usearch 2.21.3 → 2.25.0

package/numkong/c/dispatch_u32.c

@@ -0,0 +1,118 @@
+/**
+ *  @brief Dispatch Initialization for U32 Data Types.
+ *  @file c/dispatch_u32.c
+ *  @author Ash Vardanian
+ *  @date February 3, 2026
+ */
+#include "dispatch.h"
+
+void nk_dispatch_u32_find_(nk_capability_t v, nk_kernel_kind_t k, nk_kernel_punned_t *m, nk_capability_t *c) {
+    typedef nk_kernel_punned_t m_t;
+#if NK_TARGET_V128RELAXED
+    if (v & nk_cap_v128relaxed_k) switch (k) {
+        case nk_kernel_jaccard_k: *m = (m_t)&nk_jaccard_u32_v128relaxed, *c = nk_cap_v128relaxed_k; return;
+        case nk_kernel_reduce_moments_k:
+            *m = (m_t)&nk_reduce_moments_u32_v128relaxed, *c = nk_cap_v128relaxed_k;
+            return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u32_v128relaxed, *c = nk_cap_v128relaxed_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_SVE2
+    if (v & nk_cap_sve2_k) switch (k) {
+        case nk_kernel_sparse_intersect_k: *m = (m_t)&nk_sparse_intersect_u32_sve2, *c = nk_cap_sve2_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_SVE
+    if (v & nk_cap_sve_k) switch (k) {
+        case nk_kernel_jaccard_k: *m = (m_t)&nk_jaccard_u32_sve, *c = nk_cap_sve_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_NEON
+    if (v & nk_cap_neon_k) switch (k) {
+        case nk_kernel_sparse_intersect_k: *m = (m_t)&nk_sparse_intersect_u32_neon, *c = nk_cap_neon_k; return;
+        case nk_kernel_each_fma_k: *m = (m_t)&nk_each_fma_u32_neon, *c = nk_cap_neon_k; return;
+        case nk_kernel_each_scale_k: *m = (m_t)&nk_each_scale_u32_neon, *c = nk_cap_neon_k; return;
+        case nk_kernel_each_sum_k: *m = (m_t)&nk_each_sum_u32_neon, *c = nk_cap_neon_k; return;
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u32_neon, *c = nk_cap_neon_k; return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u32_neon, *c = nk_cap_neon_k; return;
+        case nk_kernel_jaccard_k: *m = (m_t)&nk_jaccard_u32_neon, *c = nk_cap_neon_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_TURIN
+    if (v & nk_cap_turin_k) switch (k) {
+        case nk_kernel_sparse_intersect_k: *m = (m_t)&nk_sparse_intersect_u32_turin, *c = nk_cap_skylake_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_ICELAKE
+    if (v & nk_cap_icelake_k) switch (k) {
+        case nk_kernel_sparse_intersect_k: *m = (m_t)&nk_sparse_intersect_u32_icelake, *c = nk_cap_skylake_k; return;
+        case nk_kernel_each_sum_k: *m = (m_t)&nk_each_sum_u32_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_jaccard_k: *m = (m_t)&nk_jaccard_u32_icelake, *c = nk_cap_icelake_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_SKYLAKE
+    if (v & nk_cap_skylake_k) switch (k) {
+        case nk_kernel_each_fma_k: *m = (m_t)&nk_each_fma_u32_skylake, *c = nk_cap_skylake_k; return;
+        case nk_kernel_each_scale_k: *m = (m_t)&nk_each_scale_u32_skylake, *c = nk_cap_skylake_k; return;
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u32_skylake, *c = nk_cap_skylake_k; return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u32_skylake, *c = nk_cap_skylake_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_HASWELL
+    if (v & nk_cap_haswell_k) switch (k) {
+        case nk_kernel_each_fma_k: *m = (m_t)&nk_each_fma_u32_haswell, *c = nk_cap_haswell_k; return;
+        case nk_kernel_each_scale_k: *m = (m_t)&nk_each_scale_u32_haswell, *c = nk_cap_haswell_k; return;
+        case nk_kernel_each_sum_k: *m = (m_t)&nk_each_sum_u32_haswell, *c = nk_cap_haswell_k; return;
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u32_haswell, *c = nk_cap_haswell_k; return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u32_haswell, *c = nk_cap_haswell_k; return;
+        case nk_kernel_jaccard_k: *m = (m_t)&nk_jaccard_u32_haswell, *c = nk_cap_haswell_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_RVV
+    if (v & nk_cap_rvv_k) switch (k) {
+        case nk_kernel_jaccard_k: *m = (m_t)&nk_jaccard_u32_rvv, *c = nk_cap_rvv_k; return;
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u32_rvv, *c = nk_cap_rvv_k; return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u32_rvv, *c = nk_cap_rvv_k; return;
+        case nk_kernel_each_fma_k: *m = (m_t)&nk_each_fma_u32_rvv, *c = nk_cap_rvv_k; return;
+        case nk_kernel_each_scale_k: *m = (m_t)&nk_each_scale_u32_rvv, *c = nk_cap_rvv_k; return;
+        case nk_kernel_each_sum_k: *m = (m_t)&nk_each_sum_u32_rvv, *c = nk_cap_rvv_k; return;
+        default: break;
+        }
+#endif
+    if (v & nk_cap_serial_k) switch (k) {
+        case nk_kernel_sparse_intersect_k: *m = (m_t)&nk_sparse_intersect_u32_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_jaccard_k: *m = (m_t)&nk_jaccard_u32_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_each_fma_k: *m = (m_t)&nk_each_fma_u32_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_each_scale_k: *m = (m_t)&nk_each_scale_u32_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_each_sum_k: *m = (m_t)&nk_each_sum_u32_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_each_blend_k: *m = (m_t)&nk_each_blend_u32_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u32_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u32_serial, *c = nk_cap_serial_k; return;
+        default: break;
+        }
+
+    // Error fallback - zero capability signals lookup failure
+    *m = (m_t)nk_error_dense_, *c = 0;
+}
+
+void nk_dispatch_u32_init_(nk_capability_t caps) {
+    nk_implementations_t *t = &nk_dispatch_table;
+    nk_capability_t used;
+
+    nk_dispatch_u32_find_(caps, nk_kernel_jaccard_k, (nk_kernel_punned_t *)&t->jaccard_u32, &used);
+    nk_dispatch_u32_find_(caps, nk_kernel_sparse_intersect_k, (nk_kernel_punned_t *)&t->sparse_intersect_u32, &used);
+    nk_dispatch_u32_find_(caps, nk_kernel_each_scale_k, (nk_kernel_punned_t *)&t->each_scale_u32, &used);
+    nk_dispatch_u32_find_(caps, nk_kernel_each_sum_k, (nk_kernel_punned_t *)&t->each_sum_u32, &used);
+    nk_dispatch_u32_find_(caps, nk_kernel_each_blend_k, (nk_kernel_punned_t *)&t->each_blend_u32, &used);
+    nk_dispatch_u32_find_(caps, nk_kernel_each_fma_k, (nk_kernel_punned_t *)&t->each_fma_u32, &used);
+    nk_dispatch_u32_find_(caps, nk_kernel_reduce_moments_k, (nk_kernel_punned_t *)&t->reduce_moments_u32, &used);
+    nk_dispatch_u32_find_(caps, nk_kernel_reduce_minmax_k, (nk_kernel_punned_t *)&t->reduce_minmax_u32, &used);
+}

package/numkong/c/dispatch_u4.c

@@ -0,0 +1,153 @@
+/**
+ *  @brief Dispatch Initialization for U4 Data Types.
+ *  @file c/dispatch_u4.c
+ *  @author Ash Vardanian
+ *  @date February 3, 2026
+ */
+#include "dispatch.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void nk_dispatch_u4_find_(nk_capability_t v, nk_kernel_kind_t k, nk_kernel_punned_t *m, nk_capability_t *c) {
+    typedef nk_kernel_punned_t m_t;
+#if NK_TARGET_V128RELAXED
+    if (v & nk_cap_v128relaxed_k) switch (k) {
+        case nk_kernel_dot_k: *m = (m_t)&nk_dot_u4_v128relaxed, *c = nk_cap_v128relaxed_k; return;
+        case nk_kernel_dots_packed_size_k:
+            *m = (m_t)&nk_dots_packed_size_u4_v128relaxed, *c = nk_cap_v128relaxed_k;
+            return;
+        case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_u4_v128relaxed, *c = nk_cap_v128relaxed_k; return;
+        case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_u4_v128relaxed, *c = nk_cap_v128relaxed_k; return;
+        case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_u4_v128relaxed, *c = nk_cap_v128relaxed_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_SME
+    if (v & nk_cap_sme_k) switch (k) {
+        case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_u4_sme, *c = nk_cap_sme_k; return;
+        case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_u4_sme, *c = nk_cap_sme_k; return;
+        case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_u4_sme, *c = nk_cap_sme_k; return;
+        case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_u4_sme, *c = nk_cap_sme_k; return;
+        case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_u4_sme, *c = nk_cap_sme_k; return;
+        case nk_kernel_angulars_symmetric_k: *m = (m_t)&nk_angulars_symmetric_u4_sme, *c = nk_cap_sme_k; return;
+        case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_u4_sme, *c = nk_cap_sme_k; return;
+        case nk_kernel_euclideans_symmetric_k: *m = (m_t)&nk_euclideans_symmetric_u4_sme, *c = nk_cap_sme_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_NEONSDOT
+    if (v & nk_cap_neonsdot_k) switch (k) {
+        case nk_kernel_dot_k: *m = (m_t)&nk_dot_u4_neonsdot, *c = nk_cap_neonsdot_k; return;
+        case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_u4_neonsdot, *c = nk_cap_neonsdot_k; return;
+        case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_u4_neonsdot, *c = nk_cap_neonsdot_k; return;
+        case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_u4_neonsdot, *c = nk_cap_neonsdot_k; return;
+        case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_u4_neonsdot, *c = nk_cap_neonsdot_k; return;
+        case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_u4_neonsdot, *c = nk_cap_neonsdot_k; return;
+        case nk_kernel_angulars_symmetric_k:
+            *m = (m_t)&nk_angulars_symmetric_u4_neonsdot, *c = nk_cap_neonsdot_k;
+            return;
+        case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_u4_neonsdot, *c = nk_cap_neonsdot_k; return;
+        case nk_kernel_euclideans_symmetric_k:
+            *m = (m_t)&nk_euclideans_symmetric_u4_neonsdot, *c = nk_cap_neonsdot_k;
+            return;
+        case nk_kernel_angular_k: *m = (m_t)&nk_angular_u4_neonsdot, *c = nk_cap_neonsdot_k; return;
+        case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_u4_neonsdot, *c = nk_cap_neonsdot_k; return;
+        case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_u4_neonsdot, *c = nk_cap_neonsdot_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_ICELAKE
+    if (v & nk_cap_icelake_k) switch (k) {
+        case nk_kernel_dot_k: *m = (m_t)&nk_dot_u4_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_angular_k: *m = (m_t)&nk_angular_u4_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_u4_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_u4_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_u4_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_u4_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_u4_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_u4_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_u4_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_angulars_symmetric_k: *m = (m_t)&nk_angulars_symmetric_u4_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_u4_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_euclideans_symmetric_k:
+            *m = (m_t)&nk_euclideans_symmetric_u4_icelake, *c = nk_cap_icelake_k;
+            return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_SKYLAKE
+    if (v & nk_cap_skylake_k) switch (k) {
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u4_skylake, *c = nk_cap_skylake_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_HASWELL
+    if (v & nk_cap_haswell_k) switch (k) {
+        case nk_kernel_dot_k: *m = (m_t)&nk_dot_u4_haswell, *c = nk_cap_haswell_k; return;
+        case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_u4_haswell, *c = nk_cap_haswell_k; return;
+        case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_u4_haswell, *c = nk_cap_haswell_k; return;
+        case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_u4_haswell, *c = nk_cap_haswell_k; return;
+        case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_u4_haswell, *c = nk_cap_haswell_k; return;
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u4_haswell, *c = nk_cap_haswell_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_RVV
+    if (v & nk_cap_rvv_k) switch (k) {
+        case nk_kernel_dot_k: *m = (m_t)&nk_dot_u4_rvv, *c = nk_cap_rvv_k; return;
+        case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_u4_rvv, *c = nk_cap_rvv_k; return;
+        case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_u4_rvv, *c = nk_cap_rvv_k; return;
+        case nk_kernel_angular_k: *m = (m_t)&nk_angular_u4_rvv, *c = nk_cap_rvv_k; return;
+        default: break;
+        }
+#endif
+    if (v & nk_cap_serial_k) switch (k) {
+        case nk_kernel_dot_k: *m = (m_t)&nk_dot_u4_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_angular_k: *m = (m_t)&nk_angular_u4_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_u4_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_u4_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_u4_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_u4_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_u4_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_u4_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_u4_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_angulars_symmetric_k: *m = (m_t)&nk_angulars_symmetric_u4_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_u4_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_euclideans_symmetric_k:
+            *m = (m_t)&nk_euclideans_symmetric_u4_serial, *c = nk_cap_serial_k;
+            return;
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u4_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u4_serial, *c = nk_cap_serial_k; return;
+        default: break;
+        }
+
+    // Error fallback - zero capability signals lookup failure
+    *m = (m_t)nk_error_dense_, *c = 0;
+}
+
+void nk_dispatch_u4_init_(nk_capability_t caps) {
+    nk_implementations_t *t = &nk_dispatch_table;
+    nk_capability_t used;
+
+    nk_dispatch_u4_find_(caps, nk_kernel_dot_k, (nk_kernel_punned_t *)&t->dot_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_angular_k, (nk_kernel_punned_t *)&t->angular_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_euclidean_k, (nk_kernel_punned_t *)&t->euclidean_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_sqeuclidean_k, (nk_kernel_punned_t *)&t->sqeuclidean_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_reduce_moments_k, (nk_kernel_punned_t *)&t->reduce_moments_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_reduce_minmax_k, (nk_kernel_punned_t *)&t->reduce_minmax_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_dots_packed_size_k, (nk_kernel_punned_t *)&t->dots_packed_size_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_dots_pack_k, (nk_kernel_punned_t *)&t->dots_pack_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_dots_packed_k, (nk_kernel_punned_t *)&t->dots_packed_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_dots_symmetric_k, (nk_kernel_punned_t *)&t->dots_symmetric_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_angulars_packed_k, (nk_kernel_punned_t *)&t->angulars_packed_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_angulars_symmetric_k, (nk_kernel_punned_t *)&t->angulars_symmetric_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_euclideans_packed_k, (nk_kernel_punned_t *)&t->euclideans_packed_u4, &used);
+    nk_dispatch_u4_find_(caps, nk_kernel_euclideans_symmetric_k, (nk_kernel_punned_t *)&t->euclideans_symmetric_u4,
+                         &used);
+}
+
+#ifdef __cplusplus
+}
+#endif

package/numkong/c/dispatch_u64.c

@@ -0,0 +1,102 @@
+/**
+ *  @brief Dispatch Initialization for U64 Data Types.
+ *  @file c/dispatch_u64.c
+ *  @author Ash Vardanian
+ *  @date February 3, 2026
+ */
+#include "dispatch.h"
+
+void nk_dispatch_u64_find_(nk_capability_t v, nk_kernel_kind_t k, nk_kernel_punned_t *m, nk_capability_t *c) {
+    typedef nk_kernel_punned_t m_t;
+#if NK_TARGET_V128RELAXED
+    if (v & nk_cap_v128relaxed_k) switch (k) {
+        case nk_kernel_reduce_moments_k:
+            *m = (m_t)&nk_reduce_moments_u64_v128relaxed, *c = nk_cap_v128relaxed_k;
+            return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u64_v128relaxed, *c = nk_cap_v128relaxed_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_SVE2
+    if (v & nk_cap_sve2_k) switch (k) {
+        case nk_kernel_sparse_intersect_k: *m = (m_t)&nk_sparse_intersect_u64_sve2, *c = nk_cap_sve2_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_NEON
+    if (v & nk_cap_neon_k) switch (k) {
+        case nk_kernel_sparse_intersect_k: *m = (m_t)&nk_sparse_intersect_u64_neon, *c = nk_cap_neon_k; return;
+        case nk_kernel_each_fma_k: *m = (m_t)&nk_each_fma_u64_neon, *c = nk_cap_neon_k; return;
+        case nk_kernel_each_scale_k: *m = (m_t)&nk_each_scale_u64_neon, *c = nk_cap_neon_k; return;
+        case nk_kernel_each_sum_k: *m = (m_t)&nk_each_sum_u64_neon, *c = nk_cap_neon_k; return;
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u64_neon, *c = nk_cap_neon_k; return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u64_neon, *c = nk_cap_neon_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_TURIN
+    if (v & nk_cap_turin_k) switch (k) {
+        case nk_kernel_sparse_intersect_k: *m = (m_t)&nk_sparse_intersect_u64_turin, *c = nk_cap_turin_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_ICELAKE
+    if (v & nk_cap_icelake_k) switch (k) {
+        case nk_kernel_sparse_intersect_k: *m = (m_t)&nk_sparse_intersect_u64_icelake, *c = nk_cap_icelake_k; return;
+        case nk_kernel_each_sum_k: *m = (m_t)&nk_each_sum_u64_icelake, *c = nk_cap_icelake_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_SKYLAKE
+    if (v & nk_cap_skylake_k) switch (k) {
+        case nk_kernel_each_fma_k: *m = (m_t)&nk_each_fma_u64_skylake, *c = nk_cap_skylake_k; return;
+        case nk_kernel_each_scale_k: *m = (m_t)&nk_each_scale_u64_skylake, *c = nk_cap_skylake_k; return;
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u64_skylake, *c = nk_cap_skylake_k; return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u64_skylake, *c = nk_cap_skylake_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_HASWELL
+    if (v & nk_cap_haswell_k) switch (k) {
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u64_haswell, *c = nk_cap_haswell_k; return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u64_haswell, *c = nk_cap_haswell_k; return;
+        default: break;
+        }
+#endif
+#if NK_TARGET_RVV
+    if (v & nk_cap_rvv_k) switch (k) {
+        case nk_kernel_each_fma_k: *m = (m_t)&nk_each_fma_u64_rvv, *c = nk_cap_rvv_k; return;
+        case nk_kernel_each_scale_k: *m = (m_t)&nk_each_scale_u64_rvv, *c = nk_cap_rvv_k; return;
+        case nk_kernel_each_sum_k: *m = (m_t)&nk_each_sum_u64_rvv, *c = nk_cap_rvv_k; return;
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u64_rvv, *c = nk_cap_rvv_k; return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u64_rvv, *c = nk_cap_rvv_k; return;
+        default: break;
+        }
+#endif
+    if (v & nk_cap_serial_k) switch (k) {
+        case nk_kernel_sparse_intersect_k: *m = (m_t)&nk_sparse_intersect_u64_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_each_fma_k: *m = (m_t)&nk_each_fma_u64_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_each_scale_k: *m = (m_t)&nk_each_scale_u64_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_each_sum_k: *m = (m_t)&nk_each_sum_u64_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_each_blend_k: *m = (m_t)&nk_each_blend_u64_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_u64_serial, *c = nk_cap_serial_k; return;
+        case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_u64_serial, *c = nk_cap_serial_k; return;
+        default: break;
+        }
+
+    // Error fallback - zero capability signals lookup failure
+    *m = (m_t)nk_error_dense_, *c = 0;
+}
+
+void nk_dispatch_u64_init_(nk_capability_t caps) {
+    nk_implementations_t *t = &nk_dispatch_table;
+    nk_capability_t used;
+
+    nk_dispatch_u64_find_(caps, nk_kernel_sparse_intersect_k, (nk_kernel_punned_t *)&t->sparse_intersect_u64, &used);
+    nk_dispatch_u64_find_(caps, nk_kernel_each_scale_k, (nk_kernel_punned_t *)&t->each_scale_u64, &used);
+    nk_dispatch_u64_find_(caps, nk_kernel_each_sum_k, (nk_kernel_punned_t *)&t->each_sum_u64, &used);
+    nk_dispatch_u64_find_(caps, nk_kernel_each_blend_k, (nk_kernel_punned_t *)&t->each_blend_u64, &used);
+    nk_dispatch_u64_find_(caps, nk_kernel_each_fma_k, (nk_kernel_punned_t *)&t->each_fma_u64, &used);
+    nk_dispatch_u64_find_(caps, nk_kernel_reduce_moments_k, (nk_kernel_punned_t *)&t->reduce_moments_u64, &used);
+    nk_dispatch_u64_find_(caps, nk_kernel_reduce_minmax_k, (nk_kernel_punned_t *)&t->reduce_minmax_u64, &used);
+}