        // __m512i sixteen = _mm512_set1_epi32(16);

        // __m512i curr_indices_0 = _mm512_set_epi32(
        //         0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
        // __m512i curr_indices_1 = _mm512_set_epi32(
        //         16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30,
        //         31);
        // __m512i curr_indices_2 = _mm512_set_epi32(
        //         32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46,
        //         47);
        // __m512i curr_indices_3 = _mm512_set_epi32(
        //         48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62,
        //         63);

//3857.4, 9174.3, 8142.1

        //         __mmask16 m0 = mask64 & 0xFFFF;
        //         __mmask16 m1 = (mask64 >> 16) & 0xFFFF;
        //         __mmask16 m2 = (mask64 >> 32) & 0xFFFF;
        //         __mmask16 m3 = (mask64 >> 48) & 0xFFFF;

        //         size_t n0 = _mm_popcnt_u32(m0);
        //         size_t n1 = _mm_popcnt_u32(m1);
        //         size_t n2 = _mm_popcnt_u32(m2);
        //         size_t n3 = _mm_popcnt_u32(m3);

        //         _mm512_mask_compressstoreu_epi32(
        //                 indices_ptr, m0, curr_indices_0);
        //         _mm512_mask_compressstoreu_epi32(
        //                 indices_ptr + n0, m1, curr_indices_1);
        //         _mm512_mask_compressstoreu_epi32(
        //                 indices_ptr + n0 + n1, m2, curr_indices_2);
        //         _mm512_mask_compressstoreu_epi32(
        //                 indices_ptr + n0 + n1 + n2, m3, curr_indices_3);

        //         indices_ptr += num_active;


        //     // By now, batch_storage will hold the codes for m_level_width dims
        //     // and batch_size vectors. We need to vertically add LUT gathers.
        //     constexpr size_t sub_batch_size = 16;
        //     batch_storage_ptr = batch_storage;
        //     float* exact_distances_ptr = exact_distances;
        //     for (size_t sub_batch = 0; sub_batch < batch_size;
        //          sub_batch += sub_batch_size) {
        //         __m512 dist = _mm512_set1_ps(0);
        //         for (size_t m = 0; m < m_level_width; m++) {
        //             // 1: Load 16 int8 codes
        //             __m128i codes_batch =
        //                     _mm_loadu_si128((__m128i*)(batch_storage_ptr));
        //             batch_storage_ptr += sub_batch_size;

        //             // 2: Convert to int32
        //             __m512i codes_batch_int = _mm512_cvtepu8_epi32(codes_batch);

        //             // 3: Add offset (m * ksub)
        //             __m512i offset = _mm512_set1_epi32(m * pq.ksub);
        //             __m512i indices_to_read_from =
        //                     _mm512_add_epi32(codes_batch_int, offset);

        //             // 4: Gather 16 distances
        //             __m512 m_dist = _mm512_i32gather_ps(
        //                     indices_to_read_from, sim_table, sizeof(float));

        //             // 5: Vertically add to dist
        //             dist = _mm512_add_ps(dist, m_dist);
        //         }

        //         // Add to exact distances
        //         __m512 exact_distances_batch =
        //                 _mm512_loadu_ps(exact_distances_ptr);
        //         __m512 new_exact_distances =
        //                 _mm512_add_ps(exact_distances_batch, dist);
        //         _mm512_storeu_ps(exact_distances_ptr, new_exact_distances);
        //         exact_distances_ptr += sub_batch_size;
        //     }
        // }

        // OLD CODE FOR REFERENCE
        // size_t next_num_active = 0;
        // for (size_t j = 0; j < *num_active; j++) {
        //     size_t row = active_indices[j];
        //     we may have to skip rows!!
        //     if (res.skip_entry(row)) {
        //         continue;
        //     }
        //     float distance = distance_single_code<PQDecoder>(
        //                         pq.M / n_levels,
        //                         pq.nbits,
        //                         sim_table + level * pq.ksub * pq.M /
        //                         n_levels, codes + row * pq.code_size /
        //                         n_levels);

        //     float new_exact = exact_distances[j] + distance -
        //     ((cum_sums_level - ncode)[row] * (cum_sums_level - ncode)[row] -
        //     cum_sums_level[row] * cum_sums_level[row]);

        //     float cauchy_schwarz_bound = 2.0f * cum_sums_level[row] *
        //     query_cum_norm; float lower_bound = new_exact -
        //     cauchy_schwarz_bound; float upper_bound = new_exact +
        //     cauchy_schwarz_bound;

        //     res.add(row, dis0 + cum_sums_level_0[row] * cum_sums_level_0[row]
        //     + upper_bound);

        //     if (res.should_keep(dis0 + lower_bound + cum_sums_level_0[row] *
        //     cum_sums_level_0[row])) {
        //         active_indices[next_num_active] = row;
        //         exact_distances[next_num_active] = new_exact;
        //         next_num_active++;
        //     }
        // }

        // *num_active = next_num_active;
