numkong 7.0.0 → 7.4.1

package/include/numkong/maxsim/v128relaxed.h

@@ -46,7 +46,7 @@ NK_PUBLIC nk_size_t nk_maxsim_packed_size_f16_v128relaxed(nk_size_t vector_count
 }
 
 NK_PUBLIC void nk_maxsim_pack_bf16_v128relaxed( //
-    nk_bf16_t const *vectors, nk_size_t vector_count, nk_size_t depth, nk_size_t stride, void *packed) {
+    nk_bf16_t const *vectors, nk_size_t vector_count, nk_size_t depth, nk_size_t stride_in_bytes, void *packed) {
 
     nk_size_t const element_bytes = sizeof(nk_bf16_t);
     nk_size_t depth_i8_padded = nk_maxsim_packed_header_setup_(packed, vector_count, depth, 16, element_bytes);
@@ -58,7 +58,7 @@ NK_PUBLIC void nk_maxsim_pack_bf16_v128relaxed( //
     nk_size_t const original_stride = header->original_stride_bytes;
 
     for (nk_size_t vector_index = 0; vector_index < vector_count; vector_index++) {
-        char const *source_row = (char const *)vectors + vector_index * stride;
+        char const *source_row = (char const *)vectors + vector_index * stride_in_bytes;
         nk_f32_t norm_sq;
         nk_maxsim_quantize_vector_(source_row, element_bytes, depth, depth_i8_padded, 63.0f,
                                    (nk_maxsim_to_f32_t)nk_bf16_to_f32_serial,
@@ -72,7 +72,7 @@ NK_PUBLIC void nk_maxsim_pack_bf16_v128relaxed( //
 }
 
 NK_PUBLIC void nk_maxsim_pack_f32_v128relaxed( //
-    nk_f32_t const *vectors, nk_size_t vector_count, nk_size_t depth, nk_size_t stride, void *packed) {
+    nk_f32_t const *vectors, nk_size_t vector_count, nk_size_t depth, nk_size_t stride_in_bytes, void *packed) {
 
     nk_size_t const element_bytes = sizeof(nk_f32_t);
     nk_size_t depth_i8_padded = nk_maxsim_packed_header_setup_(packed, vector_count, depth, 16, element_bytes);
@@ -84,7 +84,7 @@ NK_PUBLIC void nk_maxsim_pack_f32_v128relaxed( //
     nk_size_t const original_stride = header->original_stride_bytes;
 
     for (nk_size_t vector_index = 0; vector_index < vector_count; vector_index++) {
-        char const *source_row = (char const *)vectors + vector_index * stride;
+        char const *source_row = (char const *)vectors + vector_index * stride_in_bytes;
         nk_f32_t norm_sq;
         nk_maxsim_quantize_vector_(source_row, element_bytes, depth, depth_i8_padded, 63.0f, nk_f32_to_f32_,
                                    &quantized_i8[vector_index * depth_i8_padded], &metadata[vector_index], &norm_sq);
@@ -97,7 +97,7 @@ NK_PUBLIC void nk_maxsim_pack_f32_v128relaxed( //
 }
 
 NK_PUBLIC void nk_maxsim_pack_f16_v128relaxed( //
-    nk_f16_t const *vectors, nk_size_t vector_count, nk_size_t depth, nk_size_t stride, void *packed) {
+    nk_f16_t const *vectors, nk_size_t vector_count, nk_size_t depth, nk_size_t stride_in_bytes, void *packed) {
 
     nk_size_t const element_bytes = sizeof(nk_f16_t);
     nk_size_t depth_i8_padded = nk_maxsim_packed_header_setup_(packed, vector_count, depth, 16, element_bytes);
@@ -109,7 +109,7 @@ NK_PUBLIC void nk_maxsim_pack_f16_v128relaxed( //
     nk_size_t const original_stride = header->original_stride_bytes;
 
     for (nk_size_t vector_index = 0; vector_index < vector_count; vector_index++) {
-        char const *source_row = (char const *)vectors + vector_index * stride;
+        char const *source_row = (char const *)vectors + vector_index * stride_in_bytes;
         nk_f32_t norm_sq;
         nk_maxsim_quantize_vector_(source_row, element_bytes, depth, depth_i8_padded, 63.0f,
                                    (nk_maxsim_to_f32_t)nk_f16_to_f32_serial,

package/include/numkong/mesh/README.md

@@ -6,33 +6,33 @@ Used in structural biology (protein alignment), robotics (point cloud registrati
 
 Centroid:
 
-```math
+$$
 \bar{a} = \frac{1}{n}\sum a_i
-```
+$$
 
 Cross-covariance matrix:
 
-```math
+$$
 H = \sum (a_i - \bar{a})(b_i - \bar{b})^T
-```
+$$
 
 SVD-based rotation:
 
-```math
+$$
 H = U \Sigma V^T, \quad R = V U^T
-```
+$$
 
 Umeyama scale factor:
 
-```math
+$$
 s = \frac{\text{tr}(\Sigma)}{n \cdot \sigma_a^2}
-```
+$$
 
 RMSD after alignment:
 
-```math
+$$
 \text{RMSD} = \sqrt{\frac{1}{n}\sum \|s \cdot R(a_i - \bar{a}) - (b_i - \bar{b})\|^2}
-```
+$$
 
 Reformulating as Python pseudocode:
 
@@ -87,14 +87,14 @@ 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 (det(R) = -1, reflections) after computing R = V·Uᵀ.
-If det(R) is negative, the last column of V is flipped.
-This ensures the output is always a proper rotation matrix (det = +1).
+`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 into the rotation matrix before applying to points.
-`sr[i] = scale * r[i]` is computed once and broadcast — avoiding a per-point scalar multiply.
+`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
 
@@ -142,17 +142,23 @@ Workloads that significantly degrade CPU frequencies (Intel AMX, Apple SME) run
 | `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`     |        936 mp/s, 0.3 ulp |        970 mp/s, 0.3 ulp |        426 mp/s, 0.3 ulp |
-| `nk_kabsch_f32_skylake`   |        122 mp/s, 0.7 ulp |        258 mp/s, 0.7 ulp |        290 mp/s, 0.9 ulp |
-| `nk_umeyama_f32_skylake`  |        133 mp/s, 0.2 ulp |        231 mp/s, 0.3 ulp |        285 mp/s, 0.5 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
 
@@ -176,52 +182,52 @@ Measured with Wasmtime v42 (Cranelift backend).
 | `nk_umeyama_f32_v128relaxed` |       18.3 mp/s, 0.4 ulp |       38.9 mp/s, 0.8 ulp |          ? mp/s, 1.5 ulp |
 
 
-### Apple M4
+### Apple M5
 
 #### Native
 
-| Kernel                   |                      256 |                     1024 |                     4096 |
-| :----------------------- | -----------------------: | -----------------------: | -----------------------: |
-| __f64__                  | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
-| `nk_rmsd_f64_serial`     |        447 mp/s, 1.4 ulp |        410 mp/s, 2.6 ulp |        412 mp/s, 5.3 ulp |
-| `nk_kabsch_f64_serial`   |       95.2 mp/s, 1.4 ulp |        169 mp/s, 2.6 ulp |        214 mp/s, 5.4 ulp |
-| `nk_umeyama_f64_serial`  |       89.2 mp/s, 1.0 ulp |        157 mp/s, 1.9 ulp |        195 mp/s, 3.7 ulp |
-| `nk_rmsd_f64_neon`       |        823 mp/s, 0.4 ulp |        761 mp/s, 0.7 ulp |        702 mp/s, 1.3 ulp |
-| `nk_kabsch_f64_neon`     |        105 mp/s, 0.8 ulp |        213 mp/s, 1.3 ulp |        287 mp/s, 2.2 ulp |
-| `nk_umeyama_f64_neon`    |        106 mp/s, 0.4 ulp |        214 mp/s, 0.8 ulp |        297 mp/s, 1.6 ulp |
-| __f32__                  | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
-| `nk_rmsd_f32_serial`     |        554 mp/s, 1.4 ulp |        566 mp/s, 2.6 ulp |        532 mp/s, 5.2 ulp |
-| `nk_kabsch_f32_serial`   |        110 mp/s, 1.4 ulp |        214 mp/s, 2.7 ulp |        264 mp/s, 5.0 ulp |
-| `nk_umeyama_f32_serial`  |        104 mp/s, 0.9 ulp |        197 mp/s, 1.8 ulp |        240 mp/s, 3.5 ulp |
-| `nk_rmsd_f32_neon`       |      1,580 mp/s, 0.3 ulp |      1,560 mp/s, 0.4 ulp |      1,200 mp/s, 0.8 ulp |
-| `nk_kabsch_f32_neon`     |        139 mp/s, 0.7 ulp |        336 mp/s, 0.9 ulp |        485 mp/s, 1.4 ulp |
-| `nk_umeyama_f32_neon`    |        137 mp/s, 0.3 ulp |        325 mp/s, 0.4 ulp |        470 mp/s, 0.8 ulp |
-| __bf16__                 | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
-| `nk_rmsd_bf16_serial`    |      1,740 mp/s, 0.5 ulp |      1,880 mp/s, 6.0 ulp |     1,860 mp/s, 10.0 ulp |
-| `nk_kabsch_bf16_serial`  |        137 mp/s, 0.7 ulp |        335 mp/s, 0.9 ulp |        527 mp/s, 1.3 ulp |
-| `nk_umeyama_bf16_serial` |        135 mp/s, 0.2 ulp |        329 mp/s, 0.4 ulp |        510 mp/s, 0.8 ulp |
-| __f16__                  | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
-| `nk_rmsd_f16_serial`     |      1,840 mp/s, 0.4 ulp |      1,900 mp/s, 1.7 ulp |      1,860 mp/s, 4.6 ulp |
-| `nk_kabsch_f16_serial`   |        140 mp/s, 0.9 ulp |        349 mp/s, 1.3 ulp |        547 mp/s, 2.4 ulp |
-| `nk_umeyama_f16_serial`  |        135 mp/s, 0.4 ulp |        316 mp/s, 0.8 ulp |        474 mp/s, 1.5 ulp |
+| Kernel                      |                      256 |                     1024 |                     4096 |
+| :-------------------------- | -----------------------: | -----------------------: | -----------------------: |
+| __f64__                     | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_rmsd_f64_serial`        |        120 mp/s, 1.4 ulp |        118 mp/s, 2.6 ulp |        121 mp/s, 5.3 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`          |      1,418 mp/s, 0.4 ulp |      1,338 mp/s, 0.7 ulp |      1,419 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`        |        122 mp/s, 1.4 ulp |        123 mp/s, 2.6 ulp |        125 mp/s, 5.2 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`          |      1,337 mp/s, 0.3 ulp |      1,377 mp/s, 0.4 ulp |      1,261 mp/s, 0.8 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`    |      2,342 mp/s, 0.5 ulp |      2,378 mp/s, 6.0 ulp |     2,416 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`      |      2,315 mp/s, 0.4 ulp |      2,372 mp/s, 1.7 ulp |      2,423 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 v42 (Cranelift backend).
+Measured with Wasmtime v43 (Cranelift backend).
 
 | Kernel                       |                      256 |                     1024 |                     4096 |
 | :--------------------------- | -----------------------: | -----------------------: | -----------------------: |
 | __f64__                      | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
-| `nk_rmsd_f64_serial`         |        219 mp/s, 2.6 ulp |        202 mp/s, 2.6 ulp |        255 mp/s, 2.6 ulp |
-| `nk_rmsd_f64_v128relaxed`    |        434 mp/s, 0.8 ulp |        363 mp/s, 0.8 ulp |        586 mp/s, 0.8 ulp |
-| `nk_kabsch_f64_serial`       |       42.8 mp/s, 2.7 ulp |       76.0 mp/s, 2.7 ulp |        110 mp/s, 2.7 ulp |
-| `nk_kabsch_f64_v128relaxed`  |       55.2 mp/s, 2.2 ulp |        110 mp/s, 2.2 ulp |        202 mp/s, 2.2 ulp |
-| `nk_umeyama_f64_serial`      |       36.1 mp/s, 1.8 ulp |       58.9 mp/s, 1.8 ulp |       98.7 mp/s, 1.8 ulp |
-| `nk_umeyama_f64_v128relaxed` |       52.4 mp/s, 1.5 ulp |        103 mp/s, 1.5 ulp |        183 mp/s, 1.5 ulp |
+| `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`         |        218 mp/s, 2.7 ulp |        223 mp/s, 2.7 ulp |        271 mp/s, 2.7 ulp |
-| `nk_rmsd_f32_v128relaxed`    |        626 mp/s, 0.5 ulp |        626 mp/s, 0.5 ulp |        687 mp/s, 0.5 ulp |
-| `nk_kabsch_f32_serial`       |       45.5 mp/s, 2.6 ulp |       77.0 mp/s, 2.6 ulp |        112 mp/s, 2.6 ulp |
-| `nk_kabsch_f32_v128relaxed`  |       68.6 mp/s, 1.3 ulp |        160 mp/s, 1.3 ulp |        273 mp/s, 1.3 ulp |
-| `nk_umeyama_f32_serial`      |       38.7 mp/s, 1.8 ulp |       60.0 mp/s, 1.8 ulp |       80.5 mp/s, 1.8 ulp |
-| `nk_umeyama_f32_v128relaxed` |       66.9 mp/s, 0.8 ulp |        157 mp/s, 0.8 ulp |        291 mp/s, 0.8 ulp |
+| `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 |