numkong 7.5.0 → 7.6.0

package/include/numkong/dots/serial.h

@@ -522,7 +522,7 @@ NK_INTERNAL nk_i32_t nk_dots_reduce_sum_i4_(nk_i4x2_t const *data, nk_size_t cou
                                 load_a_vec_fn, partial_load_a_vec_fn, load_b_vec_fn, partial_load_b_vec_fn,            \
                                 inner_product_fn, reduce_accumulators_fn, store_fn, partial_store_fn,                  \
                                 depth_simd_dimensions, dimensions_per_value)                                           \
-    NK_PUBLIC void nk_##api_name##_packed_##input_type_name##_##isa_suffix##_aligned_(                                 \
+    NK_INTERNAL void nk_##api_name##_packed_##input_type_name##_##isa_suffix##_aligned_(                               \
         nk_##input_value_type##_t const *a_matrix, void const *b_packed_buffer, nk_##result_value_type##_t *c_matrix,  \
         nk_size_t row_count, nk_size_t column_count, nk_size_t depth, nk_size_t a_stride_in_bytes,                     \
         nk_size_t c_stride_in_bytes) {                                                                                 \
@@ -698,7 +698,7 @@ NK_INTERNAL nk_i32_t nk_dots_reduce_sum_i4_(nk_i4x2_t const *data, nk_size_t cou
             }                                                                                                          \
         }                                                                                                              \
     }                                                                                                                  \
-    NK_PUBLIC void nk_##api_name##_packed_##input_type_name##_##isa_suffix##_1x8_aligned_(                             \
+    NK_INTERNAL void nk_##api_name##_packed_##input_type_name##_##isa_suffix##_1x8_aligned_(                           \
         nk_##input_value_type##_t const *a_matrix, void const *b_packed_buffer, nk_##result_value_type##_t *c_matrix,  \
         nk_size_t row_count, nk_size_t column_count, nk_size_t depth, nk_size_t a_stride_in_bytes,                     \
         nk_size_t c_stride_in_bytes) {                                                                                 \
@@ -1090,7 +1090,7 @@ NK_INTERNAL nk_i32_t nk_dots_reduce_sum_i4_(nk_i4x2_t const *data, nk_size_t cou
     norm_value_type, vec_type, state_type, result_vec_type, init_accumulator_fn, load_a_vec_fn, partial_load_a_vec_fn, \
     load_b_vec_fn, partial_load_b_vec_fn, inner_product_fn, compensated_finalize_fn, store_fn, partial_store_fn,       \
     load_sum_fn, partial_load_sum_fn, compute_a_sum_fn, depth_simd_dimensions, dimensions_per_value)                   \
-    NK_PUBLIC void nk_##api_name##_packed_##input_type_name##_##isa_suffix##_aligned_(                                 \
+    NK_INTERNAL void nk_##api_name##_packed_##input_type_name##_##isa_suffix##_aligned_(                               \
         nk_##input_value_type##_t const *a_matrix, void const *b_packed_buffer, nk_##result_value_type##_t *c_matrix,  \
         nk_size_t row_count, nk_size_t column_count, nk_size_t depth, nk_size_t a_stride_in_bytes,                     \
         nk_size_t c_stride_in_bytes) {                                                                                 \
@@ -1200,7 +1200,7 @@ NK_INTERNAL nk_i32_t nk_dots_reduce_sum_i4_(nk_i4x2_t const *data, nk_size_t cou
             }                                                                                                          \
         }                                                                                                              \
     }                                                                                                                  \
-    NK_PUBLIC void nk_##api_name##_packed_##input_type_name##_##isa_suffix##_1x8_aligned_(                             \
+    NK_INTERNAL void nk_##api_name##_packed_##input_type_name##_##isa_suffix##_1x8_aligned_(                           \
         nk_##input_value_type##_t const *a_matrix, void const *b_packed_buffer, nk_##result_value_type##_t *c_matrix,  \
         nk_size_t row_count, nk_size_t column_count, nk_size_t depth, nk_size_t a_stride_in_bytes,                     \
         nk_size_t c_stride_in_bytes) {                                                                                 \
@@ -2431,10 +2431,19 @@ NK_INTERNAL nk_i32_t nk_dots_reduce_sum_i4_(nk_i4x2_t const *data, nk_size_t cou
         }                                                                                                              \
     }
 
-/* Optimize serial GEMM instantiations for size rather than speed.
- * These fallback kernels are only used when no SIMD backend is available, so aggressive inlining/unrolling from -O3
- * wastes over 1 MB of binary space with negligible performance benefit on the serial path.
- */
+/*  Keep the serial instantiations below actually scalar, regardless of build type.
+ *  Without this, -O3 + LTO can vectorize or clone the serial kernels under AVX-512
+ *  callers in dispatch_*.c, which wastes ~1 MB of binary and — more importantly —
+ *  breaks the nk_*_serial-as-scalar-oracle contract that tests and the numerical-
+ *  stability docs in this header rely on. */
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((noinline)), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC optimize("no-tree-vectorize", "no-tree-slp-vectorize", "no-ipa-cp-clone", "no-inline")
+#endif
+
+/*  Size bias for release. Gated on NDEBUG so Debug builds keep -O0 for stepping. */
 #if defined(NDEBUG)
 #if defined(_MSC_VER)
 #pragma optimize("s", on)
@@ -2689,6 +2698,12 @@ nk_define_cross_packed_(dots, u1, serial, u1x8, u1x8, u32, nk_b128_vec_t, nk_dot
 #endif
 #endif
 
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
 /*  BF16 compact: truncate F32 → BF16 in-place.
  *  Reads F32 matrix with c_stride_in_bytes, writes BF16 tightly packed (stride_in_bytes = column_count × sizeof(bf16)).
  */

package/include/numkong/dots/skylake.h

@@ -114,45 +114,50 @@ nk_define_cross_packed_(dots, f16, skylake, f16, f32, f32, nk_b512_vec_t, nk_dot
                         nk_dot_through_f32_finalize_skylake_, nk_store_b128_haswell_, nk_partial_store_b32x4_skylake_,
                         /*depth_simd_dimensions=*/16, /*dimensions_per_value=*/1)
 
-/* E4M3 GEMM: depth_simd_dimensions=16 (16 e4m3s = 16 bytes = quarter cache line), F32 accumulator */
-nk_define_cross_pack_size_(dots, e4m3, skylake, e4m3, f32, /*norm_value_type=*/f32, /*depth_simd_dimensions=*/16,
+/* E4M3 GEMM: F16-pack with asymmetric A/B representations at compute time. Pack converts
+ * E4M3 → F16 once (~10 ops/16 elements, 2 bytes/elt stored). A-stream uses the Giesen E4M3→F32
+ * cast (identical cost to F32-pack path). B-loader widens F16 → F32 inline (1 vcvtph2ps per 16
+ * lanes). Update takes both as F32 → plain fmadd. Saves 2 bytes/elt vs F32-pack; inner loop
+ * adds one cvtph2ps per B-read. Symmetric uses E4M3→F32 for both sides (no pack involved). */
+nk_define_cross_pack_size_(dots, e4m3, skylake, e4m3, f16, /*norm_value_type=*/f32, /*depth_simd_dimensions=*/16,
                            /*dimensions_per_value=*/1)
-nk_define_cross_pack_(dots, e4m3, skylake, e4m3, f32, nk_b512_vec_t, nk_load_e4m3x16_to_f32x16_skylake_,
-                      nk_partial_load_e4m3x16_to_f32x16_skylake_, nk_store_b512_skylake_,
-                      nk_partial_store_b32x16_skylake_, /*simd_width=*/16, /*norm_value_type=*/f32,
-                      nk_dots_reduce_sumsq_e4m3_, /*depth_simd_dimensions=*/16, /*dimensions_per_value=*/1)
+nk_define_cross_pack_(dots, e4m3, skylake, e4m3, f16, nk_b256_vec_t, nk_load_e4m3x16_to_f16x16_skylake_,
+                      nk_partial_load_e4m3x16_to_f16x16_skylake_, nk_store_b256_haswell_,
+                      nk_partial_store_b16x16_serial_,
+                      /*simd_width=*/16, /*norm_value_type=*/f32, nk_dots_reduce_sumsq_e4m3_,
+                      /*depth_simd_dimensions=*/16, /*dimensions_per_value=*/1)
 nk_define_cross_symmetric_(dots, e4m3, skylake, e4m3, f32, nk_b512_vec_t, nk_dot_through_f32_state_skylake_t_,
                            nk_b128_vec_t, nk_dot_through_f32_init_skylake_, nk_load_e4m3x16_to_f32x16_skylake_,
                            nk_partial_load_e4m3x16_to_f32x16_skylake_, nk_dot_through_f32_update_skylake_,
                            nk_dot_through_f32_finalize_skylake_, nk_store_b128_haswell_,
                            nk_partial_store_b32x4_skylake_,
                            /*depth_simd_dimensions=*/16, /*dimensions_per_value=*/1)
-nk_define_cross_packed_(dots, e4m3, skylake, e4m3, f32, f32, nk_b512_vec_t, nk_dot_through_f32_state_skylake_t_,
+nk_define_cross_packed_(dots, e4m3, skylake, e4m3, f16, f32, nk_b512_vec_t, nk_dot_through_f32_state_skylake_t_,
                         nk_b128_vec_t, nk_dot_through_f32_init_skylake_, nk_load_e4m3x16_to_f32x16_skylake_,
-                        nk_partial_load_e4m3x16_to_f32x16_skylake_, nk_load_b512_skylake_,
-                        nk_partial_load_b32x16_skylake_, nk_dot_through_f32_update_skylake_,
+                        nk_partial_load_e4m3x16_to_f32x16_skylake_, nk_load_f16x16_to_f32x16_skylake_,
+                        nk_partial_load_f16x16_to_f32x16_skylake_, nk_dot_through_f32_update_skylake_,
                         nk_dot_through_f32_finalize_skylake_, nk_store_b128_haswell_, nk_partial_store_b32x4_skylake_,
                         /*depth_simd_dimensions=*/16, /*dimensions_per_value=*/1)
 
-/* E5M2 GEMM: depth_simd_dimensions=16 (16 e5m2s = 16 bytes = quarter cache line), F32 accumulator */
-nk_define_cross_pack_size_(dots, e5m2, skylake, e5m2, f32, /*norm_value_type=*/f32, /*depth_simd_dimensions=*/16,
+/* E5M2 GEMM: depth_simd_dimensions=64 (byte-level batch; widen inside the update helper) */
+nk_define_cross_pack_size_(dots, e5m2, skylake, e5m2, f32, /*norm_value_type=*/f32, /*depth_simd_dimensions=*/64,
                            /*dimensions_per_value=*/1)
-nk_define_cross_pack_(dots, e5m2, skylake, e5m2, f32, nk_b512_vec_t, nk_load_e5m2x16_to_f32x16_skylake_,
-                      nk_partial_load_e5m2x16_to_f32x16_skylake_, nk_store_b512_skylake_,
-                      nk_partial_store_b32x16_skylake_, /*simd_width=*/16, /*norm_value_type=*/f32,
-                      nk_dots_reduce_sumsq_e5m2_, /*depth_simd_dimensions=*/16, /*dimensions_per_value=*/1)
+nk_define_cross_pack_(dots, e5m2, skylake, e5m2, f32, nk_b512_vec_t, nk_load_b512_skylake_,
+                      nk_partial_load_b8x64_skylake_, nk_store_b512_skylake_, nk_partial_store_b8x64_skylake_,
+                      /*simd_width=*/64, /*norm_value_type=*/f32, nk_dots_reduce_sumsq_e5m2_,
+                      /*depth_simd_dimensions=*/64, /*dimensions_per_value=*/1)
 nk_define_cross_symmetric_(dots, e5m2, skylake, e5m2, f32, nk_b512_vec_t, nk_dot_through_f32_state_skylake_t_,
-                           nk_b128_vec_t, nk_dot_through_f32_init_skylake_, nk_load_e5m2x16_to_f32x16_skylake_,
-                           nk_partial_load_e5m2x16_to_f32x16_skylake_, nk_dot_through_f32_update_skylake_,
+                           nk_b128_vec_t, nk_dot_through_f32_init_skylake_, nk_load_b512_skylake_,
+                           nk_partial_load_b8x64_skylake_, nk_dot_e5m2x64_update_skylake_,
                            nk_dot_through_f32_finalize_skylake_, nk_store_b128_haswell_,
                            nk_partial_store_b32x4_skylake_,
-                           /*depth_simd_dimensions=*/16, /*dimensions_per_value=*/1)
+                           /*depth_simd_dimensions=*/64, /*dimensions_per_value=*/1)
 nk_define_cross_packed_(dots, e5m2, skylake, e5m2, f32, f32, nk_b512_vec_t, nk_dot_through_f32_state_skylake_t_,
-                        nk_b128_vec_t, nk_dot_through_f32_init_skylake_, nk_load_e5m2x16_to_f32x16_skylake_,
-                        nk_partial_load_e5m2x16_to_f32x16_skylake_, nk_load_b512_skylake_,
-                        nk_partial_load_b32x16_skylake_, nk_dot_through_f32_update_skylake_,
-                        nk_dot_through_f32_finalize_skylake_, nk_store_b128_haswell_, nk_partial_store_b32x4_skylake_,
-                        /*depth_simd_dimensions=*/16, /*dimensions_per_value=*/1)
+                        nk_b128_vec_t, nk_dot_through_f32_init_skylake_, nk_load_b512_skylake_,
+                        nk_partial_load_b8x64_skylake_, nk_load_b512_skylake_, nk_partial_load_b8x64_skylake_,
+                        nk_dot_e5m2x64_update_skylake_, nk_dot_through_f32_finalize_skylake_, nk_store_b128_haswell_,
+                        nk_partial_store_b32x4_skylake_,
+                        /*depth_simd_dimensions=*/64, /*dimensions_per_value=*/1)
 
 /* E2M3 GEMM: integer LUT path, depth_simd_dimensions=64 (64 e2m3s = 64 bytes = AVX-512 register width) */
 nk_define_cross_pack_size_(dots, e2m3, skylake, e2m3, e2m3, /*norm_value_type=*/f32, /*depth_simd_dimensions=*/64,

package/include/numkong/dots.h

@@ -698,6 +698,18 @@ NK_PUBLIC void nk_dots_packed_f16_graniteamx(nk_f16_t const *a, void const *b_pa
 NK_PUBLIC void nk_dots_symmetric_f16_graniteamx(nk_f16_t const *vectors, nk_size_t vectors_count, nk_size_t depth,
                                                 nk_size_t stride, nk_f32_t *result, nk_size_t result_stride,
                                                 nk_size_t row_start, nk_size_t row_count);
+/** @copydoc nk_dots_packed_size_f16 */
+NK_PUBLIC nk_size_t nk_dots_packed_size_e5m2_graniteamx(nk_size_t width, nk_size_t depth);
+/** @copydoc nk_dots_pack_f16 */
+NK_PUBLIC void nk_dots_pack_e5m2_graniteamx(nk_e5m2_t const *b, nk_size_t width, nk_size_t depth, nk_size_t b_stride,
+                                            void *b_packed);
+/** @copydoc nk_dots_packed_f16 */
+NK_PUBLIC void nk_dots_packed_e5m2_graniteamx(nk_e5m2_t const *a, void const *b_packed, nk_f32_t *c, nk_size_t height,
+                                              nk_size_t width, nk_size_t depth, nk_size_t a_stride, nk_size_t c_stride);
+/** @copydoc nk_dots_symmetric_f16 */
+NK_PUBLIC void nk_dots_symmetric_e5m2_graniteamx(nk_e5m2_t const *vectors, nk_size_t vectors_count, nk_size_t depth,
+                                                 nk_size_t stride, nk_f32_t *result, nk_size_t result_stride,
+                                                 nk_size_t row_start, nk_size_t row_count);
 #endif // NK_TARGET_GRANITEAMX
 
 /*  ARM SME backends using Scalable Matrix Extension.

package/include/numkong/each/serial.h

@@ -76,7 +76,18 @@ extern "C" {
         }                                                                                                             \
     }
 
-/* Optimize serial fallbacks for size — see dots/serial.h for rationale. */
+/*  Keep the serial instantiations below actually scalar, regardless of build type.
+ *  Without this, -O3 + LTO can vectorize or clone the serial kernels under AVX-512
+ *  callers in dispatch_*.c, which wastes binary and breaks the nk_*_serial-as-scalar-oracle
+ *  contract. See dots/serial.h. */
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((noinline)), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC optimize("no-tree-vectorize", "no-tree-slp-vectorize", "no-ipa-cp-clone", "no-inline")
+#endif
+
+/* Size bias for release. Gated on NDEBUG so Debug builds keep -O0 for stepping. */
 #if defined(NDEBUG)
 #if defined(_MSC_VER)
 #pragma optimize("s", on)
@@ -275,6 +286,12 @@ NK_PUBLIC void nk_each_fma_f64c_serial(nk_f64c_t const *a, nk_f64c_t const *b, n
 #endif
 #endif
 
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
 #if defined(__cplusplus)
 } // extern "C"
 #endif

package/include/numkong/geospatial/serial.h

@@ -17,9 +17,14 @@
 extern "C" {
 #endif
 
-/*  Serial implementations of geospatial distance functions.
- *  These use the trigonometric functions from trigonometry.h for sin, cos, and atan2.
- */
+/*  Keep the serial instantiations below actually scalar, regardless of build type.
+ *  See dots/serial.h for rationale. */
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((noinline)), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC optimize("no-tree-vectorize", "no-tree-slp-vectorize", "no-ipa-cp-clone", "no-inline")
+#endif
 
 NK_PUBLIC void nk_haversine_f64_serial(             //
     nk_f64_t const *a_lats, nk_f64_t const *a_lons, //
@@ -302,6 +307,12 @@ NK_PUBLIC void nk_vincenty_f32_serial(              //
     }
 }
 
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
 #if defined(__cplusplus)
 } // extern "C"
 #endif

package/include/numkong/maxsim/serial.h

@@ -71,6 +71,15 @@ NK_STATIC_ASSERT(sizeof(nk_maxsim_vector_metadata_t) == 12, nk_maxsim_vector_met
  */
 typedef void (*nk_maxsim_to_f32_t)(void const *source, nk_f32_t *destination);
 
+/*  Keep the serial instantiations below actually scalar, regardless of build type.
+ *  See dots/serial.h for rationale. */
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((noinline)), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC optimize("no-tree-vectorize", "no-tree-slp-vectorize", "no-ipa-cp-clone", "no-inline")
+#endif
+
 /** @brief Identity conversion for f32 sources — just a typed memcpy. */
 NK_INTERNAL void nk_f32_to_f32_(void const *source, nk_f32_t *destination) { *destination = *(nk_f32_t const *)source; }
 
@@ -483,6 +492,12 @@ NK_PUBLIC void nk_maxsim_packed_f16_serial( //
     *result = (nk_f32_t)total_angular_distance;
 }
 
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
 #if defined(__cplusplus)
 } // extern "C"
 #endif

package/include/numkong/mesh/README.md

@@ -105,67 +105,73 @@ 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
+### Intel Granite Rapids
+
+Xeon 6776P, 2.3 GHz base, `cpu_scaling_enabled=false`.
+Serial kernels compiled with `-fno-tree-vectorize`.
 
 #### 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 |
+| `nk_rmsd_f64_serial`      |       93.7 mp/s, 0.5 ulp |       87.4 mp/s, 0.5 ulp |       69.8 mp/s, 0.5 ulp |
+| `nk_kabsch_f64_serial`    |       11.8 mp/s, 0.8 ulp |       13.6 mp/s, 0.8 ulp |       12.8 mp/s, 0.8 ulp |
+| `nk_umeyama_f64_serial`   |       10.4 mp/s, 0.3 ulp |       11.7 mp/s, 0.3 ulp |       11.5 mp/s, 0.3 ulp |
+| `nk_rmsd_f64_haswell`     |        523 mp/s, 0.3 ulp |        564 mp/s, 0.4 ulp |        449 mp/s, 0.8 ulp |
+| `nk_kabsch_f64_haswell`   |       65.3 mp/s, 0.5 ulp |        203 mp/s, 0.9 ulp |        326 mp/s, 1.5 ulp |
+| `nk_umeyama_f64_haswell`  |       68.0 mp/s, 0.5 ulp |        200 mp/s, 0.8 ulp |        324 mp/s, 1.5 ulp |
+| `nk_rmsd_f64_skylake`     |        546 mp/s, 0.2 ulp |        587 mp/s, 0.3 ulp |        583 mp/s, 0.4 ulp |
+| `nk_kabsch_f64_skylake`   |       34.5 mp/s, 0.4 ulp |        107 mp/s, 0.5 ulp |        261 mp/s, 0.8 ulp |
+| `nk_umeyama_f64_skylake`  |       24.3 mp/s, 0.3 ulp |       82.7 mp/s, 0.5 ulp |        201 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 |
+| `nk_rmsd_f32_serial`      |       68.9 mp/s, 0.5 ulp |       70.7 mp/s, 0.5 ulp |       72.1 mp/s, 0.5 ulp |
+| `nk_kabsch_f32_serial`    |       11.2 mp/s, 0.8 ulp |       12.8 mp/s, 0.8 ulp |       14.0 mp/s, 0.9 ulp |
+| `nk_umeyama_f32_serial`   |       10.1 mp/s, 0.3 ulp |       11.2 mp/s, 0.3 ulp |       12.1 mp/s, 0.4 ulp |
+| `nk_rmsd_f32_haswell`     |        686 mp/s, 0.3 ulp |        848 mp/s, 0.5 ulp |        841 mp/s, 0.9 ulp |
+| `nk_kabsch_f32_haswell`   |       90.4 mp/s, 0.9 ulp |        250 mp/s, 1.3 ulp |        455 mp/s, 7.6 ulp |
+| `nk_umeyama_f32_haswell`  |       87.7 mp/s, 0.3 ulp |        250 mp/s, 0.4 ulp |        374 mp/s, 0.7 ulp |
+| `nk_rmsd_f32_skylake`     |      1,016 mp/s, 1.2 ulp |      1,112 mp/s, 1.2 ulp |      1,042 mp/s, 4.3 ulp |
+| `nk_kabsch_f32_skylake`   |       81.8 mp/s, 0.9 ulp |        241 mp/s, 4.1 ulp |        549 mp/s, 3.1 ulp |
+| `nk_umeyama_f32_skylake`  |       58.0 mp/s, 0.6 ulp |        168 mp/s, 2.9 ulp |        459 mp/s, 2.1 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 |
+| `nk_rmsd_bf16_haswell`    |        284 mp/s, 0.3 ulp |        281 mp/s, 3.5 ulp |       273 mp/s, 12.8 ulp |
+| `nk_kabsch_bf16_haswell`  |       36.2 mp/s, 0.4 ulp |        106 mp/s, 7.6 ulp |       186 mp/s, 33.0 ulp |
+| `nk_umeyama_bf16_haswell` |       34.5 mp/s, 0.3 ulp |        102 mp/s, 5.3 ulp |       186 mp/s, 23.1 ulp |
+| `nk_rmsd_bf16_skylake`    |      1,837 mp/s, 0.4 ulp |      2,357 mp/s, 5.4 ulp |     2,422 mp/s, 11.8 ulp |
+| `nk_kabsch_bf16_skylake`  |       34.1 mp/s, 0.3 ulp |        131 mp/s, 3.2 ulp |       487 mp/s, 20.4 ulp |
+| `nk_umeyama_bf16_skylake` |       34.6 mp/s, 0.3 ulp |        130 mp/s, 2.2 ulp |       394 mp/s, 14.3 ulp |
+| `nk_rmsd_bf16_genoa`      |      1,743 mp/s, 0.3 ulp |      2,323 mp/s, 3.1 ulp |     2,066 mp/s, 20.2 ulp |
+| `nk_kabsch_bf16_genoa`    |       33.4 mp/s, 0.3 ulp |        133 mp/s, 3.2 ulp |       405 mp/s, 20.3 ulp |
+| `nk_umeyama_bf16_genoa`   |       33.2 mp/s, 0.3 ulp |        129 mp/s, 2.2 ulp |       439 mp/s, 14.3 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 |
+| `nk_rmsd_f16_haswell`     |        273 mp/s, 0.2 ulp |        274 mp/s, 0.7 ulp |        291 mp/s, 2.5 ulp |
+| `nk_kabsch_f16_haswell`   |       34.4 mp/s, 0.5 ulp |       98.0 mp/s, 1.8 ulp |        197 mp/s, 8.2 ulp |
+| `nk_umeyama_f16_haswell`  |       35.5 mp/s, 0.4 ulp |       97.9 mp/s, 1.2 ulp |        196 mp/s, 5.7 ulp |
+| `nk_rmsd_f16_skylake`     |      1,834 mp/s, 0.3 ulp |      2,341 mp/s, 1.3 ulp |      2,418 mp/s, 3.9 ulp |
+| `nk_kabsch_f16_skylake`   |       34.0 mp/s, 0.7 ulp |        132 mp/s, 0.5 ulp |        480 mp/s, 4.7 ulp |
+| `nk_umeyama_f16_skylake`  |       33.8 mp/s, 0.5 ulp |        127 mp/s, 0.4 ulp |        481 mp/s, 3.3 ulp |
 
 #### WASM
 
-Measured with Wasmtime v42 (Cranelift backend).
+Measured with Wasmtime v43 (Cranelift backend), WASI-SDK 24, `-msimd128 -mrelaxed-simd`.
 
 | 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 |
+| `nk_rmsd_f64_serial`         |       89.9 mp/s, 0.5 ulp |       86.1 mp/s, 0.5 ulp |       73.4 mp/s, 0.5 ulp |
+| `nk_rmsd_f64_v128relaxed`    |        485 mp/s, 0.4 ulp |        552 mp/s, 0.7 ulp |        412 mp/s, 1.3 ulp |
+| `nk_kabsch_f64_serial`       |       12.1 mp/s, 0.8 ulp |       13.9 mp/s, 0.8 ulp |       14.0 mp/s, 0.9 ulp |
+| `nk_kabsch_f64_v128relaxed`  |       66.0 mp/s, 0.9 ulp |        188 mp/s, 1.7 ulp |        177 mp/s, 3.1 ulp |
+| `nk_umeyama_f64_serial`      |       10.8 mp/s, 0.3 ulp |       12.3 mp/s, 0.3 ulp |       12.2 mp/s, 0.4 ulp |
+| `nk_umeyama_f64_v128relaxed` |       64.0 mp/s, 0.8 ulp |        187 mp/s, 1.6 ulp |        178 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 |
+| `nk_rmsd_f32_serial`         |       80.6 mp/s, 0.5 ulp |       82.7 mp/s, 0.5 ulp |       70.3 mp/s, 0.5 ulp |
+| `nk_rmsd_f32_v128relaxed`    |        452 mp/s, 1.5 ulp |        416 mp/s, 1.3 ulp |        399 mp/s, 4.8 ulp |
+| `nk_kabsch_f32_serial`       |       11.4 mp/s, 0.8 ulp |       12.8 mp/s, 0.9 ulp |       12.7 mp/s, 0.8 ulp |
+| `nk_kabsch_f32_v128relaxed`  |       79.5 mp/s, 4.2 ulp |        132 mp/s, 3.9 ulp |       177 mp/s, 14.3 ulp |
+| `nk_umeyama_f32_serial`      |       10.1 mp/s, 0.3 ulp |       11.2 mp/s, 0.3 ulp |       11.2 mp/s, 0.3 ulp |
+| `nk_umeyama_f32_v128relaxed` |       79.4 mp/s, 2.8 ulp |        138 mp/s, 2.8 ulp |       194 mp/s, 10.1 ulp |
 
 
 ### Apple M5