numkong 7.0.0 → 7.4.1

package/include/numkong/mesh/rvv.h

@@ -15,7 +15,7 @@
  *
  *  Fused helpers minimize data passes:
  *
- *  - `nk_bicentroid_*_rvv_`: both centroids in a single pass (used by RMSD)
+ *  - RMSD: fully fused single-pass (centroids + squared diffs), no separate helper
  *  - `nk_centroid_and_cross_covariance_*_rvv_`: centroids + H in one pass (Kabsch)
  *  - `nk_centroid_and_cross_covariance_and_variance_*_rvv_`: + variance (Umeyama)
  *
@@ -89,104 +89,6 @@ NK_INTERNAL void nk_accumulate_product_f64m1_rvv_(vfloat64m1_t *sum_f64m1, vfloa
                                                     vector_length);
 }
 
-/**
- *  @brief Compute centroids of two f32 point clouds in a single pass.
- *
- *  Reads both clouds simultaneously, accumulating 6 sums (3 per cloud) in f64.
- *  Reduces RMSD from 3 passes to 2 (bicentroid + SSD).
- *  Uses per-lane `vfwadd_wv` accumulation with deferred `vfredusum` after the loop.
- */
-NK_INTERNAL void nk_bicentroid_f32_rvv_(               //
-    nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, //
-    nk_f64_t *ca_x, nk_f64_t *ca_y, nk_f64_t *ca_z,    //
-    nk_f64_t *cb_x, nk_f64_t *cb_y, nk_f64_t *cb_z) {
-    nk_size_t vlmax = __riscv_vsetvlmax_e64m2();
-    vfloat64m2_t sum_a_x_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t sum_a_y_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t sum_a_z_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t sum_b_x_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t sum_b_y_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t sum_b_z_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    nk_f32_t const *a_ptr = a, *b_ptr = b;
-    nk_size_t remaining = n;
-    for (nk_size_t vector_length; remaining > 0;
-         remaining -= vector_length, a_ptr += vector_length * 3, b_ptr += vector_length * 3) {
-        vector_length = __riscv_vsetvl_e32m1(remaining);
-        vfloat32m1x3_t a_f32m1x3 = __riscv_vlseg3e32_v_f32m1x3(a_ptr, vector_length);
-        sum_a_x_f64m2 = __riscv_vfwadd_wv_f64m2_tu(sum_a_x_f64m2, sum_a_x_f64m2,
-                                                   __riscv_vget_v_f32m1x3_f32m1(a_f32m1x3, 0), vector_length);
-        sum_a_y_f64m2 = __riscv_vfwadd_wv_f64m2_tu(sum_a_y_f64m2, sum_a_y_f64m2,
-                                                   __riscv_vget_v_f32m1x3_f32m1(a_f32m1x3, 1), vector_length);
-        sum_a_z_f64m2 = __riscv_vfwadd_wv_f64m2_tu(sum_a_z_f64m2, sum_a_z_f64m2,
-                                                   __riscv_vget_v_f32m1x3_f32m1(a_f32m1x3, 2), vector_length);
-        vfloat32m1x3_t b_f32m1x3 = __riscv_vlseg3e32_v_f32m1x3(b_ptr, vector_length);
-        sum_b_x_f64m2 = __riscv_vfwadd_wv_f64m2_tu(sum_b_x_f64m2, sum_b_x_f64m2,
-                                                   __riscv_vget_v_f32m1x3_f32m1(b_f32m1x3, 0), vector_length);
-        sum_b_y_f64m2 = __riscv_vfwadd_wv_f64m2_tu(sum_b_y_f64m2, sum_b_y_f64m2,
-                                                   __riscv_vget_v_f32m1x3_f32m1(b_f32m1x3, 1), vector_length);
-        sum_b_z_f64m2 = __riscv_vfwadd_wv_f64m2_tu(sum_b_z_f64m2, sum_b_z_f64m2,
-                                                   __riscv_vget_v_f32m1x3_f32m1(b_f32m1x3, 2), vector_length);
-    }
-    vfloat64m1_t zero_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, 1);
-    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
-    *ca_x = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_a_x_f64m2, zero_f64m1, vlmax)) * inv_n;
-    *ca_y = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_a_y_f64m2, zero_f64m1, vlmax)) * inv_n;
-    *ca_z = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_a_z_f64m2, zero_f64m1, vlmax)) * inv_n;
-    *cb_x = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_b_x_f64m2, zero_f64m1, vlmax)) * inv_n;
-    *cb_y = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_b_y_f64m2, zero_f64m1, vlmax)) * inv_n;
-    *cb_z = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_b_z_f64m2, zero_f64m1, vlmax)) * inv_n;
-}
-
-/**
- *  @brief Compute centroids of two f64 point clouds in a single pass.
- *  Uses per-lane `vfadd_vv` accumulation with deferred `vfredusum` after the loop.
- */
-NK_INTERNAL void nk_bicentroid_f64_rvv_(               //
-    nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, //
-    nk_f64_t *ca_x, nk_f64_t *ca_y, nk_f64_t *ca_z,    //
-    nk_f64_t *cb_x, nk_f64_t *cb_y, nk_f64_t *cb_z) {
-    nk_size_t vlmax = __riscv_vsetvlmax_e64m1();
-    vfloat64m1_t sum_a_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t sum_a_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t sum_a_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t sum_b_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t sum_b_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t sum_b_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_a_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_a_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_a_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_b_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_b_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_b_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    nk_f64_t const *a_ptr = a, *b_ptr = b;
-    nk_size_t remaining = n;
-    for (nk_size_t vector_length; remaining > 0;
-         remaining -= vector_length, a_ptr += vector_length * 3, b_ptr += vector_length * 3) {
-        vector_length = __riscv_vsetvl_e64m1(remaining);
-        vfloat64m1x3_t a_f64m1x3 = __riscv_vlseg3e64_v_f64m1x3(a_ptr, vector_length);
-        nk_accumulate_sum_f64m1_rvv_(&sum_a_x_f64m1, &compensation_a_x_f64m1,
-                                     __riscv_vget_v_f64m1x3_f64m1(a_f64m1x3, 0), vector_length);
-        nk_accumulate_sum_f64m1_rvv_(&sum_a_y_f64m1, &compensation_a_y_f64m1,
-                                     __riscv_vget_v_f64m1x3_f64m1(a_f64m1x3, 1), vector_length);
-        nk_accumulate_sum_f64m1_rvv_(&sum_a_z_f64m1, &compensation_a_z_f64m1,
-                                     __riscv_vget_v_f64m1x3_f64m1(a_f64m1x3, 2), vector_length);
-        vfloat64m1x3_t b_f64m1x3 = __riscv_vlseg3e64_v_f64m1x3(b_ptr, vector_length);
-        nk_accumulate_sum_f64m1_rvv_(&sum_b_x_f64m1, &compensation_b_x_f64m1,
-                                     __riscv_vget_v_f64m1x3_f64m1(b_f64m1x3, 0), vector_length);
-        nk_accumulate_sum_f64m1_rvv_(&sum_b_y_f64m1, &compensation_b_y_f64m1,
-                                     __riscv_vget_v_f64m1x3_f64m1(b_f64m1x3, 1), vector_length);
-        nk_accumulate_sum_f64m1_rvv_(&sum_b_z_f64m1, &compensation_b_z_f64m1,
-                                     __riscv_vget_v_f64m1x3_f64m1(b_f64m1x3, 2), vector_length);
-    }
-    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
-    *ca_x = nk_dot_stable_sum_f64m1_rvv_(sum_a_x_f64m1, compensation_a_x_f64m1) * inv_n;
-    *ca_y = nk_dot_stable_sum_f64m1_rvv_(sum_a_y_f64m1, compensation_a_y_f64m1) * inv_n;
-    *ca_z = nk_dot_stable_sum_f64m1_rvv_(sum_a_z_f64m1, compensation_a_z_f64m1) * inv_n;
-    *cb_x = nk_dot_stable_sum_f64m1_rvv_(sum_b_x_f64m1, compensation_b_x_f64m1) * inv_n;
-    *cb_y = nk_dot_stable_sum_f64m1_rvv_(sum_b_y_f64m1, compensation_b_y_f64m1) * inv_n;
-    *cb_z = nk_dot_stable_sum_f64m1_rvv_(sum_b_z_f64m1, compensation_b_z_f64m1) * inv_n;
-}
-
 /**
  *  @brief Compute centroids and cross-covariance matrix in a single pass (f32).
  *
@@ -198,27 +100,29 @@ NK_INTERNAL void nk_bicentroid_f64_rvv_(               //
  *  Cross-products use per-lane `vfwmacc_vv` accumulation (vfloat64m2_t) with
  *  deferred `vfredusum` after the loop — eliminates 9 reductions per iteration.
  */
-NK_INTERNAL void nk_centroid_and_cross_covariance_f32_rvv_( //
-    nk_f32_t const *a, nk_f32_t const *b, nk_size_t n,      //
-    nk_f64_t *ca_x, nk_f64_t *ca_y, nk_f64_t *ca_z,         //
-    nk_f64_t *cb_x, nk_f64_t *cb_y, nk_f64_t *cb_z,         //
+NK_INTERNAL void nk_centroid_and_cross_covariance_f32_rvv_(                 //
+    nk_f32_t const *a, nk_f32_t const *b, nk_size_t points_count,           //
+    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, //
     nk_f64_t h[9]) {
-    nk_size_t vlmax = __riscv_vsetvlmax_e64m2();
-    vfloat64m2_t sum_a_x_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax), sum_a_y_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t sum_a_z_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t sum_b_x_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax), sum_b_y_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t sum_b_z_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t cross_00_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax),
-                 cross_01_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t cross_02_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax),
-                 cross_10_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t cross_11_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax),
-                 cross_12_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t cross_20_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax),
-                 cross_21_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t cross_22_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
+    nk_size_t max_vector_length = __riscv_vsetvlmax_e64m2();
+    vfloat64m2_t sum_a_x_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length),
+                 sum_a_y_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_a_z_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_b_x_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length),
+                 sum_b_y_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_b_z_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t cross_00_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length),
+                 cross_01_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t cross_02_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length),
+                 cross_10_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t cross_11_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length),
+                 cross_12_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t cross_20_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length),
+                 cross_21_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t cross_22_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
     nk_f32_t const *a_ptr = a, *b_ptr = b;
-    nk_size_t remaining = n;
+    nk_size_t remaining = points_count;
     for (nk_size_t vector_length; remaining > 0;
          remaining -= vector_length, a_ptr += vector_length * 3, b_ptr += vector_length * 3) {
         vector_length = __riscv_vsetvl_e32m1(remaining);
@@ -248,45 +152,51 @@ NK_INTERNAL void nk_centroid_and_cross_covariance_f32_rvv_( //
     }
     vfloat64m1_t zero_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, 1);
     // Compute centroids
-    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
-    nk_f64_t ca_x_ = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_a_x_f64m2, zero_f64m1, vlmax)) *
-                     inv_n;
-    nk_f64_t ca_y_ = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_a_y_f64m2, zero_f64m1, vlmax)) *
-                     inv_n;
-    nk_f64_t ca_z_ = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_a_z_f64m2, zero_f64m1, vlmax)) *
-                     inv_n;
-    nk_f64_t cb_x_ = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_b_x_f64m2, zero_f64m1, vlmax)) *
-                     inv_n;
-    nk_f64_t cb_y_ = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_b_y_f64m2, zero_f64m1, vlmax)) *
-                     inv_n;
-    nk_f64_t cb_z_ = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_b_z_f64m2, zero_f64m1, vlmax)) *
-                     inv_n;
-    *ca_x = ca_x_;
-    *ca_y = ca_y_;
-    *ca_z = ca_z_;
-    *cb_x = cb_x_;
-    *cb_y = cb_y_;
-    *cb_z = cb_z_;
-    // Fix up: H[i][j] = raw[i][j] - n * ca[i] * cb[j]
-    nk_f64_t n_f64 = (nk_f64_t)n;
-    h[0] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_00_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_x_ * cb_x_;
-    h[1] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_01_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_x_ * cb_y_;
-    h[2] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_02_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_x_ * cb_z_;
-    h[3] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_10_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_y_ * cb_x_;
-    h[4] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_11_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_y_ * cb_y_;
-    h[5] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_12_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_y_ * cb_z_;
-    h[6] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_20_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_z_ * cb_x_;
-    h[7] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_21_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_z_ * cb_y_;
-    h[8] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_22_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_z_ * cb_z_;
+    nk_f64_t inv_points_count = 1.0 / (nk_f64_t)points_count;
+    nk_f64_t centroid_a_x_f64 = __riscv_vfmv_f_s_f64m1_f64(
+                                    __riscv_vfredusum_vs_f64m2_f64m1(sum_a_x_f64m2, zero_f64m1, max_vector_length)) *
+                                inv_points_count;
+    nk_f64_t centroid_a_y_f64 = __riscv_vfmv_f_s_f64m1_f64(
+                                    __riscv_vfredusum_vs_f64m2_f64m1(sum_a_y_f64m2, zero_f64m1, max_vector_length)) *
+                                inv_points_count;
+    nk_f64_t centroid_a_z_f64 = __riscv_vfmv_f_s_f64m1_f64(
+                                    __riscv_vfredusum_vs_f64m2_f64m1(sum_a_z_f64m2, zero_f64m1, max_vector_length)) *
+                                inv_points_count;
+    nk_f64_t centroid_b_x_f64 = __riscv_vfmv_f_s_f64m1_f64(
+                                    __riscv_vfredusum_vs_f64m2_f64m1(sum_b_x_f64m2, zero_f64m1, max_vector_length)) *
+                                inv_points_count;
+    nk_f64_t centroid_b_y_f64 = __riscv_vfmv_f_s_f64m1_f64(
+                                    __riscv_vfredusum_vs_f64m2_f64m1(sum_b_y_f64m2, zero_f64m1, max_vector_length)) *
+                                inv_points_count;
+    nk_f64_t centroid_b_z_f64 = __riscv_vfmv_f_s_f64m1_f64(
+                                    __riscv_vfredusum_vs_f64m2_f64m1(sum_b_z_f64m2, zero_f64m1, max_vector_length)) *
+                                inv_points_count;
+    *centroid_a_x = centroid_a_x_f64;
+    *centroid_a_y = centroid_a_y_f64;
+    *centroid_a_z = centroid_a_z_f64;
+    *centroid_b_x = centroid_b_x_f64;
+    *centroid_b_y = centroid_b_y_f64;
+    *centroid_b_z = centroid_b_z_f64;
+    // Fix up: H[i][j] = raw[i][j] - points_count * ca[i] * cb[j]
+    nk_f64_t n_f64 = (nk_f64_t)points_count;
+    h[0] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_00_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_x_f64 * centroid_b_x_f64;
+    h[1] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_01_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_x_f64 * centroid_b_y_f64;
+    h[2] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_02_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_x_f64 * centroid_b_z_f64;
+    h[3] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_10_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_y_f64 * centroid_b_x_f64;
+    h[4] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_11_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_y_f64 * centroid_b_y_f64;
+    h[5] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_12_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_y_f64 * centroid_b_z_f64;
+    h[6] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_20_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_z_f64 * centroid_b_x_f64;
+    h[7] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_21_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_z_f64 * centroid_b_y_f64;
+    h[8] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_22_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_z_f64 * centroid_b_z_f64;
 }
 
 /**
@@ -295,42 +205,44 @@ NK_INTERNAL void nk_centroid_and_cross_covariance_f32_rvv_( //
  *  Per-lane `vfadd_vv`/`vfmacc_vv` accumulation with deferred `vfredusum` after the loop
  *  — eliminates 15 horizontal reductions per iteration.
  */
-NK_INTERNAL void nk_centroid_and_cross_covariance_f64_rvv_( //
-    nk_f64_t const *a, nk_f64_t const *b, nk_size_t n,      //
-    nk_f64_t *ca_x, nk_f64_t *ca_y, nk_f64_t *ca_z,         //
-    nk_f64_t *cb_x, nk_f64_t *cb_y, nk_f64_t *cb_z,         //
+NK_INTERNAL void nk_centroid_and_cross_covariance_f64_rvv_(                 //
+    nk_f64_t const *a, nk_f64_t const *b, nk_size_t points_count,           //
+    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, //
     nk_f64_t h[9]) {
-    nk_size_t vlmax = __riscv_vsetvlmax_e64m1();
-    vfloat64m1_t sum_a_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax), sum_a_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t sum_a_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t sum_b_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax), sum_b_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t sum_b_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_a_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_a_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_a_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_b_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_b_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_b_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t cross_00_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax),
-                 cross_01_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t cross_02_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax),
-                 cross_10_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t cross_11_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax),
-                 cross_12_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t cross_20_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax),
-                 cross_21_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t cross_22_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_00_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_01_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_02_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_10_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_11_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_12_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_20_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_21_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_22_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
+    nk_size_t max_vector_length = __riscv_vsetvlmax_e64m1();
+    vfloat64m1_t sum_a_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length),
+                 sum_a_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_a_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_b_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length),
+                 sum_b_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_b_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_a_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_a_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_a_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_b_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_b_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_b_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t cross_00_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length),
+                 cross_01_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t cross_02_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length),
+                 cross_10_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t cross_11_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length),
+                 cross_12_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t cross_20_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length),
+                 cross_21_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t cross_22_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_00_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_01_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_02_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_10_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_11_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_12_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_20_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_21_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_22_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
     nk_f64_t const *a_ptr = a, *b_ptr = b;
-    nk_size_t remaining = n;
+    nk_size_t remaining = points_count;
     for (nk_size_t vector_length; remaining > 0;
          remaining -= vector_length, a_ptr += vector_length * 3, b_ptr += vector_length * 3) {
         vector_length = __riscv_vsetvl_e64m1(remaining);
@@ -359,29 +271,38 @@ NK_INTERNAL void nk_centroid_and_cross_covariance_f64_rvv_( //
         nk_accumulate_product_f64m1_rvv_(&cross_22_f64m1, &compensation_22_f64m1, a_z_f64m1, b_z_f64m1, vector_length);
     }
     // Compute centroids.
-    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
-    nk_f64_t ca_x_ = nk_dot_stable_sum_f64m1_rvv_(sum_a_x_f64m1, compensation_a_x_f64m1) * inv_n;
-    nk_f64_t ca_y_ = nk_dot_stable_sum_f64m1_rvv_(sum_a_y_f64m1, compensation_a_y_f64m1) * inv_n;
-    nk_f64_t ca_z_ = nk_dot_stable_sum_f64m1_rvv_(sum_a_z_f64m1, compensation_a_z_f64m1) * inv_n;
-    nk_f64_t cb_x_ = nk_dot_stable_sum_f64m1_rvv_(sum_b_x_f64m1, compensation_b_x_f64m1) * inv_n;
-    nk_f64_t cb_y_ = nk_dot_stable_sum_f64m1_rvv_(sum_b_y_f64m1, compensation_b_y_f64m1) * inv_n;
-    nk_f64_t cb_z_ = nk_dot_stable_sum_f64m1_rvv_(sum_b_z_f64m1, compensation_b_z_f64m1) * inv_n;
-    *ca_x = ca_x_;
-    *ca_y = ca_y_;
-    *ca_z = ca_z_;
-    *cb_x = cb_x_;
-    *cb_y = cb_y_;
-    *cb_z = cb_z_;
-    nk_f64_t n_f64 = (nk_f64_t)n;
-    h[0] = nk_dot_stable_sum_f64m1_rvv_(cross_00_f64m1, compensation_00_f64m1) - n_f64 * ca_x_ * cb_x_;
-    h[1] = nk_dot_stable_sum_f64m1_rvv_(cross_01_f64m1, compensation_01_f64m1) - n_f64 * ca_x_ * cb_y_;
-    h[2] = nk_dot_stable_sum_f64m1_rvv_(cross_02_f64m1, compensation_02_f64m1) - n_f64 * ca_x_ * cb_z_;
-    h[3] = nk_dot_stable_sum_f64m1_rvv_(cross_10_f64m1, compensation_10_f64m1) - n_f64 * ca_y_ * cb_x_;
-    h[4] = nk_dot_stable_sum_f64m1_rvv_(cross_11_f64m1, compensation_11_f64m1) - n_f64 * ca_y_ * cb_y_;
-    h[5] = nk_dot_stable_sum_f64m1_rvv_(cross_12_f64m1, compensation_12_f64m1) - n_f64 * ca_y_ * cb_z_;
-    h[6] = nk_dot_stable_sum_f64m1_rvv_(cross_20_f64m1, compensation_20_f64m1) - n_f64 * ca_z_ * cb_x_;
-    h[7] = nk_dot_stable_sum_f64m1_rvv_(cross_21_f64m1, compensation_21_f64m1) - n_f64 * ca_z_ * cb_y_;
-    h[8] = nk_dot_stable_sum_f64m1_rvv_(cross_22_f64m1, compensation_22_f64m1) - n_f64 * ca_z_ * cb_z_;
+    nk_f64_t inv_points_count = 1.0 / (nk_f64_t)points_count;
+    nk_f64_t centroid_a_x_f64 = nk_dot_stable_sum_f64m1_rvv_(sum_a_x_f64m1, compensation_a_x_f64m1) * inv_points_count;
+    nk_f64_t centroid_a_y_f64 = nk_dot_stable_sum_f64m1_rvv_(sum_a_y_f64m1, compensation_a_y_f64m1) * inv_points_count;
+    nk_f64_t centroid_a_z_f64 = nk_dot_stable_sum_f64m1_rvv_(sum_a_z_f64m1, compensation_a_z_f64m1) * inv_points_count;
+    nk_f64_t centroid_b_x_f64 = nk_dot_stable_sum_f64m1_rvv_(sum_b_x_f64m1, compensation_b_x_f64m1) * inv_points_count;
+    nk_f64_t centroid_b_y_f64 = nk_dot_stable_sum_f64m1_rvv_(sum_b_y_f64m1, compensation_b_y_f64m1) * inv_points_count;
+    nk_f64_t centroid_b_z_f64 = nk_dot_stable_sum_f64m1_rvv_(sum_b_z_f64m1, compensation_b_z_f64m1) * inv_points_count;
+    *centroid_a_x = centroid_a_x_f64;
+    *centroid_a_y = centroid_a_y_f64;
+    *centroid_a_z = centroid_a_z_f64;
+    *centroid_b_x = centroid_b_x_f64;
+    *centroid_b_y = centroid_b_y_f64;
+    *centroid_b_z = centroid_b_z_f64;
+    nk_f64_t n_f64 = (nk_f64_t)points_count;
+    h[0] = nk_dot_stable_sum_f64m1_rvv_(cross_00_f64m1, compensation_00_f64m1) -
+           n_f64 * centroid_a_x_f64 * centroid_b_x_f64;
+    h[1] = nk_dot_stable_sum_f64m1_rvv_(cross_01_f64m1, compensation_01_f64m1) -
+           n_f64 * centroid_a_x_f64 * centroid_b_y_f64;
+    h[2] = nk_dot_stable_sum_f64m1_rvv_(cross_02_f64m1, compensation_02_f64m1) -
+           n_f64 * centroid_a_x_f64 * centroid_b_z_f64;
+    h[3] = nk_dot_stable_sum_f64m1_rvv_(cross_10_f64m1, compensation_10_f64m1) -
+           n_f64 * centroid_a_y_f64 * centroid_b_x_f64;
+    h[4] = nk_dot_stable_sum_f64m1_rvv_(cross_11_f64m1, compensation_11_f64m1) -
+           n_f64 * centroid_a_y_f64 * centroid_b_y_f64;
+    h[5] = nk_dot_stable_sum_f64m1_rvv_(cross_12_f64m1, compensation_12_f64m1) -
+           n_f64 * centroid_a_y_f64 * centroid_b_z_f64;
+    h[6] = nk_dot_stable_sum_f64m1_rvv_(cross_20_f64m1, compensation_20_f64m1) -
+           n_f64 * centroid_a_z_f64 * centroid_b_x_f64;
+    h[7] = nk_dot_stable_sum_f64m1_rvv_(cross_21_f64m1, compensation_21_f64m1) -
+           n_f64 * centroid_a_z_f64 * centroid_b_y_f64;
+    h[8] = nk_dot_stable_sum_f64m1_rvv_(cross_22_f64m1, compensation_22_f64m1) -
+           n_f64 * centroid_a_z_f64 * centroid_b_z_f64;
 }
 
 /**
@@ -394,28 +315,30 @@ NK_INTERNAL void nk_centroid_and_cross_covariance_f64_rvv_( //
  *  Cross-products use per-lane `vfwmacc_vv` accumulation (vfloat64m2_t) with
  *  deferred `vfredusum` after the loop — eliminates 9 reductions per iteration.
  */
-NK_INTERNAL void nk_centroid_and_cross_covariance_and_variance_f32_rvv_( //
-    nk_f32_t const *a, nk_f32_t const *b, nk_size_t n,                   //
-    nk_f64_t *ca_x, nk_f64_t *ca_y, nk_f64_t *ca_z,                      //
-    nk_f64_t *cb_x, nk_f64_t *cb_y, nk_f64_t *cb_z,                      //
+NK_INTERNAL void nk_centroid_and_cross_covariance_and_variance_f32_rvv_(    //
+    nk_f32_t const *a, nk_f32_t const *b, nk_size_t points_count,           //
+    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, //
     nk_f64_t h[9], nk_f64_t *variance_a) {
-    nk_size_t vlmax = __riscv_vsetvlmax_e64m2();
-    vfloat64m2_t sum_a_x_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax), sum_a_y_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t sum_a_z_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t sum_b_x_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax), sum_b_y_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t sum_b_z_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t cross_00_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax),
-                 cross_01_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t cross_02_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax),
-                 cross_10_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t cross_11_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax),
-                 cross_12_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t cross_20_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax),
-                 cross_21_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t cross_22_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
-    vfloat64m2_t sum_norm_squared_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
+    nk_size_t max_vector_length = __riscv_vsetvlmax_e64m2();
+    vfloat64m2_t sum_a_x_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length),
+                 sum_a_y_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_a_z_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_b_x_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length),
+                 sum_b_y_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_b_z_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t cross_00_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length),
+                 cross_01_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t cross_02_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length),
+                 cross_10_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t cross_11_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length),
+                 cross_12_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t cross_20_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length),
+                 cross_21_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t cross_22_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_norm_squared_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
     nk_f32_t const *a_ptr = a, *b_ptr = b;
-    nk_size_t remaining = n;
+    nk_size_t remaining = points_count;
     for (nk_size_t vector_length; remaining > 0;
          remaining -= vector_length, a_ptr += vector_length * 3, b_ptr += vector_length * 3) {
         vector_length = __riscv_vsetvl_e32m1(remaining);
@@ -450,49 +373,56 @@ NK_INTERNAL void nk_centroid_and_cross_covariance_and_variance_f32_rvv_( //
                                                            norm_squared_f64m2, vector_length);
     }
     vfloat64m1_t zero_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, 1);
-    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
-    nk_f64_t ca_x_ = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_a_x_f64m2, zero_f64m1, vlmax)) *
-                     inv_n;
-    nk_f64_t ca_y_ = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_a_y_f64m2, zero_f64m1, vlmax)) *
-                     inv_n;
-    nk_f64_t ca_z_ = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_a_z_f64m2, zero_f64m1, vlmax)) *
-                     inv_n;
-    nk_f64_t cb_x_ = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_b_x_f64m2, zero_f64m1, vlmax)) *
-                     inv_n;
-    nk_f64_t cb_y_ = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_b_y_f64m2, zero_f64m1, vlmax)) *
-                     inv_n;
-    nk_f64_t cb_z_ = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_b_z_f64m2, zero_f64m1, vlmax)) *
-                     inv_n;
-    *ca_x = ca_x_;
-    *ca_y = ca_y_;
-    *ca_z = ca_z_;
-    *cb_x = cb_x_;
-    *cb_y = cb_y_;
-    *cb_z = cb_z_;
-    nk_f64_t n_f64 = (nk_f64_t)n;
-    h[0] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_00_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_x_ * cb_x_;
-    h[1] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_01_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_x_ * cb_y_;
-    h[2] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_02_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_x_ * cb_z_;
-    h[3] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_10_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_y_ * cb_x_;
-    h[4] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_11_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_y_ * cb_y_;
-    h[5] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_12_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_y_ * cb_z_;
-    h[6] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_20_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_z_ * cb_x_;
-    h[7] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_21_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_z_ * cb_y_;
-    h[8] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_22_f64m2, zero_f64m1, vlmax)) -
-           n_f64 * ca_z_ * cb_z_;
-    // variance_a = (1/n) * (Σ ||a[i]||² - n * ||ca||²)
+    nk_f64_t inv_points_count = 1.0 / (nk_f64_t)points_count;
+    nk_f64_t centroid_a_x_f64 = __riscv_vfmv_f_s_f64m1_f64(
+                                    __riscv_vfredusum_vs_f64m2_f64m1(sum_a_x_f64m2, zero_f64m1, max_vector_length)) *
+                                inv_points_count;
+    nk_f64_t centroid_a_y_f64 = __riscv_vfmv_f_s_f64m1_f64(
+                                    __riscv_vfredusum_vs_f64m2_f64m1(sum_a_y_f64m2, zero_f64m1, max_vector_length)) *
+                                inv_points_count;
+    nk_f64_t centroid_a_z_f64 = __riscv_vfmv_f_s_f64m1_f64(
+                                    __riscv_vfredusum_vs_f64m2_f64m1(sum_a_z_f64m2, zero_f64m1, max_vector_length)) *
+                                inv_points_count;
+    nk_f64_t centroid_b_x_f64 = __riscv_vfmv_f_s_f64m1_f64(
+                                    __riscv_vfredusum_vs_f64m2_f64m1(sum_b_x_f64m2, zero_f64m1, max_vector_length)) *
+                                inv_points_count;
+    nk_f64_t centroid_b_y_f64 = __riscv_vfmv_f_s_f64m1_f64(
+                                    __riscv_vfredusum_vs_f64m2_f64m1(sum_b_y_f64m2, zero_f64m1, max_vector_length)) *
+                                inv_points_count;
+    nk_f64_t centroid_b_z_f64 = __riscv_vfmv_f_s_f64m1_f64(
+                                    __riscv_vfredusum_vs_f64m2_f64m1(sum_b_z_f64m2, zero_f64m1, max_vector_length)) *
+                                inv_points_count;
+    *centroid_a_x = centroid_a_x_f64;
+    *centroid_a_y = centroid_a_y_f64;
+    *centroid_a_z = centroid_a_z_f64;
+    *centroid_b_x = centroid_b_x_f64;
+    *centroid_b_y = centroid_b_y_f64;
+    *centroid_b_z = centroid_b_z_f64;
+    nk_f64_t n_f64 = (nk_f64_t)points_count;
+    h[0] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_00_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_x_f64 * centroid_b_x_f64;
+    h[1] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_01_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_x_f64 * centroid_b_y_f64;
+    h[2] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_02_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_x_f64 * centroid_b_z_f64;
+    h[3] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_10_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_y_f64 * centroid_b_x_f64;
+    h[4] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_11_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_y_f64 * centroid_b_y_f64;
+    h[5] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_12_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_y_f64 * centroid_b_z_f64;
+    h[6] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_20_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_z_f64 * centroid_b_x_f64;
+    h[7] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_21_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_z_f64 * centroid_b_y_f64;
+    h[8] = __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(cross_22_f64m2, zero_f64m1, max_vector_length)) -
+           n_f64 * centroid_a_z_f64 * centroid_b_z_f64;
+    // variance_a = (1/points_count) * (Σ ||a[i]||² - points_count * ||ca||²)
     *variance_a = __riscv_vfmv_f_s_f64m1_f64(
-                      __riscv_vfredusum_vs_f64m2_f64m1(sum_norm_squared_f64m2, zero_f64m1, vlmax)) *
-                      inv_n -
-                  (ca_x_ * ca_x_ + ca_y_ * ca_y_ + ca_z_ * ca_z_);
+                      __riscv_vfredusum_vs_f64m2_f64m1(sum_norm_squared_f64m2, zero_f64m1, max_vector_length)) *
+                      inv_points_count -
+                  (centroid_a_x_f64 * centroid_a_x_f64 + centroid_a_y_f64 * centroid_a_y_f64 +
+                   centroid_a_z_f64 * centroid_a_z_f64);
 }
 
 /**
@@ -501,44 +431,46 @@ NK_INTERNAL void nk_centroid_and_cross_covariance_and_variance_f32_rvv_( //
  *  Per-lane `vfadd_vv`/`vfmacc_vv` accumulation with deferred `vfredusum` after the loop
  *  — eliminates 16 horizontal reductions per iteration.
  */
-NK_INTERNAL void nk_centroid_and_cross_covariance_and_variance_f64_rvv_( //
-    nk_f64_t const *a, nk_f64_t const *b, nk_size_t n,                   //
-    nk_f64_t *ca_x, nk_f64_t *ca_y, nk_f64_t *ca_z,                      //
-    nk_f64_t *cb_x, nk_f64_t *cb_y, nk_f64_t *cb_z,                      //
+NK_INTERNAL void nk_centroid_and_cross_covariance_and_variance_f64_rvv_(    //
+    nk_f64_t const *a, nk_f64_t const *b, nk_size_t points_count,           //
+    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, //
     nk_f64_t h[9], nk_f64_t *variance_a) {
-    nk_size_t vlmax = __riscv_vsetvlmax_e64m1();
-    vfloat64m1_t sum_a_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax), sum_a_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t sum_a_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t sum_b_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax), sum_b_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t sum_b_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_a_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_a_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_a_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_b_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_b_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_b_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t cross_00_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax),
-                 cross_01_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t cross_02_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax),
-                 cross_10_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t cross_11_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax),
-                 cross_12_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t cross_20_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax),
-                 cross_21_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t cross_22_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_00_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_01_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_02_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_10_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_11_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_12_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_20_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_21_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_22_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t sum_norm_squared_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_norm_squared_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
+    nk_size_t max_vector_length = __riscv_vsetvlmax_e64m1();
+    vfloat64m1_t sum_a_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length),
+                 sum_a_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_a_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_b_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length),
+                 sum_b_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_b_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_a_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_a_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_a_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_b_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_b_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_b_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t cross_00_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length),
+                 cross_01_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t cross_02_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length),
+                 cross_10_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t cross_11_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length),
+                 cross_12_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t cross_20_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length),
+                 cross_21_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t cross_22_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_00_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_01_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_02_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_10_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_11_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_12_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_20_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_21_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_22_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_norm_squared_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_norm_squared_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
     nk_f64_t const *a_ptr = a, *b_ptr = b;
-    nk_size_t remaining = n;
+    nk_size_t remaining = points_count;
     for (nk_size_t vector_length; remaining > 0;
          remaining -= vector_length, a_ptr += vector_length * 3, b_ptr += vector_length * 3) {
         vector_length = __riscv_vsetvl_e64m1(remaining);
@@ -571,56 +503,70 @@ NK_INTERNAL void nk_centroid_and_cross_covariance_and_variance_f64_rvv_( //
         nk_accumulate_sum_f64m1_rvv_(&sum_norm_squared_f64m1, &compensation_norm_squared_f64m1, norm_squared_f64m1,
                                      vector_length);
     }
-    nk_f64_t inv_n = 1.0 / (nk_f64_t)n;
-    nk_f64_t ca_x_ = nk_dot_stable_sum_f64m1_rvv_(sum_a_x_f64m1, compensation_a_x_f64m1) * inv_n;
-    nk_f64_t ca_y_ = nk_dot_stable_sum_f64m1_rvv_(sum_a_y_f64m1, compensation_a_y_f64m1) * inv_n;
-    nk_f64_t ca_z_ = nk_dot_stable_sum_f64m1_rvv_(sum_a_z_f64m1, compensation_a_z_f64m1) * inv_n;
-    nk_f64_t cb_x_ = nk_dot_stable_sum_f64m1_rvv_(sum_b_x_f64m1, compensation_b_x_f64m1) * inv_n;
-    nk_f64_t cb_y_ = nk_dot_stable_sum_f64m1_rvv_(sum_b_y_f64m1, compensation_b_y_f64m1) * inv_n;
-    nk_f64_t cb_z_ = nk_dot_stable_sum_f64m1_rvv_(sum_b_z_f64m1, compensation_b_z_f64m1) * inv_n;
-    *ca_x = ca_x_;
-    *ca_y = ca_y_;
-    *ca_z = ca_z_;
-    *cb_x = cb_x_;
-    *cb_y = cb_y_;
-    *cb_z = cb_z_;
-    nk_f64_t n_f64 = (nk_f64_t)n;
-    h[0] = nk_dot_stable_sum_f64m1_rvv_(cross_00_f64m1, compensation_00_f64m1) - n_f64 * ca_x_ * cb_x_;
-    h[1] = nk_dot_stable_sum_f64m1_rvv_(cross_01_f64m1, compensation_01_f64m1) - n_f64 * ca_x_ * cb_y_;
-    h[2] = nk_dot_stable_sum_f64m1_rvv_(cross_02_f64m1, compensation_02_f64m1) - n_f64 * ca_x_ * cb_z_;
-    h[3] = nk_dot_stable_sum_f64m1_rvv_(cross_10_f64m1, compensation_10_f64m1) - n_f64 * ca_y_ * cb_x_;
-    h[4] = nk_dot_stable_sum_f64m1_rvv_(cross_11_f64m1, compensation_11_f64m1) - n_f64 * ca_y_ * cb_y_;
-    h[5] = nk_dot_stable_sum_f64m1_rvv_(cross_12_f64m1, compensation_12_f64m1) - n_f64 * ca_y_ * cb_z_;
-    h[6] = nk_dot_stable_sum_f64m1_rvv_(cross_20_f64m1, compensation_20_f64m1) - n_f64 * ca_z_ * cb_x_;
-    h[7] = nk_dot_stable_sum_f64m1_rvv_(cross_21_f64m1, compensation_21_f64m1) - n_f64 * ca_z_ * cb_y_;
-    h[8] = nk_dot_stable_sum_f64m1_rvv_(cross_22_f64m1, compensation_22_f64m1) - n_f64 * ca_z_ * cb_z_;
-    *variance_a = nk_dot_stable_sum_f64m1_rvv_(sum_norm_squared_f64m1, compensation_norm_squared_f64m1) * inv_n -
-                  (ca_x_ * ca_x_ + ca_y_ * ca_y_ + ca_z_ * ca_z_);
+    nk_f64_t inv_points_count = 1.0 / (nk_f64_t)points_count;
+    nk_f64_t centroid_a_x_f64 = nk_dot_stable_sum_f64m1_rvv_(sum_a_x_f64m1, compensation_a_x_f64m1) * inv_points_count;
+    nk_f64_t centroid_a_y_f64 = nk_dot_stable_sum_f64m1_rvv_(sum_a_y_f64m1, compensation_a_y_f64m1) * inv_points_count;
+    nk_f64_t centroid_a_z_f64 = nk_dot_stable_sum_f64m1_rvv_(sum_a_z_f64m1, compensation_a_z_f64m1) * inv_points_count;
+    nk_f64_t centroid_b_x_f64 = nk_dot_stable_sum_f64m1_rvv_(sum_b_x_f64m1, compensation_b_x_f64m1) * inv_points_count;
+    nk_f64_t centroid_b_y_f64 = nk_dot_stable_sum_f64m1_rvv_(sum_b_y_f64m1, compensation_b_y_f64m1) * inv_points_count;
+    nk_f64_t centroid_b_z_f64 = nk_dot_stable_sum_f64m1_rvv_(sum_b_z_f64m1, compensation_b_z_f64m1) * inv_points_count;
+    *centroid_a_x = centroid_a_x_f64;
+    *centroid_a_y = centroid_a_y_f64;
+    *centroid_a_z = centroid_a_z_f64;
+    *centroid_b_x = centroid_b_x_f64;
+    *centroid_b_y = centroid_b_y_f64;
+    *centroid_b_z = centroid_b_z_f64;
+    nk_f64_t n_f64 = (nk_f64_t)points_count;
+    h[0] = nk_dot_stable_sum_f64m1_rvv_(cross_00_f64m1, compensation_00_f64m1) -
+           n_f64 * centroid_a_x_f64 * centroid_b_x_f64;
+    h[1] = nk_dot_stable_sum_f64m1_rvv_(cross_01_f64m1, compensation_01_f64m1) -
+           n_f64 * centroid_a_x_f64 * centroid_b_y_f64;
+    h[2] = nk_dot_stable_sum_f64m1_rvv_(cross_02_f64m1, compensation_02_f64m1) -
+           n_f64 * centroid_a_x_f64 * centroid_b_z_f64;
+    h[3] = nk_dot_stable_sum_f64m1_rvv_(cross_10_f64m1, compensation_10_f64m1) -
+           n_f64 * centroid_a_y_f64 * centroid_b_x_f64;
+    h[4] = nk_dot_stable_sum_f64m1_rvv_(cross_11_f64m1, compensation_11_f64m1) -
+           n_f64 * centroid_a_y_f64 * centroid_b_y_f64;
+    h[5] = nk_dot_stable_sum_f64m1_rvv_(cross_12_f64m1, compensation_12_f64m1) -
+           n_f64 * centroid_a_y_f64 * centroid_b_z_f64;
+    h[6] = nk_dot_stable_sum_f64m1_rvv_(cross_20_f64m1, compensation_20_f64m1) -
+           n_f64 * centroid_a_z_f64 * centroid_b_x_f64;
+    h[7] = nk_dot_stable_sum_f64m1_rvv_(cross_21_f64m1, compensation_21_f64m1) -
+           n_f64 * centroid_a_z_f64 * centroid_b_y_f64;
+    h[8] = nk_dot_stable_sum_f64m1_rvv_(cross_22_f64m1, compensation_22_f64m1) -
+           n_f64 * centroid_a_z_f64 * centroid_b_z_f64;
+    *variance_a = nk_dot_stable_sum_f64m1_rvv_(sum_norm_squared_f64m1, compensation_norm_squared_f64m1) *
+                      inv_points_count -
+                  (centroid_a_x_f64 * centroid_a_x_f64 + centroid_a_y_f64 * centroid_a_y_f64 +
+                   centroid_a_z_f64 * centroid_a_z_f64);
 }
 
-NK_INTERNAL nk_f64_t nk_transformed_ssd_f32_rvv_(      //
-    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 ca_x, nk_f64_t ca_y, nk_f64_t ca_z,       //
-    nk_f64_t cb_x, nk_f64_t cb_y, nk_f64_t cb_z) {
+NK_INTERNAL nk_f64_t nk_transformed_ssd_f32_rvv_(                        //
+    nk_f32_t const *a, nk_f32_t const *b, nk_size_t points_count,        //
+    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) {
     nk_f64_t scaled_rotation_x_x = scale * r[0], scaled_rotation_x_y = scale * r[1], scaled_rotation_x_z = scale * r[2];
     nk_f64_t scaled_rotation_y_x = scale * r[3], scaled_rotation_y_y = scale * r[4], scaled_rotation_y_z = scale * r[5];
     nk_f64_t scaled_rotation_z_x = scale * r[6], scaled_rotation_z_y = scale * r[7], scaled_rotation_z_z = scale * r[8];
-    nk_size_t vlmax = __riscv_vsetvlmax_e64m2();
-    vfloat64m2_t sum_distance_squared_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, vlmax);
+    nk_size_t max_vector_length = __riscv_vsetvlmax_e64m2();
+    vfloat64m2_t sum_distance_squared_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
     vfloat64m1_t zero_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, 1);
     nk_f32_t const *a_ptr = a, *b_ptr = b;
-    nk_size_t remaining = n;
+    nk_size_t remaining = points_count;
     for (nk_size_t vector_length; remaining > 0;
          remaining -= vector_length, a_ptr += vector_length * 3, b_ptr += vector_length * 3) {
         vector_length = __riscv_vsetvl_e32m1(remaining);
         vfloat32m1x3_t a_f32m1x3 = __riscv_vlseg3e32_v_f32m1x3(a_ptr, vector_length);
         vfloat64m2_t centered_a_x_f64m2 = __riscv_vfsub_vf_f64m2(
-            __riscv_vfwcvt_f_f_v_f64m2(__riscv_vget_v_f32m1x3_f32m1(a_f32m1x3, 0), vector_length), ca_x, vector_length);
+            __riscv_vfwcvt_f_f_v_f64m2(__riscv_vget_v_f32m1x3_f32m1(a_f32m1x3, 0), vector_length), centroid_a_x,
+            vector_length);
         vfloat64m2_t centered_a_y_f64m2 = __riscv_vfsub_vf_f64m2(
-            __riscv_vfwcvt_f_f_v_f64m2(__riscv_vget_v_f32m1x3_f32m1(a_f32m1x3, 1), vector_length), ca_y, vector_length);
+            __riscv_vfwcvt_f_f_v_f64m2(__riscv_vget_v_f32m1x3_f32m1(a_f32m1x3, 1), vector_length), centroid_a_y,
+            vector_length);
         vfloat64m2_t centered_a_z_f64m2 = __riscv_vfsub_vf_f64m2(
-            __riscv_vfwcvt_f_f_v_f64m2(__riscv_vget_v_f32m1x3_f32m1(a_f32m1x3, 2), vector_length), ca_z, vector_length);
+            __riscv_vfwcvt_f_f_v_f64m2(__riscv_vget_v_f32m1x3_f32m1(a_f32m1x3, 2), vector_length), centroid_a_z,
+            vector_length);
         vfloat64m2_t rotated_a_x_f64m2 = __riscv_vfmul_vf_f64m2(centered_a_x_f64m2, scaled_rotation_x_x, vector_length);
         rotated_a_x_f64m2 = __riscv_vfmacc_vf_f64m2(rotated_a_x_f64m2, scaled_rotation_x_y, centered_a_y_f64m2,
                                                     vector_length);
@@ -638,11 +584,14 @@ NK_INTERNAL nk_f64_t nk_transformed_ssd_f32_rvv_(      //
                                                     vector_length);
         vfloat32m1x3_t b_f32m1x3 = __riscv_vlseg3e32_v_f32m1x3(b_ptr, vector_length);
         vfloat64m2_t centered_b_x_f64m2 = __riscv_vfsub_vf_f64m2(
-            __riscv_vfwcvt_f_f_v_f64m2(__riscv_vget_v_f32m1x3_f32m1(b_f32m1x3, 0), vector_length), cb_x, vector_length);
+            __riscv_vfwcvt_f_f_v_f64m2(__riscv_vget_v_f32m1x3_f32m1(b_f32m1x3, 0), vector_length), centroid_b_x,
+            vector_length);
         vfloat64m2_t centered_b_y_f64m2 = __riscv_vfsub_vf_f64m2(
-            __riscv_vfwcvt_f_f_v_f64m2(__riscv_vget_v_f32m1x3_f32m1(b_f32m1x3, 1), vector_length), cb_y, vector_length);
+            __riscv_vfwcvt_f_f_v_f64m2(__riscv_vget_v_f32m1x3_f32m1(b_f32m1x3, 1), vector_length), centroid_b_y,
+            vector_length);
         vfloat64m2_t centered_b_z_f64m2 = __riscv_vfsub_vf_f64m2(
-            __riscv_vfwcvt_f_f_v_f64m2(__riscv_vget_v_f32m1x3_f32m1(b_f32m1x3, 2), vector_length), cb_z, vector_length);
+            __riscv_vfwcvt_f_f_v_f64m2(__riscv_vget_v_f32m1x3_f32m1(b_f32m1x3, 2), vector_length), centroid_b_z,
+            vector_length);
         vfloat64m2_t delta_x_f64m2 = __riscv_vfsub_vv_f64m2(rotated_a_x_f64m2, centered_b_x_f64m2, vector_length);
         vfloat64m2_t delta_y_f64m2 = __riscv_vfsub_vv_f64m2(rotated_a_y_f64m2, centered_b_y_f64m2, vector_length);
         vfloat64m2_t delta_z_f64m2 = __riscv_vfsub_vv_f64m2(rotated_a_z_f64m2, centered_b_z_f64m2, vector_length);
@@ -653,32 +602,33 @@ NK_INTERNAL nk_f64_t nk_transformed_ssd_f32_rvv_(      //
         sum_distance_squared_f64m2 = __riscv_vfmacc_vv_f64m2_tu(sum_distance_squared_f64m2, delta_z_f64m2,
                                                                 delta_z_f64m2, vector_length);
     }
-    return __riscv_vfmv_f_s_f64m1_f64(__riscv_vfredusum_vs_f64m2_f64m1(sum_distance_squared_f64m2, zero_f64m1, vlmax));
+    return __riscv_vfmv_f_s_f64m1_f64(
+        __riscv_vfredusum_vs_f64m2_f64m1(sum_distance_squared_f64m2, zero_f64m1, max_vector_length));
 }
 
-NK_INTERNAL nk_f64_t nk_transformed_ssd_f64_rvv_(      //
-    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 ca_x, nk_f64_t ca_y, nk_f64_t ca_z,       //
-    nk_f64_t cb_x, nk_f64_t cb_y, nk_f64_t cb_z) {
+NK_INTERNAL nk_f64_t nk_transformed_ssd_f64_rvv_(                        //
+    nk_f64_t const *a, nk_f64_t const *b, nk_size_t points_count,        //
+    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) {
     nk_f64_t scaled_rotation_x_x = scale * r[0], scaled_rotation_x_y = scale * r[1], scaled_rotation_x_z = scale * r[2];
     nk_f64_t scaled_rotation_y_x = scale * r[3], scaled_rotation_y_y = scale * r[4], scaled_rotation_y_z = scale * r[5];
     nk_f64_t scaled_rotation_z_x = scale * r[6], scaled_rotation_z_y = scale * r[7], scaled_rotation_z_z = scale * r[8];
-    nk_size_t vlmax = __riscv_vsetvlmax_e64m1();
-    vfloat64m1_t sum_distance_squared_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
-    vfloat64m1_t compensation_distance_squared_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, vlmax);
+    nk_size_t max_vector_length = __riscv_vsetvlmax_e64m1();
+    vfloat64m1_t sum_distance_squared_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_distance_squared_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
     nk_f64_t const *a_ptr = a, *b_ptr = b;
-    nk_size_t remaining = n;
+    nk_size_t remaining = points_count;
     for (nk_size_t vector_length; remaining > 0;
          remaining -= vector_length, a_ptr += vector_length * 3, b_ptr += vector_length * 3) {
         vector_length = __riscv_vsetvl_e64m1(remaining);
         vfloat64m1x3_t a_f64m1x3 = __riscv_vlseg3e64_v_f64m1x3(a_ptr, vector_length);
-        vfloat64m1_t centered_a_x_f64m1 = __riscv_vfsub_vf_f64m1(__riscv_vget_v_f64m1x3_f64m1(a_f64m1x3, 0), ca_x,
-                                                                 vector_length);
-        vfloat64m1_t centered_a_y_f64m1 = __riscv_vfsub_vf_f64m1(__riscv_vget_v_f64m1x3_f64m1(a_f64m1x3, 1), ca_y,
-                                                                 vector_length);
-        vfloat64m1_t centered_a_z_f64m1 = __riscv_vfsub_vf_f64m1(__riscv_vget_v_f64m1x3_f64m1(a_f64m1x3, 2), ca_z,
-                                                                 vector_length);
+        vfloat64m1_t centered_a_x_f64m1 = __riscv_vfsub_vf_f64m1(__riscv_vget_v_f64m1x3_f64m1(a_f64m1x3, 0),
+                                                                 centroid_a_x, vector_length);
+        vfloat64m1_t centered_a_y_f64m1 = __riscv_vfsub_vf_f64m1(__riscv_vget_v_f64m1x3_f64m1(a_f64m1x3, 1),
+                                                                 centroid_a_y, vector_length);
+        vfloat64m1_t centered_a_z_f64m1 = __riscv_vfsub_vf_f64m1(__riscv_vget_v_f64m1x3_f64m1(a_f64m1x3, 2),
+                                                                 centroid_a_z, vector_length);
         vfloat64m1_t rotated_a_x_f64m1 = __riscv_vfmul_vf_f64m1(centered_a_x_f64m1, scaled_rotation_x_x, vector_length);
         rotated_a_x_f64m1 = __riscv_vfmacc_vf_f64m1(rotated_a_x_f64m1, scaled_rotation_x_y, centered_a_y_f64m1,
                                                     vector_length);
@@ -695,12 +645,12 @@ NK_INTERNAL nk_f64_t nk_transformed_ssd_f64_rvv_(      //
         rotated_a_z_f64m1 = __riscv_vfmacc_vf_f64m1(rotated_a_z_f64m1, scaled_rotation_z_z, centered_a_z_f64m1,
                                                     vector_length);
         vfloat64m1x3_t b_f64m1x3 = __riscv_vlseg3e64_v_f64m1x3(b_ptr, vector_length);
-        vfloat64m1_t centered_b_x_f64m1 = __riscv_vfsub_vf_f64m1(__riscv_vget_v_f64m1x3_f64m1(b_f64m1x3, 0), cb_x,
-                                                                 vector_length);
-        vfloat64m1_t centered_b_y_f64m1 = __riscv_vfsub_vf_f64m1(__riscv_vget_v_f64m1x3_f64m1(b_f64m1x3, 1), cb_y,
-                                                                 vector_length);
-        vfloat64m1_t centered_b_z_f64m1 = __riscv_vfsub_vf_f64m1(__riscv_vget_v_f64m1x3_f64m1(b_f64m1x3, 2), cb_z,
-                                                                 vector_length);
+        vfloat64m1_t centered_b_x_f64m1 = __riscv_vfsub_vf_f64m1(__riscv_vget_v_f64m1x3_f64m1(b_f64m1x3, 0),
+                                                                 centroid_b_x, vector_length);
+        vfloat64m1_t centered_b_y_f64m1 = __riscv_vfsub_vf_f64m1(__riscv_vget_v_f64m1x3_f64m1(b_f64m1x3, 1),
+                                                                 centroid_b_y, vector_length);
+        vfloat64m1_t centered_b_z_f64m1 = __riscv_vfsub_vf_f64m1(__riscv_vget_v_f64m1x3_f64m1(b_f64m1x3, 2),
+                                                                 centroid_b_z, vector_length);
         vfloat64m1_t delta_x_f64m1 = __riscv_vfsub_vv_f64m1(rotated_a_x_f64m1, centered_b_x_f64m1, vector_length);
         vfloat64m1_t delta_y_f64m1 = __riscv_vfsub_vv_f64m1(rotated_a_y_f64m1, centered_b_y_f64m1, vector_length);
         vfloat64m1_t delta_z_f64m1 = __riscv_vfsub_vv_f64m1(rotated_a_z_f64m1, centered_b_z_f64m1, vector_length);
@@ -745,42 +695,176 @@ NK_INTERNAL void nk_rotation_from_svd_f64_rvv_( //
     nk_rotation_from_svd_f64_serial_(svd_u, svd_v, r);
 }
 
-NK_PUBLIC void nk_rmsd_f32_rvv(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+NK_PUBLIC void nk_rmsd_f32_rvv(nk_f32_t const *a, nk_f32_t const *b, nk_size_t points_count, nk_f32_t *a_centroid,
                                nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f64_t *result) {
-    nk_f64_t identity[9] = {1, 0, 0, 0, 1, 0, 0, 0, 1};
     if (rotation)
-        for (int j = 0; j < 9; ++j) rotation[j] = identity[j];
+        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;
-    nk_f64_t ca_x, ca_y, ca_z, cb_x, cb_y, cb_z;
-    nk_bicentroid_f32_rvv_(a, b, n, &ca_x, &ca_y, &ca_z, &cb_x, &cb_y, &cb_z);
-    if (a_centroid) a_centroid[0] = (nk_f32_t)ca_x, a_centroid[1] = (nk_f32_t)ca_y, a_centroid[2] = (nk_f32_t)ca_z;
-    if (b_centroid) b_centroid[0] = (nk_f32_t)cb_x, b_centroid[1] = (nk_f32_t)cb_y, b_centroid[2] = (nk_f32_t)cb_z;
-    nk_f64_t ssd = nk_transformed_ssd_f32_rvv_(a, b, n, identity, 1.0, ca_x, ca_y, ca_z, cb_x, cb_y, cb_z);
-    *result = nk_f64_sqrt_rvv(ssd / (nk_f64_t)n);
+
+    // Fused single-pass: accumulate centroids and squared differences simultaneously.
+    // RMSD = √(E[(a−b)²] − (ā − b̄)²)
+    nk_size_t max_vector_length = __riscv_vsetvlmax_e64m2();
+    vfloat64m2_t sum_a_x_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_a_y_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_a_z_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_b_x_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_b_y_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_b_z_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    vfloat64m2_t sum_squared_f64m2 = __riscv_vfmv_v_f_f64m2(0.0, max_vector_length);
+    nk_f32_t const *a_ptr = a, *b_ptr = b;
+    nk_size_t remaining = points_count;
+    for (nk_size_t vector_length; remaining > 0;
+         remaining -= vector_length, a_ptr += vector_length * 3, b_ptr += vector_length * 3) {
+        vector_length = __riscv_vsetvl_e32m1(remaining);
+        vfloat32m1x3_t a_f32m1x3 = __riscv_vlseg3e32_v_f32m1x3(a_ptr, vector_length);
+        vfloat32m1_t a_x_f32m1 = __riscv_vget_v_f32m1x3_f32m1(a_f32m1x3, 0);
+        vfloat32m1_t a_y_f32m1 = __riscv_vget_v_f32m1x3_f32m1(a_f32m1x3, 1);
+        vfloat32m1_t a_z_f32m1 = __riscv_vget_v_f32m1x3_f32m1(a_f32m1x3, 2);
+        vfloat32m1x3_t b_f32m1x3 = __riscv_vlseg3e32_v_f32m1x3(b_ptr, vector_length);
+        vfloat32m1_t b_x_f32m1 = __riscv_vget_v_f32m1x3_f32m1(b_f32m1x3, 0);
+        vfloat32m1_t b_y_f32m1 = __riscv_vget_v_f32m1x3_f32m1(b_f32m1x3, 1);
+        vfloat32m1_t b_z_f32m1 = __riscv_vget_v_f32m1x3_f32m1(b_f32m1x3, 2);
+        // Accumulate centroids in f64.
+        sum_a_x_f64m2 = __riscv_vfwadd_wv_f64m2_tu(sum_a_x_f64m2, sum_a_x_f64m2, a_x_f32m1, vector_length);
+        sum_a_y_f64m2 = __riscv_vfwadd_wv_f64m2_tu(sum_a_y_f64m2, sum_a_y_f64m2, a_y_f32m1, vector_length);
+        sum_a_z_f64m2 = __riscv_vfwadd_wv_f64m2_tu(sum_a_z_f64m2, sum_a_z_f64m2, a_z_f32m1, vector_length);
+        sum_b_x_f64m2 = __riscv_vfwadd_wv_f64m2_tu(sum_b_x_f64m2, sum_b_x_f64m2, b_x_f32m1, vector_length);
+        sum_b_y_f64m2 = __riscv_vfwadd_wv_f64m2_tu(sum_b_y_f64m2, sum_b_y_f64m2, b_y_f32m1, vector_length);
+        sum_b_z_f64m2 = __riscv_vfwadd_wv_f64m2_tu(sum_b_z_f64m2, sum_b_z_f64m2, b_z_f32m1, vector_length);
+        // Accumulate (a−b)² per component. Widen a,b to f64 before subtracting to avoid f32
+        // cancellation in the single-pass formula RMSD = √(E[(a−b)²] − (ā − b̄)²).
+        vfloat64m2_t a_x_f64m2 = __riscv_vfwcvt_f_f_v_f64m2(a_x_f32m1, vector_length);
+        vfloat64m2_t b_x_f64m2 = __riscv_vfwcvt_f_f_v_f64m2(b_x_f32m1, vector_length);
+        vfloat64m2_t a_y_f64m2 = __riscv_vfwcvt_f_f_v_f64m2(a_y_f32m1, vector_length);
+        vfloat64m2_t b_y_f64m2 = __riscv_vfwcvt_f_f_v_f64m2(b_y_f32m1, vector_length);
+        vfloat64m2_t a_z_f64m2 = __riscv_vfwcvt_f_f_v_f64m2(a_z_f32m1, vector_length);
+        vfloat64m2_t b_z_f64m2 = __riscv_vfwcvt_f_f_v_f64m2(b_z_f32m1, vector_length);
+        vfloat64m2_t delta_x_f64m2 = __riscv_vfsub_vv_f64m2(a_x_f64m2, b_x_f64m2, vector_length);
+        vfloat64m2_t delta_y_f64m2 = __riscv_vfsub_vv_f64m2(a_y_f64m2, b_y_f64m2, vector_length);
+        vfloat64m2_t delta_z_f64m2 = __riscv_vfsub_vv_f64m2(a_z_f64m2, b_z_f64m2, vector_length);
+        sum_squared_f64m2 = __riscv_vfmacc_vv_f64m2_tu(sum_squared_f64m2, delta_x_f64m2, delta_x_f64m2, vector_length);
+        sum_squared_f64m2 = __riscv_vfmacc_vv_f64m2_tu(sum_squared_f64m2, delta_y_f64m2, delta_y_f64m2, vector_length);
+        sum_squared_f64m2 = __riscv_vfmacc_vv_f64m2_tu(sum_squared_f64m2, delta_z_f64m2, delta_z_f64m2, vector_length);
+    }
+    vfloat64m1_t zero_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, 1);
+    nk_f64_t inv_points_count = 1.0 / (nk_f64_t)points_count;
+    nk_f64_t centroid_a_x = __riscv_vfmv_f_s_f64m1_f64(
+                                __riscv_vfredusum_vs_f64m2_f64m1(sum_a_x_f64m2, zero_f64m1, max_vector_length)) *
+                            inv_points_count;
+    nk_f64_t centroid_a_y = __riscv_vfmv_f_s_f64m1_f64(
+                                __riscv_vfredusum_vs_f64m2_f64m1(sum_a_y_f64m2, zero_f64m1, max_vector_length)) *
+                            inv_points_count;
+    nk_f64_t centroid_a_z = __riscv_vfmv_f_s_f64m1_f64(
+                                __riscv_vfredusum_vs_f64m2_f64m1(sum_a_z_f64m2, zero_f64m1, max_vector_length)) *
+                            inv_points_count;
+    nk_f64_t centroid_b_x = __riscv_vfmv_f_s_f64m1_f64(
+                                __riscv_vfredusum_vs_f64m2_f64m1(sum_b_x_f64m2, zero_f64m1, max_vector_length)) *
+                            inv_points_count;
+    nk_f64_t centroid_b_y = __riscv_vfmv_f_s_f64m1_f64(
+                                __riscv_vfredusum_vs_f64m2_f64m1(sum_b_y_f64m2, zero_f64m1, max_vector_length)) *
+                            inv_points_count;
+    nk_f64_t centroid_b_z = __riscv_vfmv_f_s_f64m1_f64(
+                                __riscv_vfredusum_vs_f64m2_f64m1(sum_b_z_f64m2, zero_f64m1, max_vector_length)) *
+                            inv_points_count;
+    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;
+
+    nk_f64_t sum_squared = __riscv_vfmv_f_s_f64m1_f64(
+        __riscv_vfredusum_vs_f64m2_f64m1(sum_squared_f64m2, zero_f64m1, max_vector_length));
+    nk_f64_t mean_diff_x = centroid_a_x - centroid_b_x, mean_diff_y = centroid_a_y - centroid_b_y,
+             mean_diff_z = centroid_a_z - centroid_b_z;
+    nk_f64_t mean_diff_sq = mean_diff_x * mean_diff_x + mean_diff_y * mean_diff_y + mean_diff_z * mean_diff_z;
+    *result = nk_f64_sqrt_rvv(sum_squared * inv_points_count - mean_diff_sq);
 }
 
-NK_PUBLIC void nk_rmsd_f64_rvv(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *a_centroid,
+NK_PUBLIC void nk_rmsd_f64_rvv(nk_f64_t const *a, nk_f64_t const *b, nk_size_t points_count, nk_f64_t *a_centroid,
                                nk_f64_t *b_centroid, nk_f64_t *rotation, nk_f64_t *scale, nk_f64_t *result) {
-    nk_f64_t identity[9] = {1, 0, 0, 0, 1, 0, 0, 0, 1};
     if (rotation)
-        for (int j = 0; j < 9; ++j) rotation[j] = identity[j];
+        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;
-    nk_f64_t ca_x, ca_y, ca_z, cb_x, cb_y, cb_z;
-    nk_bicentroid_f64_rvv_(a, b, n, &ca_x, &ca_y, &ca_z, &cb_x, &cb_y, &cb_z);
-    if (a_centroid) a_centroid[0] = ca_x, a_centroid[1] = ca_y, a_centroid[2] = ca_z;
-    if (b_centroid) b_centroid[0] = cb_x, b_centroid[1] = cb_y, b_centroid[2] = cb_z;
-    nk_f64_t ssd = nk_transformed_ssd_f64_rvv_(a, b, n, identity, 1.0, ca_x, ca_y, ca_z, cb_x, cb_y, cb_z);
-    *result = nk_f64_sqrt_rvv(ssd / (nk_f64_t)n);
+
+    // Fused single-pass: accumulate centroids and squared differences simultaneously.
+    // RMSD = √(E[(a−b)²] − (ā − b̄)²)
+    nk_size_t max_vector_length = __riscv_vsetvlmax_e64m1();
+    vfloat64m1_t sum_a_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_a_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_a_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_b_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_b_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_b_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_a_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_a_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_a_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_b_x_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_b_y_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_b_z_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t sum_squared_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    vfloat64m1_t compensation_squared_f64m1 = __riscv_vfmv_v_f_f64m1(0.0, max_vector_length);
+    nk_f64_t const *a_ptr = a, *b_ptr = b;
+    nk_size_t remaining = points_count;
+    for (nk_size_t vector_length; remaining > 0;
+         remaining -= vector_length, a_ptr += vector_length * 3, b_ptr += vector_length * 3) {
+        vector_length = __riscv_vsetvl_e64m1(remaining);
+        vfloat64m1x3_t a_f64m1x3 = __riscv_vlseg3e64_v_f64m1x3(a_ptr, vector_length);
+        vfloat64m1_t a_x_f64m1 = __riscv_vget_v_f64m1x3_f64m1(a_f64m1x3, 0);
+        vfloat64m1_t a_y_f64m1 = __riscv_vget_v_f64m1x3_f64m1(a_f64m1x3, 1);
+        vfloat64m1_t a_z_f64m1 = __riscv_vget_v_f64m1x3_f64m1(a_f64m1x3, 2);
+        vfloat64m1x3_t b_f64m1x3 = __riscv_vlseg3e64_v_f64m1x3(b_ptr, vector_length);
+        vfloat64m1_t b_x_f64m1 = __riscv_vget_v_f64m1x3_f64m1(b_f64m1x3, 0);
+        vfloat64m1_t b_y_f64m1 = __riscv_vget_v_f64m1x3_f64m1(b_f64m1x3, 1);
+        vfloat64m1_t b_z_f64m1 = __riscv_vget_v_f64m1x3_f64m1(b_f64m1x3, 2);
+        // Accumulate centroids with Kahan compensation.
+        nk_accumulate_sum_f64m1_rvv_(&sum_a_x_f64m1, &compensation_a_x_f64m1, a_x_f64m1, vector_length);
+        nk_accumulate_sum_f64m1_rvv_(&sum_a_y_f64m1, &compensation_a_y_f64m1, a_y_f64m1, vector_length);
+        nk_accumulate_sum_f64m1_rvv_(&sum_a_z_f64m1, &compensation_a_z_f64m1, a_z_f64m1, vector_length);
+        nk_accumulate_sum_f64m1_rvv_(&sum_b_x_f64m1, &compensation_b_x_f64m1, b_x_f64m1, vector_length);
+        nk_accumulate_sum_f64m1_rvv_(&sum_b_y_f64m1, &compensation_b_y_f64m1, b_y_f64m1, vector_length);
+        nk_accumulate_sum_f64m1_rvv_(&sum_b_z_f64m1, &compensation_b_z_f64m1, b_z_f64m1, vector_length);
+        // Accumulate (a-b)^2 per component.
+        vfloat64m1_t delta_x_f64m1 = __riscv_vfsub_vv_f64m1(a_x_f64m1, b_x_f64m1, vector_length);
+        vfloat64m1_t delta_y_f64m1 = __riscv_vfsub_vv_f64m1(a_y_f64m1, b_y_f64m1, vector_length);
+        vfloat64m1_t delta_z_f64m1 = __riscv_vfsub_vv_f64m1(a_z_f64m1, b_z_f64m1, vector_length);
+        vfloat64m1_t dist_sq_f64m1 = __riscv_vfmul_vv_f64m1(delta_x_f64m1, delta_x_f64m1, vector_length);
+        dist_sq_f64m1 = __riscv_vfmacc_vv_f64m1(dist_sq_f64m1, delta_y_f64m1, delta_y_f64m1, vector_length);
+        dist_sq_f64m1 = __riscv_vfmacc_vv_f64m1(dist_sq_f64m1, delta_z_f64m1, delta_z_f64m1, vector_length);
+        nk_accumulate_sum_f64m1_rvv_(&sum_squared_f64m1, &compensation_squared_f64m1, dist_sq_f64m1, vector_length);
+    }
+    nk_f64_t inv_points_count = 1.0 / (nk_f64_t)points_count;
+    nk_f64_t centroid_a_x = nk_dot_stable_sum_f64m1_rvv_(sum_a_x_f64m1, compensation_a_x_f64m1) * inv_points_count;
+    nk_f64_t centroid_a_y = nk_dot_stable_sum_f64m1_rvv_(sum_a_y_f64m1, compensation_a_y_f64m1) * inv_points_count;
+    nk_f64_t centroid_a_z = nk_dot_stable_sum_f64m1_rvv_(sum_a_z_f64m1, compensation_a_z_f64m1) * inv_points_count;
+    nk_f64_t centroid_b_x = nk_dot_stable_sum_f64m1_rvv_(sum_b_x_f64m1, compensation_b_x_f64m1) * inv_points_count;
+    nk_f64_t centroid_b_y = nk_dot_stable_sum_f64m1_rvv_(sum_b_y_f64m1, compensation_b_y_f64m1) * inv_points_count;
+    nk_f64_t centroid_b_z = nk_dot_stable_sum_f64m1_rvv_(sum_b_z_f64m1, compensation_b_z_f64m1) * inv_points_count;
+    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;
+
+    nk_f64_t sum_squared = nk_dot_stable_sum_f64m1_rvv_(sum_squared_f64m1, compensation_squared_f64m1);
+    nk_f64_t mean_diff_x = centroid_a_x - centroid_b_x, mean_diff_y = centroid_a_y - centroid_b_y,
+             mean_diff_z = centroid_a_z - centroid_b_z;
+    nk_f64_t mean_diff_sq = mean_diff_x * mean_diff_x + mean_diff_y * mean_diff_y + mean_diff_z * mean_diff_z;
+    *result = nk_f64_sqrt_rvv(sum_squared * inv_points_count - mean_diff_sq);
 }
 
-NK_PUBLIC void nk_kabsch_f32_rvv(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+NK_PUBLIC void nk_kabsch_f32_rvv(nk_f32_t const *a, nk_f32_t const *b, nk_size_t points_count, nk_f32_t *a_centroid,
                                  nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f64_t *result) {
     if (scale) *scale = 1.0f;
-    nk_f64_t ca_x, ca_y, ca_z, cb_x, cb_y, cb_z;
+    nk_f64_t centroid_a_x, centroid_a_y, centroid_a_z, centroid_b_x, centroid_b_y, centroid_b_z;
     nk_f64_t h[9];
-    nk_centroid_and_cross_covariance_f32_rvv_(a, b, n, &ca_x, &ca_y, &ca_z, &cb_x, &cb_y, &cb_z, h);
-    if (a_centroid) a_centroid[0] = (nk_f32_t)ca_x, a_centroid[1] = (nk_f32_t)ca_y, a_centroid[2] = (nk_f32_t)ca_z;
-    if (b_centroid) b_centroid[0] = (nk_f32_t)cb_x, b_centroid[1] = (nk_f32_t)cb_y, b_centroid[2] = (nk_f32_t)cb_z;
+    nk_centroid_and_cross_covariance_f32_rvv_(a, b, points_count, &centroid_a_x, &centroid_a_y, &centroid_a_z,
+                                              &centroid_b_x, &centroid_b_y, &centroid_b_z, h);
+    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;
     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];
@@ -791,18 +875,20 @@ NK_PUBLIC void nk_kabsch_f32_rvv(nk_f32_t const *a, nk_f32_t const *b, nk_size_t
     }
     if (rotation)
         for (int j = 0; j < 9; ++j) rotation[j] = (nk_f32_t)r[j];
-    nk_f64_t ssd = nk_transformed_ssd_f32_rvv_(a, b, n, r, 1.0, ca_x, ca_y, ca_z, cb_x, cb_y, cb_z);
-    *result = nk_f64_sqrt_rvv(ssd / (nk_f64_t)n);
+    nk_f64_t ssd = nk_transformed_ssd_f32_rvv_(a, b, points_count, 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_rvv(ssd / (nk_f64_t)points_count);
 }
 
-NK_PUBLIC void nk_kabsch_f64_rvv(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *a_centroid,
+NK_PUBLIC void nk_kabsch_f64_rvv(nk_f64_t const *a, nk_f64_t const *b, nk_size_t points_count, nk_f64_t *a_centroid,
                                  nk_f64_t *b_centroid, nk_f64_t *rotation, nk_f64_t *scale, nk_f64_t *result) {
     if (scale) *scale = 1.0;
-    nk_f64_t ca_x, ca_y, ca_z, cb_x, cb_y, cb_z;
+    nk_f64_t centroid_a_x, centroid_a_y, centroid_a_z, centroid_b_x, centroid_b_y, centroid_b_z;
     nk_f64_t h[9];
-    nk_centroid_and_cross_covariance_f64_rvv_(a, b, n, &ca_x, &ca_y, &ca_z, &cb_x, &cb_y, &cb_z, h);
-    if (a_centroid) a_centroid[0] = ca_x, a_centroid[1] = ca_y, a_centroid[2] = ca_z;
-    if (b_centroid) b_centroid[0] = cb_x, b_centroid[1] = cb_y, b_centroid[2] = cb_z;
+    nk_centroid_and_cross_covariance_f64_rvv_(a, b, points_count, &centroid_a_x, &centroid_a_y, &centroid_a_z,
+                                              &centroid_b_x, &centroid_b_y, &centroid_b_z, h);
+    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;
     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];
@@ -813,18 +899,24 @@ NK_PUBLIC void nk_kabsch_f64_rvv(nk_f64_t const *a, nk_f64_t const *b, nk_size_t
     }
     if (rotation)
         for (int j = 0; j < 9; ++j) rotation[j] = r[j];
-    nk_f64_t ssd = nk_transformed_ssd_f64_rvv_(a, b, n, r, 1.0, ca_x, ca_y, ca_z, cb_x, cb_y, cb_z);
-    *result = nk_f64_sqrt_rvv(ssd / (nk_f64_t)n);
+    nk_f64_t ssd = nk_transformed_ssd_f64_rvv_(a, b, points_count, 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_rvv(ssd / (nk_f64_t)points_count);
 }
 
-NK_PUBLIC void nk_umeyama_f32_rvv(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+NK_PUBLIC void nk_umeyama_f32_rvv(nk_f32_t const *a, nk_f32_t const *b, nk_size_t points_count, nk_f32_t *a_centroid,
                                   nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f64_t *result) {
-    nk_f64_t ca_x, ca_y, ca_z, cb_x, cb_y, cb_z;
+    nk_f64_t centroid_a_x, centroid_a_y, centroid_a_z, centroid_b_x, centroid_b_y, centroid_b_z;
     nk_f64_t h[9], variance_a;
-    nk_centroid_and_cross_covariance_and_variance_f32_rvv_(a, b, n, &ca_x, &ca_y, &ca_z, &cb_x, &cb_y, &cb_z, h,
-                                                           &variance_a);
-    if (a_centroid) a_centroid[0] = (nk_f32_t)ca_x, a_centroid[1] = (nk_f32_t)ca_y, a_centroid[2] = (nk_f32_t)ca_z;
-    if (b_centroid) b_centroid[0] = (nk_f32_t)cb_x, b_centroid[1] = (nk_f32_t)cb_y, b_centroid[2] = (nk_f32_t)cb_z;
+    nk_centroid_and_cross_covariance_and_variance_f32_rvv_(a, b, points_count, &centroid_a_x, &centroid_a_y,
+                                                           &centroid_a_z, &centroid_b_x, &centroid_b_y, &centroid_b_z,
+                                                           h, &variance_a);
+    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;
     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];
@@ -832,7 +924,7 @@ NK_PUBLIC void nk_umeyama_f32_rvv(nk_f32_t const *a, nk_f32_t const *b, nk_size_
     nk_f64_t det = nk_det3x3_f64_(r);
     nk_f64_t sign_det = det < 0 ? -1.0 : 1.0;
     nk_f64_t trace_ds = nk_sum_three_products_f64_(svd_s[0], 1.0, svd_s[4], 1.0, svd_s[8], sign_det);
-    nk_f64_t scale_factor = trace_ds / ((nk_f64_t)n * variance_a);
+    nk_f64_t scale_factor = trace_ds / ((nk_f64_t)points_count * variance_a);
     if (scale) *scale = (nk_f32_t)scale_factor;
     if (det < 0) {
         svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
@@ -840,18 +932,20 @@ NK_PUBLIC void nk_umeyama_f32_rvv(nk_f32_t const *a, nk_f32_t const *b, nk_size_
     }
     if (rotation)
         for (int j = 0; j < 9; ++j) rotation[j] = (nk_f32_t)r[j];
-    nk_f64_t ssd = nk_transformed_ssd_f32_rvv_(a, b, n, r, scale_factor, ca_x, ca_y, ca_z, cb_x, cb_y, cb_z);
-    *result = nk_f64_sqrt_rvv(ssd / (nk_f64_t)n);
+    nk_f64_t ssd = nk_transformed_ssd_f32_rvv_(a, b, points_count, r, scale_factor, centroid_a_x, centroid_a_y,
+                                               centroid_a_z, centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f64_sqrt_rvv(ssd / (nk_f64_t)points_count);
 }
 
-NK_PUBLIC void nk_umeyama_f64_rvv(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *a_centroid,
+NK_PUBLIC void nk_umeyama_f64_rvv(nk_f64_t const *a, nk_f64_t const *b, nk_size_t points_count, nk_f64_t *a_centroid,
                                   nk_f64_t *b_centroid, nk_f64_t *rotation, nk_f64_t *scale, nk_f64_t *result) {
-    nk_f64_t ca_x, ca_y, ca_z, cb_x, cb_y, cb_z;
+    nk_f64_t centroid_a_x, centroid_a_y, centroid_a_z, centroid_b_x, centroid_b_y, centroid_b_z;
     nk_f64_t h[9], variance_a;
-    nk_centroid_and_cross_covariance_and_variance_f64_rvv_(a, b, n, &ca_x, &ca_y, &ca_z, &cb_x, &cb_y, &cb_z, h,
-                                                           &variance_a);
-    if (a_centroid) a_centroid[0] = ca_x, a_centroid[1] = ca_y, a_centroid[2] = ca_z;
-    if (b_centroid) b_centroid[0] = cb_x, b_centroid[1] = cb_y, b_centroid[2] = cb_z;
+    nk_centroid_and_cross_covariance_and_variance_f64_rvv_(a, b, points_count, &centroid_a_x, &centroid_a_y,
+                                                           &centroid_a_z, &centroid_b_x, &centroid_b_y, &centroid_b_z,
+                                                           h, &variance_a);
+    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;
     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];
@@ -859,7 +953,7 @@ NK_PUBLIC void nk_umeyama_f64_rvv(nk_f64_t const *a, nk_f64_t const *b, nk_size_
     nk_f64_t det = nk_det3x3_f64_(r);
     nk_f64_t sign_det = det < 0 ? -1.0 : 1.0;
     nk_f64_t trace_ds = nk_sum_three_products_f64_(svd_s[0], 1.0, svd_s[4], 1.0, svd_s[8], sign_det);
-    nk_f64_t scale_factor = trace_ds / ((nk_f64_t)n * variance_a);
+    nk_f64_t scale_factor = trace_ds / ((nk_f64_t)points_count * variance_a);
     if (scale) *scale = scale_factor;
     if (det < 0) {
         svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];
@@ -867,38 +961,39 @@ NK_PUBLIC void nk_umeyama_f64_rvv(nk_f64_t const *a, nk_f64_t const *b, nk_size_
     }
     if (rotation)
         for (int j = 0; j < 9; ++j) rotation[j] = r[j];
-    nk_f64_t ssd = nk_transformed_ssd_f64_rvv_(a, b, n, r, scale_factor, ca_x, ca_y, ca_z, cb_x, cb_y, cb_z);
-    *result = nk_f64_sqrt_rvv(ssd / (nk_f64_t)n);
+    nk_f64_t ssd = nk_transformed_ssd_f64_rvv_(a, b, points_count, r, scale_factor, centroid_a_x, centroid_a_y,
+                                               centroid_a_z, centroid_b_x, centroid_b_y, centroid_b_z);
+    *result = nk_f64_sqrt_rvv(ssd / (nk_f64_t)points_count);
 }
 
-NK_PUBLIC void nk_rmsd_f16_rvv(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+NK_PUBLIC void nk_rmsd_f16_rvv(nk_f16_t const *a, nk_f16_t const *b, nk_size_t points_count, nk_f32_t *a_centroid,
                                nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
-    nk_rmsd_f16_serial(a, b, n, a_centroid, b_centroid, rotation, scale, result);
+    nk_rmsd_f16_serial(a, b, points_count, a_centroid, b_centroid, rotation, scale, result);
 }
 
-NK_PUBLIC void nk_kabsch_f16_rvv(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+NK_PUBLIC void nk_kabsch_f16_rvv(nk_f16_t const *a, nk_f16_t const *b, nk_size_t points_count, nk_f32_t *a_centroid,
                                  nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
-    nk_kabsch_f16_serial(a, b, n, a_centroid, b_centroid, rotation, scale, result);
+    nk_kabsch_f16_serial(a, b, points_count, a_centroid, b_centroid, rotation, scale, result);
 }
 
-NK_PUBLIC void nk_umeyama_f16_rvv(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+NK_PUBLIC void nk_umeyama_f16_rvv(nk_f16_t const *a, nk_f16_t const *b, nk_size_t points_count, nk_f32_t *a_centroid,
                                   nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
-    nk_umeyama_f16_serial(a, b, n, a_centroid, b_centroid, rotation, scale, result);
+    nk_umeyama_f16_serial(a, b, points_count, a_centroid, b_centroid, rotation, scale, result);
 }
 
-NK_PUBLIC void nk_rmsd_bf16_rvv(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+NK_PUBLIC void nk_rmsd_bf16_rvv(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t points_count, nk_f32_t *a_centroid,
                                 nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
-    nk_rmsd_bf16_serial(a, b, n, a_centroid, b_centroid, rotation, scale, result);
+    nk_rmsd_bf16_serial(a, b, points_count, a_centroid, b_centroid, rotation, scale, result);
 }
 
-NK_PUBLIC void nk_kabsch_bf16_rvv(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+NK_PUBLIC void nk_kabsch_bf16_rvv(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t points_count, nk_f32_t *a_centroid,
                                   nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
-    nk_kabsch_bf16_serial(a, b, n, a_centroid, b_centroid, rotation, scale, result);
+    nk_kabsch_bf16_serial(a, b, points_count, a_centroid, b_centroid, rotation, scale, result);
 }
 
-NK_PUBLIC void nk_umeyama_bf16_rvv(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *a_centroid,
+NK_PUBLIC void nk_umeyama_bf16_rvv(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t points_count, nk_f32_t *a_centroid,
                                    nk_f32_t *b_centroid, nk_f32_t *rotation, nk_f32_t *scale, nk_f32_t *result) {
-    nk_umeyama_bf16_serial(a, b, n, a_centroid, b_centroid, rotation, scale, result);
+    nk_umeyama_bf16_serial(a, b, points_count, a_centroid, b_centroid, rotation, scale, result);
 }
 
 #if defined(__cplusplus)