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+        r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+        r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+        r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+        r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+    }
+
+    if (rotation)
+        for (int j = 0; j < 9; ++j) rotation[j] = r[j];
+    if (scale) *scale = 1.0f;
+
+    // Compute RMSD after rotation
+    nk_f32_t sum_squared = nk_transformed_ssd_f16_neon_(a, b, n, r, 1.0f, centroid_a_x, centroid_a_y, centroid_a_z,
+                                                        centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f32_sqrt_neon(sum_squared * inv_n);
+}
+
+NK_PUBLIC void nk_umeyama_f16_neon(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+                                   nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
+    // Fused single-pass: load f16, convert to f32, compute centroids, covariance, and variance
+    float32x4_t const zeros_f32x4 = vdupq_n_f32(0);
+
+    float32x4_t sum_a_x_f32x4 = zeros_f32x4, sum_a_y_f32x4 = zeros_f32x4, sum_a_z_f32x4 = zeros_f32x4;
+    float32x4_t sum_b_x_f32x4 = zeros_f32x4, sum_b_y_f32x4 = zeros_f32x4, sum_b_z_f32x4 = zeros_f32x4;
+    float32x4_t cov_xx_f32x4 = zeros_f32x4, cov_xy_f32x4 = zeros_f32x4, cov_xz_f32x4 = zeros_f32x4;
+    float32x4_t cov_yx_f32x4 = zeros_f32x4, cov_yy_f32x4 = zeros_f32x4, cov_yz_f32x4 = zeros_f32x4;
+    float32x4_t cov_zx_f32x4 = zeros_f32x4, cov_zy_f32x4 = zeros_f32x4, cov_zz_f32x4 = zeros_f32x4;
+    float32x4_t variance_a_f32x4 = zeros_f32x4;
+
+    nk_size_t i = 0;
+    float32x4_t a_x_low_f32x4, a_x_high_f32x4, a_y_low_f32x4, a_y_high_f32x4, a_z_low_f32x4, a_z_high_f32x4;
+    float32x4_t b_x_low_f32x4, b_x_high_f32x4, b_y_low_f32x4, b_y_high_f32x4, b_z_low_f32x4, b_z_high_f32x4;
+
+    for (; i + 8 <= n; i += 8) {
+        nk_deinterleave_f16x8_to_f32x4x2_neon_(a + i * 3, &a_x_low_f32x4, &a_x_high_f32x4, &a_y_low_f32x4,
+                                               &a_y_high_f32x4, &a_z_low_f32x4, &a_z_high_f32x4);
+        nk_deinterleave_f16x8_to_f32x4x2_neon_(b + i * 3, &b_x_low_f32x4, &b_x_high_f32x4, &b_y_low_f32x4,
+                                               &b_y_high_f32x4, &b_z_low_f32x4, &b_z_high_f32x4);
+
+        sum_a_x_f32x4 = vaddq_f32(vaddq_f32(sum_a_x_f32x4, a_x_low_f32x4), a_x_high_f32x4);
+        sum_a_y_f32x4 = vaddq_f32(vaddq_f32(sum_a_y_f32x4, a_y_low_f32x4), a_y_high_f32x4);
+        sum_a_z_f32x4 = vaddq_f32(vaddq_f32(sum_a_z_f32x4, a_z_low_f32x4), a_z_high_f32x4);
+        sum_b_x_f32x4 = vaddq_f32(vaddq_f32(sum_b_x_f32x4, b_x_low_f32x4), b_x_high_f32x4);
+        sum_b_y_f32x4 = vaddq_f32(vaddq_f32(sum_b_y_f32x4, b_y_low_f32x4), b_y_high_f32x4);
+        sum_b_z_f32x4 = vaddq_f32(vaddq_f32(sum_b_z_f32x4, b_z_low_f32x4), b_z_high_f32x4);
+
+        cov_xx_f32x4 = vfmaq_f32(vfmaq_f32(cov_xx_f32x4, a_x_low_f32x4, b_x_low_f32x4), a_x_high_f32x4, b_x_high_f32x4);
+        cov_xy_f32x4 = vfmaq_f32(vfmaq_f32(cov_xy_f32x4, a_x_low_f32x4, b_y_low_f32x4), a_x_high_f32x4, b_y_high_f32x4);
+        cov_xz_f32x4 = vfmaq_f32(vfmaq_f32(cov_xz_f32x4, a_x_low_f32x4, b_z_low_f32x4), a_x_high_f32x4, b_z_high_f32x4);
+        cov_yx_f32x4 = vfmaq_f32(vfmaq_f32(cov_yx_f32x4, a_y_low_f32x4, b_x_low_f32x4), a_y_high_f32x4, b_x_high_f32x4);
+        cov_yy_f32x4 = vfmaq_f32(vfmaq_f32(cov_yy_f32x4, a_y_low_f32x4, b_y_low_f32x4), a_y_high_f32x4, b_y_high_f32x4);
+        cov_yz_f32x4 = vfmaq_f32(vfmaq_f32(cov_yz_f32x4, a_y_low_f32x4, b_z_low_f32x4), a_y_high_f32x4, b_z_high_f32x4);
+        cov_zx_f32x4 = vfmaq_f32(vfmaq_f32(cov_zx_f32x4, a_z_low_f32x4, b_x_low_f32x4), a_z_high_f32x4, b_x_high_f32x4);
+        cov_zy_f32x4 = vfmaq_f32(vfmaq_f32(cov_zy_f32x4, a_z_low_f32x4, b_y_low_f32x4), a_z_high_f32x4, b_y_high_f32x4);
+        cov_zz_f32x4 = vfmaq_f32(vfmaq_f32(cov_zz_f32x4, a_z_low_f32x4, b_z_low_f32x4), a_z_high_f32x4, b_z_high_f32x4);
+
+        variance_a_f32x4 = vfmaq_f32(variance_a_f32x4, a_x_low_f32x4, a_x_low_f32x4);
+        variance_a_f32x4 = vfmaq_f32(variance_a_f32x4, a_y_low_f32x4, a_y_low_f32x4);
+        variance_a_f32x4 = vfmaq_f32(variance_a_f32x4, a_z_low_f32x4, a_z_low_f32x4);
+        variance_a_f32x4 = vfmaq_f32(variance_a_f32x4, a_x_high_f32x4, a_x_high_f32x4);
+        variance_a_f32x4 = vfmaq_f32(variance_a_f32x4, a_y_high_f32x4, a_y_high_f32x4);
+        variance_a_f32x4 = vfmaq_f32(variance_a_f32x4, a_z_high_f32x4, a_z_high_f32x4);
+    }
+
+    if (i < n) {
+        nk_partial_deinterleave_f16_to_f32x4x2_neon_(a + i * 3, n - i, &a_x_low_f32x4, &a_x_high_f32x4, &a_y_low_f32x4,
+                                                     &a_y_high_f32x4, &a_z_low_f32x4, &a_z_high_f32x4);
+        nk_partial_deinterleave_f16_to_f32x4x2_neon_(b + i * 3, n - i, &b_x_low_f32x4, &b_x_high_f32x4, &b_y_low_f32x4,
+                                                     &b_y_high_f32x4, &b_z_low_f32x4, &b_z_high_f32x4);
+
+        sum_a_x_f32x4 = vaddq_f32(vaddq_f32(sum_a_x_f32x4, a_x_low_f32x4), a_x_high_f32x4);
+        sum_a_y_f32x4 = vaddq_f32(vaddq_f32(sum_a_y_f32x4, a_y_low_f32x4), a_y_high_f32x4);
+        sum_a_z_f32x4 = vaddq_f32(vaddq_f32(sum_a_z_f32x4, a_z_low_f32x4), a_z_high_f32x4);
+        sum_b_x_f32x4 = vaddq_f32(vaddq_f32(sum_b_x_f32x4, b_x_low_f32x4), b_x_high_f32x4);
+        sum_b_y_f32x4 = vaddq_f32(vaddq_f32(sum_b_y_f32x4, b_y_low_f32x4), b_y_high_f32x4);
+        sum_b_z_f32x4 = vaddq_f32(vaddq_f32(sum_b_z_f32x4, b_z_low_f32x4), b_z_high_f32x4);
+
+        cov_xx_f32x4 = vfmaq_f32(vfmaq_f32(cov_xx_f32x4, a_x_low_f32x4, b_x_low_f32x4), a_x_high_f32x4, b_x_high_f32x4);
+        cov_xy_f32x4 = vfmaq_f32(vfmaq_f32(cov_xy_f32x4, a_x_low_f32x4, b_y_low_f32x4), a_x_high_f32x4, b_y_high_f32x4);
+        cov_xz_f32x4 = vfmaq_f32(vfmaq_f32(cov_xz_f32x4, a_x_low_f32x4, b_z_low_f32x4), a_x_high_f32x4, b_z_high_f32x4);
+        cov_yx_f32x4 = vfmaq_f32(vfmaq_f32(cov_yx_f32x4, a_y_low_f32x4, b_x_low_f32x4), a_y_high_f32x4, b_x_high_f32x4);
+        cov_yy_f32x4 = vfmaq_f32(vfmaq_f32(cov_yy_f32x4, a_y_low_f32x4, b_y_low_f32x4), a_y_high_f32x4, b_y_high_f32x4);
+        cov_yz_f32x4 = vfmaq_f32(vfmaq_f32(cov_yz_f32x4, a_y_low_f32x4, b_z_low_f32x4), a_y_high_f32x4, b_z_high_f32x4);
+        cov_zx_f32x4 = vfmaq_f32(vfmaq_f32(cov_zx_f32x4, a_z_low_f32x4, b_x_low_f32x4), a_z_high_f32x4, b_x_high_f32x4);
+        cov_zy_f32x4 = vfmaq_f32(vfmaq_f32(cov_zy_f32x4, a_z_low_f32x4, b_y_low_f32x4), a_z_high_f32x4, b_y_high_f32x4);
+        cov_zz_f32x4 = vfmaq_f32(vfmaq_f32(cov_zz_f32x4, a_z_low_f32x4, b_z_low_f32x4), a_z_high_f32x4, b_z_high_f32x4);
+
+        variance_a_f32x4 = vfmaq_f32(variance_a_f32x4, a_x_low_f32x4, a_x_low_f32x4);
+        variance_a_f32x4 = vfmaq_f32(variance_a_f32x4, a_y_low_f32x4, a_y_low_f32x4);
+        variance_a_f32x4 = vfmaq_f32(variance_a_f32x4, a_z_low_f32x4, a_z_low_f32x4);
+        variance_a_f32x4 = vfmaq_f32(variance_a_f32x4, a_x_high_f32x4, a_x_high_f32x4);
+        variance_a_f32x4 = vfmaq_f32(variance_a_f32x4, a_y_high_f32x4, a_y_high_f32x4);
+        variance_a_f32x4 = vfmaq_f32(variance_a_f32x4, a_z_high_f32x4, a_z_high_f32x4);
+    }
+
+    // Reduce vector accumulators
+    nk_f32_t sum_a_x = vaddvq_f32(sum_a_x_f32x4);
+    nk_f32_t sum_a_y = vaddvq_f32(sum_a_y_f32x4);
+    nk_f32_t sum_a_z = vaddvq_f32(sum_a_z_f32x4);
+    nk_f32_t sum_b_x = vaddvq_f32(sum_b_x_f32x4);
+    nk_f32_t sum_b_y = vaddvq_f32(sum_b_y_f32x4);
+    nk_f32_t sum_b_z = vaddvq_f32(sum_b_z_f32x4);
+    nk_f32_t covariance_x_x = vaddvq_f32(cov_xx_f32x4);
+    nk_f32_t covariance_x_y = vaddvq_f32(cov_xy_f32x4);
+    nk_f32_t covariance_x_z = vaddvq_f32(cov_xz_f32x4);
+    nk_f32_t covariance_y_x = vaddvq_f32(cov_yx_f32x4);
+    nk_f32_t covariance_y_y = vaddvq_f32(cov_yy_f32x4);
+    nk_f32_t covariance_y_z = vaddvq_f32(cov_yz_f32x4);
+    nk_f32_t covariance_z_x = vaddvq_f32(cov_zx_f32x4);
+    nk_f32_t covariance_z_y = vaddvq_f32(cov_zy_f32x4);
+    nk_f32_t covariance_z_z = vaddvq_f32(cov_zz_f32x4);
+    nk_f32_t variance_a_sum = vaddvq_f32(variance_a_f32x4);
+
+    // Compute centroids
+    nk_f32_t inv_n = 1.0f / (nk_f32_t)n;
+    nk_f32_t centroid_a_x = sum_a_x * inv_n, centroid_a_y = sum_a_y * inv_n, centroid_a_z = sum_a_z * inv_n;
+    nk_f32_t centroid_b_x = sum_b_x * inv_n, centroid_b_y = sum_b_y * inv_n, centroid_b_z = sum_b_z * inv_n;
+
+    if (a_centroid) a_centroid[0] = centroid_a_x, a_centroid[1] = centroid_a_y, a_centroid[2] = centroid_a_z;
+    if (b_centroid) b_centroid[0] = centroid_b_x, b_centroid[1] = centroid_b_y, b_centroid[2] = centroid_b_z;
+
+    // Compute centered covariance and variance
+    nk_f32_t variance_a = variance_a_sum * inv_n -
+                          (centroid_a_x * centroid_a_x + centroid_a_y * centroid_a_y + centroid_a_z * centroid_a_z);
+
+    nk_f32_t h[9];
+    h[0] = covariance_x_x - (nk_f32_t)n * centroid_a_x * centroid_b_x;
+    h[1] = covariance_x_y - (nk_f32_t)n * centroid_a_x * centroid_b_y;
+    h[2] = covariance_x_z - (nk_f32_t)n * centroid_a_x * centroid_b_z;
+    h[3] = covariance_y_x - (nk_f32_t)n * centroid_a_y * centroid_b_x;
+    h[4] = covariance_y_y - (nk_f32_t)n * centroid_a_y * centroid_b_y;
+    h[5] = covariance_y_z - (nk_f32_t)n * centroid_a_y * centroid_b_z;
+    h[6] = covariance_z_x - (nk_f32_t)n * centroid_a_z * centroid_b_x;
+    h[7] = covariance_z_y - (nk_f32_t)n * centroid_a_z * centroid_b_y;
+    h[8] = covariance_z_z - (nk_f32_t)n * centroid_a_z * centroid_b_z;
+
+    // SVD of H = U * S * Vᵀ
+    nk_f32_t svd_u[9], svd_s[9], svd_v[9];
+    nk_svd3x3_f32_(h, svd_u, svd_s, svd_v);
+
+    // R = V * Uᵀ
+    nk_f32_t r[9];
+    r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+    r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+    r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+    r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+    r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+    r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+    r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+    r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+    r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+
+    // Handle reflection and compute scale: c = trace(D × S) / variance(a)
+    nk_f32_t det_r = nk_det3x3_f32_(r);
+    nk_f32_t sign_det = det_r < 0 ? -1.0f : 1.0f;
+    nk_f32_t trace_scaled_s = svd_s[0] + svd_s[4] + sign_det * svd_s[8];
+    nk_f32_t scale_factor = trace_scaled_s / ((nk_f32_t)n * variance_a);
+    if (scale) *scale = scale_factor;
+
+    if (det_r < 0) {
+        svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
+        r[0] = svd_v[0] * svd_u[0] + svd_v[1] * svd_u[1] + svd_v[2] * svd_u[2];
+        r[1] = svd_v[0] * svd_u[3] + svd_v[1] * svd_u[4] + svd_v[2] * svd_u[5];
+        r[2] = svd_v[0] * svd_u[6] + svd_v[1] * svd_u[7] + svd_v[2] * svd_u[8];
+        r[3] = svd_v[3] * svd_u[0] + svd_v[4] * svd_u[1] + svd_v[5] * svd_u[2];
+        r[4] = svd_v[3] * svd_u[3] + svd_v[4] * svd_u[4] + svd_v[5] * svd_u[5];
+        r[5] = svd_v[3] * svd_u[6] + svd_v[4] * svd_u[7] + svd_v[5] * svd_u[8];
+        r[6] = svd_v[6] * svd_u[0] + svd_v[7] * svd_u[1] + svd_v[8] * svd_u[2];
+        r[7] = svd_v[6] * svd_u[3] + svd_v[7] * svd_u[4] + svd_v[8] * svd_u[5];
+        r[8] = svd_v[6] * svd_u[6] + svd_v[7] * svd_u[7] + svd_v[8] * svd_u[8];
+    }
+
+    if (rotation)
+        for (int j = 0; j < 9; ++j) rotation[j] = r[j];
+
+    // Compute RMSD after similarity transform
+    nk_f32_t sum_squared = nk_transformed_ssd_f16_neon_(a, b, n, r, scale_factor, centroid_a_x, centroid_a_y,
+                                                        centroid_a_z, centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f32_sqrt_neon(sum_squared * inv_n);
+}
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_NEON
+#endif // NK_TARGET_ARM64_
+#endif // NK_MESH_NEON_H