numkong 7.0.0

package/include/numkong/mesh/rvv.h

@@ -0,0 +1,916 @@
+/**
+ *  @brief SIMD-accelerated Mesh Operations for RISC-V.
+ *  @file include/numkong/mesh/rvv.h
+ *  @author Ash Vardanian
+ *  @date February 6, 2026
+ *
+ *  @sa include/numkong/mesh.h
+ *
+ *  RVV mesh operations leverage:
+ *
+ *  - `vlseg3e32`/`vlseg3e64`: deinterleave xyz triplets in hardware
+ *  - `vfwcvt`/`vfwmacc`: widening FMA for f32→f64 accumulation
+ *  - `vfredusum`: single-instruction horizontal reduction
+ *  - Serial SVD/determinant from mesh/serial.h for fixed 3×3 matrix operations
+ *
+ *  Fused helpers minimize data passes:
+ *
+ *  - `nk_bicentroid_*_rvv_`: both centroids in a single pass (used by RMSD)
+ *  - `nk_centroid_and_cross_covariance_*_rvv_`: centroids + H in one pass (Kabsch)
+ *  - `nk_centroid_and_cross_covariance_and_variance_*_rvv_`: + variance (Umeyama)
+ *
+ *  Math for fused centroid+covariance:
+ *    H[i][j] = Σ (a[i] - ca[i]) * (b[j] - cb[j])
+ *            = Σ a[i] * b[j] - n * ca[i] * cb[j]
+ *  So we accumulate raw Σ a[i] * b[j] in the loop, then fix up after.
+ *
+ *  Key RVV-specific optimizations (vs. scalar or x86 backends):
+ *
+ *  - Deferred horizontal reduction in bicentroid: per-lane `vfwadd_wv` (f32)
+ *    or `vfadd_vv` (f64) accumulation across loop iterations, with a single
+ *    `vfredusum` after the loop — eliminates 6 `vfredusum` per iteration.
+ *  - `vfwmacc_vv` in f32 SSD: accumulates widened squared distances per-lane
+ *    (dx²+dy²+dz²) before a single reduction — saves 2 `vfredusum` per iteration.
+ *  - Vectorized R = V×Uᵀ via `vfmul_vf`/`vfmacc_vf`: each output row computed
+ *    as a 3-element vector dot product — 15 vector ops vs 45 scalar ops.
+ *  - `vfncvt_f_f_w` for f64→f32 narrowing of H matrix before SVD.
+ */
+#ifndef NK_MESH_RVV_H
+#define NK_MESH_RVV_H
+
+#if NK_TARGET_RISCV_
+#if NK_TARGET_RVV
+
+#include "numkong/types.h"
+#include "numkong/dot/rvv.h"
+#include "numkong/spatial/rvv.h" // `nk_f32_sqrt_rvv`, `nk_f64_sqrt_rvv`
+#include "numkong/mesh/serial.h" // `nk_svd3x3_f32_`, `nk_svd3x3_f64_`, `nk_det3x3_f32_`, `nk_det3x3_f64_`
+
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((target("arch=+v"))), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("arch=+v")
+#endif
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+NK_INTERNAL void nk_accumulate_sum_f64m1_rvv_(vfloat64m1_t *sum_f64m1, vfloat64m1_t *compensation_f64m1,
+                                              vfloat64m1_t addend_f64m1, nk_size_t vector_length) {
+    vfloat64m1_t tentative_sum_f64m1 = __riscv_vfadd_vv_f64m1(*sum_f64m1, addend_f64m1, vector_length);
+    vfloat64m1_t virtual_addend_f64m1 = __riscv_vfsub_vv_f64m1(tentative_sum_f64m1, *sum_f64m1, vector_length);
+    vfloat64m1_t sum_error_f64m1 = __riscv_vfadd_vv_f64m1(
+        __riscv_vfsub_vv_f64m1(*sum_f64m1,
+                               __riscv_vfsub_vv_f64m1(tentative_sum_f64m1, virtual_addend_f64m1, vector_length),
+                               vector_length),
+        __riscv_vfsub_vv_f64m1(addend_f64m1, virtual_addend_f64m1, vector_length), vector_length);
+    *sum_f64m1 = __riscv_vslideup_vx_f64m1_tu(*sum_f64m1, tentative_sum_f64m1, 0, vector_length);
+    *compensation_f64m1 = __riscv_vfadd_vv_f64m1_tu(*compensation_f64m1, *compensation_f64m1, sum_error_f64m1,
+                                                    vector_length);
+}
+
+NK_INTERNAL void nk_accumulate_product_f64m1_rvv_(vfloat64m1_t *sum_f64m1, vfloat64m1_t *compensation_f64m1,
+                                                  vfloat64m1_t left_f64m1, vfloat64m1_t right_f64m1,
+                                                  nk_size_t vector_length) {
+    vfloat64m1_t product_f64m1 = __riscv_vfmul_vv_f64m1(left_f64m1, right_f64m1, vector_length);
+    vfloat64m1_t product_error_f64m1 = __riscv_vfmsac_vv_f64m1(product_f64m1, left_f64m1, right_f64m1, vector_length);
+    vfloat64m1_t tentative_sum_f64m1 = __riscv_vfadd_vv_f64m1(*sum_f64m1, product_f64m1, vector_length);
+    vfloat64m1_t virtual_addend_f64m1 = __riscv_vfsub_vv_f64m1(tentative_sum_f64m1, *sum_f64m1, vector_length);
+    vfloat64m1_t sum_error_f64m1 = __riscv_vfadd_vv_f64m1(
+        __riscv_vfsub_vv_f64m1(*sum_f64m1,
+                               __riscv_vfsub_vv_f64m1(tentative_sum_f64m1, virtual_addend_f64m1, vector_length),
+                               vector_length),
+        __riscv_vfsub_vv_f64m1(product_f64m1, virtual_addend_f64m1, vector_length), vector_length);
+    *sum_f64m1 = __riscv_vslideup_vx_f64m1_tu(*sum_f64m1, tentative_sum_f64m1, 0, vector_length);
+    vfloat64m1_t total_error_f64m1 = __riscv_vfadd_vv_f64m1(sum_error_f64m1, product_error_f64m1, vector_length);
+    *compensation_f64m1 = __riscv_vfadd_vv_f64m1_tu(*compensation_f64m1, *compensation_f64m1, total_error_f64m1,
+                                                    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).
+ *
+ *  Accumulates raw Σ a[i]*b[j] and Σ a[i], Σ b[j] simultaneously, then:
+ *    ca = Σa / n,  cb = Σb / n
+ *    H[i][j] = raw[i][j] - n * ca[i] * cb[j]
+ *
+ *  Reduces Kabsch from 4 passes to 2 (fused centroid+covariance + SSD).
+ *  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_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_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);
+        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);
+        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);
+        cross_00_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_00_f64m2, a_x_f32m1, b_x_f32m1, vector_length);
+        cross_01_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_01_f64m2, a_x_f32m1, b_y_f32m1, vector_length);
+        cross_02_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_02_f64m2, a_x_f32m1, b_z_f32m1, vector_length);
+        cross_10_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_10_f64m2, a_y_f32m1, b_x_f32m1, vector_length);
+        cross_11_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_11_f64m2, a_y_f32m1, b_y_f32m1, vector_length);
+        cross_12_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_12_f64m2, a_y_f32m1, b_z_f32m1, vector_length);
+        cross_20_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_20_f64m2, a_z_f32m1, b_x_f32m1, vector_length);
+        cross_21_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_21_f64m2, a_z_f32m1, b_y_f32m1, vector_length);
+        cross_22_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_22_f64m2, a_z_f32m1, b_z_f32m1, vector_length);
+    }
+    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_;
+}
+
+/**
+ *  @brief Compute centroids and cross-covariance matrix in a single pass (f64).
+ *
+ *  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_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_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);
+        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);
+        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);
+        nk_accumulate_product_f64m1_rvv_(&cross_00_f64m1, &compensation_00_f64m1, a_x_f64m1, b_x_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_01_f64m1, &compensation_01_f64m1, a_x_f64m1, b_y_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_02_f64m1, &compensation_02_f64m1, a_x_f64m1, b_z_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_10_f64m1, &compensation_10_f64m1, a_y_f64m1, b_x_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_11_f64m1, &compensation_11_f64m1, a_y_f64m1, b_y_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_12_f64m1, &compensation_12_f64m1, a_y_f64m1, b_z_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_20_f64m1, &compensation_20_f64m1, a_z_f64m1, b_x_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_21_f64m1, &compensation_21_f64m1, a_z_f64m1, b_y_f64m1, vector_length);
+        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_;
+}
+
+/**
+ *  @brief Compute centroids, cross-covariance, and variance_a in a single pass (f32).
+ *
+ *  Same as centroid_and_cross_covariance but also computes:
+ *    variance_a = (1/n) * Σ ||a[i] - ca||²
+ *               = (1/n) * (Σ ||a[i]||² - n * ||ca||²)
+ *
+ *  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_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_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);
+        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);
+        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);
+        cross_00_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_00_f64m2, a_x_f32m1, b_x_f32m1, vector_length);
+        cross_01_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_01_f64m2, a_x_f32m1, b_y_f32m1, vector_length);
+        cross_02_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_02_f64m2, a_x_f32m1, b_z_f32m1, vector_length);
+        cross_10_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_10_f64m2, a_y_f32m1, b_x_f32m1, vector_length);
+        cross_11_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_11_f64m2, a_y_f32m1, b_y_f32m1, vector_length);
+        cross_12_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_12_f64m2, a_y_f32m1, b_z_f32m1, vector_length);
+        cross_20_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_20_f64m2, a_z_f32m1, b_x_f32m1, vector_length);
+        cross_21_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_21_f64m2, a_z_f32m1, b_y_f32m1, vector_length);
+        cross_22_f64m2 = __riscv_vfwmacc_vv_f64m2_tu(cross_22_f64m2, a_z_f32m1, b_z_f32m1, vector_length);
+        // Variance: Σ (a_x² + a_y² + a_z²) — raw, not centered.
+        vfloat64m2_t norm_squared_f64m2 = __riscv_vfwmul_vv_f64m2(a_x_f32m1, a_x_f32m1, vector_length);
+        norm_squared_f64m2 = __riscv_vfwmacc_vv_f64m2(norm_squared_f64m2, a_y_f32m1, a_y_f32m1, vector_length);
+        norm_squared_f64m2 = __riscv_vfwmacc_vv_f64m2(norm_squared_f64m2, a_z_f32m1, a_z_f32m1, vector_length);
+        sum_norm_squared_f64m2 = __riscv_vfadd_vv_f64m2_tu(sum_norm_squared_f64m2, sum_norm_squared_f64m2,
+                                                           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)
+    *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_);
+}
+
+/**
+ *  @brief Compute centroids, cross-covariance, and variance_a in a single pass (f64).
+ *
+ *  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_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_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);
+        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);
+        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);
+        nk_accumulate_product_f64m1_rvv_(&cross_00_f64m1, &compensation_00_f64m1, a_x_f64m1, b_x_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_01_f64m1, &compensation_01_f64m1, a_x_f64m1, b_y_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_02_f64m1, &compensation_02_f64m1, a_x_f64m1, b_z_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_10_f64m1, &compensation_10_f64m1, a_y_f64m1, b_x_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_11_f64m1, &compensation_11_f64m1, a_y_f64m1, b_y_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_12_f64m1, &compensation_12_f64m1, a_y_f64m1, b_z_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_20_f64m1, &compensation_20_f64m1, a_z_f64m1, b_x_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_21_f64m1, &compensation_21_f64m1, a_z_f64m1, b_y_f64m1, vector_length);
+        nk_accumulate_product_f64m1_rvv_(&cross_22_f64m1, &compensation_22_f64m1, a_z_f64m1, b_z_f64m1, vector_length);
+        vfloat64m1_t norm_squared_f64m1 = __riscv_vfmul_vv_f64m1(a_x_f64m1, a_x_f64m1, vector_length);
+        norm_squared_f64m1 = __riscv_vfmacc_vv_f64m1(norm_squared_f64m1, a_y_f64m1, a_y_f64m1, vector_length);
+        norm_squared_f64m1 = __riscv_vfmacc_vv_f64m1(norm_squared_f64m1, a_z_f64m1, a_z_f64m1, vector_length);
+        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_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_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);
+    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;
+    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);
+        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);
+        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);
+        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);
+        rotated_a_x_f64m2 = __riscv_vfmacc_vf_f64m2(rotated_a_x_f64m2, scaled_rotation_x_z, centered_a_z_f64m2,
+                                                    vector_length);
+        vfloat64m2_t rotated_a_y_f64m2 = __riscv_vfmul_vf_f64m2(centered_a_x_f64m2, scaled_rotation_y_x, vector_length);
+        rotated_a_y_f64m2 = __riscv_vfmacc_vf_f64m2(rotated_a_y_f64m2, scaled_rotation_y_y, centered_a_y_f64m2,
+                                                    vector_length);
+        rotated_a_y_f64m2 = __riscv_vfmacc_vf_f64m2(rotated_a_y_f64m2, scaled_rotation_y_z, centered_a_z_f64m2,
+                                                    vector_length);
+        vfloat64m2_t rotated_a_z_f64m2 = __riscv_vfmul_vf_f64m2(centered_a_x_f64m2, scaled_rotation_z_x, vector_length);
+        rotated_a_z_f64m2 = __riscv_vfmacc_vf_f64m2(rotated_a_z_f64m2, scaled_rotation_z_y, centered_a_y_f64m2,
+                                                    vector_length);
+        rotated_a_z_f64m2 = __riscv_vfmacc_vf_f64m2(rotated_a_z_f64m2, scaled_rotation_z_z, centered_a_z_f64m2,
+                                                    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);
+        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);
+        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);
+        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);
+        sum_distance_squared_f64m2 = __riscv_vfmacc_vv_f64m2_tu(sum_distance_squared_f64m2, delta_x_f64m2,
+                                                                delta_x_f64m2, vector_length);
+        sum_distance_squared_f64m2 = __riscv_vfmacc_vv_f64m2_tu(sum_distance_squared_f64m2, delta_y_f64m2,
+                                                                delta_y_f64m2, vector_length);
+        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));
+}
+
+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_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_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);
+        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 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);
+        rotated_a_x_f64m1 = __riscv_vfmacc_vf_f64m1(rotated_a_x_f64m1, scaled_rotation_x_z, centered_a_z_f64m1,
+                                                    vector_length);
+        vfloat64m1_t rotated_a_y_f64m1 = __riscv_vfmul_vf_f64m1(centered_a_x_f64m1, scaled_rotation_y_x, vector_length);
+        rotated_a_y_f64m1 = __riscv_vfmacc_vf_f64m1(rotated_a_y_f64m1, scaled_rotation_y_y, centered_a_y_f64m1,
+                                                    vector_length);
+        rotated_a_y_f64m1 = __riscv_vfmacc_vf_f64m1(rotated_a_y_f64m1, scaled_rotation_y_z, centered_a_z_f64m1,
+                                                    vector_length);
+        vfloat64m1_t rotated_a_z_f64m1 = __riscv_vfmul_vf_f64m1(centered_a_x_f64m1, scaled_rotation_z_x, vector_length);
+        rotated_a_z_f64m1 = __riscv_vfmacc_vf_f64m1(rotated_a_z_f64m1, scaled_rotation_z_y, centered_a_y_f64m1,
+                                                    vector_length);
+        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 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);
+        vfloat64m1_t distance_squared_f64m1 = __riscv_vfmul_vv_f64m1(delta_x_f64m1, delta_x_f64m1, vector_length);
+        distance_squared_f64m1 = __riscv_vfmacc_vv_f64m1(distance_squared_f64m1, delta_y_f64m1, delta_y_f64m1,
+                                                         vector_length);
+        distance_squared_f64m1 = __riscv_vfmacc_vv_f64m1(distance_squared_f64m1, delta_z_f64m1, delta_z_f64m1,
+                                                         vector_length);
+        nk_accumulate_sum_f64m1_rvv_(&sum_distance_squared_f64m1, &compensation_distance_squared_f64m1,
+                                     distance_squared_f64m1, vector_length);
+    }
+    return nk_dot_stable_sum_f64m1_rvv_(sum_distance_squared_f64m1, compensation_distance_squared_f64m1);
+}
+
+/** @brief Compute R = V * Uᵀ from SVD factors (f32), vectorized with `vfmul_vf`/`vfmacc_vf`. */
+NK_INTERNAL void nk_rotation_from_svd_f32_rvv_( //
+    nk_f32_t *svd_u, nk_f32_t *svd_v, nk_f32_t r[9]) {
+    nk_size_t vl3 = __riscv_vsetvl_e32m1(3);
+    vfloat32m1_t u_row0_f32m1 = __riscv_vle32_v_f32m1(svd_u + 0, vl3);
+    vfloat32m1_t u_row1_f32m1 = __riscv_vle32_v_f32m1(svd_u + 3, vl3);
+    vfloat32m1_t u_row2_f32m1 = __riscv_vle32_v_f32m1(svd_u + 6, vl3);
+    // Row 0: R[0..2] = V[0]*U_row0 + V[1]*U_row1 + V[2]*U_row2
+    vfloat32m1_t rotation_row_f32m1 = __riscv_vfmul_vf_f32m1(u_row0_f32m1, svd_v[0], vl3);
+    rotation_row_f32m1 = __riscv_vfmacc_vf_f32m1(rotation_row_f32m1, svd_v[1], u_row1_f32m1, vl3);
+    rotation_row_f32m1 = __riscv_vfmacc_vf_f32m1(rotation_row_f32m1, svd_v[2], u_row2_f32m1, vl3);
+    __riscv_vse32_v_f32m1(r + 0, rotation_row_f32m1, vl3);
+    // Row 1: R[3..5]
+    rotation_row_f32m1 = __riscv_vfmul_vf_f32m1(u_row0_f32m1, svd_v[3], vl3);
+    rotation_row_f32m1 = __riscv_vfmacc_vf_f32m1(rotation_row_f32m1, svd_v[4], u_row1_f32m1, vl3);
+    rotation_row_f32m1 = __riscv_vfmacc_vf_f32m1(rotation_row_f32m1, svd_v[5], u_row2_f32m1, vl3);
+    __riscv_vse32_v_f32m1(r + 3, rotation_row_f32m1, vl3);
+    // Row 2: R[6..8]
+    rotation_row_f32m1 = __riscv_vfmul_vf_f32m1(u_row0_f32m1, svd_v[6], vl3);
+    rotation_row_f32m1 = __riscv_vfmacc_vf_f32m1(rotation_row_f32m1, svd_v[7], u_row1_f32m1, vl3);
+    rotation_row_f32m1 = __riscv_vfmacc_vf_f32m1(rotation_row_f32m1, svd_v[8], u_row2_f32m1, vl3);
+    __riscv_vse32_v_f32m1(r + 6, rotation_row_f32m1, vl3);
+}
+
+/** @brief Compute R = V * Uᵀ from SVD factors (f64), vectorized with `vfmul_vf`/`vfmacc_vf`. */
+NK_INTERNAL void nk_rotation_from_svd_f64_rvv_( //
+    nk_f64_t *svd_u, nk_f64_t *svd_v, nk_f64_t r[9]) {
+    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_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];
+    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);
+}
+
+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_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];
+    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);
+}
+
+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_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 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_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];
+    nk_rotation_from_svd_f64_rvv_(svd_u, svd_v, 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_rvv_(svd_u, svd_v, r);
+    }
+    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_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_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 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_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];
+    nk_rotation_from_svd_f64_rvv_(svd_u, svd_v, 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_rvv_(svd_u, svd_v, r);
+    }
+    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_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_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 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_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];
+    nk_rotation_from_svd_f64_rvv_(svd_u, svd_v, r);
+    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);
+    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];
+        nk_rotation_from_svd_f64_rvv_(svd_u, svd_v, r);
+    }
+    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_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_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 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_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];
+    nk_rotation_from_svd_f64_rvv_(svd_u, svd_v, r);
+    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);
+    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];
+        nk_rotation_from_svd_f64_rvv_(svd_u, svd_v, r);
+    }
+    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_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_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_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_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_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_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_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_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_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_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_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_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);
+}
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#endif // NK_TARGET_RVV
+#endif // NK_TARGET_RISCV_
+#endif // NK_MESH_RVV_H