usearch 2.21.3 → 2.25.0

package/numkong/include/numkong/spatial.h

@@ -0,0 +1,1681 @@
+/**
+ *  @brief SIMD-accelerated Spatial Similarity Measures.
+ *  @file include/numkong/spatial.h
+ *  @author Ash Vardanian
+ *  @date March 14, 2023
+ *
+ *  Contains following similarity measures:
+ *
+ *  - L2 (Euclidean) regular and squared distance
+ *  - Cosine (Angular) distance - @b not similarity!
+ *
+ *  For dtypes:
+ *
+ *  - f64: 64-bit IEEE floating point numbers → 64-bit floats
+ *  - f32: 32-bit IEEE floating point numbers → 64-bit floats
+ *  - f16: 16-bit IEEE floating point numbers → 32-bit floats
+ *  - bf16: 16-bit brain floating point numbers → 32-bit floats
+ *  - e4m3: 8-bit e4m3 floating point numbers → 32-bit floats
+ *  - e5m2: 8-bit e5m2 floating point numbers → 32-bit floats
+ *  - e2m3: 8-bit e2m3 floating point numbers (MX) → 32-bit floats
+ *  - e3m2: 8-bit e3m2 floating point numbers (MX) → 32-bit floats
+ *  - i8: 8-bit signed integers → 32-bit floats
+ *  - u8: 8-bit unsigned integers → 32-bit floats
+ *  - i4: 4-bit signed integers (packed pairs) → 32-bit floats
+ *  - u4: 4-bit unsigned integers (packed pairs) → 32-bit floats
+ *
+ *  For hardware architectures:
+ *
+ *  - Arm: NEON, NEON+F16, NEON+BF16, NEON+SDOT, SVE, SVE+F16, SVE+BF16
+ *  - x86: Haswell, Skylake, Ice Lake, Genoa, Sapphire Rapids, Sierra Forest
+ *  - RISC-V: RVV, RVV+BF16, RVV+HALF
+ *  - WASM: V128Relaxed
+ *
+ *  @section numerical_stability Numerical Stability
+ *
+ *  Serial kernels use compensated summation for dot, a_norm_sq, b_norm_sq — O(1) error growth regardless of vector
+ *  dimension. `f32` public outputs widen to `f64`, so widened paths use `f64` arithmetic and `sqrt64`.
+ *  Angular finalization uses rsqrt via magic constant + 3 Newton-Raphson iterations (f32,
+ *  ~34.9 correct bits) or 4 iterations (f64, ~69.3 correct bits), then clamps result ≥ 0.
+ *  L2 uses conditional `dist_sq > 0 ? sqrt(dist_sq) : 0` to avoid NaN from rounding.
+ *  Integer types (i8/u8/i4/u4) accumulate squared differences in i32 — overflows at
+ *  n > 2^31/65,025 ≈ 33K for i8 (max diff² = 255²). Output is cast to f32.
+ *
+ *  @section streaming_api Streaming API
+ *
+ *  Angular and L2 distances can be computed from a single dot-product stream and precomputed magnitudes.
+ *  The streaming helpers operate on 512-bit blocks (`nk_b512_vec_t`) and only accumulate $A*B$.
+ *  Finalization takes the magnitudes of the full vectors (L2 norms) and computes the distance.
+ *  Let the following be computed over the full vectors:
+ *
+ *      ab   = Σᵢ (aᵢ × bᵢ)
+ *      ‖a‖ = √(Σᵢ aᵢ²)
+ *      ‖b‖ = √(Σᵢ bᵢ²)
+ *
+ *  Finalization formulas:
+ *
+ *      angular(a, b) = 1 − ab / (‖a‖ × ‖b‖)
+ *      l2(a, b)      = √( ‖a‖² + ‖b‖² − 2 × ab )
+ *
+ *  The angular distance is clamped to ≥ 0, with a 0 result when both norms are zero and a 1 result when $ab$ is zero.
+ *  L2 clamps the argument of the square root at 0 to avoid negative values from rounding.
+ *
+ *  @code{.c}
+ *  nk_b512_vec_t a_block, b_block;
+ *  nk_f32_t a_norm = ..., b_norm = ...; // Precomputed L2 norms of full vectors
+ *  nk_angular_f32x8_state_haswell_t state; // Often equivalent to dot-product state
+ *  nk_angular_f32x8_init_haswell(&state);
+ *  nk_angular_f32x8_update_haswell(&state, a_block, b_block);
+ *  nk_angular_f32x8_finalize_haswell(&state, a_norm, b_norm, &distance);
+ *  @endcode
+ *
+ *  @section rsqrt_notes Reciprocal Square Root and Newton-Raphson Notes
+ *
+ *  Angular distance normalization uses reciprocal square roots to avoid the
+ *  latency of full sqrt/div pipelines. We refine the rsqrt estimate with one
+ *  (x86) or two (Arm NEON) Newton-Raphson iterations to reduce error.
+ *
+ *  Relevant instructions and caveats:
+ *
+ *      Intrinsic                Instruction      Notes
+ *      _mm_rsqrt_ps             VRSQRTPS         fast approx; refine with NR
+ *      _mm_maskz_rsqrt14_pd     VRSQRT14PD       higher-precision approx; MSVC masked-only
+ *      _mm_sqrt_ps/_mm_sqrt_pd  VSQRTPS/VSQRTPD  higher latency, sqrt/div unit
+ *
+ *  Latency/port notes (rule of thumb):
+ *  - On Intel client cores, sqrt/rsqrt execute on the divide/sqrt unit (often
+ *    port 0) and can bottleneck tight loops.
+ *  - NR refinement uses mul/FMA ports and amortizes well when `ab` is reduced
+ *    to a scalar and reused for finalization.
+ *  - Arm NEON `rsqrt` accuracy is coarse; we apply two refinement steps to keep
+ *    angular distance error bounded.
+ *
+ *  @section x86_instructions Relevant x86 Instructions
+ *
+ *  AVX2 lacks signed 8-bit dot products, so Haswell widens to i16 and uses VPMADDWD.
+ *  AVX-512 VNNI replaces that with VPDPWSSD. BF16 uses VDPBF16PS where available to avoid
+ *  convert+FMA sequences; if the ISA lacks it, we fall back to f32 FMA in the AVX2/serial:
+ *
+ *      Intrinsic             Instruction                   Icelake    Genoa
+ *      _mm256_fmadd_ps       VFMADD231PS (YMM, YMM, YMM)   4cy @ p01  4cy @ p01
+ *      _mm256_fmadd_pd       VFMADD231PD (YMM, YMM, YMM)   4cy @ p01  4cy @ p01
+ *      _mm256_madd_epi16     VPMADDWD (YMM, YMM, YMM)      5cy @ p01  3cy @ p01
+ *      _mm512_dpwssd_epi32   VPDPWSSD (ZMM, K, ZMM, ZMM)   5cy @ p05  4cy @ p01
+ *      _mm512_dpbf16_ps      VDPBF16PS (ZMM, K, ZMM, ZMM)  n/a        6cy @ p01
+ *      _mm_rsqrt_ps          VRSQRTPS (XMM, XMM)           5cy @ p0   4cy @ p01
+ *      _mm_maskz_rsqrt14_pd  VRSQRT14PD (XMM, K, XMM)      4cy @ p0   5cy @ p01
+ *      _mm_sqrt_ps           VSQRTPS (XMM, XMM)            12cy @ p0  15cy @ p01
+ *
+ *  @section arm_instructions Relevant Arm Instructions
+ *
+ *  The NEON/SVE kernels in this header are structured around FMLA/SDOT/BFDOT loops,
+ *  which is why we avoid mul+add splits and keep reductions to scalars before square roots.
+ *  Dot-product kernels for i8/u8 are only built when the "dotprod+i8mm" target is enabled;
+ *  otherwise we rely on the serial backends. BF16 kernels are enabled only with BF16 dot
+ *  instructions skipping `vbfmlal` and `vbfmlalt` alternatives to limit shuffle overhead
+ *  and code complexity.
+ *
+ *      Intrinsic     Instruction      M1 Firestorm
+ *      vfmaq_f32     FMLA.S (vec)     4c / 4c
+ *      vfmaq_f64     FMLA.D (vec)     4c / 4c
+ *      vdotq_s32     SDOT.B (vec)     3c / 4c
+ *      vbfdotq_f32   BFDOT (vec)      n/a
+ *      vrsqrteq_f32  FRSQRTE.S (vec)  3c / 1c
+ *      vrsqrtsq_f32  FRSQRTS.S (vec)  4c / 4c
+ *      vsqrtq_f32    FSQRT.S (vec)    10c / 0.5c
+ *
+ *  @section references References
+ *
+ *  - x86 intrinsics: https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html
+ *  - Arm intrinsics: https://developer.arm.com/architectures/instruction-sets/intrinsics/
+ *
+ */
+#ifndef NK_SPATIAL_H
+#define NK_SPATIAL_H
+
+#include "numkong/types.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+/**
+ *  @brief L2 (Euclidean) distance between two vectors.
+ *
+ *  @param[in] a The first vector.
+ *  @param[in] b The second vector.
+ *  @param[in] n The number of elements in each vector.
+ *  @param[out] result The output distance value.
+ *
+ *  @note The output distance value is non-negative.
+ *  @note The output distance value is zero if and only if the two vectors are identical.
+ */
+NK_DYNAMIC void nk_euclidean_f64(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_DYNAMIC void nk_euclidean_f32(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_DYNAMIC void nk_euclidean_f16(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_DYNAMIC void nk_euclidean_bf16(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_DYNAMIC void nk_euclidean_e4m3(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_DYNAMIC void nk_euclidean_e5m2(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_DYNAMIC void nk_euclidean_e2m3(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_DYNAMIC void nk_euclidean_e3m2(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_DYNAMIC void nk_euclidean_i8(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_DYNAMIC void nk_euclidean_u8(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_DYNAMIC void nk_euclidean_i4(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_DYNAMIC void nk_euclidean_u4(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_f32_t *result);
+
+/**
+ *  @brief Squared L2 (Euclidean) distance between two vectors.
+ *
+ *  @param[in] a The first vector.
+ *  @param[in] b The second vector.
+ *  @param[in] n The number of elements in each vector.
+ *  @param[out] result The output distance value.
+ *
+ *  @note The output distance value is non-negative.
+ *  @note The output distance value is zero if and only if the two vectors are identical.
+ */
+NK_DYNAMIC void nk_sqeuclidean_f64(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_DYNAMIC void nk_sqeuclidean_f32(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_DYNAMIC void nk_sqeuclidean_f16(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_DYNAMIC void nk_sqeuclidean_bf16(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_DYNAMIC void nk_sqeuclidean_e4m3(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_DYNAMIC void nk_sqeuclidean_e5m2(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_DYNAMIC void nk_sqeuclidean_e2m3(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_DYNAMIC void nk_sqeuclidean_e3m2(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_DYNAMIC void nk_sqeuclidean_i8(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_DYNAMIC void nk_sqeuclidean_u8(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_DYNAMIC void nk_sqeuclidean_i4(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_DYNAMIC void nk_sqeuclidean_u4(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_u32_t *result);
+
+/**
+ *  @brief Angular (cosine) distance between two vectors.
+ *
+ *  @param[in] a The first vector.
+ *  @param[in] b The second vector.
+ *  @param[in] n The number of elements in each vector.
+ *  @param[out] result The output distance value.
+ *
+ *  @note The output distance value is non-negative.
+ *  @note The output distance value is zero if and only if the two vectors are identical.
+ */
+NK_DYNAMIC void nk_angular_f64(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_DYNAMIC void nk_angular_f32(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_DYNAMIC void nk_angular_f16(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_DYNAMIC void nk_angular_bf16(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_DYNAMIC void nk_angular_e4m3(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_DYNAMIC void nk_angular_e5m2(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_DYNAMIC void nk_angular_e2m3(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_DYNAMIC void nk_angular_e3m2(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_DYNAMIC void nk_angular_i8(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_DYNAMIC void nk_angular_u8(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_DYNAMIC void nk_angular_i4(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_DYNAMIC void nk_angular_u4(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_f32_t *result);
+
+/*  Serial backends for all numeric types.
+ *  By default they use 32-bit arithmetic, unless the arguments themselves contain 64-bit floats.
+ */
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f64_serial(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f64_serial(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f64_serial(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f32_serial(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f32_serial(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f32_serial(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f16_serial(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f16_serial(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f16_serial(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_bf16_serial(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_bf16_serial(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_bf16_serial(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e4m3_serial(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e4m3_serial(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e4m3_serial(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e5m2_serial(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e5m2_serial(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e5m2_serial(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i8_serial(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i8_serial(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i8_serial(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u8_serial(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u8_serial(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u8_serial(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i4_serial(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i4_serial(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i4_serial(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u4_serial(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u4_serial(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u4_serial(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_f32_t *result);
+
+/*  SIMD-powered backends for Arm NEON, mostly using 32-bit arithmetic over 128-bit words.
+ *  By far the most portable backend, covering most Arm v8 devices, over a billion phones, and almost all
+ *  server CPUs produced before 2023.
+ */
+#if NK_TARGET_NEON
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f64_neon(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f64_neon(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f64_neon(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f32_neon(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f32_neon(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f32_neon(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_bf16_neon(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_bf16_neon(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_bf16_neon(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f16_neon(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f16_neon(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f16_neon(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_NEON
+
+#if NK_TARGET_NEONBFDOT
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_bf16_neonbfdot(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_bf16_neonbfdot(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_bf16_neonbfdot(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_NEONBFDOT
+
+#if NK_TARGET_NEONSDOT
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i8_neonsdot(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i8_neonsdot(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i8_neonsdot(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u8_neonsdot(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u8_neonsdot(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u8_neonsdot(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i4_neonsdot(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i4_neonsdot(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i4_neonsdot(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u4_neonsdot(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u4_neonsdot(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u4_neonsdot(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_NEONSDOT
+
+#if NK_TARGET_SVESDOT
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i8_svesdot(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i8_svesdot(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i8_svesdot(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u8_svesdot(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u8_svesdot(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u8_svesdot(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_SVESDOT
+
+#if NK_TARGET_NEONFP8
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e4m3_neonfp8(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e4m3_neonfp8(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e4m3_neonfp8(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e5m2_neonfp8(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e5m2_neonfp8(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e5m2_neonfp8(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e2m3_neonfp8(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e2m3_neonfp8(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e2m3_neonfp8(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e3m2_neonfp8(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e3m2_neonfp8(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e3m2_neonfp8(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_NEONFP8
+
+/*  SIMD-powered backends for Arm SVE, mostly using 32-bit arithmetic over variable-length platform-defined word sizes.
+ *  Designed for Arm Graviton 3, Microsoft Cobalt, as well as Nvidia Grace and newer Ampere Altra CPUs.
+ */
+#if NK_TARGET_SVE
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f32_sve(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f32_sve(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f32_sve(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f64_sve(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f64_sve(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f64_sve(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+#endif // NK_TARGET_SVE
+
+#if NK_TARGET_SVEHALF
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f16_svehalf(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f16_svehalf(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f16_svehalf(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_SVEHALF
+
+#if NK_TARGET_SVEBFDOT
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_bf16_svebfdot(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_bf16_svebfdot(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_bf16_svebfdot(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_SVEBFDOT
+
+/*  SIMD-powered backends for AVX2 CPUs of Haswell generation and newer, using 32-bit arithmetic over 256-bit words.
+ *  First demonstrated in 2011, at least one Haswell-based processor was still being sold in 2022 — the Pentium G3420.
+ *  Practically all modern x86 CPUs support AVX2, FMA, and F16C, making it a perfect baseline for SIMD algorithms.
+ *  On other hand, there is no need to implement AVX2 versions of `f32` and `f64` functions, as those are
+ *  properly vectorized by recent compilers.
+ */
+#if NK_TARGET_HASWELL
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i8_haswell(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i8_haswell(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i8_haswell(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u8_haswell(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u8_haswell(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u8_haswell(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f16_haswell(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f16_haswell(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f16_haswell(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_bf16_haswell(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_bf16_haswell(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_bf16_haswell(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f32_haswell(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f32_haswell(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f32_haswell(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f64_haswell(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f64_haswell(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f64_haswell(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+#endif // NK_TARGET_HASWELL
+
+/*  SIMD-powered backends for AVX512 CPUs of Skylake generation and newer, using 32-bit arithmetic over 512-bit words.
+ *  Skylake was launched in 2015, and discontinued in 2019. Skylake had support for F, CD, VL, DQ, and BW extensions,
+ *  as well as masked operations. This is enough to supersede auto-vectorization on `f32` and `f64` types.
+ *
+ *  Sadly, we can't effectively interleave different kinds of arithmetic instructions to utilize more ports:
+ *
+ *  > Like Intel server architectures since Skylake-X, SPR cores feature two 512-bit FMA units, and organize them in a
+ *    similar fashion. > One 512-bit FMA unit is created by fusing two 256-bit ones on port 0 and port 1. The other is
+ *    added to port 5, as a server-specific > core extension. The FMA units on port 0 and 1 are configured into
+ *    2×256-bit or 1×512-bit mode depending on whether 512-bit FMA > instructions are present in the scheduler. That
+ *    means a mix of 256-bit and 512-bit FMA instructions will not achieve higher IPC > than executing 512-bit
+ *    instructions alone.
+ *
+ *  Source: https://chipsandcheese.com/p/a-peek-at-sapphire-rapids
+ */
+#if NK_TARGET_SKYLAKE
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f32_skylake(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f32_skylake(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f32_skylake(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f64_skylake(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f64_skylake(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f64_skylake(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f16_skylake(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f16_skylake(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f16_skylake(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e4m3_skylake(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e4m3_skylake(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e4m3_skylake(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e5m2_skylake(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e5m2_skylake(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e5m2_skylake(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_SKYLAKE
+
+/*  SIMD-powered backends for AVX512 CPUs of Ice Lake generation and newer, using mixed arithmetic over 512-bit words.
+ *  Ice Lake added VNNI, VPOPCNTDQ, IFMA, VBMI, VAES, GFNI, VBMI2, BITALG, VPCLMULQDQ, and other extensions for integral
+ *  operations. Sapphire Rapids added tiled matrix operations, but we are most interested in the new mixed-precision FMA
+ *  instructions.
+ */
+#if NK_TARGET_ICELAKE
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i4_icelake(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i4_icelake(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i4_icelake(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u4_icelake(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u4_icelake(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u4_icelake(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i8_icelake(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i8_icelake(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i8_icelake(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u8_icelake(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u8_icelake(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u8_icelake(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e4m3_icelake(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e4m3_icelake(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e4m3_icelake(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e2m3_icelake(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e2m3_icelake(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e2m3_icelake(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e3m2_icelake(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e3m2_icelake(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e3m2_icelake(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_ICELAKE
+
+#if NK_TARGET_GENOA
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_bf16_genoa(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_bf16_genoa(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_bf16_genoa(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e5m2_genoa(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e5m2_genoa(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e5m2_genoa(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_GENOA
+
+#if NK_TARGET_DIAMOND
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f16_diamond(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f16_diamond(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f16_diamond(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e4m3_diamond(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e4m3_diamond(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e4m3_diamond(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e5m2_diamond(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e5m2_diamond(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e5m2_diamond(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_DIAMOND
+
+/*  SIMD-powered backends for AVX-INT8-VNNI extensions on Xeon 6 CPUs, including Sierra Forest and Granite Rapids.
+ *  The packs many "efficiency" cores into a single socket, avoiding heavy 512-bit operations, and focusing on
+ *  256-bit ones.
+ */
+#if NK_TARGET_SIERRA
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i8_sierra(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i8_sierra(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i8_sierra(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u8_sierra(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u8_sierra(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u8_sierra(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e2m3_sierra(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e2m3_sierra(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e2m3_sierra(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e3m2_sierra(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e3m2_sierra(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e3m2_sierra(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_SIERRA
+
+#if NK_TARGET_ALDER
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i8_alder(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i8_alder(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i8_alder(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u8_alder(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u8_alder(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u8_alder(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e2m3_alder(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e2m3_alder(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e2m3_alder(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e3m2_alder(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e3m2_alder(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e3m2_alder(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_ALDER
+
+#if NK_TARGET_V128RELAXED
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f32_v128relaxed(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f64_v128relaxed(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f32_v128relaxed(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f64_v128relaxed(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f32_v128relaxed(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f64_v128relaxed(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f16_v128relaxed(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_bf16_v128relaxed(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f16_v128relaxed(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_bf16_v128relaxed(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f16_v128relaxed(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_bf16_v128relaxed(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u8_v128relaxed(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u8_v128relaxed(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u8_v128relaxed(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i8_v128relaxed(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i8_v128relaxed(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i8_v128relaxed(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e4m3_v128relaxed(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e4m3_v128relaxed(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e4m3_v128relaxed(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e5m2_v128relaxed(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e5m2_v128relaxed(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e5m2_v128relaxed(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e2m3_v128relaxed(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e2m3_v128relaxed(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e2m3_v128relaxed(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e3m2_v128relaxed(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e3m2_v128relaxed(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e3m2_v128relaxed(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_V128RELAXED
+
+/*  SIMD-powered backends for RISC-V Vector extension, using scalable vector arithmetic.
+ *  Designed for SiFive, T-Head, and other RISC-V processors with the V extension.
+ */
+#if NK_TARGET_RVV
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f64_rvv(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f64_rvv(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f64_rvv(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f32_rvv(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f32_rvv(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f32_rvv(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f16_rvv(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f16_rvv(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f16_rvv(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_bf16_rvv(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_bf16_rvv(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_bf16_rvv(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e4m3_rvv(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e4m3_rvv(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e4m3_rvv(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_e5m2_rvv(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_e5m2_rvv(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_e5m2_rvv(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i8_rvv(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i8_rvv(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i8_rvv(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u8_rvv(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u8_rvv(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u8_rvv(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_i4_rvv(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_i4_rvv(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_i4_rvv(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_u4_rvv(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_u4_rvv(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_u32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_u4_rvv(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_RVV
+
+#if NK_TARGET_RVVHALF
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_f16_rvvhalf(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_f16_rvvhalf(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_f16_rvvhalf(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_RVVHALF
+
+#if NK_TARGET_RVVBF16
+/** @copydoc nk_euclidean_f64 */
+NK_PUBLIC void nk_euclidean_bf16_rvvbf16(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_sqeuclidean_f64 */
+NK_PUBLIC void nk_sqeuclidean_bf16_rvvbf16(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+/** @copydoc nk_angular_f64 */
+NK_PUBLIC void nk_angular_bf16_rvvbf16(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result);
+#endif // NK_TARGET_RVVBF16
+
+/** @brief Returns the output dtype for L2 (Euclidean) distance. */
+NK_INTERNAL nk_dtype_t nk_euclidean_output_dtype(nk_dtype_t dtype) {
+    switch (dtype) {
+    case nk_f64_k: return nk_f64_k;
+    case nk_f32_k: return nk_f64_k;
+    case nk_f16_k: return nk_f32_k;
+    case nk_bf16_k: return nk_f32_k;
+    case nk_e4m3_k: return nk_f32_k;
+    case nk_e5m2_k: return nk_f32_k;
+    case nk_e2m3_k: return nk_f32_k;
+    case nk_e3m2_k: return nk_f32_k;
+    case nk_i8_k: return nk_f32_k;
+    case nk_u8_k: return nk_f32_k;
+    case nk_i4_k: return nk_f32_k;
+    case nk_u4_k: return nk_f32_k;
+    default: return nk_dtype_unknown_k;
+    }
+}
+
+/** @brief Returns the output dtype for L2 squared distance. */
+NK_INTERNAL nk_dtype_t nk_sqeuclidean_output_dtype(nk_dtype_t dtype) {
+    switch (dtype) {
+    case nk_f64_k: return nk_f64_k;
+    case nk_f32_k: return nk_f64_k;
+    case nk_f16_k: return nk_f32_k;
+    case nk_bf16_k: return nk_f32_k;
+    case nk_e4m3_k: return nk_f32_k;
+    case nk_e5m2_k: return nk_f32_k;
+    case nk_e2m3_k: return nk_f32_k;
+    case nk_e3m2_k: return nk_f32_k;
+    case nk_i8_k: return nk_u32_k;
+    case nk_u8_k: return nk_u32_k;
+    case nk_i4_k: return nk_u32_k;
+    case nk_u4_k: return nk_u32_k;
+    default: return nk_dtype_unknown_k;
+    }
+}
+
+/** @brief Returns the output dtype for angular/cosine distance. */
+NK_INTERNAL nk_dtype_t nk_angular_output_dtype(nk_dtype_t dtype) {
+    switch (dtype) {
+    case nk_f64_k: return nk_f64_k;
+    case nk_f32_k: return nk_f64_k;
+    case nk_f16_k: return nk_f32_k;
+    case nk_bf16_k: return nk_f32_k;
+    case nk_e4m3_k: return nk_f32_k;
+    case nk_e5m2_k: return nk_f32_k;
+    case nk_e2m3_k: return nk_f32_k;
+    case nk_e3m2_k: return nk_f32_k;
+    case nk_i8_k: return nk_f32_k;
+    case nk_u8_k: return nk_f32_k;
+    case nk_i4_k: return nk_f32_k;
+    case nk_u4_k: return nk_f32_k;
+    default: return nk_dtype_unknown_k;
+    }
+}
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#include "numkong/spatial/serial.h"
+#include "numkong/spatial/neon.h"
+#include "numkong/spatial/neonbfdot.h"
+#include "numkong/spatial/neonsdot.h"
+#include "numkong/spatial/sve.h"
+#include "numkong/spatial/svehalf.h"
+#include "numkong/spatial/svebfdot.h"
+#include "numkong/spatial/svesdot.h"
+#include "numkong/spatial/neonfp8.h"
+#include "numkong/spatial/haswell.h"
+#include "numkong/spatial/skylake.h"
+#include "numkong/spatial/genoa.h"
+#include "numkong/spatial/diamond.h"
+#include "numkong/spatial/icelake.h"
+#include "numkong/spatial/alder.h"
+#include "numkong/spatial/sierra.h"
+#include "numkong/spatial/rvv.h"
+#include "numkong/spatial/rvvhalf.h"
+#include "numkong/spatial/rvvbf16.h"
+#include "numkong/spatial/v128relaxed.h"
+#include "numkong/spatial/powervsx.h"
+#include "numkong/spatial/loongsonasx.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#if !NK_DYNAMIC_DISPATCH
+
+NK_PUBLIC void nk_euclidean_f64(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result) {
+#if NK_TARGET_V128RELAXED
+    nk_euclidean_f64_v128relaxed(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_euclidean_f64_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_euclidean_f64_loongsonasx(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_euclidean_f64_rvv(a, b, n, result);
+#elif NK_TARGET_SVE
+    nk_euclidean_f64_sve(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_euclidean_f64_neon(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_euclidean_f64_skylake(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_euclidean_f64_haswell(a, b, n, result);
+#else
+    nk_euclidean_f64_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_sqeuclidean_f64(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result) {
+#if NK_TARGET_V128RELAXED
+    nk_sqeuclidean_f64_v128relaxed(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_sqeuclidean_f64_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_sqeuclidean_f64_loongsonasx(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_sqeuclidean_f64_rvv(a, b, n, result);
+#elif NK_TARGET_SVE
+    nk_sqeuclidean_f64_sve(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_sqeuclidean_f64_neon(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_sqeuclidean_f64_skylake(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_sqeuclidean_f64_haswell(a, b, n, result);
+#else
+    nk_sqeuclidean_f64_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_angular_f64(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result) {
+#if NK_TARGET_V128RELAXED
+    nk_angular_f64_v128relaxed(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_angular_f64_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_angular_f64_loongsonasx(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_angular_f64_rvv(a, b, n, result);
+#elif NK_TARGET_SVE
+    nk_angular_f64_sve(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_angular_f64_neon(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_angular_f64_skylake(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_angular_f64_haswell(a, b, n, result);
+#else
+    nk_angular_f64_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_euclidean_f32(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result) {
+#if NK_TARGET_V128RELAXED
+    nk_euclidean_f32_v128relaxed(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_euclidean_f32_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_euclidean_f32_loongsonasx(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_euclidean_f32_rvv(a, b, n, result);
+#elif NK_TARGET_SVE
+    nk_euclidean_f32_sve(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_euclidean_f32_neon(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_euclidean_f32_skylake(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_euclidean_f32_haswell(a, b, n, result);
+#else
+    nk_euclidean_f32_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_sqeuclidean_f32(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result) {
+#if NK_TARGET_V128RELAXED
+    nk_sqeuclidean_f32_v128relaxed(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_sqeuclidean_f32_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_sqeuclidean_f32_loongsonasx(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_sqeuclidean_f32_rvv(a, b, n, result);
+#elif NK_TARGET_SVE
+    nk_sqeuclidean_f32_sve(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_sqeuclidean_f32_neon(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_sqeuclidean_f32_skylake(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_sqeuclidean_f32_haswell(a, b, n, result);
+#else
+    nk_sqeuclidean_f32_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_angular_f32(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result) {
+#if NK_TARGET_V128RELAXED
+    nk_angular_f32_v128relaxed(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_angular_f32_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_angular_f32_loongsonasx(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_angular_f32_rvv(a, b, n, result);
+#elif NK_TARGET_SVE
+    nk_angular_f32_sve(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_angular_f32_neon(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_angular_f32_skylake(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_angular_f32_haswell(a, b, n, result);
+#else
+    nk_angular_f32_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_euclidean_f16(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_V128RELAXED
+    nk_euclidean_f16_v128relaxed(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_euclidean_f16_powervsx(a, b, n, result);
+#elif NK_TARGET_RVVHALF
+    nk_euclidean_f16_rvvhalf(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_euclidean_f16_rvv(a, b, n, result);
+#elif NK_TARGET_SVEHALF
+    nk_euclidean_f16_svehalf(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_euclidean_f16_neon(a, b, n, result);
+#elif NK_TARGET_DIAMOND
+    nk_euclidean_f16_diamond(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_euclidean_f16_skylake(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_euclidean_f16_haswell(a, b, n, result);
+#else
+    nk_euclidean_f16_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_sqeuclidean_f16(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_V128RELAXED
+    nk_sqeuclidean_f16_v128relaxed(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_sqeuclidean_f16_powervsx(a, b, n, result);
+#elif NK_TARGET_RVVHALF
+    nk_sqeuclidean_f16_rvvhalf(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_sqeuclidean_f16_rvv(a, b, n, result);
+#elif NK_TARGET_SVEHALF
+    nk_sqeuclidean_f16_svehalf(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_sqeuclidean_f16_neon(a, b, n, result);
+#elif NK_TARGET_DIAMOND
+    nk_sqeuclidean_f16_diamond(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_sqeuclidean_f16_skylake(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_sqeuclidean_f16_haswell(a, b, n, result);
+#else
+    nk_sqeuclidean_f16_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_angular_f16(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_V128RELAXED
+    nk_angular_f16_v128relaxed(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_angular_f16_powervsx(a, b, n, result);
+#elif NK_TARGET_RVVHALF
+    nk_angular_f16_rvvhalf(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_angular_f16_rvv(a, b, n, result);
+#elif NK_TARGET_SVEHALF
+    nk_angular_f16_svehalf(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_angular_f16_neon(a, b, n, result);
+#elif NK_TARGET_DIAMOND
+    nk_angular_f16_diamond(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_angular_f16_skylake(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_angular_f16_haswell(a, b, n, result);
+#else
+    nk_angular_f16_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_euclidean_bf16(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_V128RELAXED
+    nk_euclidean_bf16_v128relaxed(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_euclidean_bf16_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_euclidean_bf16_loongsonasx(a, b, n, result);
+#elif NK_TARGET_RVVBF16
+    nk_euclidean_bf16_rvvbf16(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_euclidean_bf16_rvv(a, b, n, result);
+#elif NK_TARGET_SVEBFDOT
+    nk_euclidean_bf16_svebfdot(a, b, n, result);
+#elif NK_TARGET_NEONBFDOT
+    nk_euclidean_bf16_neonbfdot(a, b, n, result);
+#elif NK_TARGET_GENOA
+    nk_euclidean_bf16_genoa(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_euclidean_bf16_haswell(a, b, n, result);
+#else
+    nk_euclidean_bf16_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_sqeuclidean_bf16(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_V128RELAXED
+    nk_sqeuclidean_bf16_v128relaxed(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_sqeuclidean_bf16_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_sqeuclidean_bf16_loongsonasx(a, b, n, result);
+#elif NK_TARGET_RVVBF16
+    nk_sqeuclidean_bf16_rvvbf16(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_sqeuclidean_bf16_rvv(a, b, n, result);
+#elif NK_TARGET_SVEBFDOT
+    nk_sqeuclidean_bf16_svebfdot(a, b, n, result);
+#elif NK_TARGET_NEONBFDOT
+    nk_sqeuclidean_bf16_neonbfdot(a, b, n, result);
+#elif NK_TARGET_GENOA
+    nk_sqeuclidean_bf16_genoa(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_sqeuclidean_bf16_haswell(a, b, n, result);
+#else
+    nk_sqeuclidean_bf16_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_angular_bf16(nk_bf16_t const *a, nk_bf16_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_V128RELAXED
+    nk_angular_bf16_v128relaxed(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_angular_bf16_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_angular_bf16_loongsonasx(a, b, n, result);
+#elif NK_TARGET_RVVBF16
+    nk_angular_bf16_rvvbf16(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_angular_bf16_rvv(a, b, n, result);
+#elif NK_TARGET_SVEBFDOT
+    nk_angular_bf16_svebfdot(a, b, n, result);
+#elif NK_TARGET_NEONBFDOT
+    nk_angular_bf16_neonbfdot(a, b, n, result);
+#elif NK_TARGET_GENOA
+    nk_angular_bf16_genoa(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_angular_bf16_haswell(a, b, n, result);
+#else
+    nk_angular_bf16_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_euclidean_e4m3(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_NEONFP8
+    nk_euclidean_e4m3_neonfp8(a, b, n, result);
+#elif NK_TARGET_DIAMOND
+    nk_euclidean_e4m3_diamond(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_euclidean_e4m3_icelake(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_euclidean_e4m3_skylake(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_euclidean_e4m3_rvv(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_euclidean_e4m3_v128relaxed(a, b, n, result);
+#else
+    nk_euclidean_e4m3_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_sqeuclidean_e4m3(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_NEONFP8
+    nk_sqeuclidean_e4m3_neonfp8(a, b, n, result);
+#elif NK_TARGET_DIAMOND
+    nk_sqeuclidean_e4m3_diamond(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_sqeuclidean_e4m3_icelake(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_sqeuclidean_e4m3_skylake(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_sqeuclidean_e4m3_rvv(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_sqeuclidean_e4m3_v128relaxed(a, b, n, result);
+#else
+    nk_sqeuclidean_e4m3_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_angular_e4m3(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_NEONFP8
+    nk_angular_e4m3_neonfp8(a, b, n, result);
+#elif NK_TARGET_DIAMOND
+    nk_angular_e4m3_diamond(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_angular_e4m3_icelake(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_angular_e4m3_skylake(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_angular_e4m3_rvv(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_angular_e4m3_v128relaxed(a, b, n, result);
+#else
+    nk_angular_e4m3_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_euclidean_e5m2(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_NEONFP8
+    nk_euclidean_e5m2_neonfp8(a, b, n, result);
+#elif NK_TARGET_DIAMOND
+    nk_euclidean_e5m2_diamond(a, b, n, result);
+#elif NK_TARGET_GENOA
+    nk_euclidean_e5m2_genoa(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_euclidean_e5m2_skylake(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_euclidean_e5m2_rvv(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_euclidean_e5m2_v128relaxed(a, b, n, result);
+#else
+    nk_euclidean_e5m2_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_sqeuclidean_e5m2(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_NEONFP8
+    nk_sqeuclidean_e5m2_neonfp8(a, b, n, result);
+#elif NK_TARGET_DIAMOND
+    nk_sqeuclidean_e5m2_diamond(a, b, n, result);
+#elif NK_TARGET_GENOA
+    nk_sqeuclidean_e5m2_genoa(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_sqeuclidean_e5m2_skylake(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_sqeuclidean_e5m2_rvv(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_sqeuclidean_e5m2_v128relaxed(a, b, n, result);
+#else
+    nk_sqeuclidean_e5m2_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_angular_e5m2(nk_e5m2_t const *a, nk_e5m2_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_NEONFP8
+    nk_angular_e5m2_neonfp8(a, b, n, result);
+#elif NK_TARGET_DIAMOND
+    nk_angular_e5m2_diamond(a, b, n, result);
+#elif NK_TARGET_GENOA
+    nk_angular_e5m2_genoa(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_angular_e5m2_skylake(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_angular_e5m2_rvv(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_angular_e5m2_v128relaxed(a, b, n, result);
+#else
+    nk_angular_e5m2_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_euclidean_e2m3(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_NEONFP8
+    nk_euclidean_e2m3_neonfp8(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_euclidean_e2m3_icelake(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_euclidean_e2m3_skylake(a, b, n, result);
+#elif NK_TARGET_SIERRA
+    nk_euclidean_e2m3_sierra(a, b, n, result);
+#elif NK_TARGET_ALDER
+    nk_euclidean_e2m3_alder(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_euclidean_e2m3_haswell(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_euclidean_e2m3_neon(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_euclidean_e2m3_v128relaxed(a, b, n, result);
+#else
+    nk_euclidean_e2m3_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_sqeuclidean_e2m3(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_NEONFP8
+    nk_sqeuclidean_e2m3_neonfp8(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_sqeuclidean_e2m3_icelake(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_sqeuclidean_e2m3_skylake(a, b, n, result);
+#elif NK_TARGET_SIERRA
+    nk_sqeuclidean_e2m3_sierra(a, b, n, result);
+#elif NK_TARGET_ALDER
+    nk_sqeuclidean_e2m3_alder(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_sqeuclidean_e2m3_haswell(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_sqeuclidean_e2m3_neon(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_sqeuclidean_e2m3_v128relaxed(a, b, n, result);
+#else
+    nk_sqeuclidean_e2m3_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_angular_e2m3(nk_e2m3_t const *a, nk_e2m3_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_NEONFP8
+    nk_angular_e2m3_neonfp8(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_angular_e2m3_icelake(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_angular_e2m3_skylake(a, b, n, result);
+#elif NK_TARGET_SIERRA
+    nk_angular_e2m3_sierra(a, b, n, result);
+#elif NK_TARGET_ALDER
+    nk_angular_e2m3_alder(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_angular_e2m3_haswell(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_angular_e2m3_neon(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_angular_e2m3_v128relaxed(a, b, n, result);
+#else
+    nk_angular_e2m3_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_euclidean_e3m2(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_NEONFP8
+    nk_euclidean_e3m2_neonfp8(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_euclidean_e3m2_icelake(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_euclidean_e3m2_skylake(a, b, n, result);
+#elif NK_TARGET_SIERRA
+    nk_euclidean_e3m2_sierra(a, b, n, result);
+#elif NK_TARGET_ALDER
+    nk_euclidean_e3m2_alder(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_euclidean_e3m2_haswell(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_euclidean_e3m2_neon(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_euclidean_e3m2_v128relaxed(a, b, n, result);
+#else
+    nk_euclidean_e3m2_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_sqeuclidean_e3m2(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_NEONFP8
+    nk_sqeuclidean_e3m2_neonfp8(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_sqeuclidean_e3m2_icelake(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_sqeuclidean_e3m2_skylake(a, b, n, result);
+#elif NK_TARGET_SIERRA
+    nk_sqeuclidean_e3m2_sierra(a, b, n, result);
+#elif NK_TARGET_ALDER
+    nk_sqeuclidean_e3m2_alder(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_sqeuclidean_e3m2_haswell(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_sqeuclidean_e3m2_neon(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_sqeuclidean_e3m2_v128relaxed(a, b, n, result);
+#else
+    nk_sqeuclidean_e3m2_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_angular_e3m2(nk_e3m2_t const *a, nk_e3m2_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_NEONFP8
+    nk_angular_e3m2_neonfp8(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_angular_e3m2_icelake(a, b, n, result);
+#elif NK_TARGET_SKYLAKE
+    nk_angular_e3m2_skylake(a, b, n, result);
+#elif NK_TARGET_SIERRA
+    nk_angular_e3m2_sierra(a, b, n, result);
+#elif NK_TARGET_ALDER
+    nk_angular_e3m2_alder(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_angular_e3m2_haswell(a, b, n, result);
+#elif NK_TARGET_NEON
+    nk_angular_e3m2_neon(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_angular_e3m2_v128relaxed(a, b, n, result);
+#else
+    nk_angular_e3m2_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_euclidean_i8(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_RVV
+    nk_euclidean_i8_rvv(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_euclidean_i8_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_euclidean_i8_loongsonasx(a, b, n, result);
+#elif NK_TARGET_SVESDOT
+    nk_euclidean_i8_svesdot(a, b, n, result);
+#elif NK_TARGET_NEONSDOT
+    nk_euclidean_i8_neonsdot(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_euclidean_i8_icelake(a, b, n, result);
+#elif NK_TARGET_SIERRA
+    nk_euclidean_i8_sierra(a, b, n, result);
+#elif NK_TARGET_ALDER
+    nk_euclidean_i8_alder(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_euclidean_i8_haswell(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_euclidean_i8_v128relaxed(a, b, n, result);
+#else
+    nk_euclidean_i8_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_sqeuclidean_i8(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_u32_t *result) {
+#if NK_TARGET_RVV
+    nk_sqeuclidean_i8_rvv(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_sqeuclidean_i8_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_sqeuclidean_i8_loongsonasx(a, b, n, result);
+#elif NK_TARGET_SVESDOT
+    nk_sqeuclidean_i8_svesdot(a, b, n, result);
+#elif NK_TARGET_NEONSDOT
+    nk_sqeuclidean_i8_neonsdot(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_sqeuclidean_i8_icelake(a, b, n, result);
+#elif NK_TARGET_SIERRA
+    nk_sqeuclidean_i8_sierra(a, b, n, result);
+#elif NK_TARGET_ALDER
+    nk_sqeuclidean_i8_alder(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_sqeuclidean_i8_haswell(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_sqeuclidean_i8_v128relaxed(a, b, n, result);
+#else
+    nk_sqeuclidean_i8_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_angular_i8(nk_i8_t const *a, nk_i8_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_RVV
+    nk_angular_i8_rvv(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_angular_i8_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_angular_i8_loongsonasx(a, b, n, result);
+#elif NK_TARGET_SVESDOT
+    nk_angular_i8_svesdot(a, b, n, result);
+#elif NK_TARGET_NEONSDOT
+    nk_angular_i8_neonsdot(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_angular_i8_icelake(a, b, n, result);
+#elif NK_TARGET_SIERRA
+    nk_angular_i8_sierra(a, b, n, result);
+#elif NK_TARGET_ALDER
+    nk_angular_i8_alder(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_angular_i8_haswell(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_angular_i8_v128relaxed(a, b, n, result);
+#else
+    nk_angular_i8_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_euclidean_u8(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_RVV
+    nk_euclidean_u8_rvv(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_euclidean_u8_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_euclidean_u8_loongsonasx(a, b, n, result);
+#elif NK_TARGET_SVESDOT
+    nk_euclidean_u8_svesdot(a, b, n, result);
+#elif NK_TARGET_NEONSDOT
+    nk_euclidean_u8_neonsdot(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_euclidean_u8_icelake(a, b, n, result);
+#elif NK_TARGET_SIERRA
+    nk_euclidean_u8_sierra(a, b, n, result);
+#elif NK_TARGET_ALDER
+    nk_euclidean_u8_alder(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_euclidean_u8_haswell(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_euclidean_u8_v128relaxed(a, b, n, result);
+#else
+    nk_euclidean_u8_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_sqeuclidean_u8(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result) {
+#if NK_TARGET_RVV
+    nk_sqeuclidean_u8_rvv(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_sqeuclidean_u8_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_sqeuclidean_u8_loongsonasx(a, b, n, result);
+#elif NK_TARGET_SVESDOT
+    nk_sqeuclidean_u8_svesdot(a, b, n, result);
+#elif NK_TARGET_NEONSDOT
+    nk_sqeuclidean_u8_neonsdot(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_sqeuclidean_u8_icelake(a, b, n, result);
+#elif NK_TARGET_SIERRA
+    nk_sqeuclidean_u8_sierra(a, b, n, result);
+#elif NK_TARGET_ALDER
+    nk_sqeuclidean_u8_alder(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_sqeuclidean_u8_haswell(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_sqeuclidean_u8_v128relaxed(a, b, n, result);
+#else
+    nk_sqeuclidean_u8_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_angular_u8(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_RVV
+    nk_angular_u8_rvv(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_angular_u8_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_angular_u8_loongsonasx(a, b, n, result);
+#elif NK_TARGET_SVESDOT
+    nk_angular_u8_svesdot(a, b, n, result);
+#elif NK_TARGET_NEONSDOT
+    nk_angular_u8_neonsdot(a, b, n, result);
+#elif NK_TARGET_ICELAKE
+    nk_angular_u8_icelake(a, b, n, result);
+#elif NK_TARGET_SIERRA
+    nk_angular_u8_sierra(a, b, n, result);
+#elif NK_TARGET_ALDER
+    nk_angular_u8_alder(a, b, n, result);
+#elif NK_TARGET_HASWELL
+    nk_angular_u8_haswell(a, b, n, result);
+#elif NK_TARGET_V128RELAXED
+    nk_angular_u8_v128relaxed(a, b, n, result);
+#else
+    nk_angular_u8_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_euclidean_i4(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_ICELAKE
+    nk_euclidean_i4_icelake(a, b, n, result);
+#elif NK_TARGET_NEONSDOT
+    nk_euclidean_i4_neonsdot(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_euclidean_i4_rvv(a, b, n, result);
+#else
+    nk_euclidean_i4_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_sqeuclidean_i4(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_u32_t *result) {
+#if NK_TARGET_ICELAKE
+    nk_sqeuclidean_i4_icelake(a, b, n, result);
+#elif NK_TARGET_NEONSDOT
+    nk_sqeuclidean_i4_neonsdot(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_sqeuclidean_i4_rvv(a, b, n, result);
+#else
+    nk_sqeuclidean_i4_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_angular_i4(nk_i4x2_t const *a, nk_i4x2_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_ICELAKE
+    nk_angular_i4_icelake(a, b, n, result);
+#elif NK_TARGET_NEONSDOT
+    nk_angular_i4_neonsdot(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_angular_i4_rvv(a, b, n, result);
+#else
+    nk_angular_i4_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_euclidean_u4(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_ICELAKE
+    nk_euclidean_u4_icelake(a, b, n, result);
+#elif NK_TARGET_NEONSDOT
+    nk_euclidean_u4_neonsdot(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_euclidean_u4_rvv(a, b, n, result);
+#else
+    nk_euclidean_u4_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_sqeuclidean_u4(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_u32_t *result) {
+#if NK_TARGET_ICELAKE
+    nk_sqeuclidean_u4_icelake(a, b, n, result);
+#elif NK_TARGET_NEONSDOT
+    nk_sqeuclidean_u4_neonsdot(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_sqeuclidean_u4_rvv(a, b, n, result);
+#else
+    nk_sqeuclidean_u4_serial(a, b, n, result);
+#endif
+}
+
+NK_PUBLIC void nk_angular_u4(nk_u4x2_t const *a, nk_u4x2_t const *b, nk_size_t n, nk_f32_t *result) {
+#if NK_TARGET_ICELAKE
+    nk_angular_u4_icelake(a, b, n, result);
+#elif NK_TARGET_NEONSDOT
+    nk_angular_u4_neonsdot(a, b, n, result);
+#elif NK_TARGET_RVV
+    nk_angular_u4_rvv(a, b, n, result);
+#else
+    nk_angular_u4_serial(a, b, n, result);
+#endif
+}
+
+#endif // !NK_DYNAMIC_DISPATCH
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif