usearch 2.21.4 → 2.25.0

package/numkong/include/numkong/spatial/diamond.h

@@ -0,0 +1,240 @@
+/**
+ *  @brief SIMD-accelerated Spatial Similarity Measures for Diamond Rapids.
+ *  @file include/numkong/spatial/diamond.h
+ *  @author Ash Vardanian
+ *  @date March 23, 2026
+ *
+ *  @sa include/numkong/spatial.h
+ *
+ *  For L2 distance, uses the identity: (a−b)² = a² + b² − 2 × a × b,
+ *  with VCVTHF82PH/VCVTBF82PH for 1-instruction FP8→FP16 conversion and
+ *  VDPPHPS for FP16-pair dot products accumulating into FP32.
+ */
+#ifndef NK_SPATIAL_DIAMOND_H
+#define NK_SPATIAL_DIAMOND_H
+
+#if NK_TARGET_X8664_
+#if NK_TARGET_DIAMOND
+
+#include "numkong/types.h"
+#include "numkong/spatial/haswell.h" // `nk_angular_normalize_f32_haswell_`, `nk_f32_sqrt_haswell`
+#include "numkong/reduce/skylake.h"  // `nk_reduce_add_f32x16_skylake_`
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute push(                                                                                    \
+    __attribute__((target("avx2,avx512f,avx512vl,avx512bw,avx512dq,avx512fp16,avx10.2-512,f16c,fma,bmi,bmi2"))), \
+    apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("avx2", "avx512f", "avx512vl", "avx512bw", "avx512dq", "avx512fp16", "avx10.2-512", "f16c", "fma", \
+                   "bmi", "bmi2")
+#endif
+
+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) {
+    __m512 a_sq_f32x16 = _mm512_setzero_ps();
+    __m512 b_sq_f32x16 = _mm512_setzero_ps();
+    __m512 ab_f32x16 = _mm512_setzero_ps();
+    __m256i a_e4m3x32, b_e4m3x32;
+
+nk_sqeuclidean_e4m3_diamond_cycle:
+    if (n < 32) {
+        __mmask32 mask = (__mmask32)_bzhi_u32(0xFFFFFFFF, n);
+        a_e4m3x32 = _mm256_maskz_loadu_epi8(mask, a);
+        b_e4m3x32 = _mm256_maskz_loadu_epi8(mask, b);
+        n = 0;
+    }
+    else {
+        a_e4m3x32 = _mm256_loadu_epi8(a);
+        b_e4m3x32 = _mm256_loadu_epi8(b);
+        a += 32, b += 32, n -= 32;
+    }
+    __m512h a_f16x32 = _mm512_cvthf8_ph(a_e4m3x32);
+    __m512h b_f16x32 = _mm512_cvthf8_ph(b_e4m3x32);
+    a_sq_f32x16 = _mm512_dpph_ps(a_sq_f32x16, a_f16x32, a_f16x32);
+    b_sq_f32x16 = _mm512_dpph_ps(b_sq_f32x16, b_f16x32, b_f16x32);
+    ab_f32x16 = _mm512_dpph_ps(ab_f32x16, a_f16x32, b_f16x32);
+    if (n) goto nk_sqeuclidean_e4m3_diamond_cycle;
+
+    __m512 sum_sq_f32x16 = _mm512_add_ps(a_sq_f32x16, b_sq_f32x16);
+    *result = nk_reduce_add_f32x16_skylake_(_mm512_fnmadd_ps(_mm512_set1_ps(2.0f), ab_f32x16, sum_sq_f32x16));
+}
+
+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) {
+    nk_sqeuclidean_e4m3_diamond(a, b, n, result);
+    *result = nk_f32_sqrt_haswell(*result);
+}
+
+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) {
+    __m512 dot_f32x16 = _mm512_setzero_ps();
+    __m512 a_norm_sq_f32x16 = _mm512_setzero_ps();
+    __m512 b_norm_sq_f32x16 = _mm512_setzero_ps();
+    __m256i a_e4m3x32, b_e4m3x32;
+
+nk_angular_e4m3_diamond_cycle:
+    if (n < 32) {
+        __mmask32 mask = (__mmask32)_bzhi_u32(0xFFFFFFFF, n);
+        a_e4m3x32 = _mm256_maskz_loadu_epi8(mask, a);
+        b_e4m3x32 = _mm256_maskz_loadu_epi8(mask, b);
+        n = 0;
+    }
+    else {
+        a_e4m3x32 = _mm256_loadu_epi8(a);
+        b_e4m3x32 = _mm256_loadu_epi8(b);
+        a += 32, b += 32, n -= 32;
+    }
+    __m512h a_f16x32 = _mm512_cvthf8_ph(a_e4m3x32);
+    __m512h b_f16x32 = _mm512_cvthf8_ph(b_e4m3x32);
+    dot_f32x16 = _mm512_dpph_ps(dot_f32x16, a_f16x32, b_f16x32);
+    a_norm_sq_f32x16 = _mm512_dpph_ps(a_norm_sq_f32x16, a_f16x32, a_f16x32);
+    b_norm_sq_f32x16 = _mm512_dpph_ps(b_norm_sq_f32x16, b_f16x32, b_f16x32);
+    if (n) goto nk_angular_e4m3_diamond_cycle;
+
+    nk_f32_t dot_f32 = nk_reduce_add_f32x16_skylake_(dot_f32x16);
+    nk_f32_t a_norm_sq_f32 = nk_reduce_add_f32x16_skylake_(a_norm_sq_f32x16);
+    nk_f32_t b_norm_sq_f32 = nk_reduce_add_f32x16_skylake_(b_norm_sq_f32x16);
+    *result = nk_angular_normalize_f32_haswell_(dot_f32, a_norm_sq_f32, b_norm_sq_f32);
+}
+
+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) {
+    __m512 a_sq_f32x16 = _mm512_setzero_ps();
+    __m512 b_sq_f32x16 = _mm512_setzero_ps();
+    __m512 ab_f32x16 = _mm512_setzero_ps();
+    __m256i a_e5m2x32, b_e5m2x32;
+
+nk_sqeuclidean_e5m2_diamond_cycle:
+    if (n < 32) {
+        __mmask32 mask = (__mmask32)_bzhi_u32(0xFFFFFFFF, n);
+        a_e5m2x32 = _mm256_maskz_loadu_epi8(mask, a);
+        b_e5m2x32 = _mm256_maskz_loadu_epi8(mask, b);
+        n = 0;
+    }
+    else {
+        a_e5m2x32 = _mm256_loadu_epi8(a);
+        b_e5m2x32 = _mm256_loadu_epi8(b);
+        a += 32, b += 32, n -= 32;
+    }
+    __m512h a_f16x32 = _mm512_cvtbf8_ph(a_e5m2x32);
+    __m512h b_f16x32 = _mm512_cvtbf8_ph(b_e5m2x32);
+    a_sq_f32x16 = _mm512_dpph_ps(a_sq_f32x16, a_f16x32, a_f16x32);
+    b_sq_f32x16 = _mm512_dpph_ps(b_sq_f32x16, b_f16x32, b_f16x32);
+    ab_f32x16 = _mm512_dpph_ps(ab_f32x16, a_f16x32, b_f16x32);
+    if (n) goto nk_sqeuclidean_e5m2_diamond_cycle;
+
+    __m512 sum_sq_f32x16 = _mm512_add_ps(a_sq_f32x16, b_sq_f32x16);
+    *result = nk_reduce_add_f32x16_skylake_(_mm512_fnmadd_ps(_mm512_set1_ps(2.0f), ab_f32x16, sum_sq_f32x16));
+}
+
+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) {
+    nk_sqeuclidean_e5m2_diamond(a, b, n, result);
+    *result = nk_f32_sqrt_haswell(*result);
+}
+
+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) {
+    __m512 dot_f32x16 = _mm512_setzero_ps();
+    __m512 a_norm_sq_f32x16 = _mm512_setzero_ps();
+    __m512 b_norm_sq_f32x16 = _mm512_setzero_ps();
+    __m256i a_e5m2x32, b_e5m2x32;
+
+nk_angular_e5m2_diamond_cycle:
+    if (n < 32) {
+        __mmask32 mask = (__mmask32)_bzhi_u32(0xFFFFFFFF, n);
+        a_e5m2x32 = _mm256_maskz_loadu_epi8(mask, a);
+        b_e5m2x32 = _mm256_maskz_loadu_epi8(mask, b);
+        n = 0;
+    }
+    else {
+        a_e5m2x32 = _mm256_loadu_epi8(a);
+        b_e5m2x32 = _mm256_loadu_epi8(b);
+        a += 32, b += 32, n -= 32;
+    }
+    __m512h a_f16x32 = _mm512_cvtbf8_ph(a_e5m2x32);
+    __m512h b_f16x32 = _mm512_cvtbf8_ph(b_e5m2x32);
+    dot_f32x16 = _mm512_dpph_ps(dot_f32x16, a_f16x32, b_f16x32);
+    a_norm_sq_f32x16 = _mm512_dpph_ps(a_norm_sq_f32x16, a_f16x32, a_f16x32);
+    b_norm_sq_f32x16 = _mm512_dpph_ps(b_norm_sq_f32x16, b_f16x32, b_f16x32);
+    if (n) goto nk_angular_e5m2_diamond_cycle;
+
+    nk_f32_t dot_f32 = nk_reduce_add_f32x16_skylake_(dot_f32x16);
+    nk_f32_t a_norm_sq_f32 = nk_reduce_add_f32x16_skylake_(a_norm_sq_f32x16);
+    nk_f32_t b_norm_sq_f32 = nk_reduce_add_f32x16_skylake_(b_norm_sq_f32x16);
+    *result = nk_angular_normalize_f32_haswell_(dot_f32, a_norm_sq_f32, b_norm_sq_f32);
+}
+
+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) {
+    __m512 a_sq_f32x16 = _mm512_setzero_ps();
+    __m512 b_sq_f32x16 = _mm512_setzero_ps();
+    __m512 ab_f32x16 = _mm512_setzero_ps();
+    __m512h a_f16x32, b_f16x32;
+
+nk_sqeuclidean_f16_diamond_cycle:
+    if (n < 32) {
+        __mmask32 mask = (__mmask32)_bzhi_u32(0xFFFFFFFF, n);
+        a_f16x32 = _mm512_castsi512_ph(_mm512_maskz_loadu_epi16(mask, a));
+        b_f16x32 = _mm512_castsi512_ph(_mm512_maskz_loadu_epi16(mask, b));
+        n = 0;
+    }
+    else {
+        a_f16x32 = _mm512_castsi512_ph(_mm512_loadu_epi16(a));
+        b_f16x32 = _mm512_castsi512_ph(_mm512_loadu_epi16(b));
+        a += 32, b += 32, n -= 32;
+    }
+    a_sq_f32x16 = _mm512_dpph_ps(a_sq_f32x16, a_f16x32, a_f16x32);
+    b_sq_f32x16 = _mm512_dpph_ps(b_sq_f32x16, b_f16x32, b_f16x32);
+    ab_f32x16 = _mm512_dpph_ps(ab_f32x16, a_f16x32, b_f16x32);
+    if (n) goto nk_sqeuclidean_f16_diamond_cycle;
+
+    __m512 sum_sq_f32x16 = _mm512_add_ps(a_sq_f32x16, b_sq_f32x16);
+    *result = nk_reduce_add_f32x16_skylake_(_mm512_fnmadd_ps(_mm512_set1_ps(2.0f), ab_f32x16, sum_sq_f32x16));
+}
+
+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) {
+    nk_sqeuclidean_f16_diamond(a, b, n, result);
+    *result = nk_f32_sqrt_haswell(*result);
+}
+
+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) {
+    __m512 dot_f32x16 = _mm512_setzero_ps();
+    __m512 a_norm_sq_f32x16 = _mm512_setzero_ps();
+    __m512 b_norm_sq_f32x16 = _mm512_setzero_ps();
+    __m512h a_f16x32, b_f16x32;
+
+nk_angular_f16_diamond_cycle:
+    if (n < 32) {
+        __mmask32 mask = (__mmask32)_bzhi_u32(0xFFFFFFFF, n);
+        a_f16x32 = _mm512_castsi512_ph(_mm512_maskz_loadu_epi16(mask, a));
+        b_f16x32 = _mm512_castsi512_ph(_mm512_maskz_loadu_epi16(mask, b));
+        n = 0;
+    }
+    else {
+        a_f16x32 = _mm512_castsi512_ph(_mm512_loadu_epi16(a));
+        b_f16x32 = _mm512_castsi512_ph(_mm512_loadu_epi16(b));
+        a += 32, b += 32, n -= 32;
+    }
+    dot_f32x16 = _mm512_dpph_ps(dot_f32x16, a_f16x32, b_f16x32);
+    a_norm_sq_f32x16 = _mm512_dpph_ps(a_norm_sq_f32x16, a_f16x32, a_f16x32);
+    b_norm_sq_f32x16 = _mm512_dpph_ps(b_norm_sq_f32x16, b_f16x32, b_f16x32);
+    if (n) goto nk_angular_f16_diamond_cycle;
+
+    nk_f32_t dot_f32 = nk_reduce_add_f32x16_skylake_(dot_f32x16);
+    nk_f32_t a_norm_sq_f32 = nk_reduce_add_f32x16_skylake_(a_norm_sq_f32x16);
+    nk_f32_t b_norm_sq_f32 = nk_reduce_add_f32x16_skylake_(b_norm_sq_f32x16);
+    *result = nk_angular_normalize_f32_haswell_(dot_f32, a_norm_sq_f32, b_norm_sq_f32);
+}
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_DIAMOND
+#endif // NK_TARGET_X8664_
+#endif // NK_SPATIAL_DIAMOND_H

package/numkong/include/numkong/spatial/genoa.h

@@ -0,0 +1,222 @@
+/**
+ *  @brief SIMD-accelerated Spatial Similarity Measures for Genoa.
+ *  @file include/numkong/spatial/genoa.h
+ *  @author Ash Vardanian
+ *  @date December 27, 2025
+ *
+ *  @sa include/numkong/spatial.h
+ */
+#ifndef NK_SPATIAL_GENOA_H
+#define NK_SPATIAL_GENOA_H
+
+#if NK_TARGET_X8664_
+#if NK_TARGET_GENOA
+
+#include "numkong/types.h"
+#include "numkong/spatial/haswell.h" // `nk_angular_normalize_f32_haswell_`, `nk_f32_sqrt_haswell`
+#include "numkong/reduce/skylake.h"  // `nk_reduce_add_f32x16_skylake_`
+#include "numkong/cast/icelake.h"    // `nk_e4m3x32_to_bf16x32_icelake_`
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute push(                                                                        \
+    __attribute__((target("avx2,avx512f,avx512vl,avx512bw,avx512dq,avx512bf16,f16c,fma,bmi,bmi2"))), \
+    apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("avx2", "avx512f", "avx512vl", "avx512bw", "avx512dq", "avx512bf16", "f16c", "fma", "bmi", "bmi2")
+#endif
+
+NK_INTERNAL __m512i nk_substract_bf16x32_genoa_(__m512i a_i16, __m512i b_i16) {
+
+    nk_b512_vec_t d, a_f32_even, b_f32_even, d_f32_even, a_f32_odd, b_f32_odd, d_f32_odd;
+
+    // There are several approaches to perform subtraction in `bf16`. The first one is:
+    //
+    //      Perform a couple of casts - each is a bitshift. To convert `bf16` to `f32`,
+    //      expand it to 32-bit integers, then shift the bits by 16 to the left.
+    //      Then subtract as floats, and shift back. During expansion, we will double the space,
+    //      and should use separate registers for top and bottom halves.
+    //      Some compilers don't have `_mm512_extracti32x8_epi32`, so we use `_mm512_extracti64x4_epi64`:
+    //
+    //          a_f32_bot.fvec = _mm512_castsi512_ps(_mm512_slli_epi32(
+    //              _mm512_cvtepu16_epi32(_mm512_castsi512_si256(a_i16)), 16));
+    //          b_f32_bot.fvec = _mm512_castsi512_ps(_mm512_slli_epi32(
+    //              _mm512_cvtepu16_epi32(_mm512_castsi512_si256(b_i16)), 16));
+    //          a_f32_top.fvec =_mm512_castsi512_ps(
+    //              _mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm512_extracti64x4_epi64(a_i16, 1)), 16));
+    //          b_f32_top.fvec =_mm512_castsi512_ps(
+    //              _mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm512_extracti64x4_epi64(b_i16, 1)), 16));
+    //          d_f32_top.fvec = _mm512_sub_ps(a_f32_top.fvec, b_f32_top.fvec);
+    //          d_f32_bot.fvec = _mm512_sub_ps(a_f32_bot.fvec, b_f32_bot.fvec);
+    //          d.ivec = _mm512_castsi256_si512(_mm512_cvtepi32_epi16(
+    //              _mm512_srli_epi32(_mm512_castps_si512(d_f32_bot.fvec), 16)));
+    //          d.ivec = _mm512_inserti64x4(d.ivec, _mm512_cvtepi32_epi16(
+    //              _mm512_srli_epi32(_mm512_castps_si512(d_f32_top.fvec), 16)), 1);
+    //
+    // Instead of using multple shifts and an insertion, we can achieve similar result with fewer expensive
+    // calls to `_mm512_permutex2var_epi16`, or a cheap `_mm512_mask_shuffle_epi8` and blend:
+    //
+    a_f32_odd.zmm = _mm512_and_si512(a_i16, _mm512_set1_epi32(0xFFFF0000));
+    a_f32_even.zmm = _mm512_slli_epi32(a_i16, 16);
+    b_f32_odd.zmm = _mm512_and_si512(b_i16, _mm512_set1_epi32(0xFFFF0000));
+    b_f32_even.zmm = _mm512_slli_epi32(b_i16, 16);
+
+    d_f32_odd.zmm_ps = _mm512_sub_ps(a_f32_odd.zmm_ps, b_f32_odd.zmm_ps);
+    d_f32_even.zmm_ps = _mm512_sub_ps(a_f32_even.zmm_ps, b_f32_even.zmm_ps);
+
+    d_f32_even.zmm = _mm512_srli_epi32(d_f32_even.zmm, 16);
+    d.zmm = _mm512_mask_blend_epi16(0x55555555, d_f32_odd.zmm, d_f32_even.zmm);
+
+    return d.zmm;
+}
+
+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) {
+    __m512 a_sq_f32x16 = _mm512_setzero_ps();
+    __m512 b_sq_f32x16 = _mm512_setzero_ps();
+    __m512 ab_f32x16 = _mm512_setzero_ps();
+    __m512i a_bf16x32, b_bf16x32;
+
+nk_sqeuclidean_bf16_genoa_cycle:
+    if (n < 32) {
+        __mmask32 mask = (__mmask32)_bzhi_u32(0xFFFFFFFF, n);
+        a_bf16x32 = _mm512_maskz_loadu_epi16(mask, a);
+        b_bf16x32 = _mm512_maskz_loadu_epi16(mask, b);
+        n = 0;
+    }
+    else {
+        a_bf16x32 = _mm512_loadu_epi16(a);
+        b_bf16x32 = _mm512_loadu_epi16(b);
+        a += 32, b += 32, n -= 32;
+    }
+    a_sq_f32x16 = _mm512_dpbf16_ps(a_sq_f32x16, nk_m512bh_from_m512i_(a_bf16x32), nk_m512bh_from_m512i_(a_bf16x32));
+    b_sq_f32x16 = _mm512_dpbf16_ps(b_sq_f32x16, nk_m512bh_from_m512i_(b_bf16x32), nk_m512bh_from_m512i_(b_bf16x32));
+    ab_f32x16 = _mm512_dpbf16_ps(ab_f32x16, nk_m512bh_from_m512i_(a_bf16x32), nk_m512bh_from_m512i_(b_bf16x32));
+    if (n) goto nk_sqeuclidean_bf16_genoa_cycle;
+
+    // (a-b)² = a² + b² - 2ab
+    __m512 sum_sq_f32x16 = _mm512_add_ps(a_sq_f32x16, b_sq_f32x16);
+    *result = nk_reduce_add_f32x16_skylake_(_mm512_fnmadd_ps(_mm512_set1_ps(2.0f), ab_f32x16, sum_sq_f32x16));
+}
+
+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) {
+    nk_sqeuclidean_bf16_genoa(a, b, n, result);
+    *result = nk_f32_sqrt_haswell(*result);
+}
+
+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) {
+    __m512 dot_product_f32x16 = _mm512_setzero_ps();
+    __m512 a_norm_sq_f32x16 = _mm512_setzero_ps();
+    __m512 b_norm_sq_f32x16 = _mm512_setzero_ps();
+    __m512i a_bf16x32, b_bf16x32;
+
+nk_angular_bf16_genoa_cycle:
+    if (n < 32) {
+        __mmask32 mask = (__mmask32)_bzhi_u32(0xFFFFFFFF, n);
+        a_bf16x32 = _mm512_maskz_loadu_epi16(mask, a);
+        b_bf16x32 = _mm512_maskz_loadu_epi16(mask, b);
+        n = 0;
+    }
+    else {
+        a_bf16x32 = _mm512_loadu_epi16(a);
+        b_bf16x32 = _mm512_loadu_epi16(b);
+        a += 32, b += 32, n -= 32;
+    }
+    dot_product_f32x16 = _mm512_dpbf16_ps(dot_product_f32x16, nk_m512bh_from_m512i_(a_bf16x32),
+                                          nk_m512bh_from_m512i_(b_bf16x32));
+    a_norm_sq_f32x16 = _mm512_dpbf16_ps(a_norm_sq_f32x16, nk_m512bh_from_m512i_(a_bf16x32),
+                                        nk_m512bh_from_m512i_(a_bf16x32));
+    b_norm_sq_f32x16 = _mm512_dpbf16_ps(b_norm_sq_f32x16, nk_m512bh_from_m512i_(b_bf16x32),
+                                        nk_m512bh_from_m512i_(b_bf16x32));
+    if (n) goto nk_angular_bf16_genoa_cycle;
+
+    nk_f32_t dot_product_f32 = nk_reduce_add_f32x16_skylake_(dot_product_f32x16);
+    nk_f32_t a_norm_sq_f32 = nk_reduce_add_f32x16_skylake_(a_norm_sq_f32x16);
+    nk_f32_t b_norm_sq_f32 = nk_reduce_add_f32x16_skylake_(b_norm_sq_f32x16);
+    *result = nk_angular_normalize_f32_haswell_(dot_product_f32, a_norm_sq_f32, b_norm_sq_f32);
+}
+
+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) {
+    __m512 a_sq_f32x16 = _mm512_setzero_ps();
+    __m512 b_sq_f32x16 = _mm512_setzero_ps();
+    __m512 ab_f32x16 = _mm512_setzero_ps();
+    __m256i a_e5m2x32, b_e5m2x32;
+
+nk_sqeuclidean_e5m2_genoa_cycle:
+    if (n < 32) {
+        __mmask32 mask = (__mmask32)_bzhi_u32(0xFFFFFFFF, n);
+        a_e5m2x32 = _mm256_maskz_loadu_epi8(mask, a);
+        b_e5m2x32 = _mm256_maskz_loadu_epi8(mask, b);
+        n = 0;
+    }
+    else {
+        a_e5m2x32 = _mm256_loadu_epi8(a);
+        b_e5m2x32 = _mm256_loadu_epi8(b);
+        a += 32, b += 32, n -= 32;
+    }
+    __m512i a_bf16x32 = nk_e5m2x32_to_bf16x32_icelake_(a_e5m2x32);
+    __m512i b_bf16x32 = nk_e5m2x32_to_bf16x32_icelake_(b_e5m2x32);
+    a_sq_f32x16 = _mm512_dpbf16_ps(a_sq_f32x16, nk_m512bh_from_m512i_(a_bf16x32), nk_m512bh_from_m512i_(a_bf16x32));
+    b_sq_f32x16 = _mm512_dpbf16_ps(b_sq_f32x16, nk_m512bh_from_m512i_(b_bf16x32), nk_m512bh_from_m512i_(b_bf16x32));
+    ab_f32x16 = _mm512_dpbf16_ps(ab_f32x16, nk_m512bh_from_m512i_(a_bf16x32), nk_m512bh_from_m512i_(b_bf16x32));
+    if (n) goto nk_sqeuclidean_e5m2_genoa_cycle;
+
+    // (a-b)² = a² + b² - 2ab
+    __m512 sum_sq_f32x16 = _mm512_add_ps(a_sq_f32x16, b_sq_f32x16);
+    *result = nk_reduce_add_f32x16_skylake_(_mm512_fnmadd_ps(_mm512_set1_ps(2.0f), ab_f32x16, sum_sq_f32x16));
+}
+
+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) {
+    nk_sqeuclidean_e5m2_genoa(a, b, n, result);
+    *result = nk_f32_sqrt_haswell(*result);
+}
+
+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) {
+    __m512 dot_f32x16 = _mm512_setzero_ps();
+    __m512 a_norm_sq_f32x16 = _mm512_setzero_ps();
+    __m512 b_norm_sq_f32x16 = _mm512_setzero_ps();
+    __m256i a_e5m2x32, b_e5m2x32;
+
+nk_angular_e5m2_genoa_cycle:
+    if (n < 32) {
+        __mmask32 mask = (__mmask32)_bzhi_u32(0xFFFFFFFF, n);
+        a_e5m2x32 = _mm256_maskz_loadu_epi8(mask, a);
+        b_e5m2x32 = _mm256_maskz_loadu_epi8(mask, b);
+        n = 0;
+    }
+    else {
+        a_e5m2x32 = _mm256_loadu_epi8(a);
+        b_e5m2x32 = _mm256_loadu_epi8(b);
+        a += 32, b += 32, n -= 32;
+    }
+    __m512i a_bf16x32 = nk_e5m2x32_to_bf16x32_icelake_(a_e5m2x32);
+    __m512i b_bf16x32 = nk_e5m2x32_to_bf16x32_icelake_(b_e5m2x32);
+    dot_f32x16 = _mm512_dpbf16_ps(dot_f32x16, nk_m512bh_from_m512i_(a_bf16x32), nk_m512bh_from_m512i_(b_bf16x32));
+    a_norm_sq_f32x16 = _mm512_dpbf16_ps(a_norm_sq_f32x16, nk_m512bh_from_m512i_(a_bf16x32),
+                                        nk_m512bh_from_m512i_(a_bf16x32));
+    b_norm_sq_f32x16 = _mm512_dpbf16_ps(b_norm_sq_f32x16, nk_m512bh_from_m512i_(b_bf16x32),
+                                        nk_m512bh_from_m512i_(b_bf16x32));
+    if (n) goto nk_angular_e5m2_genoa_cycle;
+
+    nk_f32_t dot_f32 = nk_reduce_add_f32x16_skylake_(dot_f32x16);
+    nk_f32_t a_norm_sq_f32 = nk_reduce_add_f32x16_skylake_(a_norm_sq_f32x16);
+    nk_f32_t b_norm_sq_f32 = nk_reduce_add_f32x16_skylake_(b_norm_sq_f32x16);
+    *result = nk_angular_normalize_f32_haswell_(dot_f32, a_norm_sq_f32, b_norm_sq_f32);
+}
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_GENOA
+#endif // NK_TARGET_X8664_
+#endif // NK_SPATIAL_GENOA_H