numkong 7.0.0

package/include/numkong/sets/smebi32.h

@@ -0,0 +1,1099 @@
+/**
+ *  @brief SIMD-accelerated Batched Set Distances for SME.
+ *  @file include/numkong/sets/smebi32.h
+ *  @author Ash Vardanian
+ *  @date February 6, 2026
+ *  @sa include/numkong/sets.h
+ *
+ *  Uses ARM Scalable Matrix Extension (SME) for efficient binary set operations.
+ *  Leverages streaming mode's wider vectors (512-bit on Apple M4) for fast
+ *  XOR+POPCNT operations on binary vectors.
+ *
+ *  @section smebi32_math Mathematical Foundation
+ *
+ *  Hamming distance: popcount(a XOR b) = number of differing bits
+ *
+ *  Jaccard distance using intersection:
+ *    intersection = popcount(a AND b)
+ *    union = popcount(a) + popcount(b) - intersection
+ *    jaccard = 1 - intersection / union
+ *
+ *  @section smebi32_tiles SME Dimensions (512-bit SVL)
+ *
+ *  - svcntw(): 16 (number of 32-bit elements per vector)
+ *  - svcntb(): 64 (number of bytes per SVE vector)
+ *  - Tile blocking: 16x16 output tiles for cache efficiency
+ *  - Depth processing: 64 bytes (512 bits) per iteration
+ *
+ *  @section smebi32_perf Performance Characteristics (Apple M4)
+ *
+ *  - SVL: 512 bits (64 bytes)
+ *  - Streaming mode provides dedicated register file
+ *  - Streaming mode overhead: ~50-100 cycles for SMSTART/SMSTOP
+ */
+
+#ifndef NK_SETS_SMEBI32_H
+#define NK_SETS_SMEBI32_H
+
+#if NK_TARGET_ARM_
+#if NK_TARGET_SMEBI32
+
+#include "numkong/types.h"
+#include "numkong/set/serial.h"
+#include "numkong/sets/serial.h"
+#include "numkong/dots/sme.h" // `nk_sme_zero_za32_*` constants
+#include "numkong/reduce.h"   // `nk_reduce_moments_u1`
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+/*
+ *  Binary set operations using SME BMOPA instruction.
+ *
+ *  BMOPA computes: ZA[i,j] += popcount(~(Zn[i] ^ Zm[j])) = popcount(XNOR)
+ *  This counts matching bits. Hamming = depth_bits - matching.
+ *
+ *  Tile layout (SVL=512, Apple M4):
+ *  - ZA32 output tile: 16 × 16 u32 elements (1 KB)
+ *  - Input vectors: 16 u32 elements (SVL/32)
+ *  - Each BMOPA processes 32 bits (one u32) across 16×16 pairs
+ *  - BMOPA predicates: b32 (u32 input granularity)
+ *  - Packed kernel: 4-tile path (ZA0-ZA3) for 4 B-column tiles simultaneously
+ *  - Unpacked kernel: ZA transpose (ZA0.S=staging, ZA1-3.S=accumulation, 3-tile fast path)
+ *  - Packed format: column-major u32 within each tile
+ */
+
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((target("sme2,sve2"))), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("+sme2")
+#endif
+
+/*  Read SVL in bytes from non-streaming context using RDSVL instruction. */
+NK_INTERNAL nk_size_t nk_smebi32_svl_bytes_(void) {
+    nk_size_t svl_bytes;
+    __asm__ volatile("rdsvl %0, #1" : "=r"(svl_bytes));
+    return svl_bytes;
+}
+
+/*  Get ZA32 tile dimension (number of f32/u32 elements per row). */
+NK_INTERNAL nk_size_t nk_smebi32_tile_dim_(void) { return nk_smebi32_svl_bytes_() / sizeof(nk_u32_t); }
+
+typedef struct {
+    nk_u32_t row_tile_count;   // ceiling(rows / tile_dim)
+    nk_u32_t depth_tile_count; // ceiling(depth_bits / depth_tile_bits)
+    nk_u32_t rows;             // actual row count
+    nk_u32_t depth_bits;       // actual depth in bits
+    nk_u32_t svl_bytes;        // SVL at pack time for validation
+    nk_u32_t norms_offset;     // byte offset to norms (0 if none)
+    nk_u32_t reserved[10];     // padding to 64 bytes
+} nk_sets_smebi32_packed_header_t;
+
+/** Count total set bits across a byte vector using streaming SVE.
+ *  Accumulates per-byte popcounts into u32 lanes via svdot; single horizontal reduction at end. */
+NK_PUBLIC nk_u32_t nk_sets_reduce_sumsq_u1_streaming_(nk_u1x8_t const *data,
+                                                      nk_size_t n_bytes) NK_STREAMING_COMPATIBLE_ {
+    svuint32_t acc_u32x = svdup_u32(0);
+    svuint8_t const ones_u8x = svdup_u8(1);
+    for (nk_size_t offset = 0; offset < n_bytes; offset += svcntb()) {
+        svbool_t predicate_u8x = svwhilelt_b8_u64(offset, n_bytes);
+        acc_u32x = svdot_u32(acc_u32x, svcnt_u8_z(predicate_u8x, svld1_u8(predicate_u8x, data + offset)), ones_u8x);
+    }
+    return (nk_u32_t)svaddv_u32(svptrue_b32(), acc_u32x);
+}
+
+#pragma region Hamming Distance
+
+NK_PUBLIC nk_size_t nk_dots_packed_size_u1_smebi32(nk_size_t row_count, nk_size_t depth_bits) {
+    nk_size_t const tile_dim = nk_smebi32_tile_dim_();        // 16 rows per tile
+    nk_size_t const depth_tile_size = nk_smebi32_tile_dim_(); // 16 u32 per depth tile = 512 bits
+
+    nk_size_t const depth_u32 = nk_size_divide_round_up_(depth_bits, 32);
+    nk_size_t const row_tile_count = nk_size_divide_round_up_(row_count, tile_dim);
+    nk_size_t const depth_tile_count = nk_size_divide_round_up_(depth_u32, depth_tile_size);
+
+    nk_size_t const tile_elements = tile_dim * depth_tile_size; // 256 u32 per tile
+    nk_size_t size = sizeof(nk_sets_smebi32_packed_header_t);
+    size += row_tile_count * depth_tile_count * tile_elements * sizeof(nk_u32_t);
+    size += row_count * sizeof(nk_u32_t); // per-row population counts
+
+    return size;
+}
+
+NK_PUBLIC void nk_dots_pack_u1_smebi32(nk_u1x8_t const *b, nk_size_t row_count, nk_size_t depth_bits,
+                                       nk_size_t b_stride_in_bytes, void *b_packed) {
+    nk_size_t const svl_bytes = nk_smebi32_svl_bytes_();
+    nk_size_t const tile_dim = nk_smebi32_tile_dim_();        // 16 rows per tile
+    nk_size_t const depth_tile_size = nk_smebi32_tile_dim_(); // 16 u32 per depth tile
+    nk_size_t const tile_elements = tile_dim * depth_tile_size;
+    nk_size_t const depth_in_bytes = nk_size_divide_round_up_(depth_bits, NK_BITS_PER_BYTE);
+
+    nk_size_t const depth_u32_total = nk_size_divide_round_up_(depth_bits, 32);
+    nk_size_t const row_tile_count = nk_size_divide_round_up_(row_count, tile_dim);
+    nk_size_t const depth_tile_count = nk_size_divide_round_up_(depth_u32_total, depth_tile_size);
+    nk_size_t const total_tiles = row_tile_count * depth_tile_count;
+    nk_size_t const data_size = total_tiles * tile_elements * sizeof(nk_u32_t);
+
+    nk_sets_smebi32_packed_header_t *header = (nk_sets_smebi32_packed_header_t *)b_packed;
+    header->row_tile_count = (nk_u32_t)row_tile_count;
+    header->depth_tile_count = (nk_u32_t)depth_tile_count;
+    header->rows = (nk_u32_t)row_count;
+    header->depth_bits = (nk_u32_t)depth_bits;
+    header->svl_bytes = (nk_u32_t)svl_bytes;
+    header->norms_offset = (nk_u32_t)(sizeof(nk_sets_smebi32_packed_header_t) + data_size);
+
+    nk_u32_t *tiles_ptr = (nk_u32_t *)((char *)b_packed + sizeof(nk_sets_smebi32_packed_header_t));
+    nk_u32_t *norms_ptr = (nk_u32_t *)((char *)b_packed + header->norms_offset);
+
+    // Zero-initialize all tiles (partial tiles stay zero-padded for predicated loads)
+    for (nk_size_t i = 0; i < total_tiles * tile_elements; i++) tiles_ptr[i] = 0;
+
+    // Pack tiles: column-major u32 within each tile for efficient SVE loads
+    for (nk_size_t row_tile = 0; row_tile < row_tile_count; row_tile++) {
+        for (nk_size_t depth_tile = 0; depth_tile < depth_tile_count; depth_tile++) {
+            nk_size_t const tile_index = row_tile * depth_tile_count + depth_tile;
+            nk_u32_t *tile_output = tiles_ptr + tile_index * tile_elements;
+
+            nk_size_t const src_row_start = row_tile * tile_dim;
+            nk_size_t const src_u32_start = depth_tile * depth_tile_size;
+            nk_size_t const rows_to_pack = (src_row_start + tile_dim <= row_count) ? tile_dim
+                                                                                   : (row_count - src_row_start);
+            nk_size_t const u32s_to_pack = (src_u32_start + depth_tile_size <= depth_u32_total)
+                                               ? depth_tile_size
+                                               : (depth_u32_total > src_u32_start ? depth_u32_total - src_u32_start
+                                                                                  : 0);
+
+            // Column-major packing: tile_output[col * tile_dim + row]
+            for (nk_size_t row = 0; row < rows_to_pack; row++) {
+                nk_u32_t const *src_row = (nk_u32_t const *)((char const *)b +
+                                                             (src_row_start + row) * b_stride_in_bytes);
+                for (nk_size_t col = 0; col < u32s_to_pack; col++) {
+                    nk_size_t const dst_idx = col * tile_dim + row; // Column-major!
+                    tile_output[dst_idx] = src_row[src_u32_start + col];
+                }
+            }
+        }
+    }
+
+    // Compute per-row population counts
+    for (nk_size_t row = 0; row < row_count; row++) {
+        nk_u1x8_t const *src_row = (nk_u1x8_t const *)((char const *)b + row * b_stride_in_bytes);
+        {
+            nk_u64_t nk_local_sum_, nk_local_sumsq_;
+            nk_reduce_moments_u1(src_row, depth_in_bytes * 8, sizeof(nk_u1x8_t), &nk_local_sum_, &nk_local_sumsq_);
+            norms_ptr[row] = (nk_u32_t)nk_local_sum_;
+        }
+    }
+}
+
+/**
+ *  SME Hamming kernel using ZA transpose for unpacked A.
+ *  ZA0.S = staging (A rows loaded horizontally, read vertically for BMOPA).
+ *  ZA1-3.S = BMOPA accumulation (3 B column tiles in fast path).
+ *
+ *  Each ZA0.S batch covers 16 depth u32 steps (one full depth tile).
+ *  BMOPA expansion=1 for u32: each u32 contributes 32 bits via XNOR+POPCNT.
+ */
+__arm_locally_streaming __arm_new("za") static void nk_hammings_packed_u1_smebi32_streaming_(
+    nk_u1x8_t const *a, void const *b_packed, nk_u32_t *c, nk_size_t row_count_a, nk_size_t row_count_b,
+    nk_size_t depth_bits, nk_size_t a_stride_in_bytes, nk_size_t c_stride_in_bytes) {
+
+    nk_sets_smebi32_packed_header_t const *header = (nk_sets_smebi32_packed_header_t const *)b_packed;
+    nk_size_t const row_tile_count_b = header->row_tile_count;
+    nk_size_t const depth_tile_count = header->depth_tile_count;
+
+    nk_size_t const tile_dim = svcntw();        // 16 for 512-bit SVL
+    nk_size_t const depth_tile_size = svcntw(); // 16 u32 per depth tile
+    nk_size_t const tile_elements = tile_dim * depth_tile_size;
+    nk_size_t const depth_u32_total = nk_size_divide_round_up_(depth_bits, 32);
+
+    nk_u32_t const *b_tiles = (nk_u32_t const *)((char const *)b_packed + sizeof(nk_sets_smebi32_packed_header_t));
+
+    svbool_t const predicate_all_u32x = svptrue_b32();
+    svuint32_t const depth_u32x = svdup_u32((nk_u32_t)depth_bits);
+    nk_size_t const row_tile_count_a = nk_size_divide_round_up_(row_count_a, tile_dim);
+
+    for (nk_size_t row_tile_a = 0; row_tile_a < row_tile_count_a; row_tile_a++) {
+        nk_size_t const row_start_a = row_tile_a * tile_dim;
+        nk_size_t const rows_a_remaining = (row_start_a + tile_dim <= row_count_a) ? tile_dim
+                                                                                   : (row_count_a - row_start_a);
+        svbool_t const row_predicate_u32x = svwhilelt_b32_u64(0u, rows_a_remaining);
+
+        // Fast path: 3 B column tiles using ZA1-ZA3 (ZA0.S = staging)
+        nk_size_t row_tile_b = 0;
+        for (; row_tile_b + 3 <= row_tile_count_b; row_tile_b += 3) {
+            svzero_mask_za(nk_sme_zero_za32_tiles_123_);
+
+            for (nk_size_t d_tile = 0; d_tile < depth_tile_count; d_tile++) {
+                nk_size_t const d_start_u32 = d_tile * depth_tile_size;
+                nk_size_t const u32s_this_tile = (d_start_u32 + depth_tile_size <= depth_u32_total)
+                                                     ? depth_tile_size
+                                                     : (depth_u32_total > d_start_u32 ? depth_u32_total - d_start_u32
+                                                                                      : 0);
+                if (u32s_this_tile == 0) break;
+
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+
+                svbool_t const batch_predicate_u32x = svwhilelt_b32_u64(0u, u32s_this_tile);
+
+                // Load A rows into ZA0.S horizontally as u32 words
+                for (nk_size_t row_in_tile = 0; row_in_tile < rows_a_remaining; row_in_tile++) {
+                    nk_u32_t const *a_row_u32 = (nk_u32_t const *)((char const *)a +
+                                                                   (row_start_a + row_in_tile) * a_stride_in_bytes) +
+                                                d_start_u32;
+                    svld1_hor_za32(0, row_in_tile, batch_predicate_u32x, a_row_u32);
+                }
+
+                // B tile pointers for 3 column tiles
+                nk_u32_t const *b_tile0 = b_tiles + ((row_tile_b + 0) * depth_tile_count + d_tile) * tile_elements;
+                nk_u32_t const *b_tile1 = b_tiles + ((row_tile_b + 1) * depth_tile_count + d_tile) * tile_elements;
+                nk_u32_t const *b_tile2 = b_tiles + ((row_tile_b + 2) * depth_tile_count + d_tile) * tile_elements;
+
+                // Vertical read + BMOPA for each depth step
+                for (nk_size_t step = 0; step < u32s_this_tile; step++) {
+                    svuint32_t a_column_u32x = svread_ver_za32_u32_m(svdup_u32(0), row_predicate_u32x, 0, step);
+
+                    svbmopa_za32_u32_m(1, row_predicate_u32x, predicate_all_u32x, a_column_u32x,
+                                       svld1_u32(predicate_all_u32x, b_tile0 + step * tile_dim));
+                    svbmopa_za32_u32_m(2, row_predicate_u32x, predicate_all_u32x, a_column_u32x,
+                                       svld1_u32(predicate_all_u32x, b_tile1 + step * tile_dim));
+                    svbmopa_za32_u32_m(3, row_predicate_u32x, predicate_all_u32x, a_column_u32x,
+                                       svld1_u32(predicate_all_u32x, b_tile2 + step * tile_dim));
+                }
+            }
+
+            // Extract from ZA1-3: Hamming = depth_bits - matching_bits
+            for (nk_size_t row = 0; row < rows_a_remaining; row++) {
+                nk_u32_t *c_row = (nk_u32_t *)((char *)c + (row_start_a + row) * c_stride_in_bytes);
+
+                svuint32_t za1_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_u32x, 1, row);
+                svuint32_t za2_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_u32x, 2, row);
+                svuint32_t za3_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_u32x, 3, row);
+
+                svst1_u32(predicate_all_u32x, c_row + (row_tile_b + 0) * tile_dim,
+                          svsub_u32_x(predicate_all_u32x, depth_u32x, za1_u32x));
+                svst1_u32(predicate_all_u32x, c_row + (row_tile_b + 1) * tile_dim,
+                          svsub_u32_x(predicate_all_u32x, depth_u32x, za2_u32x));
+                svst1_u32(predicate_all_u32x, c_row + (row_tile_b + 2) * tile_dim,
+                          svsub_u32_x(predicate_all_u32x, depth_u32x, za3_u32x));
+            }
+        }
+
+        // Remainder: 1 B column tile at a time using ZA1
+        for (; row_tile_b < row_tile_count_b; row_tile_b++) {
+            nk_size_t const row_start_b = row_tile_b * tile_dim;
+            nk_size_t const rows_b_remaining = (row_start_b + tile_dim <= row_count_b) ? tile_dim
+                                                                                       : (row_count_b - row_start_b);
+            svbool_t const column_predicate_u32x = svwhilelt_b32_u64(0u, rows_b_remaining);
+
+            svzero_mask_za(nk_sme_zero_za32_tile_1_);
+
+            for (nk_size_t d_tile = 0; d_tile < depth_tile_count; d_tile++) {
+                nk_size_t const d_start_u32 = d_tile * depth_tile_size;
+                nk_size_t const u32s_this_tile = (d_start_u32 + depth_tile_size <= depth_u32_total)
+                                                     ? depth_tile_size
+                                                     : (depth_u32_total > d_start_u32 ? depth_u32_total - d_start_u32
+                                                                                      : 0);
+                if (u32s_this_tile == 0) break;
+
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+
+                svbool_t const batch_predicate_u32x = svwhilelt_b32_u64(0u, u32s_this_tile);
+
+                // Load A rows into ZA0.S horizontally
+                for (nk_size_t row_in_tile = 0; row_in_tile < rows_a_remaining; row_in_tile++) {
+                    nk_u32_t const *a_row_u32 = (nk_u32_t const *)((char const *)a +
+                                                                   (row_start_a + row_in_tile) * a_stride_in_bytes) +
+                                                d_start_u32;
+                    svld1_hor_za32(0, row_in_tile, batch_predicate_u32x, a_row_u32);
+                }
+
+                nk_u32_t const *b_tile = b_tiles + (row_tile_b * depth_tile_count + d_tile) * tile_elements;
+
+                // Vertical read + BMOPA
+                for (nk_size_t step = 0; step < u32s_this_tile; step++) {
+                    svuint32_t a_column_u32x = svread_ver_za32_u32_m(svdup_u32(0), row_predicate_u32x, 0, step);
+                    svuint32_t b_u32x = svld1_u32(predicate_all_u32x, b_tile + step * tile_dim);
+                    svbmopa_za32_u32_m(1, row_predicate_u32x, column_predicate_u32x, a_column_u32x, b_u32x);
+                }
+            }
+
+            // Extract from ZA1: Hamming = depth_bits - matching_bits
+            for (nk_size_t row = 0; row < rows_a_remaining; row++) {
+                svuint32_t za1_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_u32x, 1, row);
+                svuint32_t hamming_u32x = svsub_u32_x(predicate_all_u32x, depth_u32x, za1_u32x);
+                nk_u32_t *c_row = (nk_u32_t *)((char *)c + (row_start_a + row) * c_stride_in_bytes);
+                svst1_u32(column_predicate_u32x, c_row + row_start_b, hamming_u32x);
+            }
+        }
+    }
+}
+
+NK_PUBLIC void nk_hammings_packed_u1_smebi32(nk_u1x8_t const *a, void const *b_packed, nk_u32_t *c,
+                                             nk_size_t row_count_a, nk_size_t row_count_b, nk_size_t depth_bits,
+                                             nk_size_t a_stride_in_bytes, nk_size_t c_stride_in_bytes) {
+    nk_hammings_packed_u1_smebi32_streaming_(a, b_packed, c, row_count_a, row_count_b, depth_bits, a_stride_in_bytes,
+                                             c_stride_in_bytes);
+}
+
+/**
+ *  Symmetric Hamming using ZA0 time-sharing + 3-tile fast path.
+ *  ZA0.S = staging (A rows loaded horizontally, read vertically for BMOPA).
+ *  ZA1-3.S = BMOPA accumulators (3 B column tiles in fast path).
+ *  Mirrors the unpacked kernel nk_hammings_packed_u1_smebi32_streaming_ pattern.
+ */
+__arm_locally_streaming __arm_new("za") static void nk_hammings_symmetric_u1_smebi32_streaming_(
+    nk_u1x8_t const *vectors, nk_size_t n_vectors, nk_size_t depth_bits, nk_size_t stride, nk_u32_t *result,
+    nk_size_t result_stride, nk_size_t row_start, nk_size_t row_count) {
+
+    nk_size_t const tile_dim = svcntw();        // 16 for 512-bit SVL
+    nk_size_t const depth_tile_size = svcntw(); // 16 u32 per depth tile
+    nk_size_t const depth_u32_total = nk_size_divide_round_up_(depth_bits, 32);
+    nk_size_t const depth_tile_count = nk_size_divide_round_up_(depth_u32_total, depth_tile_size);
+
+    svbool_t const predicate_all_u32x = svptrue_b32();
+    svuint32_t const depth_u32x = svdup_u32((nk_u32_t)depth_bits);
+
+    NK_ALIGN64 nk_u32_t a_buffer[16][16]; // Stack buffer for A column save
+
+    nk_size_t const row_end = row_start + row_count;
+    nk_size_t const column_tile_count = nk_size_divide_round_up_(n_vectors, tile_dim);
+
+    for (nk_size_t row_tile_start = row_start; row_tile_start < row_end && row_tile_start < n_vectors;
+         row_tile_start += tile_dim) {
+        nk_size_t const rows_remaining = (row_tile_start + tile_dim <= row_end) ? tile_dim : (row_end - row_tile_start);
+        nk_size_t const rows_clamped = (row_tile_start + rows_remaining <= n_vectors) ? rows_remaining
+                                                                                      : (n_vectors - row_tile_start);
+        svbool_t const row_predicate_u32x = svwhilelt_b32_u64(0u, rows_clamped);
+
+        nk_size_t column_tile_index = 0;
+
+        // Fast path: 3 column tiles using ZA1-ZA3 (ZA0 = staging)
+        for (; column_tile_index + 3 <= column_tile_count; column_tile_index += 3) {
+            svzero_mask_za(nk_sme_zero_za32_tiles_123_);
+
+            for (nk_size_t d_tile = 0; d_tile < depth_tile_count; d_tile++) {
+                nk_size_t const d_start_u32 = d_tile * depth_tile_size;
+                nk_size_t const u32s_this_tile = (d_start_u32 + depth_tile_size <= depth_u32_total)
+                                                     ? depth_tile_size
+                                                     : (depth_u32_total > d_start_u32 ? depth_u32_total - d_start_u32
+                                                                                      : 0);
+                if (u32s_this_tile == 0) break;
+
+                // Load A rows into ZA0 horizontally
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+                svbool_t const batch_predicate_u32x = svwhilelt_b32_u64(0u, u32s_this_tile);
+
+                for (nk_size_t row_in_tile = 0; row_in_tile < rows_clamped; row_in_tile++) {
+                    nk_u32_t const *a_row_u32 = (nk_u32_t const *)((char const *)vectors +
+                                                                   (row_tile_start + row_in_tile) * stride) +
+                                                d_start_u32;
+                    svld1_hor_za32(0, row_in_tile, batch_predicate_u32x, a_row_u32);
+                }
+
+                // Save A columns from ZA0 to stack buffer
+                for (nk_size_t s = 0; s < u32s_this_tile; s++)
+                    svst1_u32(predicate_all_u32x, a_buffer[s],
+                              svread_ver_za32_u32_m(svdup_u32(0), row_predicate_u32x, 0, s));
+
+                // B column tile 0
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+                for (nk_size_t col = 0; col < tile_dim; col++) {
+                    nk_size_t const col_abs = (column_tile_index + 0) * tile_dim + col;
+                    if (col_abs < n_vectors) {
+                        nk_u32_t const *b_row = (nk_u32_t const *)((char const *)vectors + col_abs * stride) +
+                                                d_start_u32;
+                        svld1_hor_za32(0, col, batch_predicate_u32x, b_row);
+                    }
+                }
+                for (nk_size_t step = 0; step < u32s_this_tile; step++) {
+                    svuint32_t a_u32x = svld1_u32(predicate_all_u32x, a_buffer[step]);
+                    svuint32_t b_u32x = svread_ver_za32_u32_m(svdup_u32(0), predicate_all_u32x, 0, step);
+                    svbmopa_za32_u32_m(1, row_predicate_u32x, predicate_all_u32x, a_u32x, b_u32x);
+                }
+
+                // B column tile 1
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+                for (nk_size_t col = 0; col < tile_dim; col++) {
+                    nk_size_t const col_abs = (column_tile_index + 1) * tile_dim + col;
+                    if (col_abs < n_vectors) {
+                        nk_u32_t const *b_row = (nk_u32_t const *)((char const *)vectors + col_abs * stride) +
+                                                d_start_u32;
+                        svld1_hor_za32(0, col, batch_predicate_u32x, b_row);
+                    }
+                }
+                for (nk_size_t step = 0; step < u32s_this_tile; step++) {
+                    svuint32_t a_u32x = svld1_u32(predicate_all_u32x, a_buffer[step]);
+                    svuint32_t b_u32x = svread_ver_za32_u32_m(svdup_u32(0), predicate_all_u32x, 0, step);
+                    svbmopa_za32_u32_m(2, row_predicate_u32x, predicate_all_u32x, a_u32x, b_u32x);
+                }
+
+                // B column tile 2
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+                for (nk_size_t col = 0; col < tile_dim; col++) {
+                    nk_size_t const col_abs = (column_tile_index + 2) * tile_dim + col;
+                    if (col_abs < n_vectors) {
+                        nk_u32_t const *b_row = (nk_u32_t const *)((char const *)vectors + col_abs * stride) +
+                                                d_start_u32;
+                        svld1_hor_za32(0, col, batch_predicate_u32x, b_row);
+                    }
+                }
+                for (nk_size_t step = 0; step < u32s_this_tile; step++) {
+                    svuint32_t a_u32x = svld1_u32(predicate_all_u32x, a_buffer[step]);
+                    svuint32_t b_u32x = svread_ver_za32_u32_m(svdup_u32(0), predicate_all_u32x, 0, step);
+                    svbmopa_za32_u32_m(3, row_predicate_u32x, predicate_all_u32x, a_u32x, b_u32x);
+                }
+            }
+
+            // Extract ZA1-3: hamming = depth_bits - ZA[i][j]
+            for (nk_size_t row = 0; row < rows_clamped; row++) {
+                nk_u32_t *c_row = (nk_u32_t *)((char *)result + (row_tile_start + row) * result_stride);
+
+                svuint32_t za1_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_u32x, 1, row);
+                svuint32_t za2_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_u32x, 2, row);
+                svuint32_t za3_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_u32x, 3, row);
+
+                svst1_u32(predicate_all_u32x, c_row + (column_tile_index + 0) * tile_dim,
+                          svsub_u32_x(predicate_all_u32x, depth_u32x, za1_u32x));
+                svst1_u32(predicate_all_u32x, c_row + (column_tile_index + 1) * tile_dim,
+                          svsub_u32_x(predicate_all_u32x, depth_u32x, za2_u32x));
+                svst1_u32(predicate_all_u32x, c_row + (column_tile_index + 2) * tile_dim,
+                          svsub_u32_x(predicate_all_u32x, depth_u32x, za3_u32x));
+            }
+        }
+
+        // Remainder: 1 column tile at a time using ZA1
+        for (; column_tile_index < column_tile_count; column_tile_index++) {
+            nk_size_t const col_tile_start = column_tile_index * tile_dim;
+            nk_size_t const cols_remaining = (col_tile_start + tile_dim <= n_vectors) ? tile_dim
+                                                                                      : (n_vectors - col_tile_start);
+            svbool_t const column_predicate_u32x = svwhilelt_b32_u64(0u, cols_remaining);
+
+            svzero_mask_za(nk_sme_zero_za32_tile_1_);
+
+            for (nk_size_t d_tile = 0; d_tile < depth_tile_count; d_tile++) {
+                nk_size_t const d_start_u32 = d_tile * depth_tile_size;
+                nk_size_t const u32s_this_tile = (d_start_u32 + depth_tile_size <= depth_u32_total)
+                                                     ? depth_tile_size
+                                                     : (depth_u32_total > d_start_u32 ? depth_u32_total - d_start_u32
+                                                                                      : 0);
+                if (u32s_this_tile == 0) break;
+
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+                svbool_t const batch_predicate_u32x = svwhilelt_b32_u64(0u, u32s_this_tile);
+
+                // Load A rows into ZA0 horizontally
+                for (nk_size_t row_in_tile = 0; row_in_tile < rows_clamped; row_in_tile++) {
+                    nk_u32_t const *a_row_u32 = (nk_u32_t const *)((char const *)vectors +
+                                                                   (row_tile_start + row_in_tile) * stride) +
+                                                d_start_u32;
+                    svld1_hor_za32(0, row_in_tile, batch_predicate_u32x, a_row_u32);
+                }
+
+                // Save A columns from ZA0 to stack buffer
+                for (nk_size_t s = 0; s < u32s_this_tile; s++)
+                    svst1_u32(predicate_all_u32x, a_buffer[s],
+                              svread_ver_za32_u32_m(svdup_u32(0), row_predicate_u32x, 0, s));
+
+                // Load B column tile into ZA0
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+                for (nk_size_t col = 0; col < tile_dim; col++) {
+                    nk_size_t const col_abs = col_tile_start + col;
+                    if (col_abs < n_vectors) {
+                        nk_u32_t const *b_row = (nk_u32_t const *)((char const *)vectors + col_abs * stride) +
+                                                d_start_u32;
+                        svld1_hor_za32(0, col, batch_predicate_u32x, b_row);
+                    }
+                }
+                for (nk_size_t step = 0; step < u32s_this_tile; step++) {
+                    svuint32_t a_u32x = svld1_u32(predicate_all_u32x, a_buffer[step]);
+                    svuint32_t b_u32x = svread_ver_za32_u32_m(svdup_u32(0), column_predicate_u32x, 0, step);
+                    svbmopa_za32_u32_m(1, row_predicate_u32x, column_predicate_u32x, a_u32x, b_u32x);
+                }
+            }
+
+            for (nk_size_t row = 0; row < rows_clamped; row++) {
+                svuint32_t za1_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_u32x, 1, row);
+                svuint32_t hamming_u32x = svsub_u32_x(predicate_all_u32x, depth_u32x, za1_u32x);
+                nk_u32_t *c_row = (nk_u32_t *)((char *)result + (row_tile_start + row) * result_stride);
+                svst1_u32(column_predicate_u32x, c_row + col_tile_start, hamming_u32x);
+            }
+        }
+    }
+}
+
+NK_PUBLIC void nk_hammings_symmetric_u1_smebi32(nk_u1x8_t const *vectors, nk_size_t n_vectors, nk_size_t depth_bits,
+                                                nk_size_t stride, nk_u32_t *result, nk_size_t result_stride,
+                                                nk_size_t row_start, nk_size_t row_count) {
+    nk_hammings_symmetric_u1_smebi32_streaming_(vectors, n_vectors, depth_bits, stride, result, result_stride,
+                                                row_start, row_count);
+}
+
+#pragma endregion // Hamming Distance
+
+/*
+ *  Jaccard distance via BMOPA matching counts + algebraic normalization.
+ *
+ *  BMOPA gives: matching = popcount(XNOR(a,b))
+ *  Then:
+ *    hamming      = depth_bits - matching
+ *    intersection = (norm_a + norm_b - hamming) / 2  =  (norm_a + norm_b - depth_bits + matching) / 2
+ *    union        = (norm_a + norm_b + hamming) / 2  =  sum_norms - intersection
+ *    jaccard      = 1 - intersection / union          (1.0 when union == 0)
+ *
+ *  Inner BMOPA loop is identical to Hamming; only the extraction phase differs.
+ *  Packed format shares the Hamming tile layout for B operand, plus per-row norms.
+ */
+
+#pragma region Jaccard Distance
+
+/**
+ *  SME Jaccard kernel using BMOPA for matching-bit counts.
+ *  Mirrors nk_hammings_packed_u1_smebi32_streaming_ exactly in structure,
+ *  but derives intersection/union algebraically from the matching counts:
+ *    matching      = popcount(XNOR(a,b))          (from BMOPA)
+ *    hamming       = depth_bits - matching
+ *    intersection  = (norm_a + norm_b - hamming) / 2
+ *    union         = (norm_a + norm_b + hamming) / 2
+ *    jaccard       = 1 - intersection / union      (1.0 when union == 0)
+ */
+__arm_locally_streaming __arm_new("za") static void nk_jaccards_packed_u1_smebi32_streaming_(
+    nk_u1x8_t const *a, void const *b_packed, nk_f32_t *c, nk_size_t row_count_a, nk_size_t row_count_b,
+    nk_size_t depth_bits, nk_size_t a_stride_in_bytes, nk_size_t c_stride_in_bytes) {
+
+    nk_sets_smebi32_packed_header_t const *header = (nk_sets_smebi32_packed_header_t const *)b_packed;
+    nk_size_t const row_tile_count_b = header->row_tile_count;
+    nk_size_t const depth_tile_count = header->depth_tile_count;
+
+    nk_size_t const tile_dim = svcntw();        // 16 for 512-bit SVL
+    nk_size_t const depth_tile_size = svcntw(); // 16 u32 per depth tile
+    nk_size_t const tile_elements = tile_dim * depth_tile_size;
+    nk_size_t const depth_u32_total = nk_size_divide_round_up_(depth_bits, 32);
+
+    nk_u32_t const *b_tiles = (nk_u32_t const *)((char const *)b_packed + sizeof(nk_sets_smebi32_packed_header_t));
+    nk_u32_t const *b_norms = header->norms_offset ? (nk_u32_t const *)((char const *)b_packed + header->norms_offset)
+                                                   : (nk_u32_t const *)0;
+
+    svbool_t const predicate_all_f32x = svptrue_b32();
+    svfloat32_t const depth_f32x = svdup_f32((nk_f32_t)depth_bits);
+    svfloat32_t const half_f32x = svdup_f32(0.5f);
+    svfloat32_t const one_f32x = svdup_f32(1.0f);
+    svfloat32_t const zero_f32x = svdup_f32(0.0f);
+    nk_size_t const depth_in_bytes = nk_size_divide_round_up_(depth_bits, 8);
+    nk_size_t const row_tile_count_a = nk_size_divide_round_up_(row_count_a, tile_dim);
+
+    for (nk_size_t row_tile_a = 0; row_tile_a < row_tile_count_a; row_tile_a++) {
+        nk_size_t const row_start_a = row_tile_a * tile_dim;
+        nk_size_t const rows_a_remaining = (row_start_a + tile_dim <= row_count_a) ? tile_dim
+                                                                                   : (row_count_a - row_start_a);
+        svbool_t const row_predicate_f32x = svwhilelt_b32_u64(0u, rows_a_remaining);
+
+        // Compute A tile norms using streaming SVE popcount
+        NK_ALIGN64 nk_f32_t a_tile_norms[16];
+        for (nk_size_t r = 0; r < rows_a_remaining; r++) {
+            nk_u1x8_t const *a_row = (nk_u1x8_t const *)((char const *)a + (row_start_a + r) * a_stride_in_bytes);
+            a_tile_norms[r] = (nk_f32_t)nk_sets_reduce_sumsq_u1_streaming_(a_row, depth_in_bytes);
+        }
+
+        // Fast path: 3 B column tiles using ZA1-ZA3 (ZA0.S = staging)
+        nk_size_t row_tile_b = 0;
+        for (; row_tile_b + 3 <= row_tile_count_b; row_tile_b += 3) {
+            svzero_mask_za(nk_sme_zero_za32_tiles_123_);
+
+            for (nk_size_t d_tile = 0; d_tile < depth_tile_count; d_tile++) {
+                nk_size_t const d_start_u32 = d_tile * depth_tile_size;
+                nk_size_t const u32s_this_tile = (d_start_u32 + depth_tile_size <= depth_u32_total)
+                                                     ? depth_tile_size
+                                                     : (depth_u32_total > d_start_u32 ? depth_u32_total - d_start_u32
+                                                                                      : 0);
+                if (u32s_this_tile == 0) break;
+
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+
+                svbool_t const batch_predicate_f32x = svwhilelt_b32_u64(0u, u32s_this_tile);
+
+                // Load A rows into ZA0.S horizontally as u32 words
+                for (nk_size_t row_in_tile = 0; row_in_tile < rows_a_remaining; row_in_tile++) {
+                    nk_u32_t const *a_row_u32 = (nk_u32_t const *)((char const *)a +
+                                                                   (row_start_a + row_in_tile) * a_stride_in_bytes) +
+                                                d_start_u32;
+                    svld1_hor_za32(0, row_in_tile, batch_predicate_f32x, a_row_u32);
+                }
+
+                // B tile pointers for 3 column tiles
+                nk_u32_t const *b_tile0 = b_tiles + ((row_tile_b + 0) * depth_tile_count + d_tile) * tile_elements;
+                nk_u32_t const *b_tile1 = b_tiles + ((row_tile_b + 1) * depth_tile_count + d_tile) * tile_elements;
+                nk_u32_t const *b_tile2 = b_tiles + ((row_tile_b + 2) * depth_tile_count + d_tile) * tile_elements;
+
+                // Vertical read + BMOPA for each depth step
+                for (nk_size_t step = 0; step < u32s_this_tile; step++) {
+                    svuint32_t a_column_u32x = svread_ver_za32_u32_m(svdup_u32(0), row_predicate_f32x, 0, step);
+
+                    svbmopa_za32_u32_m(1, row_predicate_f32x, predicate_all_f32x, a_column_u32x,
+                                       svld1_u32(predicate_all_f32x, b_tile0 + step * tile_dim));
+                    svbmopa_za32_u32_m(2, row_predicate_f32x, predicate_all_f32x, a_column_u32x,
+                                       svld1_u32(predicate_all_f32x, b_tile1 + step * tile_dim));
+                    svbmopa_za32_u32_m(3, row_predicate_f32x, predicate_all_f32x, a_column_u32x,
+                                       svld1_u32(predicate_all_f32x, b_tile2 + step * tile_dim));
+                }
+            }
+
+            // Extract from ZA1-3: Jaccard normalization via streaming SVE
+            // Hoist B norms outside row loop (same for all A rows in this tile-pair)
+            svfloat32_t b_norms_0_f32x = svcvt_f32_u32_x(
+                predicate_all_f32x, svld1_u32(predicate_all_f32x, b_norms + (row_tile_b + 0) * tile_dim));
+            svfloat32_t b_norms_1_f32x = svcvt_f32_u32_x(
+                predicate_all_f32x, svld1_u32(predicate_all_f32x, b_norms + (row_tile_b + 1) * tile_dim));
+            svfloat32_t b_norms_2_f32x = svcvt_f32_u32_x(
+                predicate_all_f32x, svld1_u32(predicate_all_f32x, b_norms + (row_tile_b + 2) * tile_dim));
+
+            for (nk_size_t row = 0; row < rows_a_remaining; row++) {
+                nk_f32_t *c_row = (nk_f32_t *)((char *)c + (row_start_a + row) * c_stride_in_bytes);
+                svfloat32_t norm_a_f32x = svdup_f32(a_tile_norms[row]);
+
+                // ZA1
+                {
+                    svuint32_t za1_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_f32x, 1, row);
+                    svfloat32_t matching_f32x = svcvt_f32_u32_x(predicate_all_f32x, za1_u32x);
+                    svfloat32_t sum_norms_f32x = svadd_f32_x(predicate_all_f32x, norm_a_f32x, b_norms_0_f32x);
+                    svfloat32_t intersection_f32x = svmul_f32_x(
+                        predicate_all_f32x,
+                        svadd_f32_x(predicate_all_f32x, svsub_f32_x(predicate_all_f32x, sum_norms_f32x, depth_f32x),
+                                    matching_f32x),
+                        half_f32x);
+                    svfloat32_t union_val_f32x = svsub_f32_x(predicate_all_f32x, sum_norms_f32x, intersection_f32x);
+                    svbool_t nonzero_f32x = svcmpne_f32(predicate_all_f32x, union_val_f32x, zero_f32x);
+                    svfloat32_t ratio_f32x = svdiv_f32_x(predicate_all_f32x, intersection_f32x, union_val_f32x);
+                    svfloat32_t jaccard_f32x = svsel_f32(
+                        nonzero_f32x, svsub_f32_x(predicate_all_f32x, one_f32x, ratio_f32x), one_f32x);
+                    svst1_f32(predicate_all_f32x, c_row + (row_tile_b + 0) * tile_dim, jaccard_f32x);
+                }
+                // ZA2
+                {
+                    svuint32_t za2_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_f32x, 2, row);
+                    svfloat32_t matching_f32x = svcvt_f32_u32_x(predicate_all_f32x, za2_u32x);
+                    svfloat32_t sum_norms_f32x = svadd_f32_x(predicate_all_f32x, norm_a_f32x, b_norms_1_f32x);
+                    svfloat32_t intersection_f32x = svmul_f32_x(
+                        predicate_all_f32x,
+                        svadd_f32_x(predicate_all_f32x, svsub_f32_x(predicate_all_f32x, sum_norms_f32x, depth_f32x),
+                                    matching_f32x),
+                        half_f32x);
+                    svfloat32_t union_val_f32x = svsub_f32_x(predicate_all_f32x, sum_norms_f32x, intersection_f32x);
+                    svbool_t nonzero_f32x = svcmpne_f32(predicate_all_f32x, union_val_f32x, zero_f32x);
+                    svfloat32_t ratio_f32x = svdiv_f32_x(predicate_all_f32x, intersection_f32x, union_val_f32x);
+                    svfloat32_t jaccard_f32x = svsel_f32(
+                        nonzero_f32x, svsub_f32_x(predicate_all_f32x, one_f32x, ratio_f32x), one_f32x);
+                    svst1_f32(predicate_all_f32x, c_row + (row_tile_b + 1) * tile_dim, jaccard_f32x);
+                }
+                // ZA3
+                {
+                    svuint32_t za3_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_f32x, 3, row);
+                    svfloat32_t matching_f32x = svcvt_f32_u32_x(predicate_all_f32x, za3_u32x);
+                    svfloat32_t sum_norms_f32x = svadd_f32_x(predicate_all_f32x, norm_a_f32x, b_norms_2_f32x);
+                    svfloat32_t intersection_f32x = svmul_f32_x(
+                        predicate_all_f32x,
+                        svadd_f32_x(predicate_all_f32x, svsub_f32_x(predicate_all_f32x, sum_norms_f32x, depth_f32x),
+                                    matching_f32x),
+                        half_f32x);
+                    svfloat32_t union_val_f32x = svsub_f32_x(predicate_all_f32x, sum_norms_f32x, intersection_f32x);
+                    svbool_t nonzero_f32x = svcmpne_f32(predicate_all_f32x, union_val_f32x, zero_f32x);
+                    svfloat32_t ratio_f32x = svdiv_f32_x(predicate_all_f32x, intersection_f32x, union_val_f32x);
+                    svfloat32_t jaccard_f32x = svsel_f32(
+                        nonzero_f32x, svsub_f32_x(predicate_all_f32x, one_f32x, ratio_f32x), one_f32x);
+                    svst1_f32(predicate_all_f32x, c_row + (row_tile_b + 2) * tile_dim, jaccard_f32x);
+                }
+            }
+        }
+
+        // Remainder: 1 B column tile at a time using ZA1
+        for (; row_tile_b < row_tile_count_b; row_tile_b++) {
+            nk_size_t const row_start_b = row_tile_b * tile_dim;
+            nk_size_t const rows_b_remaining = (row_start_b + tile_dim <= row_count_b) ? tile_dim
+                                                                                       : (row_count_b - row_start_b);
+            svbool_t const column_predicate_f32x = svwhilelt_b32_u64(0u, rows_b_remaining);
+
+            svzero_mask_za(nk_sme_zero_za32_tile_1_);
+
+            for (nk_size_t d_tile = 0; d_tile < depth_tile_count; d_tile++) {
+                nk_size_t const d_start_u32 = d_tile * depth_tile_size;
+                nk_size_t const u32s_this_tile = (d_start_u32 + depth_tile_size <= depth_u32_total)
+                                                     ? depth_tile_size
+                                                     : (depth_u32_total > d_start_u32 ? depth_u32_total - d_start_u32
+                                                                                      : 0);
+                if (u32s_this_tile == 0) break;
+
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+
+                svbool_t const batch_predicate_f32x = svwhilelt_b32_u64(0u, u32s_this_tile);
+
+                // Load A rows into ZA0.S horizontally
+                for (nk_size_t row_in_tile = 0; row_in_tile < rows_a_remaining; row_in_tile++) {
+                    nk_u32_t const *a_row_u32 = (nk_u32_t const *)((char const *)a +
+                                                                   (row_start_a + row_in_tile) * a_stride_in_bytes) +
+                                                d_start_u32;
+                    svld1_hor_za32(0, row_in_tile, batch_predicate_f32x, a_row_u32);
+                }
+
+                nk_u32_t const *b_tile = b_tiles + (row_tile_b * depth_tile_count + d_tile) * tile_elements;
+
+                // Vertical read + BMOPA
+                for (nk_size_t step = 0; step < u32s_this_tile; step++) {
+                    svuint32_t a_column_u32x = svread_ver_za32_u32_m(svdup_u32(0), row_predicate_f32x, 0, step);
+                    svuint32_t b_u32x = svld1_u32(predicate_all_f32x, b_tile + step * tile_dim);
+                    svbmopa_za32_u32_m(1, row_predicate_f32x, column_predicate_f32x, a_column_u32x, b_u32x);
+                }
+            }
+
+            // Extract from ZA1: Jaccard normalization
+            svfloat32_t b_norms_f32x = svcvt_f32_u32_x(predicate_all_f32x,
+                                                       svld1_u32(predicate_all_f32x, b_norms + row_start_b));
+            for (nk_size_t row = 0; row < rows_a_remaining; row++) {
+                svuint32_t za1_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_f32x, 1, row);
+                svfloat32_t matching_f32x = svcvt_f32_u32_x(predicate_all_f32x, za1_u32x);
+                svfloat32_t norm_a_f32x = svdup_f32(a_tile_norms[row]);
+                svfloat32_t sum_norms_f32x = svadd_f32_x(predicate_all_f32x, norm_a_f32x, b_norms_f32x);
+                svfloat32_t intersection_f32x = svmul_f32_x(
+                    predicate_all_f32x,
+                    svadd_f32_x(predicate_all_f32x, svsub_f32_x(predicate_all_f32x, sum_norms_f32x, depth_f32x),
+                                matching_f32x),
+                    half_f32x);
+                svfloat32_t union_val_f32x = svsub_f32_x(predicate_all_f32x, sum_norms_f32x, intersection_f32x);
+                svbool_t nonzero_f32x = svcmpne_f32(predicate_all_f32x, union_val_f32x, zero_f32x);
+                svfloat32_t ratio_f32x = svdiv_f32_x(predicate_all_f32x, intersection_f32x, union_val_f32x);
+                svfloat32_t jaccard_f32x = svsel_f32(nonzero_f32x,
+                                                     svsub_f32_x(predicate_all_f32x, one_f32x, ratio_f32x), one_f32x);
+                nk_f32_t *c_row = (nk_f32_t *)((char *)c + (row_start_a + row) * c_stride_in_bytes);
+                svst1_f32(column_predicate_f32x, c_row + row_start_b, jaccard_f32x);
+            }
+        }
+    }
+}
+
+NK_PUBLIC void nk_jaccards_packed_u1_smebi32(nk_u1x8_t const *a, void const *b_packed, nk_f32_t *c,
+                                             nk_size_t row_count_a, nk_size_t row_count_b, nk_size_t depth_bits,
+                                             nk_size_t a_stride_in_bytes, nk_size_t c_stride_in_bytes) {
+    nk_jaccards_packed_u1_smebi32_streaming_(a, b_packed, c, row_count_a, row_count_b, depth_bits, a_stride_in_bytes,
+                                             c_stride_in_bytes);
+}
+
+/**
+ *  Symmetric Jaccard kernel using ZA0 time-sharing + 3-tile fast path.
+ *  Fills upper triangle only (column_tile >= row_tile); caller sees result[i][j] for j >= i.
+ *  Norms computed on-the-fly using streaming SVE popcount.
+ */
+__arm_locally_streaming __arm_new("za") static void nk_jaccards_symmetric_u1_smebi32_streaming_(
+    nk_u1x8_t const *vectors, nk_size_t n_vectors, nk_size_t depth_bits, nk_size_t stride, nk_f32_t *result,
+    nk_size_t result_stride, nk_size_t row_start, nk_size_t row_count) {
+
+    nk_size_t const tile_dim = svcntw();        // 16 for 512-bit SVL
+    nk_size_t const depth_tile_size = svcntw(); // 16 u32 per depth tile
+    nk_size_t const depth_u32_total = nk_size_divide_round_up_(depth_bits, 32);
+    nk_size_t const depth_tile_count = nk_size_divide_round_up_(depth_u32_total, depth_tile_size);
+    nk_size_t const depth_in_bytes = nk_size_divide_round_up_(depth_bits, NK_BITS_PER_BYTE);
+
+    svbool_t const predicate_all_f32x = svptrue_b32();
+    svfloat32_t const depth_f32x = svdup_f32((nk_f32_t)depth_bits);
+    svfloat32_t const half_f32x = svdup_f32(0.5f);
+    svfloat32_t const one_f32x = svdup_f32(1.0f);
+    svfloat32_t const zero_f32x = svdup_f32(0.0f);
+
+    NK_ALIGN64 nk_u32_t a_buffer[16][16]; // Stack buffer for A column save
+
+    nk_size_t const row_end = row_start + row_count;
+    nk_size_t const column_tile_count = nk_size_divide_round_up_(n_vectors, tile_dim);
+
+    for (nk_size_t row_tile_start = row_start; row_tile_start < row_end && row_tile_start < n_vectors;
+         row_tile_start += tile_dim) {
+        nk_size_t const rows_remaining = (row_tile_start + tile_dim <= row_end) ? tile_dim : (row_end - row_tile_start);
+        nk_size_t const rows_clamped = (row_tile_start + rows_remaining <= n_vectors) ? rows_remaining
+                                                                                      : (n_vectors - row_tile_start);
+        svbool_t const row_predicate_f32x = svwhilelt_b32_u64(0u, rows_clamped);
+
+        // Compute A tile norms
+        NK_ALIGN64 nk_f32_t a_tile_norms[16];
+        for (nk_size_t r = 0; r < rows_clamped; r++) {
+            nk_u1x8_t const *a_row = (nk_u1x8_t const *)((char const *)vectors + (row_tile_start + r) * stride);
+            a_tile_norms[r] = (nk_f32_t)nk_sets_reduce_sumsq_u1_streaming_(a_row, depth_in_bytes);
+        }
+        for (nk_size_t r = rows_clamped; r < tile_dim; r++) a_tile_norms[r] = 0.0f;
+
+        // Upper triangle: start from this row tile's column
+        nk_size_t column_tile_index = row_tile_start / tile_dim;
+
+        // Fast path: 3 column tiles using ZA1-ZA3 (ZA0 = staging)
+        for (; column_tile_index + 3 <= column_tile_count; column_tile_index += 3) {
+            svzero_mask_za(nk_sme_zero_za32_tiles_123_);
+
+            for (nk_size_t d_tile = 0; d_tile < depth_tile_count; d_tile++) {
+                nk_size_t const d_start_u32 = d_tile * depth_tile_size;
+                nk_size_t const u32s_this_tile = (d_start_u32 + depth_tile_size <= depth_u32_total)
+                                                     ? depth_tile_size
+                                                     : (depth_u32_total > d_start_u32 ? depth_u32_total - d_start_u32
+                                                                                      : 0);
+                if (u32s_this_tile == 0) break;
+
+                // Load A rows into ZA0 horizontally
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+                svbool_t const batch_predicate_f32x = svwhilelt_b32_u64(0u, u32s_this_tile);
+
+                for (nk_size_t row_in_tile = 0; row_in_tile < rows_clamped; row_in_tile++) {
+                    nk_u32_t const *a_row_u32 = (nk_u32_t const *)((char const *)vectors +
+                                                                   (row_tile_start + row_in_tile) * stride) +
+                                                d_start_u32;
+                    svld1_hor_za32(0, row_in_tile, batch_predicate_f32x, a_row_u32);
+                }
+
+                // Save A columns from ZA0 to stack buffer
+                for (nk_size_t s = 0; s < u32s_this_tile; s++)
+                    svst1_u32(predicate_all_f32x, a_buffer[s],
+                              svread_ver_za32_u32_m(svdup_u32(0), row_predicate_f32x, 0, s));
+
+                // B column tile 0
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+                for (nk_size_t col = 0; col < tile_dim; col++) {
+                    nk_size_t const col_abs = (column_tile_index + 0) * tile_dim + col;
+                    if (col_abs < n_vectors) {
+                        nk_u32_t const *b_row = (nk_u32_t const *)((char const *)vectors + col_abs * stride) +
+                                                d_start_u32;
+                        svld1_hor_za32(0, col, batch_predicate_f32x, b_row);
+                    }
+                }
+                for (nk_size_t step = 0; step < u32s_this_tile; step++) {
+                    svuint32_t a_u32x = svld1_u32(predicate_all_f32x, a_buffer[step]);
+                    svuint32_t b_u32x = svread_ver_za32_u32_m(svdup_u32(0), predicate_all_f32x, 0, step);
+                    svbmopa_za32_u32_m(1, row_predicate_f32x, predicate_all_f32x, a_u32x, b_u32x);
+                }
+
+                // B column tile 1
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+                for (nk_size_t col = 0; col < tile_dim; col++) {
+                    nk_size_t const col_abs = (column_tile_index + 1) * tile_dim + col;
+                    if (col_abs < n_vectors) {
+                        nk_u32_t const *b_row = (nk_u32_t const *)((char const *)vectors + col_abs * stride) +
+                                                d_start_u32;
+                        svld1_hor_za32(0, col, batch_predicate_f32x, b_row);
+                    }
+                }
+                for (nk_size_t step = 0; step < u32s_this_tile; step++) {
+                    svuint32_t a_u32x = svld1_u32(predicate_all_f32x, a_buffer[step]);
+                    svuint32_t b_u32x = svread_ver_za32_u32_m(svdup_u32(0), predicate_all_f32x, 0, step);
+                    svbmopa_za32_u32_m(2, row_predicate_f32x, predicate_all_f32x, a_u32x, b_u32x);
+                }
+
+                // B column tile 2
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+                for (nk_size_t col = 0; col < tile_dim; col++) {
+                    nk_size_t const col_abs = (column_tile_index + 2) * tile_dim + col;
+                    if (col_abs < n_vectors) {
+                        nk_u32_t const *b_row = (nk_u32_t const *)((char const *)vectors + col_abs * stride) +
+                                                d_start_u32;
+                        svld1_hor_za32(0, col, batch_predicate_f32x, b_row);
+                    }
+                }
+                for (nk_size_t step = 0; step < u32s_this_tile; step++) {
+                    svuint32_t a_u32x = svld1_u32(predicate_all_f32x, a_buffer[step]);
+                    svuint32_t b_u32x = svread_ver_za32_u32_m(svdup_u32(0), predicate_all_f32x, 0, step);
+                    svbmopa_za32_u32_m(3, row_predicate_f32x, predicate_all_f32x, a_u32x, b_u32x);
+                }
+            }
+
+            // Compute B tile norms for 3 column tiles
+            NK_ALIGN64 nk_u32_t b_tile_norms_0[16];
+            NK_ALIGN64 nk_u32_t b_tile_norms_1[16];
+            NK_ALIGN64 nk_u32_t b_tile_norms_2[16];
+            for (nk_size_t col = 0; col < tile_dim; col++) {
+                nk_size_t const col_abs_0 = (column_tile_index + 0) * tile_dim + col;
+                nk_size_t const col_abs_1 = (column_tile_index + 1) * tile_dim + col;
+                nk_size_t const col_abs_2 = (column_tile_index + 2) * tile_dim + col;
+                b_tile_norms_0[col] = (col_abs_0 < n_vectors)
+                                          ? nk_sets_reduce_sumsq_u1_streaming_(
+                                                (nk_u1x8_t const *)((char const *)vectors + col_abs_0 * stride),
+                                                depth_in_bytes)
+                                          : 0;
+                b_tile_norms_1[col] = (col_abs_1 < n_vectors)
+                                          ? nk_sets_reduce_sumsq_u1_streaming_(
+                                                (nk_u1x8_t const *)((char const *)vectors + col_abs_1 * stride),
+                                                depth_in_bytes)
+                                          : 0;
+                b_tile_norms_2[col] = (col_abs_2 < n_vectors)
+                                          ? nk_sets_reduce_sumsq_u1_streaming_(
+                                                (nk_u1x8_t const *)((char const *)vectors + col_abs_2 * stride),
+                                                depth_in_bytes)
+                                          : 0;
+            }
+
+            // Extract ZA1-3: Jaccard normalization
+            svfloat32_t b_norms_0_f32x = svcvt_f32_u32_x(predicate_all_f32x,
+                                                         svld1_u32(predicate_all_f32x, b_tile_norms_0));
+            svfloat32_t b_norms_1_f32x = svcvt_f32_u32_x(predicate_all_f32x,
+                                                         svld1_u32(predicate_all_f32x, b_tile_norms_1));
+            svfloat32_t b_norms_2_f32x = svcvt_f32_u32_x(predicate_all_f32x,
+                                                         svld1_u32(predicate_all_f32x, b_tile_norms_2));
+
+            for (nk_size_t row = 0; row < rows_clamped; row++) {
+                nk_f32_t *c_row = (nk_f32_t *)((char *)result + (row_tile_start + row) * result_stride);
+                svfloat32_t norm_a_f32x = svdup_f32(a_tile_norms[row]);
+
+                // ZA1
+                {
+                    svuint32_t za1_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_f32x, 1, row);
+                    svfloat32_t matching_f32x = svcvt_f32_u32_x(predicate_all_f32x, za1_u32x);
+                    svfloat32_t sum_norms_f32x = svadd_f32_x(predicate_all_f32x, norm_a_f32x, b_norms_0_f32x);
+                    svfloat32_t intersection_f32x = svmul_f32_x(
+                        predicate_all_f32x,
+                        svadd_f32_x(predicate_all_f32x, svsub_f32_x(predicate_all_f32x, sum_norms_f32x, depth_f32x),
+                                    matching_f32x),
+                        half_f32x);
+                    svfloat32_t union_val_f32x = svsub_f32_x(predicate_all_f32x, sum_norms_f32x, intersection_f32x);
+                    svbool_t nonzero_f32x = svcmpne_f32(predicate_all_f32x, union_val_f32x, zero_f32x);
+                    svfloat32_t ratio_f32x = svdiv_f32_x(predicate_all_f32x, intersection_f32x, union_val_f32x);
+                    svfloat32_t jaccard_f32x = svsel_f32(
+                        nonzero_f32x, svsub_f32_x(predicate_all_f32x, one_f32x, ratio_f32x), one_f32x);
+                    svst1_f32(predicate_all_f32x, c_row + (column_tile_index + 0) * tile_dim, jaccard_f32x);
+                }
+                // ZA2
+                {
+                    svuint32_t za2_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_f32x, 2, row);
+                    svfloat32_t matching_f32x = svcvt_f32_u32_x(predicate_all_f32x, za2_u32x);
+                    svfloat32_t sum_norms_f32x = svadd_f32_x(predicate_all_f32x, norm_a_f32x, b_norms_1_f32x);
+                    svfloat32_t intersection_f32x = svmul_f32_x(
+                        predicate_all_f32x,
+                        svadd_f32_x(predicate_all_f32x, svsub_f32_x(predicate_all_f32x, sum_norms_f32x, depth_f32x),
+                                    matching_f32x),
+                        half_f32x);
+                    svfloat32_t union_val_f32x = svsub_f32_x(predicate_all_f32x, sum_norms_f32x, intersection_f32x);
+                    svbool_t nonzero_f32x = svcmpne_f32(predicate_all_f32x, union_val_f32x, zero_f32x);
+                    svfloat32_t ratio_f32x = svdiv_f32_x(predicate_all_f32x, intersection_f32x, union_val_f32x);
+                    svfloat32_t jaccard_f32x = svsel_f32(
+                        nonzero_f32x, svsub_f32_x(predicate_all_f32x, one_f32x, ratio_f32x), one_f32x);
+                    svst1_f32(predicate_all_f32x, c_row + (column_tile_index + 1) * tile_dim, jaccard_f32x);
+                }
+                // ZA3
+                {
+                    svuint32_t za3_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_f32x, 3, row);
+                    svfloat32_t matching_f32x = svcvt_f32_u32_x(predicate_all_f32x, za3_u32x);
+                    svfloat32_t sum_norms_f32x = svadd_f32_x(predicate_all_f32x, norm_a_f32x, b_norms_2_f32x);
+                    svfloat32_t intersection_f32x = svmul_f32_x(
+                        predicate_all_f32x,
+                        svadd_f32_x(predicate_all_f32x, svsub_f32_x(predicate_all_f32x, sum_norms_f32x, depth_f32x),
+                                    matching_f32x),
+                        half_f32x);
+                    svfloat32_t union_val_f32x = svsub_f32_x(predicate_all_f32x, sum_norms_f32x, intersection_f32x);
+                    svbool_t nonzero_f32x = svcmpne_f32(predicate_all_f32x, union_val_f32x, zero_f32x);
+                    svfloat32_t ratio_f32x = svdiv_f32_x(predicate_all_f32x, intersection_f32x, union_val_f32x);
+                    svfloat32_t jaccard_f32x = svsel_f32(
+                        nonzero_f32x, svsub_f32_x(predicate_all_f32x, one_f32x, ratio_f32x), one_f32x);
+                    svst1_f32(predicate_all_f32x, c_row + (column_tile_index + 2) * tile_dim, jaccard_f32x);
+                }
+            }
+        }
+
+        // Remainder: 1 column tile at a time using ZA1
+        for (; column_tile_index < column_tile_count; column_tile_index++) {
+            nk_size_t const col_tile_start = column_tile_index * tile_dim;
+            nk_size_t const cols_remaining = (col_tile_start + tile_dim <= n_vectors) ? tile_dim
+                                                                                      : (n_vectors - col_tile_start);
+            svbool_t const column_predicate_f32x = svwhilelt_b32_u64(0u, cols_remaining);
+
+            svzero_mask_za(nk_sme_zero_za32_tile_1_);
+
+            for (nk_size_t d_tile = 0; d_tile < depth_tile_count; d_tile++) {
+                nk_size_t const d_start_u32 = d_tile * depth_tile_size;
+                nk_size_t const u32s_this_tile = (d_start_u32 + depth_tile_size <= depth_u32_total)
+                                                     ? depth_tile_size
+                                                     : (depth_u32_total > d_start_u32 ? depth_u32_total - d_start_u32
+                                                                                      : 0);
+                if (u32s_this_tile == 0) break;
+
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+                svbool_t const batch_predicate_f32x = svwhilelt_b32_u64(0u, u32s_this_tile);
+
+                // Load A rows into ZA0 horizontally
+                for (nk_size_t row_in_tile = 0; row_in_tile < rows_clamped; row_in_tile++) {
+                    nk_u32_t const *a_row_u32 = (nk_u32_t const *)((char const *)vectors +
+                                                                   (row_tile_start + row_in_tile) * stride) +
+                                                d_start_u32;
+                    svld1_hor_za32(0, row_in_tile, batch_predicate_f32x, a_row_u32);
+                }
+
+                // Save A columns from ZA0 to stack buffer
+                for (nk_size_t s = 0; s < u32s_this_tile; s++)
+                    svst1_u32(predicate_all_f32x, a_buffer[s],
+                              svread_ver_za32_u32_m(svdup_u32(0), row_predicate_f32x, 0, s));
+
+                // Load B column tile into ZA0
+                svzero_mask_za(nk_sme_zero_za32_tile_0_);
+                for (nk_size_t col = 0; col < tile_dim; col++) {
+                    nk_size_t const col_abs = col_tile_start + col;
+                    if (col_abs < n_vectors) {
+                        nk_u32_t const *b_row = (nk_u32_t const *)((char const *)vectors + col_abs * stride) +
+                                                d_start_u32;
+                        svld1_hor_za32(0, col, batch_predicate_f32x, b_row);
+                    }
+                }
+                for (nk_size_t step = 0; step < u32s_this_tile; step++) {
+                    svuint32_t a_u32x = svld1_u32(predicate_all_f32x, a_buffer[step]);
+                    svuint32_t b_u32x = svread_ver_za32_u32_m(svdup_u32(0), column_predicate_f32x, 0, step);
+                    svbmopa_za32_u32_m(1, row_predicate_f32x, column_predicate_f32x, a_u32x, b_u32x);
+                }
+            }
+
+            // Compute B tile norms for remainder tile
+            NK_ALIGN64 nk_u32_t b_tile_norms[16];
+            for (nk_size_t col = 0; col < tile_dim; col++) {
+                nk_size_t const col_abs = col_tile_start + col;
+                b_tile_norms[col] = (col_abs < n_vectors)
+                                        ? nk_sets_reduce_sumsq_u1_streaming_(
+                                              (nk_u1x8_t const *)((char const *)vectors + col_abs * stride),
+                                              depth_in_bytes)
+                                        : 0;
+            }
+
+            svfloat32_t b_norms_f32x = svcvt_f32_u32_x(predicate_all_f32x, svld1_u32(predicate_all_f32x, b_tile_norms));
+            for (nk_size_t row = 0; row < rows_clamped; row++) {
+                svuint32_t za1_u32x = svread_hor_za32_u32_m(svdup_u32(0), predicate_all_f32x, 1, row);
+                svfloat32_t matching_f32x = svcvt_f32_u32_x(predicate_all_f32x, za1_u32x);
+                svfloat32_t norm_a_f32x = svdup_f32(a_tile_norms[row]);
+                svfloat32_t sum_norms_f32x = svadd_f32_x(predicate_all_f32x, norm_a_f32x, b_norms_f32x);
+                svfloat32_t intersection_f32x = svmul_f32_x(
+                    predicate_all_f32x,
+                    svadd_f32_x(predicate_all_f32x, svsub_f32_x(predicate_all_f32x, sum_norms_f32x, depth_f32x),
+                                matching_f32x),
+                    half_f32x);
+                svfloat32_t union_val_f32x = svsub_f32_x(predicate_all_f32x, sum_norms_f32x, intersection_f32x);
+                svbool_t nonzero_f32x = svcmpne_f32(predicate_all_f32x, union_val_f32x, zero_f32x);
+                svfloat32_t ratio_f32x = svdiv_f32_x(predicate_all_f32x, intersection_f32x, union_val_f32x);
+                svfloat32_t jaccard_f32x = svsel_f32(nonzero_f32x,
+                                                     svsub_f32_x(predicate_all_f32x, one_f32x, ratio_f32x), one_f32x);
+                nk_f32_t *c_row = (nk_f32_t *)((char *)result + (row_tile_start + row) * result_stride);
+                svst1_f32(column_predicate_f32x, c_row + col_tile_start, jaccard_f32x);
+            }
+        }
+    }
+}
+
+NK_PUBLIC void nk_jaccards_symmetric_u1_smebi32(nk_u1x8_t const *vectors, nk_size_t n_vectors, nk_size_t depth_bits,
+                                                nk_size_t stride, nk_f32_t *result, nk_size_t result_stride,
+                                                nk_size_t row_start, nk_size_t row_count) {
+    nk_jaccards_symmetric_u1_smebi32_streaming_(vectors, n_vectors, depth_bits, stride, result, result_stride,
+                                                row_start, row_count);
+}
+
+#pragma endregion // Jaccard Distance
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_SMEBI32
+#endif // NK_TARGET_ARM_
+
+#endif // NK_SETS_SMEBI32_H