usearch 2.21.3 → 2.25.0

package/numkong/include/numkong/maxsim.hpp

@@ -0,0 +1,133 @@
+/**
+ *  @brief C++ bindings for multi-target MaxSim (ColBERT late-interaction) kernels.
+ *  @file include/numkong/maxsim.hpp
+ *  @author Ash Vardanian
+ *  @date February 28, 2026
+ */
+#ifndef NK_MAXSIM_HPP
+#define NK_MAXSIM_HPP
+
+#include <cstddef>
+#include <cstring>
+#include <limits>
+#include <type_traits>
+
+#include "numkong/maxsim.h"
+#include "numkong/types.hpp"
+#include "numkong/spatial.hpp" // angular<>
+
+namespace ashvardanian::numkong {
+
+/**
+ *  @brief Computes angular distance late-interaction on pre-packed vectors.
+ *  Returns Σᵢ minⱼ angular(qᵢ, dⱼ).
+ *  @param[in] query_packed Packed query vectors.
+ *  @param[in] document_packed Packed document vectors.
+ *  @param[in] query_count Number of query vectors.
+ *  @param[in] document_count Number of document vectors.
+ *  @param[in] depth Number of dimensions per vector.
+ *  @return Sum of per-query minimum angular distances.
+ *
+ *  @tparam in_type_ Input element type (bf16_t, f32_t, f16_t).
+ *  @tparam result_type_ Result type, defaults to `in_type_::maxsim_result_t`.
+ *  @tparam allow_simd_ Enable SIMD kernel dispatch when `prefer_simd_k`.
+ */
+template <numeric_dtype in_type_, numeric_dtype result_type_ = typename in_type_::maxsim_result_t,
+          allow_simd_t allow_simd_ = prefer_simd_k>
+NK_PUBLIC void maxsim_packed(void const *query_packed, void const *document_packed, std::size_t query_count,
+                             std::size_t document_count, std::size_t depth, result_type_ *result) {
+    constexpr bool simd = allow_simd_ == prefer_simd_k &&
+                          std::is_same_v<result_type_, typename in_type_::maxsim_result_t>;
+
+    if constexpr (std::is_same_v<in_type_, bf16_t> && simd)
+        nk_maxsim_packed_bf16(query_packed, document_packed, query_count, document_count, depth, &result->raw_);
+    else if constexpr (std::is_same_v<in_type_, f32_t> && simd)
+        nk_maxsim_packed_f32(query_packed, document_packed, query_count, document_count, depth, &result->raw_);
+    else if constexpr (std::is_same_v<in_type_, f16_t> && simd)
+        nk_maxsim_packed_f16(query_packed, document_packed, query_count, document_count, depth, &result->raw_);
+    else {
+        typename in_type_::raw_t const *q_ptr;
+        std::size_t q_stride;
+        char const *q_bytes = reinterpret_cast<char const *>(query_packed);
+        std::memcpy(&q_ptr, q_bytes, sizeof(void *));
+        std::memcpy(&q_stride, q_bytes + sizeof(void *), sizeof(std::size_t));
+
+        typename in_type_::raw_t const *d_ptr;
+        std::size_t d_stride;
+        char const *d_bytes = reinterpret_cast<char const *>(document_packed);
+        std::memcpy(&d_ptr, d_bytes, sizeof(void *));
+        std::memcpy(&d_stride, d_bytes + sizeof(void *), sizeof(std::size_t));
+
+        maxsim_reference<in_type_, result_type_>(q_ptr, query_count, q_stride, d_ptr, document_count, d_stride, depth,
+                                                 result);
+    }
+}
+
+/**
+ *  @brief Exhaustive angular reference for testing: Σᵢ minⱼ angular(qᵢ, dⱼ).
+ *  Computes all pairwise angular distances and picks the minimum per query.
+ *  Uses f64 accumulator for precision.
+ *  @param[in] queries Query vectors in row-major order.
+ *  @param[in] query_count Number of query vectors.
+ *  @param[in] query_stride Row stride in bytes for query vectors.
+ *  @param[in] documents Document vectors in row-major order.
+ *  @param[in] document_count Number of document vectors.
+ *  @param[in] document_stride Row stride in bytes for document vectors.
+ *  @param[in] depth Number of dimensions per vector.
+ *  @param[out] result Pointer to store the sum of per-query minimum angular distances.
+ *
+ *  @tparam in_type_ Input element type (bf16_t, f32_t, f16_t).
+ *  @tparam result_type_ Result type, defaults to `in_type_::angular_result_t`.
+ */
+template <numeric_dtype in_type_, numeric_dtype result_type_ = typename in_type_::angular_result_t,
+          allow_simd_t allow_simd_ = prefer_simd_k>
+NK_PUBLIC void maxsim_reference(typename in_type_::raw_t const *queries, std::size_t query_count,
+                                std::size_t query_stride, typename in_type_::raw_t const *documents,
+                                std::size_t document_count, std::size_t document_stride, std::size_t depth,
+                                result_type_ *result) {
+    result_type_ total_angular_distance {};
+
+    for (std::size_t query_index = 0; query_index < query_count; query_index++) {
+        in_type_ const *query_row = reinterpret_cast<in_type_ const *>(reinterpret_cast<char const *>(queries) +
+                                                                       query_index * query_stride);
+
+        result_type_ min_angular = result_type_::finite_max();
+
+        for (std::size_t document_index = 0; document_index < document_count; document_index++) {
+            in_type_ const *document_row = reinterpret_cast<in_type_ const *>(
+                reinterpret_cast<char const *>(documents) + document_index * document_stride);
+
+            result_type_ angular_distance {};
+            angular<in_type_, result_type_, allow_simd_>(query_row, document_row, depth, &angular_distance);
+
+            if (angular_distance < min_angular) min_angular = angular_distance;
+        }
+
+        total_angular_distance = total_angular_distance + min_angular;
+    }
+
+    *result = total_angular_distance;
+}
+
+} // namespace ashvardanian::numkong
+
+#include "numkong/matrix.hpp"
+
+namespace ashvardanian::numkong {
+
+/** @brief MaxSim: Σᵢ minⱼ angular(qᵢ, dⱼ) on pre-packed vectors. */
+template <numeric_dtype value_type_>
+typename value_type_::maxsim_result_t maxsim(packed_maxsim<value_type_> const &queries,
+                                             packed_maxsim<value_type_> const &documents) noexcept {
+    using result_t = typename value_type_::maxsim_result_t;
+    result_t result {};
+    if (queries.empty() || documents.empty()) return result;
+    if (queries.depth() != documents.depth()) return result;
+    maxsim_packed<value_type_>(queries.data(), documents.data(), queries.vector_count(), documents.vector_count(),
+                               queries.depth(), &result);
+    return result;
+}
+
+} // namespace ashvardanian::numkong
+
+#endif // NK_MAXSIM_HPP

package/numkong/include/numkong/mesh/README.md

@@ -0,0 +1,219 @@
+# Point Cloud Alignment in NumKong
+
+NumKong implements three algorithms for 3D point cloud comparison and alignment, used in structural biology (protein alignment), robotics (point cloud registration), and computer graphics (mesh registration).
+
+RMSD measures raw point-pair deviation without centering or alignment:
+
+$$
+\text{RMSD} = \sqrt{\frac{1}{n}\sum \|a_i - b_i\|^2}
+$$
+
+Kabsch finds the optimal rotation $R$ that minimizes RMSD after centering both clouds at their centroids $\bar{a}$, $\bar{b}$, recovering $R$ from the SVD of the cross-covariance matrix $H$:
+
+$$
+H = \sum (a_i - \bar{a})(b_i - \bar{b})^T = U \Sigma V^T, \quad R = V U^T
+$$
+
+Umeyama extends Kabsch with a uniform scale factor $s$ derived from the singular values and source variance $\sigma_a^2$:
+
+$$
+s = \frac{\text{tr}(\Sigma)}{n \cdot \sigma_a^2}, \quad \text{RMSD} = \sqrt{\frac{1}{n}\sum \|s \cdot R(a_i - \bar{a}) - (b_i - \bar{b})\|^2}
+$$
+
+Reformulating as Python pseudocode:
+
+```python
+import numpy as np
+
+def kabsch(a: np.ndarray, b: np.ndarray) -> np.ndarray:
+    a_c, b_c = a - a.mean(0), b - b.mean(0)
+    H = a_c.T @ b_c
+    U, S, Vt = np.linalg.svd(H)
+    d = np.sign(np.linalg.det(Vt.T @ U.T))
+    R = Vt.T @ np.diag([1, 1, d]) @ U.T
+    return R
+
+def umeyama(a: np.ndarray, b: np.ndarray) -> tuple:
+    a_c, b_c = a - a.mean(0), b - b.mean(0)
+    H = a_c.T @ b_c
+    U, S, Vt = np.linalg.svd(H)
+    d = np.sign(np.linalg.det(Vt.T @ U.T))
+    R = Vt.T @ np.diag([1, 1, d]) @ U.T
+    scale = S.sum() / (len(a) * np.var(a_c))
+    return R, scale
+
+def rmsd(a: np.ndarray, b: np.ndarray) -> float:
+    return np.sqrt(np.mean(np.sum((a - b) ** 2, axis=1)))
+```
+
+## Input & Output Types
+
+| Input Type | Output Type | Description                                    |
+| ---------- | ----------- | ---------------------------------------------- |
+| `f64`      | `f64`       | 64-bit IEEE 754 double precision               |
+| `f32`      | `f32`       | 32-bit IEEE 754 single precision               |
+| `f16`      | `f32`       | 16-bit IEEE 754 half precision, widened output |
+| `bf16`     | `f32`       | 16-bit brain float, widened output             |
+
+## Optimizations
+
+### McAdams Branching-Free 3×3 SVD
+
+`nk_kabsch_f32_serial`, `nk_kabsch_f64_haswell`, `nk_umeyama_f32_neon` use a Jacobi eigenanalysis with fixed 16 iterations (no convergence check) for deterministic behavior.
+Quaternion-accumulated rotations: each Jacobi sweep updates a 4-element quaternion instead of recomputing eigenvectors.
+Approximate Givens angles via `nk_approximate_givens_quaternion_` — a γ-threshold test selects between computed angles and precomputed cos(π/8), sin(π/8) constants.
+Cyclic permutation of matrix elements avoids explicit sorting of eigenvalues.
+
+### Stride-3 Deinterleaving
+
+Point clouds are stored interleaved as [x₀,y₀,z₀, x₁,y₁,z₁, ...].
+NEON uses `vld3q_f32` to hardware-deinterleave 4 XYZ triplets in one instruction — no gather needed.
+Haswell uses `_mm256_i32gather_ps` with indices [0,3,6,9,12,15,18,21] to load 8 x-coordinates from 8 points.
+RVV uses indexed loads with dynamic stride to adapt to variable vector length.
+
+### Reflection Correction
+
+`nk_kabsch_f32_haswell`, `nk_kabsch_f64_skylake` check for improper rotations after computing $R = V U^T$ from the SVD of the cross-covariance matrix $H = U \Sigma V^T$.
+If $\det(R) = -1$ (a reflection rather than a rotation), the last column of $V$ is negated before recomputing $R$.
+This ensures the output is always a proper rotation matrix with $\det(R) = +1$.
+
+### Pre-Scaled Rotation for Umeyama
+
+`nk_umeyama_f32_haswell`, `nk_umeyama_f64_skylake` fold the computed scale factor $s$ into the rotation matrix before applying to points.
+The Umeyama transform is $b_i = s R a_i + t$; by precomputing $R' = s R$ once, the per-point operation reduces to $b_i = R' a_i + t$, avoiding a per-point scalar multiply.
+
+### Why SME and SVE Were Removed
+
+Historical note: experimental SME variants of RMSD, Kabsch, and Umeyama were implemented in 1,052 lines across `sme.h` and `smef64.h` (commit `0e0bc30c`) and removed 4 days later (commit `f55e9a71`).
+The fundamental mismatch: the algorithm computes a 3×3 cross-covariance matrix $H = \sum (a_i - \bar{a})(b_i - \bar{b})^T$ — a sum of outer products of 3D vectors.
+SME's `FMOPA` operates on SVL-wide vectors (16+ elements at SVL=512), but the outer products here are 3×3 — the tile is 99.6% wasted (9 useful cells out of 256).
+Three approaches were explored in a design document (`sme_design.h`, 398 lines):
+(1) batched outer products — reformulates as 9 independent dot products but loses SME's outer-product strength, falling back to what NEON already does;
+(2) streaming SVE with `svld3` — hardware stride-3 deinterleaving processes 16 points per iteration vs NEON's 4, but `SMSTART`/`SMSTOP` mode transitions cost ~100 cycles and the 3×3 SVD step cannot use streaming mode at all;
+(3) SME for SVD — the 3×3 matrix cannot fill even one 16×16 tile.
+Performance estimates from the design document: NEON baseline ~2.25N cycles for N points; streaming SVE ~1.2N cycles but with ~100-cycle mode transition overhead — for typical protein alignment workloads (N = 100–500 atoms), the overhead dominates.
+Experimental SVE mesh kernels (`sve.h`, `svehalf.h`, 112 lines total) were removed in the same commit — variable vector length added complexity without clear benefit over fixed-width NEON for the 3D point cloud problem.
+
+## Performance
+
+The following performance tables are produced by manually re-running `nk_test` and `nk_bench` included internal tools to measure both accuracy and throughput at different input shapes.
+The input size is controlled by the `NK_MESH_POINTS` environment variable and set to 256, 1024, and 4096 points.
+Each alignment computes centroids, covariance, and a 3×3 SVD over $N$ point pairs, so cost is $O(N)$ per alignment with a large constant.
+The throughput is measured in mp/s as millions of 3D points aligned per second.
+Accuracy is reported as mean ULP (units in last place) unless noted otherwise — the average number of representable floating-point values between the result and the exact answer.
+Each kernel runs for at least 20 seconds per configuration.
+Benchmark threads are pinned to specific cores; on machines with heterogeneous core types (e.g., Apple P/E cores), only the fastest cores are used.
+Workloads that significantly degrade CPU frequencies (Intel AMX, Apple SME) run in separate passes to avoid affecting throughput measurements of other kernels.
+
+### Intel Sapphire Rapids
+
+#### Native
+
+| Kernel                    |                      256 |                     1024 |                     4096 |
+| :------------------------ | -----------------------: | -----------------------: | -----------------------: |
+| __f64__                   | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_f64_serial`      |        354 mp/s, 1.4 ulp |        176 mp/s, 2.7 ulp |        159 mp/s, 5.0 ulp |
+| `nk_kabsch_f64_serial`    |       71.1 mp/s, 1.4 ulp |       70.8 mp/s, 2.7 ulp |       80.3 mp/s, 5.2 ulp |
+| `nk_umeyama_f64_serial`   |       70.1 mp/s, 1.0 ulp |       75.1 mp/s, 1.8 ulp |       79.1 mp/s, 3.9 ulp |
+| `nk_rmsd_f64_haswell`     |        405 mp/s, 0.3 ulp |        260 mp/s, 0.4 ulp |        192 mp/s, 0.8 ulp |
+| `nk_kabsch_f64_haswell`   |       82.1 mp/s, 0.9 ulp |        105 mp/s, 1.3 ulp |        133 mp/s, 2.3 ulp |
+| `nk_umeyama_f64_haswell`  |       82.6 mp/s, 0.4 ulp |        119 mp/s, 0.8 ulp |        134 mp/s, 1.5 ulp |
+| `nk_rmsd_f64_skylake`     |        540 mp/s, 0.3 ulp |        219 mp/s, 0.3 ulp |        213 mp/s, 0.5 ulp |
+| `nk_kabsch_f64_skylake`   |       96.8 mp/s, 0.7 ulp |        115 mp/s, 0.9 ulp |        159 mp/s, 1.1 ulp |
+| `nk_umeyama_f64_skylake`  |        101 mp/s, 0.2 ulp |        119 mp/s, 0.4 ulp |        157 mp/s, 0.8 ulp |
+| __f32__                   | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_f32_serial`      |        480 mp/s, 1.4 ulp |        314 mp/s, 2.7 ulp |        270 mp/s, 5.4 ulp |
+| `nk_kabsch_f32_serial`    |       83.2 mp/s, 1.5 ulp |       91.6 mp/s, 2.6 ulp |        110 mp/s, 5.3 ulp |
+| `nk_umeyama_f32_serial`   |       80.4 mp/s, 1.0 ulp |        104 mp/s, 1.9 ulp |        106 mp/s, 3.7 ulp |
+| `nk_rmsd_f32_haswell`     |        447 mp/s, 0.3 ulp |        484 mp/s, 0.3 ulp |        350 mp/s, 0.4 ulp |
+| `nk_kabsch_f32_haswell`   |        101 mp/s, 0.7 ulp |        192 mp/s, 0.9 ulp |        213 mp/s, 1.3 ulp |
+| `nk_umeyama_f32_haswell`  |       97.4 mp/s, 0.3 ulp |        155 mp/s, 0.4 ulp |        207 mp/s, 0.8 ulp |
+| `nk_rmsd_f32_skylake`     |      1,000 mp/s, 0.7 ulp |        974 mp/s, 1.2 ulp |        786 mp/s, 2.4 ulp |
+| `nk_kabsch_f32_skylake`   |       97.5 mp/s, 0.7 ulp |        232 mp/s, 0.7 ulp |        332 mp/s, 0.9 ulp |
+| `nk_umeyama_f32_skylake`  |       92.5 mp/s, 0.2 ulp |        227 mp/s, 0.2 ulp |        325 mp/s, 0.3 ulp |
+| __bf16__                  | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_bf16_haswell`    |        511 mp/s, 0.3 ulp |        481 mp/s, 3.5 ulp |       497 mp/s, 12.8 ulp |
+| `nk_kabsch_bf16_haswell`  |       52.4 mp/s, 0.7 ulp |       65.3 mp/s, 0.9 ulp |       74.8 mp/s, 1.3 ulp |
+| `nk_umeyama_bf16_haswell` |       51.5 mp/s, 0.2 ulp |       69.2 mp/s, 0.4 ulp |       74.6 mp/s, 0.8 ulp |
+| `nk_rmsd_bf16_skylake`    |      1,765 mp/s, 0.3 ulp |      1,945 mp/s, 0.5 ulp |      2,056 mp/s, 6.0 ulp |
+| `nk_kabsch_bf16_skylake`  |        132 mp/s, 0.7 ulp |        370 mp/s, 0.8 ulp |        689 mp/s, 0.9 ulp |
+| `nk_umeyama_bf16_skylake` |        130 mp/s, 0.2 ulp |        366 mp/s, 0.3 ulp |        689 mp/s, 0.5 ulp |
+| __f16__                   | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_f16_haswell`     |        415 mp/s, 0.3 ulp |        497 mp/s, 0.7 ulp |        458 mp/s, 2.5 ulp |
+| `nk_kabsch_f16_haswell`   |        151 mp/s, 0.7 ulp |        222 mp/s, 0.9 ulp |        221 mp/s, 1.4 ulp |
+| `nk_umeyama_f16_haswell`  |        186 mp/s, 0.2 ulp |        232 mp/s, 0.5 ulp |        222 mp/s, 0.9 ulp |
+| `nk_rmsd_f16_skylake`     |      1,813 mp/s, 0.3 ulp |      1,982 mp/s, 0.4 ulp |      2,049 mp/s, 1.8 ulp |
+| `nk_kabsch_f16_skylake`   |        367 mp/s, 0.7 ulp |        695 mp/s, 0.7 ulp |        903 mp/s, 0.9 ulp |
+| `nk_umeyama_f16_skylake`  |        341 mp/s, 0.2 ulp |        686 mp/s, 0.2 ulp |        882 mp/s, 0.4 ulp |
+
+#### WASM
+
+Measured with Wasmtime v42 (Cranelift backend).
+
+| Kernel                       |                      256 |                     1024 |                     4096 |
+| :--------------------------- | -----------------------: | -----------------------: | -----------------------: |
+| __f64__                      | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_f64_serial`         |        178 mp/s, 1.4 ulp |        158 mp/s, 2.6 ulp |          ? mp/s, 5.3 ulp |
+| `nk_rmsd_f64_v128relaxed`    |        273 mp/s, 0.4 ulp |        307 mp/s, 0.7 ulp |          ? mp/s, 1.3 ulp |
+| `nk_kabsch_f64_serial`       |       37.7 mp/s, 1.4 ulp |       51.7 mp/s, 2.5 ulp |          ? mp/s, 5.2 ulp |
+| `nk_kabsch_f64_v128relaxed`  |       31.7 mp/s, 1.2 ulp |       56.9 mp/s, 2.3 ulp |          ? mp/s, 4.5 ulp |
+| `nk_umeyama_f64_serial`      |       36.5 mp/s, 0.9 ulp |       49.6 mp/s, 1.9 ulp |          ? mp/s, 3.6 ulp |
+| `nk_umeyama_f64_v128relaxed` |       32.6 mp/s, 0.8 ulp |       55.5 mp/s, 1.5 ulp |          ? mp/s, 3.2 ulp |
+| __f32__                      | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_f32_serial`         |        105 mp/s, 1.4 ulp |        122 mp/s, 2.7 ulp |          ? mp/s, 5.2 ulp |
+| `nk_rmsd_f32_v128relaxed`    |        213 mp/s, 0.3 ulp |        258 mp/s, 0.4 ulp |          ? mp/s, 0.8 ulp |
+| `nk_kabsch_f32_serial`       |       15.5 mp/s, 1.4 ulp |       32.8 mp/s, 2.6 ulp |          ? mp/s, 5.1 ulp |
+| `nk_kabsch_f32_v128relaxed`  |       13.5 mp/s, 0.9 ulp |       46.2 mp/s, 1.3 ulp |          ? mp/s, 2.5 ulp |
+| `nk_umeyama_f32_serial`      |       15.2 mp/s, 1.0 ulp |       37.4 mp/s, 1.8 ulp |          ? mp/s, 3.7 ulp |
+| `nk_umeyama_f32_v128relaxed` |       18.3 mp/s, 0.4 ulp |       38.9 mp/s, 0.8 ulp |          ? mp/s, 1.5 ulp |
+
+
+### Apple M5
+
+#### Native
+
+| Kernel                      |                      256 |                     1024 |                     4096 |
+| :-------------------------- | -----------------------: | -----------------------: | -----------------------: |
+| __f64__                     | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_f64_serial`        |        279 mp/s, 0.5 ulp |        267 mp/s, 0.5 ulp |        279 mp/s, 0.5 ulp |
+| `nk_kabsch_f64_serial`      |       40.4 mp/s, 1.4 ulp |       47.3 mp/s, 2.6 ulp |       50.2 mp/s, 5.4 ulp |
+| `nk_umeyama_f64_serial`     |       34.5 mp/s, 1.0 ulp |       39.2 mp/s, 1.9 ulp |       41.6 mp/s, 3.7 ulp |
+| `nk_rmsd_f64_neon`          |      1,776 mp/s, 0.4 ulp |      1,536 mp/s, 0.7 ulp |      2,037 mp/s, 1.3 ulp |
+| `nk_kabsch_f64_neon`        |        119 mp/s, 0.8 ulp |        222 mp/s, 1.3 ulp |        304 mp/s, 2.2 ulp |
+| `nk_umeyama_f64_neon`       |        115 mp/s, 0.4 ulp |        220 mp/s, 0.8 ulp |        296 mp/s, 1.6 ulp |
+| __f32__                     | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_f32_serial`        |        264 mp/s, 0.5 ulp |        264 mp/s, 0.5 ulp |        261 mp/s, 0.5 ulp |
+| `nk_kabsch_f32_serial`      |       39.4 mp/s, 1.4 ulp |       46.0 mp/s, 2.7 ulp |       49.9 mp/s, 5.0 ulp |
+| `nk_umeyama_f32_serial`     |       33.6 mp/s, 0.9 ulp |       38.8 mp/s, 1.8 ulp |       41.4 mp/s, 3.5 ulp |
+| `nk_rmsd_f32_neon`          |      1,912 mp/s, 1.5 ulp |      2,239 mp/s, 1.3 ulp |      1,966 mp/s, 4.8 ulp |
+| `nk_kabsch_f32_neon`        |        135 mp/s, 0.7 ulp |        288 mp/s, 0.9 ulp |        385 mp/s, 1.4 ulp |
+| `nk_umeyama_f32_neon`       |        130 mp/s, 0.3 ulp |        272 mp/s, 0.4 ulp |        367 mp/s, 0.8 ulp |
+| __bf16__                    | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_bf16_neonbfdot`    |      3,728 mp/s, 0.4 ulp |      3,756 mp/s, 6.0 ulp |     3,769 mp/s, 10.0 ulp |
+| `nk_kabsch_bf16_neonbfdot`  |        180 mp/s, 0.7 ulp |        448 mp/s, 0.9 ulp |        726 mp/s, 1.3 ulp |
+| `nk_umeyama_bf16_neonbfdot` |        176 mp/s, 0.2 ulp |        433 mp/s, 0.4 ulp |        705 mp/s, 0.8 ulp |
+| __f16__                     | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_f16_neonhalf`      |      2,998 mp/s, 0.4 ulp |      3,215 mp/s, 1.7 ulp |      3,216 mp/s, 4.6 ulp |
+| `nk_kabsch_f16_neonhalf`    |        178 mp/s, 0.9 ulp |        443 mp/s, 1.3 ulp |        711 mp/s, 2.4 ulp |
+| `nk_umeyama_f16_neonhalf`   |        175 mp/s, 0.4 ulp |        408 mp/s, 0.8 ulp |        620 mp/s, 1.5 ulp |
+
+#### WASM
+
+Measured with Wasmtime v43 (Cranelift backend).
+
+| Kernel                       |                      256 |                     1024 |                     4096 |
+| :--------------------------- | -----------------------: | -----------------------: | -----------------------: |
+| __f64__                      | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_f64_serial`         |        137 mp/s, 2.6 ulp |        134 mp/s, 2.6 ulp |        142 mp/s, 2.6 ulp |
+| `nk_rmsd_f64_v128relaxed`    |      1,377 mp/s, 0.8 ulp |      1,038 mp/s, 0.8 ulp |      1,566 mp/s, 0.8 ulp |
+| `nk_kabsch_f64_serial`       |       42.3 mp/s, 2.7 ulp |       50.4 mp/s, 2.7 ulp |       55.5 mp/s, 2.7 ulp |
+| `nk_kabsch_f64_v128relaxed`  |        121 mp/s, 2.2 ulp |        225 mp/s, 2.2 ulp |        345 mp/s, 2.2 ulp |
+| `nk_umeyama_f64_serial`      |       36.1 mp/s, 1.8 ulp |       41.3 mp/s, 1.8 ulp |       46.0 mp/s, 1.8 ulp |
+| `nk_umeyama_f64_v128relaxed` |        112 mp/s, 1.5 ulp |        207 mp/s, 1.5 ulp |        293 mp/s, 1.5 ulp |
+| __f32__                      | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_f32_serial`         |        120 mp/s, 2.7 ulp |        120 mp/s, 2.7 ulp |        124 mp/s, 2.7 ulp |
+| `nk_rmsd_f32_v128relaxed`    |      1,025 mp/s, 0.5 ulp |      1,038 mp/s, 0.5 ulp |      1,093 mp/s, 0.5 ulp |
+| `nk_kabsch_f32_serial`       |       39.6 mp/s, 2.6 ulp |       47.6 mp/s, 2.6 ulp |       51.4 mp/s, 2.6 ulp |
+| `nk_kabsch_f32_v128relaxed`  |        125 mp/s, 1.3 ulp |        255 mp/s, 1.3 ulp |        366 mp/s, 1.3 ulp |
+| `nk_umeyama_f32_serial`      |       30.5 mp/s, 1.8 ulp |       35.0 mp/s, 1.8 ulp |       38.9 mp/s, 1.8 ulp |
+| `nk_umeyama_f32_v128relaxed` |        118 mp/s, 0.8 ulp |        240 mp/s, 0.8 ulp |        338 mp/s, 0.8 ulp |