usearch 2.21.4 → 2.25.0

package/numkong/include/numkong/set/powervsx.h

@@ -0,0 +1,326 @@
+/**
+ *  @brief SIMD-accelerated Set Similarity Measures for Power ISA VSX.
+ *  @file include/numkong/set/powervsx.h
+ *  @author Ash Vardanian
+ *  @date March 23, 2026
+ *
+ *  @sa include/numkong/set.h
+ *
+ *  @section set_powervsx_instructions Power9 VSX Set Instructions
+ *
+ *  Key Power9 VSX instructions for binary/bitwise operations:
+ *
+ *      Intrinsic      Instruction          P9
+ *      vec_popcnt     vpopcntb/h/w/d       2cy @ 2p    element-wise popcount
+ *      vec_xor        xxlxor               1cy @ 4p
+ *      vec_and        xxland               1cy @ 4p
+ *      vec_or         xxlor                1cy @ 4p
+ *      vec_cmpne      vcmpneb/h/w          2cy @ 2p    byte/half/word not-equal
+ *      vec_xl_len     lxvll                6cy @ 1p    partial vector load
+ *
+ *  Power9 has native doubleword `vpopcntd` instruction, providing efficient SIMD popcount
+ *  with minimal data flow complexity. `vec_xl_len` enables branchless tail handling.
+ *
+ *  @section set_powervsx_stateful Stateful Streaming Logic
+ *
+ *  To build memory-optimal tiled algorithms, this file defines:
+ *
+ *  - nk_hamming_u1x128_state_powervsx_t for streaming Hamming distance
+ *  - nk_jaccard_u1x128_state_powervsx_t for streaming Jaccard similarity
+ *
+ *  @code{c}
+ *  nk_jaccard_u1x128_state_powervsx_t state_first, state_second, state_third, state_fourth;
+ *  nk_jaccard_u1x128_init_powervsx(&state_first);
+ *  // ... stream through packed binary vectors ...
+ *  nk_jaccard_u1x128_finalize_powervsx(&state_first, &state_second, &state_third, &state_fourth,
+ *      query_popcount, target_popcount_a, target_popcount_b, target_popcount_c, target_popcount_d,
+ *      total_dimensions, &results);
+ *  @endcode
+ */
+#ifndef NK_SET_POWERVSX_H
+#define NK_SET_POWERVSX_H
+
+#if NK_TARGET_POWER64_
+#if NK_TARGET_POWERVSX
+
+#include "numkong/types.h"
+#include "numkong/set/serial.h"   // `nk_u1x8_popcount_`
+#include "numkong/dot/powervsx.h" // `nk_hsum_u32x4_powervsx_`, `nk_hsum_u64x2_powervsx_`
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((target("power9-vector"))), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("power9-vector")
+#endif
+
+NK_PUBLIC void nk_hamming_u1_powervsx(nk_u1x8_t const *a, nk_u1x8_t const *b, nk_size_t n, nk_u32_t *result) {
+    nk_size_t n_bytes = nk_size_divide_round_up_(n, NK_BITS_PER_BYTE);
+    nk_vu64x2_t differences_u64x2 = vec_splats((nk_u64_t)0);
+    nk_size_t i = 0;
+    // Process 16 bytes at a time using doubleword popcount (vpopcntd)
+    for (; i + 16 <= n_bytes; i += 16) {
+        nk_vu8x16_t a_u8x16 = vec_xl(0, (nk_u8_t const *)(a + i));
+        nk_vu8x16_t b_u8x16 = vec_xl(0, (nk_u8_t const *)(b + i));
+        nk_vu8x16_t xor_u8x16 = vec_xor(a_u8x16, b_u8x16);
+        nk_vu64x2_t popcnt_u64x2 = vec_popcnt((nk_vu64x2_t)xor_u8x16);
+        differences_u64x2 = vec_add(differences_u64x2, popcnt_u64x2);
+    }
+    // Branchless tail: vec_xl_len zero-fills beyond remaining_bytes
+    nk_size_t remaining_bytes = n_bytes - i;
+    nk_vu8x16_t a_u8x16 = vec_xl_len((nk_u8_t *)(a + i), remaining_bytes);
+    nk_vu8x16_t b_u8x16 = vec_xl_len((nk_u8_t *)(b + i), remaining_bytes);
+    nk_vu8x16_t xor_u8x16 = vec_xor(a_u8x16, b_u8x16);
+    nk_vu64x2_t popcnt_u64x2 = vec_popcnt((nk_vu64x2_t)xor_u8x16);
+    differences_u64x2 = vec_add(differences_u64x2, popcnt_u64x2);
+    *result = (nk_u32_t)nk_hsum_u64x2_powervsx_(differences_u64x2);
+}
+
+NK_PUBLIC void nk_jaccard_u1_powervsx(nk_u1x8_t const *a, nk_u1x8_t const *b, nk_size_t n, nk_f32_t *result) {
+    nk_size_t n_bytes = nk_size_divide_round_up_(n, NK_BITS_PER_BYTE);
+    nk_vu64x2_t intersection_u64x2 = vec_splats((nk_u64_t)0);
+    nk_vu64x2_t union_u64x2 = vec_splats((nk_u64_t)0);
+    nk_size_t i = 0;
+    for (; i + 16 <= n_bytes; i += 16) {
+        nk_vu8x16_t a_u8x16 = vec_xl(0, (nk_u8_t const *)(a + i));
+        nk_vu8x16_t b_u8x16 = vec_xl(0, (nk_u8_t const *)(b + i));
+        nk_vu64x2_t and_popcnt_u64x2 = vec_popcnt((nk_vu64x2_t)vec_and(a_u8x16, b_u8x16));
+        nk_vu64x2_t or_popcnt_u64x2 = vec_popcnt((nk_vu64x2_t)vec_or(a_u8x16, b_u8x16));
+        intersection_u64x2 = vec_add(intersection_u64x2, and_popcnt_u64x2);
+        union_u64x2 = vec_add(union_u64x2, or_popcnt_u64x2);
+    }
+    // Branchless tail
+    nk_size_t remaining_bytes = n_bytes - i;
+    nk_vu8x16_t a_u8x16 = vec_xl_len((nk_u8_t *)(a + i), remaining_bytes);
+    nk_vu8x16_t b_u8x16 = vec_xl_len((nk_u8_t *)(b + i), remaining_bytes);
+    nk_vu64x2_t and_popcnt_u64x2 = vec_popcnt((nk_vu64x2_t)vec_and(a_u8x16, b_u8x16));
+    nk_vu64x2_t or_popcnt_u64x2 = vec_popcnt((nk_vu64x2_t)vec_or(a_u8x16, b_u8x16));
+    intersection_u64x2 = vec_add(intersection_u64x2, and_popcnt_u64x2);
+    union_u64x2 = vec_add(union_u64x2, or_popcnt_u64x2);
+    nk_u32_t intersection_count = (nk_u32_t)nk_hsum_u64x2_powervsx_(intersection_u64x2);
+    nk_u32_t union_count = (nk_u32_t)nk_hsum_u64x2_powervsx_(union_u64x2);
+    *result = (union_count != 0) ? 1.0f - (nk_f32_t)intersection_count / (nk_f32_t)union_count : 0.0f;
+}
+
+NK_PUBLIC void nk_hamming_u8_powervsx(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result) {
+    nk_vu32x4_t differences_u32x4 = vec_splats((nk_u32_t)0);
+    nk_vu8x16_t ones_u8x16 = vec_splats((nk_u8_t)1);
+    nk_size_t i = 0;
+    // Process 16 bytes at a time using vec_cmpne
+    for (; i + 16 <= n; i += 16) {
+        nk_vu8x16_t a_u8x16 = vec_xl(0, (nk_u8_t const *)(a + i));
+        nk_vu8x16_t b_u8x16 = vec_xl(0, (nk_u8_t const *)(b + i));
+        // vec_cmpne returns 0xFF for not-equal, 0x00 for equal
+        // AND with 1 to get 0x01 for not-equal, then sum groups of 4 bytes → u32
+        nk_vu8x16_t not_equal_u8x16 = vec_and((nk_vu8x16_t)vec_cmpne(a_u8x16, b_u8x16), ones_u8x16);
+        differences_u32x4 = vec_sum4s(not_equal_u8x16, differences_u32x4);
+    }
+    // Branchless tail
+    nk_size_t remaining_bytes = n - i;
+    nk_vu8x16_t a_u8x16 = vec_xl_len((nk_u8_t *)(a + i), remaining_bytes);
+    nk_vu8x16_t b_u8x16 = vec_xl_len((nk_u8_t *)(b + i), remaining_bytes);
+    nk_vu8x16_t not_equal_u8x16 = vec_and((nk_vu8x16_t)vec_cmpne(a_u8x16, b_u8x16), ones_u8x16);
+    differences_u32x4 = vec_sum4s(not_equal_u8x16, differences_u32x4);
+    *result = nk_hsum_u32x4_powervsx_(differences_u32x4);
+}
+
+typedef struct nk_hamming_u1x128_state_powervsx_t {
+    nk_vu32x4_t intersection_count_u32x4;
+} nk_hamming_u1x128_state_powervsx_t;
+
+NK_INTERNAL void nk_hamming_u1x128_init_powervsx(nk_hamming_u1x128_state_powervsx_t *state) {
+    state->intersection_count_u32x4 = vec_splats((nk_u32_t)0);
+}
+
+NK_INTERNAL void nk_hamming_u1x128_update_powervsx(nk_hamming_u1x128_state_powervsx_t *state, nk_b128_vec_t a,
+                                                   nk_b128_vec_t b, nk_size_t depth_offset,
+                                                   nk_size_t active_dimensions) {
+    nk_unused_(depth_offset);
+    nk_unused_(active_dimensions);
+
+    // Process one 128-bit chunk (native VSX register size).
+    // Uses vector accumulation → horizontal sum deferred to finalize.
+    //
+    // Power9 VSX instruction characteristics:
+    // - `vec_xor`:     xxlxor (V, V, V)             1cy, bitwise XOR
+    // - `vec_popcnt`:  vpopcntw (V.4S, V.4S)        3cy, word popcount
+    // - `vec_add`:     vadduwm (V.4S, V.4S, V.4S)   2cy, u32 add
+    // Total: ~6cy per 128-bit chunk (horizontal sum deferred to finalize)
+
+    // Step 1: Compute difference bits (A XOR B)
+    nk_vu8x16_t a_u8x16 = *(nk_vu8x16_t *)&a;
+    nk_vu8x16_t b_u8x16 = *(nk_vu8x16_t *)&b;
+    nk_vu8x16_t xor_u8x16 = vec_xor(a_u8x16, b_u8x16);
+
+    // Step 2: Word popcount → each u32 lane contains set bits for 4 bytes
+    nk_vu32x4_t popcnt_u32x4 = vec_popcnt((nk_vu32x4_t)xor_u8x16);
+
+    // Step 3: Vector accumulation (defers horizontal sum to finalize)
+    state->intersection_count_u32x4 = vec_add(state->intersection_count_u32x4, popcnt_u32x4);
+}
+
+NK_INTERNAL void nk_hamming_u1x128_finalize_powervsx( //
+    nk_hamming_u1x128_state_powervsx_t const *state_a, nk_hamming_u1x128_state_powervsx_t const *state_b,
+    nk_hamming_u1x128_state_powervsx_t const *state_c, nk_hamming_u1x128_state_powervsx_t const *state_d,
+    nk_size_t total_dimensions, nk_b128_vec_t *result) {
+    nk_unused_(total_dimensions);
+
+    nk_vu32x4_t a_u32x4 = state_a->intersection_count_u32x4, b_u32x4 = state_b->intersection_count_u32x4,
+                c_u32x4 = state_c->intersection_count_u32x4, d_u32x4 = state_d->intersection_count_u32x4;
+    nk_vu32x4_t transpose_ab_low_u32x4 = vec_mergeh(a_u32x4, b_u32x4);
+    nk_vu32x4_t transpose_cd_low_u32x4 = vec_mergeh(c_u32x4, d_u32x4);
+    nk_vu32x4_t transpose_ab_high_u32x4 = vec_mergel(a_u32x4, b_u32x4);
+    nk_vu32x4_t transpose_cd_high_u32x4 = vec_mergel(c_u32x4, d_u32x4);
+    nk_vu32x4_t sum_lane0_u32x4 = (nk_vu32x4_t)vec_xxpermdi((nk_vu64x2_t)transpose_ab_low_u32x4,
+                                                            (nk_vu64x2_t)transpose_cd_low_u32x4, 0);
+    nk_vu32x4_t sum_lane1_u32x4 = (nk_vu32x4_t)vec_xxpermdi((nk_vu64x2_t)transpose_ab_low_u32x4,
+                                                            (nk_vu64x2_t)transpose_cd_low_u32x4, 3);
+    nk_vu32x4_t sum_lane2_u32x4 = (nk_vu32x4_t)vec_xxpermdi((nk_vu64x2_t)transpose_ab_high_u32x4,
+                                                            (nk_vu64x2_t)transpose_cd_high_u32x4, 0);
+    nk_vu32x4_t sum_lane3_u32x4 = (nk_vu32x4_t)vec_xxpermdi((nk_vu64x2_t)transpose_ab_high_u32x4,
+                                                            (nk_vu64x2_t)transpose_cd_high_u32x4, 3);
+    result->vu32x4 = vec_add(vec_add(sum_lane0_u32x4, sum_lane1_u32x4), vec_add(sum_lane2_u32x4, sum_lane3_u32x4));
+}
+
+typedef struct nk_jaccard_u1x128_state_powervsx_t {
+    nk_vu32x4_t intersection_count_u32x4;
+} nk_jaccard_u1x128_state_powervsx_t;
+
+NK_INTERNAL void nk_jaccard_u1x128_init_powervsx(nk_jaccard_u1x128_state_powervsx_t *state) {
+    state->intersection_count_u32x4 = vec_splats((nk_u32_t)0);
+}
+
+NK_INTERNAL void nk_jaccard_u1x128_update_powervsx(nk_jaccard_u1x128_state_powervsx_t *state, nk_b128_vec_t a,
+                                                   nk_b128_vec_t b, nk_size_t depth_offset,
+                                                   nk_size_t active_dimensions) {
+    nk_unused_(depth_offset);
+    nk_unused_(active_dimensions);
+
+    // Process one 128-bit chunk (native VSX register size).
+    // Uses vector accumulation → horizontal sum deferred to finalize.
+    //
+    // Power9 VSX instruction characteristics:
+    // - `vec_and`:     xxland (V, V, V)              1cy, bitwise AND
+    // - `vec_popcnt`:  vpopcntw (V.4S, V.4S)         3cy, word popcount
+    // - `vec_add`:     vadduwm (V.4S, V.4S, V.4S)    2cy, u32 add
+    // Total: ~6cy per 128-bit chunk (horizontal sum deferred to finalize)
+
+    // Step 1: Compute intersection bits (A AND B)
+    nk_vu8x16_t a_u8x16 = *(nk_vu8x16_t *)&a;
+    nk_vu8x16_t b_u8x16 = *(nk_vu8x16_t *)&b;
+    nk_vu8x16_t intersection_u8x16 = vec_and(a_u8x16, b_u8x16);
+
+    // Step 2: Word popcount → each u32 lane contains set bits for 4 bytes
+    nk_vu32x4_t popcnt_u32x4 = vec_popcnt((nk_vu32x4_t)intersection_u8x16);
+
+    // Step 3: Vector accumulation (defers horizontal sum to finalize)
+    state->intersection_count_u32x4 = vec_add(state->intersection_count_u32x4, popcnt_u32x4);
+}
+
+NK_INTERNAL void nk_jaccard_u1x128_finalize_powervsx( //
+    nk_jaccard_u1x128_state_powervsx_t const *state_a, nk_jaccard_u1x128_state_powervsx_t const *state_b,
+    nk_jaccard_u1x128_state_powervsx_t const *state_c, nk_jaccard_u1x128_state_powervsx_t const *state_d,
+    nk_f32_t query_popcount, nk_f32_t target_popcount_a, nk_f32_t target_popcount_b, nk_f32_t target_popcount_c,
+    nk_f32_t target_popcount_d, nk_size_t total_dimensions, nk_b128_vec_t *result) {
+    nk_unused_(total_dimensions);
+
+    // Transpose-based 4-way horizontal sum of u32x4 intersection counts
+    nk_vu32x4_t a_u32x4 = state_a->intersection_count_u32x4, b_u32x4 = state_b->intersection_count_u32x4,
+                c_u32x4 = state_c->intersection_count_u32x4, d_u32x4 = state_d->intersection_count_u32x4;
+    nk_vu32x4_t transpose_ab_low_u32x4 = vec_mergeh(a_u32x4, b_u32x4);
+    nk_vu32x4_t transpose_cd_low_u32x4 = vec_mergeh(c_u32x4, d_u32x4);
+    nk_vu32x4_t transpose_ab_high_u32x4 = vec_mergel(a_u32x4, b_u32x4);
+    nk_vu32x4_t transpose_cd_high_u32x4 = vec_mergel(c_u32x4, d_u32x4);
+    nk_vu32x4_t sum_lane0_u32x4 = (nk_vu32x4_t)vec_xxpermdi((nk_vu64x2_t)transpose_ab_low_u32x4,
+                                                            (nk_vu64x2_t)transpose_cd_low_u32x4, 0);
+    nk_vu32x4_t sum_lane1_u32x4 = (nk_vu32x4_t)vec_xxpermdi((nk_vu64x2_t)transpose_ab_low_u32x4,
+                                                            (nk_vu64x2_t)transpose_cd_low_u32x4, 3);
+    nk_vu32x4_t sum_lane2_u32x4 = (nk_vu32x4_t)vec_xxpermdi((nk_vu64x2_t)transpose_ab_high_u32x4,
+                                                            (nk_vu64x2_t)transpose_cd_high_u32x4, 0);
+    nk_vu32x4_t sum_lane3_u32x4 = (nk_vu32x4_t)vec_xxpermdi((nk_vu64x2_t)transpose_ab_high_u32x4,
+                                                            (nk_vu64x2_t)transpose_cd_high_u32x4, 3);
+    nk_vu32x4_t intersection_u32x4 = vec_add(vec_add(sum_lane0_u32x4, sum_lane1_u32x4),
+                                             vec_add(sum_lane2_u32x4, sum_lane3_u32x4));
+    nk_vf32x4_t intersection_f32x4 = vec_ctf(intersection_u32x4, 0);
+
+    // Build target popcounts vector via vec_insert
+    nk_vf32x4_t targets_f32x4 = vec_splats(0.0f);
+    targets_f32x4 = vec_insert(target_popcount_a, targets_f32x4, 0);
+    targets_f32x4 = vec_insert(target_popcount_b, targets_f32x4, 1);
+    targets_f32x4 = vec_insert(target_popcount_c, targets_f32x4, 2);
+    targets_f32x4 = vec_insert(target_popcount_d, targets_f32x4, 3);
+    nk_vf32x4_t query_f32x4 = vec_splats(query_popcount);
+
+    // Compute union using |A union B| = |A| + |B| - |A intersection B|
+    nk_vf32x4_t union_f32x4 = vec_sub(vec_add(query_f32x4, targets_f32x4), intersection_f32x4);
+
+    // Handle zero-union edge case (empty vectors → distance = 0.0)
+    nk_vf32x4_t one_f32x4 = vec_splats(1.0f);
+    nk_vf32x4_t zero_f32x4 = vec_splats(0.0f);
+    nk_vu32x4_t zero_union_mask_u32x4 = (nk_vu32x4_t)vec_cmpeq(union_f32x4, zero_f32x4);
+    nk_vf32x4_t safe_union_f32x4 = vec_sel(union_f32x4, one_f32x4, zero_union_mask_u32x4);
+
+    // Fast reciprocal with Newton-Raphson refinement
+    nk_vf32x4_t union_reciprocal_f32x4 = vec_re(safe_union_f32x4);
+    // One Newton-Raphson step: reciprocal = reciprocal × (2 - value * reciprocal)
+    nk_vf32x4_t two_f32x4 = vec_splats(2.0f);
+    union_reciprocal_f32x4 = vec_mul(union_reciprocal_f32x4,
+                                     vec_sub(two_f32x4, vec_mul(safe_union_f32x4, union_reciprocal_f32x4)));
+
+    // Compute Jaccard distance = 1 - intersection / union
+    nk_vf32x4_t ratio_f32x4 = vec_mul(intersection_f32x4, union_reciprocal_f32x4);
+    nk_vf32x4_t jaccard_f32x4 = vec_sub(one_f32x4, ratio_f32x4);
+    result->vf32x4 = vec_sel(jaccard_f32x4, zero_f32x4, zero_union_mask_u32x4);
+}
+
+/** @brief Hamming from_dot: computes pop_a + pop_b - 2 × dot for 4 pairs (Power VSX). */
+NK_INTERNAL void nk_hamming_u32x4_from_dot_powervsx_(nk_b128_vec_t dots, nk_u32_t query_pop, nk_b128_vec_t target_pops,
+                                                     nk_b128_vec_t *results) {
+    nk_vu32x4_t dots_u32x4 = dots.vu32x4;
+    nk_vu32x4_t query_u32x4 = vec_splats(query_pop);
+    nk_vu32x4_t target_u32x4 = target_pops.vu32x4;
+    nk_vu32x4_t two_dots_u32x4 = vec_add(dots_u32x4, dots_u32x4);
+    results->vu32x4 = vec_sub(vec_add(query_u32x4, target_u32x4), two_dots_u32x4);
+}
+
+/** @brief Jaccard from_dot: computes 1 - dot / (pop_a + pop_b - dot) for 4 pairs (Power VSX). */
+NK_INTERNAL void nk_jaccard_f32x4_from_dot_powervsx_(nk_b128_vec_t dots, nk_u32_t query_pop, nk_b128_vec_t target_pops,
+                                                     nk_b128_vec_t *results) {
+    nk_vf32x4_t dot_f32x4 = vec_ctf(dots.vu32x4, 0);
+    nk_vf32x4_t query_f32x4 = vec_splats((nk_f32_t)query_pop);
+    nk_vf32x4_t target_f32x4 = vec_ctf(target_pops.vu32x4, 0);
+    nk_vf32x4_t union_f32x4 = vec_sub(vec_add(query_f32x4, target_f32x4), dot_f32x4);
+
+    nk_vf32x4_t one_f32x4 = vec_splats(1.0f);
+    nk_vf32x4_t zero_f32x4 = vec_splats(0.0f);
+    nk_vu32x4_t zero_union_mask_u32x4 = (nk_vu32x4_t)vec_cmpeq(union_f32x4, zero_f32x4);
+    nk_vf32x4_t safe_union_f32x4 = vec_sel(union_f32x4, one_f32x4, zero_union_mask_u32x4);
+
+    // Fast reciprocal with Newton-Raphson
+    nk_vf32x4_t union_reciprocal_f32x4 = vec_re(safe_union_f32x4);
+    nk_vf32x4_t two_f32x4 = vec_splats(2.0f);
+    union_reciprocal_f32x4 = vec_mul(union_reciprocal_f32x4,
+                                     vec_sub(two_f32x4, vec_mul(safe_union_f32x4, union_reciprocal_f32x4)));
+
+    nk_vf32x4_t ratio_f32x4 = vec_mul(dot_f32x4, union_reciprocal_f32x4);
+    nk_vf32x4_t jaccard_f32x4 = vec_sub(one_f32x4, ratio_f32x4);
+    results->vf32x4 = vec_sel(jaccard_f32x4, zero_f32x4, zero_union_mask_u32x4);
+}
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_POWERVSX
+#endif // NK_TARGET_POWER64_
+#endif // NK_SET_POWERVSX_H

package/numkong/include/numkong/set/rvv.h

@@ -0,0 +1,226 @@
+/**
+ *  @brief SIMD-accelerated Set Similarity Measures for RISC-V.
+ *  @file include/numkong/set/rvv.h
+ *  @author Ash Vardanian
+ *  @date January 13, 2026
+ *
+ *  @sa include/numkong/set.h
+ *
+ *  SpacemiT K1 and similar chips implement RVA22 profile with base RVV 1.0.
+ *  This does NOT include the Zvbb extension, so we lack native element-wise popcount (`vcpop.v`).
+ *
+ *  @section rvv_popcount_lut Popcount via vrgather LUT
+ *
+ *  We implement popcount using a 16-entry nibble lookup table with `vrgather`:
+ *  - Split each byte into high and low nibbles
+ *  - Use vrgather to look up popcount of each nibble (0-4)
+ *  - Sum the results (0-8 per byte)
+ *
+ *  This approach is efficient on SpacemiT X60 cores which have optimized vrgather
+ *  for small indices (LMUL=1 with indices 0-15).
+ *
+ *  @section set_rvv_instructions Key RVV Set Instructions
+ *
+ *      Intrinsic                       Purpose
+ *      vxor_vv_u8m1                    XOR for Hamming difference
+ *      vand_vv_u8m1                    AND for Jaccard intersection
+ *      vor_vv_u8m1                     OR for Jaccard union
+ *      vrgather_vv_u8m1                LUT lookup (16-entry nibble table)
+ *      vsrl_vx_u8m1                    Right shift to extract high nibble
+ *      vwaddu_vx_u16m2                 Widen u8 → u16 for accumulation
+ *      vwredsumu_vs_u16m2_u32m1        Widening reduction sum
+ */
+#ifndef NK_SET_RVV_H
+#define NK_SET_RVV_H
+
+#if NK_TARGET_RISCV64_
+#if NK_TARGET_RVV
+
+#include "numkong/types.h"
+#include "numkong/set/serial.h" // `nk_u1x8_popcount_`
+
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((target("arch=+v"))), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("arch=+v")
+#endif
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#pragma region Binary Sets
+
+/**
+ *  @brief  Compute byte-level popcount using arithmetic SWAR.
+ *
+ *  Uses parallel bit counting (Hamming weight) — no vrgather, so scales
+ *  linearly with LMUL unlike the nibble-LUT approach.
+ *  Cost: 11 ALU instructions regardless of LMUL.
+ *
+ *  @param[in] v_u8m4 Input vector of bytes
+ *  @param[in] vector_length Vector length
+ *  @return Vector where each byte contains its popcount (0-8)
+ */
+NK_INTERNAL vuint8m4_t nk_popcount_u8m4_rvv_(vuint8m4_t v_u8m4, nk_size_t vector_length) {
+    // Step 1: count pairs — v = (v & 0x55) + ((v >> 1) & 0x55)
+    vuint8m4_t t_u8m4 = __riscv_vsrl_vx_u8m4(v_u8m4, 1, vector_length);
+    t_u8m4 = __riscv_vand_vx_u8m4(t_u8m4, 0x55, vector_length);
+    v_u8m4 = __riscv_vand_vx_u8m4(v_u8m4, 0x55, vector_length);
+    v_u8m4 = __riscv_vadd_vv_u8m4(v_u8m4, t_u8m4, vector_length);
+    // Step 2: count nibbles — v = (v & 0x33) + ((v >> 2) & 0x33)
+    t_u8m4 = __riscv_vsrl_vx_u8m4(v_u8m4, 2, vector_length);
+    t_u8m4 = __riscv_vand_vx_u8m4(t_u8m4, 0x33, vector_length);
+    v_u8m4 = __riscv_vand_vx_u8m4(v_u8m4, 0x33, vector_length);
+    v_u8m4 = __riscv_vadd_vv_u8m4(v_u8m4, t_u8m4, vector_length);
+    // Step 3: count bytes — v = (v + (v >> 4)) & 0x0F
+    t_u8m4 = __riscv_vsrl_vx_u8m4(v_u8m4, 4, vector_length);
+    v_u8m4 = __riscv_vadd_vv_u8m4(v_u8m4, t_u8m4, vector_length);
+    return __riscv_vand_vx_u8m4(v_u8m4, 0x0F, vector_length);
+}
+
+NK_PUBLIC void nk_hamming_u1_rvv(nk_u1x8_t const *a, nk_u1x8_t const *b, nk_size_t n, nk_u32_t *result) {
+    nk_size_t count_bytes = nk_size_divide_round_up_(n, NK_BITS_PER_BYTE);
+
+    // Accumulator for total differences
+    vuint32m1_t sum_u32m1 = __riscv_vmv_v_x_u32m1(0, 1);
+
+    nk_size_t i = 0;
+    for (nk_size_t vector_length; i + 1 <= count_bytes; i += vector_length) {
+        vector_length = __riscv_vsetvl_e8m4(count_bytes - i);
+
+        // Load and XOR to find differing bits
+        vuint8m4_t a_u8m4 = __riscv_vle8_v_u8m4(a + i, vector_length);
+        vuint8m4_t b_u8m4 = __riscv_vle8_v_u8m4(b + i, vector_length);
+        vuint8m4_t xor_u8m4 = __riscv_vxor_vv_u8m4(a_u8m4, b_u8m4, vector_length);
+
+        // Popcount each byte (0-8 per byte) using arithmetic SWAR
+        vuint8m4_t popcount_u8m4 = nk_popcount_u8m4_rvv_(xor_u8m4, vector_length);
+
+        // Widen to u16 and accumulate via widening reduction sum
+        vuint16m8_t popcount_u16m8 = __riscv_vwaddu_vx_u16m8(popcount_u8m4, 0, vector_length);
+        sum_u32m1 = __riscv_vwredsumu_vs_u16m8_u32m1(popcount_u16m8, sum_u32m1, vector_length);
+    }
+
+    *result = __riscv_vmv_x_s_u32m1_u32(sum_u32m1);
+}
+
+NK_PUBLIC void nk_jaccard_u1_rvv(nk_u1x8_t const *a, nk_u1x8_t const *b, nk_size_t n, nk_f32_t *result) {
+    nk_size_t count_bytes = nk_size_divide_round_up_(n, NK_BITS_PER_BYTE);
+
+    // Accumulators for intersection and union counts
+    vuint32m1_t intersection_sum_u32m1 = __riscv_vmv_v_x_u32m1(0, 1);
+    vuint32m1_t union_sum_u32m1 = __riscv_vmv_v_x_u32m1(0, 1);
+
+    nk_size_t i = 0;
+    for (nk_size_t vector_length; i + 1 <= count_bytes; i += vector_length) {
+        vector_length = __riscv_vsetvl_e8m4(count_bytes - i);
+
+        // Load vectors
+        vuint8m4_t a_u8m4 = __riscv_vle8_v_u8m4(a + i, vector_length);
+        vuint8m4_t b_u8m4 = __riscv_vle8_v_u8m4(b + i, vector_length);
+
+        // Compute intersection (AND) and union (OR)
+        vuint8m4_t intersection_u8m4 = __riscv_vand_vv_u8m4(a_u8m4, b_u8m4, vector_length);
+        vuint8m4_t union_u8m4 = __riscv_vor_vv_u8m4(a_u8m4, b_u8m4, vector_length);
+
+        // Popcount each using arithmetic SWAR
+        vuint8m4_t intersection_popcount_u8m4 = nk_popcount_u8m4_rvv_(intersection_u8m4, vector_length);
+        vuint8m4_t union_popcount_u8m4 = nk_popcount_u8m4_rvv_(union_u8m4, vector_length);
+
+        // Widen and accumulate
+        vuint16m8_t intersection_popcount_u16m8 = __riscv_vwaddu_vx_u16m8(intersection_popcount_u8m4, 0, vector_length);
+        vuint16m8_t union_popcount_u16m8 = __riscv_vwaddu_vx_u16m8(union_popcount_u8m4, 0, vector_length);
+        intersection_sum_u32m1 = __riscv_vwredsumu_vs_u16m8_u32m1(intersection_popcount_u16m8, intersection_sum_u32m1,
+                                                                  vector_length);
+        union_sum_u32m1 = __riscv_vwredsumu_vs_u16m8_u32m1(union_popcount_u16m8, union_sum_u32m1, vector_length);
+    }
+
+    nk_u32_t intersection_count_u32 = __riscv_vmv_x_s_u32m1_u32(intersection_sum_u32m1);
+    nk_u32_t union_count_u32 = __riscv_vmv_x_s_u32m1_u32(union_sum_u32m1);
+    *result = (union_count_u32 != 0) ? 1.0f - (nk_f32_t)intersection_count_u32 / (nk_f32_t)union_count_u32 : 0.0f;
+}
+
+#pragma endregion Binary Sets
+
+#pragma region Integer Sets
+
+NK_PUBLIC void nk_hamming_u8_rvv(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result) {
+    vuint32m1_t difference_count_u32m1 = __riscv_vmv_v_x_u32m1(0, 1);
+
+    nk_size_t i = 0;
+    for (nk_size_t vector_length; i + 1 <= n; i += vector_length) {
+        vector_length = __riscv_vsetvl_e8m1(n - i);
+
+        vuint8m1_t a_u8m1 = __riscv_vle8_v_u8m1(a + i, vector_length);
+        vuint8m1_t b_u8m1 = __riscv_vle8_v_u8m1(b + i, vector_length);
+
+        // Compare: mask where a != b
+        vbool8_t not_equal_mask_b8 = __riscv_vmsne_vv_u8m1_b8(a_u8m1, b_u8m1, vector_length);
+
+        // Count set bits in mask via vcpop.m (this IS available in base RVV 1.0)
+        nk_u32_t difference_count_u32 = __riscv_vcpop_m_b8(not_equal_mask_b8, vector_length);
+
+        // Accumulate (scalar addition is fine here, vcpop already reduced)
+        difference_count_u32m1 = __riscv_vadd_vx_u32m1(difference_count_u32m1, difference_count_u32, 1);
+    }
+
+    *result = __riscv_vmv_x_s_u32m1_u32(difference_count_u32m1);
+}
+
+NK_PUBLIC void nk_jaccard_u32_rvv(nk_u32_t const *a, nk_u32_t const *b, nk_size_t n, nk_f32_t *result) {
+    nk_u32_t match_count_u32 = 0;
+
+    nk_size_t i = 0;
+    for (nk_size_t vector_length; i + 1 <= n; i += vector_length) {
+        vector_length = __riscv_vsetvl_e32m1(n - i);
+
+        vuint32m1_t a_u32m1 = __riscv_vle32_v_u32m1(a + i, vector_length);
+        vuint32m1_t b_u32m1 = __riscv_vle32_v_u32m1(b + i, vector_length);
+
+        // Compare: mask where a == b
+        vbool32_t equal_mask_b32 = __riscv_vmseq_vv_u32m1_b32(a_u32m1, b_u32m1, vector_length);
+
+        // Count matches via vcpop.m
+        match_count_u32 += __riscv_vcpop_m_b32(equal_mask_b32, vector_length);
+    }
+
+    *result = (n != 0) ? 1.0f - (nk_f32_t)match_count_u32 / (nk_f32_t)n : 0.0f;
+}
+
+NK_PUBLIC void nk_jaccard_u16_rvv(nk_u16_t const *a, nk_u16_t const *b, nk_size_t n, nk_f32_t *result) {
+    nk_u32_t match_count_u32 = 0;
+
+    nk_size_t i = 0;
+    for (nk_size_t vector_length; i + 1 <= n; i += vector_length) {
+        vector_length = __riscv_vsetvl_e16m1(n - i);
+
+        vuint16m1_t a_u16m1 = __riscv_vle16_v_u16m1(a + i, vector_length);
+        vuint16m1_t b_u16m1 = __riscv_vle16_v_u16m1(b + i, vector_length);
+
+        // Compare: mask where a == b
+        vbool16_t equal_mask_b16 = __riscv_vmseq_vv_u16m1_b16(a_u16m1, b_u16m1, vector_length);
+
+        // Count matches via vcpop.m
+        match_count_u32 += __riscv_vcpop_m_b16(equal_mask_b16, vector_length);
+    }
+
+    *result = (n != 0) ? 1.0f - (nk_f32_t)match_count_u32 / (nk_f32_t)n : 0.0f;
+}
+
+#pragma endregion Integer Sets
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#endif // NK_TARGET_RVV
+#endif // NK_TARGET_RISCV64_
+#endif // NK_SET_RVV_H

package/numkong/include/numkong/set/rvvbb.h

@@ -0,0 +1,117 @@
+/**
+ *  @brief SIMD-accelerated Set Similarity Measures for RISC-V with Zvbb.
+ *  @file include/numkong/set/rvvbb.h
+ *  @author Ash Vardanian
+ *  @date February 9, 2026
+ *
+ *  @sa include/numkong/set.h
+ *
+ *  Zvbb (Vector Basic Bit-manipulation) provides native per-element popcount via `vcpop.v`,
+ *  replacing the 11-instruction SWAR approach in set/rvv.h with a single instruction.
+ *
+ *  Only `nk_hamming_u1` and `nk_jaccard_u1` benefit from Zvbb (they need byte-level popcount).
+ *  Integer set operations (hamming_u8, jaccard_u16, jaccard_u32) use mask-level vcpop.m
+ *  which is already available in base RVV 1.0.
+ *
+ *  Requires: RVV 1.0 + Zvbb extension (GCC 14+ or Clang 18+)
+ */
+#ifndef NK_SET_RVVBB_H
+#define NK_SET_RVVBB_H
+
+#if NK_TARGET_RISCV64_
+#if NK_TARGET_RVVBB
+
+#include "numkong/types.h"
+
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((target("arch=+v,+zvbb"))), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("arch=+v,+zvbb")
+#endif
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+/**
+ *  @brief  Native per-element popcount using Zvbb vcpop.v (1 instruction).
+ *
+ *  Replaces the 11-instruction SWAR approach in nk_popcount_u8m4_rvv_.
+ */
+NK_INTERNAL vuint8m4_t nk_popcount_u8m4_rvvbb_(vuint8m4_t v_u8m4) {
+    vuint8m4_t result_u8m4;
+    __asm__ volatile("vcpop.v %0, %1" : "=vr"(result_u8m4) : "vr"(v_u8m4));
+    return result_u8m4;
+}
+
+NK_PUBLIC void nk_hamming_u1_rvvbb(nk_u1x8_t const *a, nk_u1x8_t const *b, nk_size_t n, nk_u32_t *result) {
+    nk_size_t count_bytes = nk_size_divide_round_up_(n, NK_BITS_PER_BYTE);
+
+    vuint32m1_t sum_u32m1 = __riscv_vmv_v_x_u32m1(0, 1);
+
+    nk_size_t i = 0;
+    for (nk_size_t vector_length; i + 1 <= count_bytes; i += vector_length) {
+        vector_length = __riscv_vsetvl_e8m4(count_bytes - i);
+
+        vuint8m4_t a_u8m4 = __riscv_vle8_v_u8m4(a + i, vector_length);
+        vuint8m4_t b_u8m4 = __riscv_vle8_v_u8m4(b + i, vector_length);
+        vuint8m4_t xor_u8m4 = __riscv_vxor_vv_u8m4(a_u8m4, b_u8m4, vector_length);
+
+        // Native per-element popcount via Zvbb (1 instruction vs 11 SWAR)
+        vuint8m4_t popcount_u8m4 = nk_popcount_u8m4_rvvbb_(xor_u8m4);
+
+        // Widen to u16 and accumulate via widening reduction sum
+        vuint16m8_t popcount_u16m8 = __riscv_vwaddu_vx_u16m8(popcount_u8m4, 0, vector_length);
+        sum_u32m1 = __riscv_vwredsumu_vs_u16m8_u32m1(popcount_u16m8, sum_u32m1, vector_length);
+    }
+
+    *result = __riscv_vmv_x_s_u32m1_u32(sum_u32m1);
+}
+
+NK_PUBLIC void nk_jaccard_u1_rvvbb(nk_u1x8_t const *a, nk_u1x8_t const *b, nk_size_t n, nk_f32_t *result) {
+    nk_size_t count_bytes = nk_size_divide_round_up_(n, NK_BITS_PER_BYTE);
+
+    vuint32m1_t intersection_sum_u32m1 = __riscv_vmv_v_x_u32m1(0, 1);
+    vuint32m1_t union_sum_u32m1 = __riscv_vmv_v_x_u32m1(0, 1);
+
+    nk_size_t i = 0;
+    for (nk_size_t vector_length; i + 1 <= count_bytes; i += vector_length) {
+        vector_length = __riscv_vsetvl_e8m4(count_bytes - i);
+
+        vuint8m4_t a_u8m4 = __riscv_vle8_v_u8m4(a + i, vector_length);
+        vuint8m4_t b_u8m4 = __riscv_vle8_v_u8m4(b + i, vector_length);
+
+        vuint8m4_t intersection_u8m4 = __riscv_vand_vv_u8m4(a_u8m4, b_u8m4, vector_length);
+        vuint8m4_t union_u8m4 = __riscv_vor_vv_u8m4(a_u8m4, b_u8m4, vector_length);
+
+        // Native per-element popcount via Zvbb
+        vuint8m4_t intersection_popcount_u8m4 = nk_popcount_u8m4_rvvbb_(intersection_u8m4);
+        vuint8m4_t union_popcount_u8m4 = nk_popcount_u8m4_rvvbb_(union_u8m4);
+
+        // Widen and accumulate
+        vuint16m8_t intersection_popcount_u16m8 = __riscv_vwaddu_vx_u16m8(intersection_popcount_u8m4, 0, vector_length);
+        vuint16m8_t union_popcount_u16m8 = __riscv_vwaddu_vx_u16m8(union_popcount_u8m4, 0, vector_length);
+        intersection_sum_u32m1 = __riscv_vwredsumu_vs_u16m8_u32m1(intersection_popcount_u16m8, intersection_sum_u32m1,
+                                                                  vector_length);
+        union_sum_u32m1 = __riscv_vwredsumu_vs_u16m8_u32m1(union_popcount_u16m8, union_sum_u32m1, vector_length);
+    }
+
+    nk_u32_t intersection_count_u32 = __riscv_vmv_x_s_u32m1_u32(intersection_sum_u32m1);
+    nk_u32_t union_count_u32 = __riscv_vmv_x_s_u32m1_u32(union_sum_u32m1);
+    *result = (union_count_u32 != 0) ? 1.0f - (nk_f32_t)intersection_count_u32 / (nk_f32_t)union_count_u32 : 0.0f;
+}
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#endif // NK_TARGET_RVVBB
+#endif // NK_TARGET_RISCV64_
+#endif // NK_SET_RVVBB_H