numkong 7.0.0

package/include/numkong/probability/serial.h

@@ -0,0 +1,169 @@
+/**
+ *  @brief Serial Probability Distribution Similarity Measures.
+ *  @file include/numkong/probability/serial.h
+ *  @author Ash Vardanian
+ *  @date February 6, 2026
+ *
+ *  @sa include/numkong/probability.h
+ */
+#ifndef NK_PROBABILITY_SERIAL_H
+#define NK_PROBABILITY_SERIAL_H
+
+#include "numkong/types.h"
+#include "numkong/cast/serial.h"    // `nk_f16_to_f32_serial`, `nk_bf16_to_f32_serial`, `nk_assign_from_to_`
+#include "numkong/spatial/serial.h" // `nk_f32_sqrt_serial`, `nk_f64_sqrt_serial`
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#define nk_define_kld_(input_type, accumulator_type, output_type, load_and_convert, epsilon, compute_log)   \
+    NK_PUBLIC void nk_kld_##input_type##_serial(nk_##input_type##_t const *a, nk_##input_type##_t const *b, \
+                                                nk_size_t n, output_type *result) {                         \
+        nk_##accumulator_type##_t d = 0, ai, bi;                                                            \
+        for (nk_size_t i = 0; i != n; ++i) {                                                                \
+            load_and_convert(a + i, &ai);                                                                   \
+            load_and_convert(b + i, &bi);                                                                   \
+            d += ai * compute_log((ai + epsilon) / (bi + epsilon));                                         \
+        }                                                                                                   \
+        *result = (output_type)d;                                                                           \
+    }
+
+#define nk_define_jsd_(input_type, accumulator_type, output_type, load_and_convert, epsilon, compute_log,   \
+                       compute_sqrt)                                                                        \
+    NK_PUBLIC void nk_jsd_##input_type##_serial(nk_##input_type##_t const *a, nk_##input_type##_t const *b, \
+                                                nk_size_t n, output_type *result) {                         \
+        nk_##accumulator_type##_t d = 0, ai, bi;                                                            \
+        for (nk_size_t i = 0; i != n; ++i) {                                                                \
+            load_and_convert(a + i, &ai);                                                                   \
+            load_and_convert(b + i, &bi);                                                                   \
+            nk_##accumulator_type##_t mi = (ai + bi) / 2;                                                   \
+            d += ai * compute_log((ai + epsilon) / (mi + epsilon));                                         \
+            d += bi * compute_log((bi + epsilon) / (mi + epsilon));                                         \
+        }                                                                                                   \
+        output_type d_half = ((output_type)d / 2);                                                          \
+        *result = d_half > 0 ? compute_sqrt(d_half) : 0;                                                    \
+    }
+
+/**
+ *  @brief  Computes `log(x)` for any positive float using IEEE 754 bit extraction
+ *          and a fast-converging series expansion.
+ *
+ *  Exploits the IEEE 754 representation to extract the exponent and mantissa:
+ *  `log(x) = log(2) * exponent + log(mantissa)`. The mantissa is reduced to the
+ *  range `[√2/2, √2]` for optimal convergence. Uses the transformation
+ *  `u = (m-1)/(m+1)` which converges much faster than the classic Mercator series,
+ *  since `u` is bounded to approximately `[-0.17, 0.17]` after range reduction.
+ *
+ *  Maximum relative error is approximately 0.00001% across all positive floats,
+ *  roughly 300,000x more accurate than the 3-term Mercator series (which also
+ *  only converges for inputs in `(0, 2)`).
+ *
+ *  https://en.wikipedia.org/wiki/Logarithm#Power_series
+ */
+NK_INTERNAL nk_f32_t nk_f32_log_serial_(nk_f32_t x) {
+    nk_fui32_t conv;
+    conv.f = x;
+    int exp = ((conv.u >> 23) & 0xFF) - 127;
+    conv.u = (conv.u & 0x007FFFFF) | 0x3F800000; // mantissa ∈ [1, 2)
+    nk_f32_t m = conv.f;
+    // Range reduction: if m > √2, halve it and increment exponent
+    if (m > 1.41421356f) m *= 0.5f, exp++;
+    // Use (m-1)/(m+1) transformation for faster convergence
+    nk_f32_t u = (m - 1.0f) / (m + 1.0f);
+    nk_f32_t u2 = u * u;
+    // log(m) = 2 × (u + u³/3 + u⁵/5 + u⁷/7)
+    nk_f32_t log_m = 2.0f * u * (1.0f + u2 * (0.3333333333f + u2 * (0.2f + u2 * 0.142857143f)));
+    return (nk_f32_t)exp * 0.6931471805599453f + log_m;
+}
+
+/**
+ *  @brief  Computes `log(x)` for any positive double using IEEE 754 bit extraction
+ *          and a fast-converging series expansion.
+ *
+ *  Exploits the IEEE 754 representation to extract the 11-bit exponent and 52-bit mantissa:
+ *  `log(x) = log(2) * exponent + log(mantissa)`. The mantissa is reduced to the
+ *  range `[√2/2, √2]` for optimal convergence. Uses the transformation
+ *  `u = (m-1)/(m+1)` which converges much faster than the classic Mercator series,
+ *  since `u` is bounded to approximately `[-0.17, 0.17]` after range reduction.
+ *
+ *  Uses more series terms than the f32 version to achieve near-full f64 precision,
+ *  with maximum relative error approximately 0.00000000000001% across all positive doubles.
+ *
+ *  https://en.wikipedia.org/wiki/Logarithm#Power_series
+ */
+NK_INTERNAL nk_f64_t nk_f64_log_serial_(nk_f64_t x) {
+    nk_fui64_t conv;
+    conv.f = x;
+    int exp = ((conv.u >> 52) & 0x7FF) - 1023;
+    conv.u = (conv.u & 0x000FFFFFFFFFFFFFULL) | 0x3FF0000000000000ULL; // mantissa ∈ [1, 2)
+    nk_f64_t m = conv.f;
+    // Range reduction: if m > √2, halve it and increment exponent
+    if (m > 1.4142135623730950488) m *= 0.5, exp++;
+    // Use (m-1)/(m+1) transformation for faster convergence
+    nk_f64_t u = (m - 1.0) / (m + 1.0);
+    nk_f64_t u2 = u * u;
+    // 14-term Horner: P(u²) = 1 + u²/3 + u⁴/5 + ... + u²⁶/27, matching SIMD
+    nk_f64_t poly = 1.0 / 27.0;
+    poly = u2 * poly + 1.0 / 25.0;
+    poly = u2 * poly + 1.0 / 23.0;
+    poly = u2 * poly + 1.0 / 21.0;
+    poly = u2 * poly + 1.0 / 19.0;
+    poly = u2 * poly + 1.0 / 17.0;
+    poly = u2 * poly + 1.0 / 15.0;
+    poly = u2 * poly + 1.0 / 13.0;
+    poly = u2 * poly + 1.0 / 11.0;
+    poly = u2 * poly + 1.0 / 9.0;
+    poly = u2 * poly + 1.0 / 7.0;
+    poly = u2 * poly + 1.0 / 5.0;
+    poly = u2 * poly + 1.0 / 3.0;
+    poly = u2 * poly + 1.0;
+    return (nk_f64_t)exp * 0.6931471805599453 + 2.0 * u * poly;
+}
+
+nk_define_kld_(f32, f64, nk_f64_t, nk_assign_from_to_, NK_F32_DIVISION_EPSILON, nk_f32_log_serial_)
+nk_define_jsd_(f32, f64, nk_f64_t, nk_assign_from_to_, NK_F32_DIVISION_EPSILON, nk_f32_log_serial_, nk_f64_sqrt_serial)
+
+nk_define_kld_(f16, f32, nk_f32_t, nk_f16_to_f32_serial, NK_F32_DIVISION_EPSILON, nk_f32_log_serial_)
+nk_define_jsd_(f16, f32, nk_f32_t, nk_f16_to_f32_serial, NK_F32_DIVISION_EPSILON, nk_f32_log_serial_,
+               nk_f32_sqrt_serial)
+
+nk_define_kld_(bf16, f32, nk_f32_t, nk_bf16_to_f32_serial, NK_F32_DIVISION_EPSILON, nk_f32_log_serial_)
+nk_define_jsd_(bf16, f32, nk_f32_t, nk_bf16_to_f32_serial, NK_F32_DIVISION_EPSILON, nk_f32_log_serial_,
+               nk_f32_sqrt_serial)
+
+NK_PUBLIC void nk_kld_f64_serial(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result) {
+    nk_f64_t sum = 0, compensation = 0;
+    for (nk_size_t i = 0; i != n; ++i) {
+        nk_f64_t ai = a[i], bi = b[i];
+        nk_f64_t term = ai * nk_f64_log_serial_((ai + NK_F64_DIVISION_EPSILON) / (bi + NK_F64_DIVISION_EPSILON));
+        nk_f64_t t = sum + term;
+        compensation += (nk_f64_abs_(sum) >= nk_f64_abs_(term)) ? ((sum - t) + term) : ((term - t) + sum);
+        sum = t;
+    }
+    *result = sum + compensation;
+}
+
+NK_PUBLIC void nk_jsd_f64_serial(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result) {
+    nk_f64_t sum = 0, compensation = 0;
+    for (nk_size_t i = 0; i != n; ++i) {
+        nk_f64_t ai = a[i], bi = b[i];
+        nk_f64_t mi = (ai + bi) / 2;
+        nk_f64_t term_a = ai * nk_f64_log_serial_((ai + NK_F64_DIVISION_EPSILON) / (mi + NK_F64_DIVISION_EPSILON));
+        nk_f64_t t = sum + term_a;
+        compensation += (nk_f64_abs_(sum) >= nk_f64_abs_(term_a)) ? ((sum - t) + term_a) : ((term_a - t) + sum);
+        sum = t;
+        nk_f64_t term_b = bi * nk_f64_log_serial_((bi + NK_F64_DIVISION_EPSILON) / (mi + NK_F64_DIVISION_EPSILON));
+        t = sum + term_b;
+        compensation += (nk_f64_abs_(sum) >= nk_f64_abs_(term_b)) ? ((sum - t) + term_b) : ((term_b - t) + sum);
+        sum = t;
+    }
+    nk_f64_t d_half = (sum + compensation) / 2;
+    *result = d_half > 0 ? nk_f64_sqrt_serial(d_half) : 0;
+}
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_PROBABILITY_SERIAL_H

package/include/numkong/probability/skylake.h

@@ -0,0 +1,324 @@
+/**
+ *  @brief Skylake-accelerated Probability Distribution Similarity Measures.
+ *  @file include/numkong/probability/skylake.h
+ *  @author Ash Vardanian
+ *  @date February 6, 2026
+ *
+ *  @sa include/numkong/probability.h
+ */
+#ifndef NK_PROBABILITY_SKYLAKE_H
+#define NK_PROBABILITY_SKYLAKE_H
+
+#if NK_TARGET_X86_
+#if NK_TARGET_SKYLAKE
+
+#include "numkong/types.h"
+#include "numkong/spatial/haswell.h" // `nk_f32_sqrt_haswell`, `nk_f64_sqrt_haswell`
+#include "numkong/spatial/skylake.h" // `nk_f32_sqrt_skylake`, `nk_f64_sqrt_skylake`
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((target("avx2,avx512f,avx512vl,avx512bw,avx512dq,f16c,fma,bmi,bmi2"))), \
+                             apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("avx2", "avx512f", "avx512vl", "avx512bw", "avx512dq", "f16c", "fma", "bmi", "bmi2")
+#endif
+
+NK_INTERNAL __m512 nk_log2_f32x16_skylake_(__m512 x) {
+    // Extract the exponent and mantissa: x = 2^exp × m, m ∈ [1, 2)
+    __m512 one_f32x16 = _mm512_set1_ps(1.0f);
+    __m512 exponent_f32x16 = _mm512_getexp_ps(x);
+    __m512 mantissa_f32x16 = _mm512_getmant_ps(x, _MM_MANT_NORM_1_2, _MM_MANT_SIGN_src);
+
+    // Compute log2(m) using the s-series: s = (m-1)/(m+1), s ∈ [0, 1/3] for m ∈ [1, 2)
+    // log2(m) = (2/ln2) × s × (1 + s²/3 + s⁴/5 + s⁶/7 + s⁸/9)
+    __m512 s_f32x16 = _mm512_div_ps(_mm512_sub_ps(mantissa_f32x16, one_f32x16),
+                                    _mm512_add_ps(mantissa_f32x16, one_f32x16));
+    __m512 s2_f32x16 = _mm512_mul_ps(s_f32x16, s_f32x16);
+    __m512 series_f32x16 = _mm512_set1_ps(0.111111111f);                                     // 1/9
+    series_f32x16 = _mm512_fmadd_ps(series_f32x16, s2_f32x16, _mm512_set1_ps(0.142857143f)); // 1/7
+    series_f32x16 = _mm512_fmadd_ps(series_f32x16, s2_f32x16, _mm512_set1_ps(0.2f));         // 1/5
+    series_f32x16 = _mm512_fmadd_ps(series_f32x16, s2_f32x16, _mm512_set1_ps(0.333333333f)); // 1/3
+    series_f32x16 = _mm512_fmadd_ps(series_f32x16, s2_f32x16, one_f32x16);                   // 1
+    // log2(m) = (2/ln2) × s × series
+    __m512 log2m_f32x16 = _mm512_mul_ps(_mm512_set1_ps(2.885390081777927f), _mm512_mul_ps(s_f32x16, series_f32x16));
+    return _mm512_add_ps(log2m_f32x16, exponent_f32x16);
+}
+
+NK_PUBLIC void nk_kld_f32_skylake(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result) {
+    __m512 sum_f32x16 = _mm512_setzero_ps();
+    nk_f32_t epsilon = NK_F32_DIVISION_EPSILON;
+    __m512 epsilon_f32x16 = _mm512_set1_ps(epsilon);
+    __m512 a_f32x16, b_f32x16;
+
+nk_kld_f32_skylake_cycle:
+    if (n < 16) {
+        __mmask16 mask = (__mmask16)_bzhi_u32(0xFFFF, n);
+        a_f32x16 = _mm512_maskz_loadu_ps(mask, a);
+        b_f32x16 = _mm512_maskz_loadu_ps(mask, b);
+        n = 0;
+    }
+    else {
+        a_f32x16 = _mm512_loadu_ps(a);
+        b_f32x16 = _mm512_loadu_ps(b);
+        a += 16, b += 16, n -= 16;
+    }
+    __m512 ratio_f32x16 = _mm512_div_ps(_mm512_add_ps(a_f32x16, epsilon_f32x16),
+                                        _mm512_add_ps(b_f32x16, epsilon_f32x16));
+    __m512 log_ratio_f32x16 = nk_log2_f32x16_skylake_(ratio_f32x16);
+    sum_f32x16 = _mm512_fmadd_ps(a_f32x16, log_ratio_f32x16, sum_f32x16);
+    if (n) goto nk_kld_f32_skylake_cycle;
+
+    __m256 lower_f32x8 = _mm512_castps512_ps256(sum_f32x16);
+    __m256 upper_f32x8 = _mm512_extractf32x8_ps(sum_f32x16, 1);
+    nk_f64_t sum = _mm512_reduce_add_pd(_mm512_cvtps_pd(lower_f32x8)) +
+                   _mm512_reduce_add_pd(_mm512_cvtps_pd(upper_f32x8));
+    nk_f64_t log2_normalizer = 0.6931471805599453;
+    *result = sum * log2_normalizer;
+}
+
+NK_PUBLIC void nk_jsd_f32_skylake(nk_f32_t const *a, nk_f32_t const *b, nk_size_t n, nk_f64_t *result) {
+    __m512 sum_f32x16 = _mm512_setzero_ps();
+    nk_f32_t epsilon = NK_F32_DIVISION_EPSILON;
+    __m512 epsilon_f32x16 = _mm512_set1_ps(epsilon);
+    __m512 a_f32x16, b_f32x16;
+
+nk_jsd_f32_skylake_cycle:
+    if (n < 16) {
+        __mmask16 mask = (__mmask16)_bzhi_u32(0xFFFF, n);
+        a_f32x16 = _mm512_maskz_loadu_ps(mask, a);
+        b_f32x16 = _mm512_maskz_loadu_ps(mask, b);
+        n = 0;
+    }
+    else {
+        a_f32x16 = _mm512_loadu_ps(a);
+        b_f32x16 = _mm512_loadu_ps(b);
+        a += 16, b += 16, n -= 16;
+    }
+    __m512 mean_f32x16 = _mm512_mul_ps(_mm512_add_ps(a_f32x16, b_f32x16), _mm512_set1_ps(0.5f));
+    __mmask16 nonzero_mask_a = _mm512_cmp_ps_mask(a_f32x16, epsilon_f32x16, _CMP_GE_OQ);
+    __mmask16 nonzero_mask_b = _mm512_cmp_ps_mask(b_f32x16, epsilon_f32x16, _CMP_GE_OQ);
+    __mmask16 nonzero_mask = nonzero_mask_a & nonzero_mask_b;
+    __m512 mean_with_epsilon_f32x16 = _mm512_add_ps(mean_f32x16, epsilon_f32x16);
+    __m512 ratio_a_f32x16 = _mm512_div_ps(_mm512_add_ps(a_f32x16, epsilon_f32x16), mean_with_epsilon_f32x16);
+    __m512 ratio_b_f32x16 = _mm512_div_ps(_mm512_add_ps(b_f32x16, epsilon_f32x16), mean_with_epsilon_f32x16);
+    __m512 log_ratio_a_f32x16 = nk_log2_f32x16_skylake_(ratio_a_f32x16);
+    __m512 log_ratio_b_f32x16 = nk_log2_f32x16_skylake_(ratio_b_f32x16);
+    __m512 contribution_a_f32x16 = _mm512_maskz_mul_ps(nonzero_mask, a_f32x16, log_ratio_a_f32x16);
+    __m512 contribution_b_f32x16 = _mm512_maskz_mul_ps(nonzero_mask, b_f32x16, log_ratio_b_f32x16);
+    sum_f32x16 = _mm512_add_ps(sum_f32x16, _mm512_add_ps(contribution_a_f32x16, contribution_b_f32x16));
+    if (n) goto nk_jsd_f32_skylake_cycle;
+
+    __m256 lower_f32x8 = _mm512_castps512_ps256(sum_f32x16);
+    __m256 upper_f32x8 = _mm512_extractf32x8_ps(sum_f32x16, 1);
+    nk_f64_t sum = (_mm512_reduce_add_pd(_mm512_cvtps_pd(lower_f32x8)) +
+                    _mm512_reduce_add_pd(_mm512_cvtps_pd(upper_f32x8))) *
+                   0.6931471805599453 / 2.0;
+    nk_f64_t log2_normalizer = 0.6931471805599453;
+    nk_unused_(log2_normalizer);
+    *result = sum > 0 ? nk_f64_sqrt_haswell(sum) : 0;
+}
+
+NK_INTERNAL __m512d nk_log2_f64x8_skylake_(__m512d x) {
+    // Extract the exponent and mantissa: x = 2^exp × m, m ∈ [1, 2)
+    __m512d one_f64x8 = _mm512_set1_pd(1.0);
+    __m512d two_f64x8 = _mm512_set1_pd(2.0);
+    __m512d exponent_f64x8 = _mm512_getexp_pd(x);
+    __m512d mantissa_f64x8 = _mm512_getmant_pd(x, _MM_MANT_NORM_1_2, _MM_MANT_SIGN_src);
+
+    // Compute log2(m) using the s-series: s = (m-1)/(m+1), s ∈ [0, 1/3] for m ∈ [1, 2)
+    // ln(m) = 2 × s × (1 + s²/3 + s⁴/5 + s⁶/7 + ...) converges fast since s² ≤ 1/9
+    // log2(m) = ln(m) × log2(e)
+    __m512d s_f64x8 = _mm512_div_pd(_mm512_sub_pd(mantissa_f64x8, one_f64x8), _mm512_add_pd(mantissa_f64x8, one_f64x8));
+    __m512d s2_f64x8 = _mm512_mul_pd(s_f64x8, s_f64x8);
+
+    // Polynomial P(s²) = 1 + s²/3 + s⁴/5 + ... using Horner's method
+    // 14 terms (k=0..13) achieves ~1 ULP accuracy for f64
+    __m512d poly_f64x8 = _mm512_set1_pd(1.0 / 27.0); // 1/(2*13+1)
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0 / 25.0));
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0 / 23.0));
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0 / 21.0));
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0 / 19.0));
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0 / 17.0));
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0 / 15.0));
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0 / 13.0));
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0 / 11.0));
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0 / 9.0));
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0 / 7.0));
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0 / 5.0));
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0 / 3.0));
+    poly_f64x8 = _mm512_fmadd_pd(s2_f64x8, poly_f64x8, _mm512_set1_pd(1.0));
+
+    // ln(m) = 2 × s × P(s²), then log2(m) = ln(m) × log2(e)
+    __m512d ln_m_f64x8 = _mm512_mul_pd(_mm512_mul_pd(two_f64x8, s_f64x8), poly_f64x8);
+    __m512d log2e_f64x8 = _mm512_set1_pd(1.4426950408889634); // 1/ln(2)
+    __m512d log2_m_f64x8 = _mm512_mul_pd(ln_m_f64x8, log2e_f64x8);
+
+    // log2(x) = exponent + log2(m)
+    return _mm512_add_pd(exponent_f64x8, log2_m_f64x8);
+}
+
+NK_PUBLIC void nk_kld_f64_skylake(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result) {
+    __m512d sum_f64x8 = _mm512_setzero_pd();
+    __m512d compensation_f64x8 = _mm512_setzero_pd();
+    nk_f64_t epsilon = NK_F64_DIVISION_EPSILON;
+    __m512d epsilon_f64x8 = _mm512_set1_pd(epsilon);
+    __m512d a_f64x8, b_f64x8;
+
+nk_kld_f64_skylake_cycle:
+    if (n < 8) {
+        __mmask8 mask = (__mmask8)_bzhi_u32(0xFF, n);
+        a_f64x8 = _mm512_maskz_loadu_pd(mask, a);
+        b_f64x8 = _mm512_maskz_loadu_pd(mask, b);
+        n = 0;
+    }
+    else {
+        a_f64x8 = _mm512_loadu_pd(a);
+        b_f64x8 = _mm512_loadu_pd(b);
+        a += 8, b += 8, n -= 8;
+    }
+    __m512d ratio_f64x8 = _mm512_div_pd(_mm512_add_pd(a_f64x8, epsilon_f64x8), _mm512_add_pd(b_f64x8, epsilon_f64x8));
+    __m512d log_ratio_f64x8 = nk_log2_f64x8_skylake_(ratio_f64x8);
+    __m512d contribution_f64x8 = _mm512_mul_pd(a_f64x8, log_ratio_f64x8);
+    // Kahan compensated summation
+    __m512d compensated_f64x8 = _mm512_sub_pd(contribution_f64x8, compensation_f64x8);
+    __m512d tentative_f64x8 = _mm512_add_pd(sum_f64x8, compensated_f64x8);
+    compensation_f64x8 = _mm512_sub_pd(_mm512_sub_pd(tentative_f64x8, sum_f64x8), compensated_f64x8);
+    sum_f64x8 = tentative_f64x8;
+    if (n) goto nk_kld_f64_skylake_cycle;
+
+    nk_f64_t log2_normalizer = 0.6931471805599453;
+    *result = _mm512_reduce_add_pd(sum_f64x8) * log2_normalizer;
+}
+
+NK_PUBLIC void nk_jsd_f64_skylake(nk_f64_t const *a, nk_f64_t const *b, nk_size_t n, nk_f64_t *result) {
+    __m512d sum_f64x8 = _mm512_setzero_pd();
+    __m512d compensation_f64x8 = _mm512_setzero_pd();
+    nk_f64_t epsilon = NK_F64_DIVISION_EPSILON;
+    __m512d epsilon_f64x8 = _mm512_set1_pd(epsilon);
+    __m512d a_f64x8, b_f64x8;
+
+nk_jsd_f64_skylake_cycle:
+    if (n < 8) {
+        __mmask8 mask = (__mmask8)_bzhi_u32(0xFF, n);
+        a_f64x8 = _mm512_maskz_loadu_pd(mask, a);
+        b_f64x8 = _mm512_maskz_loadu_pd(mask, b);
+        n = 0;
+    }
+    else {
+        a_f64x8 = _mm512_loadu_pd(a);
+        b_f64x8 = _mm512_loadu_pd(b);
+        a += 8, b += 8, n -= 8;
+    }
+    __m512d mean_f64x8 = _mm512_mul_pd(_mm512_add_pd(a_f64x8, b_f64x8), _mm512_set1_pd(0.5));
+    __mmask8 nonzero_mask_a = _mm512_cmp_pd_mask(a_f64x8, epsilon_f64x8, _CMP_GE_OQ);
+    __mmask8 nonzero_mask_b = _mm512_cmp_pd_mask(b_f64x8, epsilon_f64x8, _CMP_GE_OQ);
+    __mmask8 nonzero_mask = nonzero_mask_a & nonzero_mask_b;
+    __m512d mean_with_epsilon_f64x8 = _mm512_add_pd(mean_f64x8, epsilon_f64x8);
+    __m512d ratio_a_f64x8 = _mm512_div_pd(_mm512_add_pd(a_f64x8, epsilon_f64x8), mean_with_epsilon_f64x8);
+    __m512d ratio_b_f64x8 = _mm512_div_pd(_mm512_add_pd(b_f64x8, epsilon_f64x8), mean_with_epsilon_f64x8);
+    __m512d log_ratio_a_f64x8 = nk_log2_f64x8_skylake_(ratio_a_f64x8);
+    __m512d log_ratio_b_f64x8 = nk_log2_f64x8_skylake_(ratio_b_f64x8);
+    __m512d contribution_a_f64x8 = _mm512_maskz_mul_pd(nonzero_mask, a_f64x8, log_ratio_a_f64x8);
+    __m512d contribution_b_f64x8 = _mm512_maskz_mul_pd(nonzero_mask, b_f64x8, log_ratio_b_f64x8);
+    // Kahan compensated summation for contribution a
+    __m512d compensated_a_f64x8 = _mm512_sub_pd(contribution_a_f64x8, compensation_f64x8);
+    __m512d tentative_a_f64x8 = _mm512_add_pd(sum_f64x8, compensated_a_f64x8);
+    compensation_f64x8 = _mm512_sub_pd(_mm512_sub_pd(tentative_a_f64x8, sum_f64x8), compensated_a_f64x8);
+    sum_f64x8 = tentative_a_f64x8;
+    // Kahan compensated summation for contribution b
+    __m512d compensated_b_f64x8 = _mm512_sub_pd(contribution_b_f64x8, compensation_f64x8);
+    __m512d tentative_b_f64x8 = _mm512_add_pd(sum_f64x8, compensated_b_f64x8);
+    compensation_f64x8 = _mm512_sub_pd(_mm512_sub_pd(tentative_b_f64x8, sum_f64x8), compensated_b_f64x8);
+    sum_f64x8 = tentative_b_f64x8;
+    if (n) goto nk_jsd_f64_skylake_cycle;
+
+    nk_f64_t log2_normalizer = 0.6931471805599453;
+    nk_f64_t sum = _mm512_reduce_add_pd(sum_f64x8);
+    sum *= log2_normalizer / 2;
+    *result = sum > 0 ? nk_f64_sqrt_haswell(sum) : 0;
+}
+
+NK_PUBLIC void nk_kld_f16_skylake(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result) {
+    __m512 sum_f32x16 = _mm512_setzero_ps();
+    __m512 epsilon_f32x16 = _mm512_set1_ps(NK_F32_DIVISION_EPSILON);
+    __m512 a_f32x16, b_f32x16;
+
+nk_kld_f16_skylake_cycle:
+    if (n < 16) {
+        __mmask16 mask = (__mmask16)_bzhi_u32(0xFFFF, n);
+        a_f32x16 = _mm512_cvtph_ps(_mm256_maskz_loadu_epi16(mask, a));
+        b_f32x16 = _mm512_cvtph_ps(_mm256_maskz_loadu_epi16(mask, b));
+        n = 0;
+    }
+    else {
+        a_f32x16 = _mm512_cvtph_ps(_mm256_loadu_si256((__m256i const *)a));
+        b_f32x16 = _mm512_cvtph_ps(_mm256_loadu_si256((__m256i const *)b));
+        a += 16, b += 16, n -= 16;
+    }
+    __m512 ratio_f32x16 = _mm512_div_ps(_mm512_add_ps(a_f32x16, epsilon_f32x16),
+                                        _mm512_add_ps(b_f32x16, epsilon_f32x16));
+    __m512 log_ratio_f32x16 = nk_log2_f32x16_skylake_(ratio_f32x16);
+    __m512 contribution_f32x16 = _mm512_mul_ps(a_f32x16, log_ratio_f32x16);
+    sum_f32x16 = _mm512_add_ps(sum_f32x16, contribution_f32x16);
+    if (n) goto nk_kld_f16_skylake_cycle;
+
+    nk_f32_t log2_normalizer = 0.6931471805599453f;
+    *result = _mm512_reduce_add_ps(sum_f32x16) * log2_normalizer;
+}
+
+NK_PUBLIC void nk_jsd_f16_skylake(nk_f16_t const *a, nk_f16_t const *b, nk_size_t n, nk_f32_t *result) {
+    __m512 sum_a_f32x16 = _mm512_setzero_ps();
+    __m512 sum_b_f32x16 = _mm512_setzero_ps();
+    __m512 epsilon_f32x16 = _mm512_set1_ps(NK_F32_DIVISION_EPSILON);
+    __m512 half_f32x16 = _mm512_set1_ps(0.5f);
+    __m512 a_f32x16, b_f32x16;
+
+nk_jsd_f16_skylake_cycle:
+    if (n < 16) {
+        __mmask16 mask = (__mmask16)_bzhi_u32(0xFFFF, n);
+        a_f32x16 = _mm512_cvtph_ps(_mm256_maskz_loadu_epi16(mask, a));
+        b_f32x16 = _mm512_cvtph_ps(_mm256_maskz_loadu_epi16(mask, b));
+        n = 0;
+    }
+    else {
+        a_f32x16 = _mm512_cvtph_ps(_mm256_loadu_si256((__m256i const *)a));
+        b_f32x16 = _mm512_cvtph_ps(_mm256_loadu_si256((__m256i const *)b));
+        a += 16, b += 16, n -= 16;
+    }
+    __m512 mean_f32x16 = _mm512_mul_ps(_mm512_add_ps(a_f32x16, b_f32x16), half_f32x16);
+    __mmask16 nonzero_mask_a = _mm512_cmp_ps_mask(a_f32x16, epsilon_f32x16, _CMP_GE_OQ);
+    __mmask16 nonzero_mask_b = _mm512_cmp_ps_mask(b_f32x16, epsilon_f32x16, _CMP_GE_OQ);
+    __mmask16 nonzero_mask = nonzero_mask_a & nonzero_mask_b;
+    __m512 mean_with_epsilon_f32x16 = _mm512_add_ps(mean_f32x16, epsilon_f32x16);
+    __m512 ratio_a_f32x16 = _mm512_div_ps(_mm512_add_ps(a_f32x16, epsilon_f32x16), mean_with_epsilon_f32x16);
+    __m512 ratio_b_f32x16 = _mm512_div_ps(_mm512_add_ps(b_f32x16, epsilon_f32x16), mean_with_epsilon_f32x16);
+    __m512 log_ratio_a_f32x16 = nk_log2_f32x16_skylake_(ratio_a_f32x16);
+    __m512 log_ratio_b_f32x16 = nk_log2_f32x16_skylake_(ratio_b_f32x16);
+    sum_a_f32x16 = _mm512_mask3_fmadd_ps(a_f32x16, log_ratio_a_f32x16, sum_a_f32x16, nonzero_mask);
+    sum_b_f32x16 = _mm512_mask3_fmadd_ps(b_f32x16, log_ratio_b_f32x16, sum_b_f32x16, nonzero_mask);
+    if (n) goto nk_jsd_f16_skylake_cycle;
+
+    nk_f32_t log2_normalizer = 0.6931471805599453f;
+    nk_f32_t sum = _mm512_reduce_add_ps(_mm512_add_ps(sum_a_f32x16, sum_b_f32x16));
+    sum *= log2_normalizer / 2;
+    *result = sum > 0 ? nk_f32_sqrt_haswell(sum) : 0;
+}
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_SKYLAKE
+#endif // NK_TARGET_X86_
+#endif // NK_PROBABILITY_SKYLAKE_H