scann-core 0.2.1

Rust bindings for scann-core: ScaNN's approximate nearest-neighbour vector search core as a standalone library (derived from Google's ScaNN; not an official Google product)
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
// Copyright 2026 The Google Research Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// Modified in 2026 for scann-core (a derived work of ScaNN, not an official
// Google product). Includes changes contributed by Arm (Gerda Zsejke More,
// Li Zhang) in google-research PR #3374 and follow-ups; see NOTICE.

#include "scann/hashes/internal/asymmetric_hashing_impl.h"

#include <algorithm>
#include <array>
#include <cmath>
#include <cstdint>
#include <cstring>
#include <memory>
#include <numeric>
#include <random>
#include <type_traits>
#include <utility>

#include "absl/random/distributions.h"
#include "absl/status/status.h"
#include "scann/data_format/datapoint.h"
#include "scann/data_format/dataset.h"
#include "scann/distance_measures/one_to_one/dot_product.h"
#include "scann/distance_measures/one_to_many/one_to_many.h"
#include "scann/distance_measures/one_to_many/one_to_many_symmetric.h"
#include "scann/hashes/internal/asymmetric_hashing_impl_neon.h"
#include "scann/hashes/internal/asymmetric_hashing_postprocess.h"
#include "scann/oss_wrappers/scann_random.h"
#include "scann/oss_wrappers/scann_status.h"
#include "scann/projection/chunking_projection.h"
#include "scann/utils/common.h"
#include "scann/utils/gmm_utils.h"
#include "scann/utils/intrinsics/flags.h"
#include "scann/utils/top_n_amortized_constant.h"
#include "scann/utils/types.h"

namespace research_scann {
namespace asymmetric_hashing_internal {

template <typename T>
StatusOr<vector<DenseDataset<double>>> AhImpl<T>::TrainAsymmetricHashing(
    const TypedDataset<T>& dataset, const TrainingOptionsT& opts,
    shared_ptr<ThreadPool> pool) {
  if (dataset.empty()) {
    return InvalidArgumentError("Cannot train AH on an empty dataset.");
  }

  ChunkedDatapoint<double> chunked_vec;

  SCANN_RET_CHECK(opts.projector());
  if (opts.preprocessing_function()) {
    SCANN_ASSIGN_OR_RETURN(Datapoint<T> preprocessed,
                           opts.preprocessing_function()(dataset[0]));
    SCANN_RETURN_IF_ERROR(
        opts.projector()->ProjectInput(preprocessed.ToPtr(), &chunked_vec));
  } else {
    SCANN_RETURN_IF_ERROR(
        opts.projector()->ProjectInput(dataset[0], &chunked_vec));
  }
  int32_t num_blocks = chunked_vec.size();
  vector<DenseDataset<double>> chunked_dataset(num_blocks);

  const float sampling_fraction =
      opts.config().has_expected_sample_size()
          ? std::min(1.0,
                     static_cast<double>(opts.config().expected_sample_size()) /
                         static_cast<double>(dataset.size()))
          : opts.config().sampling_fraction();

  ConstSpan<float> weights = opts.weights();
  if (sampling_fraction == 1.0) {
    for (int32_t i = 0; i < num_blocks; ++i) {
      DenseDataset<double>& ds = chunked_dataset[i];
      ds.set_dimensionality(chunked_vec[i].dimensionality());
      ds.Reserve(dataset.size());
    }
  }

  MTRandom rng(kDeterministicSeed * (opts.config().sampling_seed() + 1));
  vector<DatapointIndex> sample;
  for (DatapointIndex i = 0; i < dataset.size(); ++i) {
    if (absl::Uniform<double>(rng, 0, 1.0) < sampling_fraction) {
      sample.push_back(i);
    }
  }

  if (sample.size() > opts.config().max_sample_size()) {
    std::shuffle(sample.begin(), sample.end(), rng);
    sample.resize(opts.config().max_sample_size());
    std::sort(sample.begin(), sample.end());
  }

  if (sample.size() < opts.config().num_clusters_per_block()) {
    return InvalidArgumentError(absl::StrCat(
        "Number of clusters per block (",
        opts.config().num_clusters_per_block(),
        ") is greater than asymmetric hashing training data size (",
        sample.size(), ")."));
  }

  vector<float> sample_weight_storage;
  if (sample.size() < dataset.size() && !weights.empty()) {
    sample_weight_storage.reserve(sample.size());
    for (DatapointIndex dp_idx : sample) {
      sample_weight_storage.push_back(weights[dp_idx]);
    }
    weights = sample_weight_storage;
  }

  auto append_chunked_blocks = [&] {
    for (size_t j = 0; j < num_blocks; ++j) {
      chunked_dataset[j].AppendOrDie(chunked_vec[j], "");
    }
  };

  if (opts.preprocessing_function()) {
    for (DatapointIndex i : sample) {
      SCANN_RETURN_IF_ERROR(VerifyAllFinite(dataset[i].values_span()));
      SCANN_ASSIGN_OR_RETURN(Datapoint<T> preprocessed,
                             opts.preprocessing_function()(dataset[i]));
      SCANN_RETURN_IF_ERROR(VerifyAllFinite(preprocessed.values()));
      SCANN_RETURN_IF_ERROR(
          opts.projector()->ProjectInput(preprocessed.ToPtr(), &chunked_vec));
      append_chunked_blocks();
    }
  } else {
    for (DatapointIndex i : sample) {
      SCANN_RETURN_IF_ERROR(VerifyAllFinite(dataset[i].values_span()));
      SCANN_RETURN_IF_ERROR(
          opts.projector()->ProjectInput(dataset[i], &chunked_vec));
      append_chunked_blocks();
    }
  }

  const auto& quantization_distance = opts.quantization_distance();
  GmmUtils::Options gmm_opts;
  gmm_opts.seed = opts.config().clustering_seed();
  gmm_opts.max_iterations = opts.config().max_clustering_iterations();
  gmm_opts.epsilon = opts.config().clustering_convergence_tolerance();
  gmm_opts.parallelization_pool = std::move(pool);
  gmm_opts.partition_assignment_type = gmm_opts.UNBALANCED_FLOAT32;
  GmmUtils gmm(quantization_distance, gmm_opts);

  vector<DenseDataset<double>> all_centers(num_blocks);
  for (size_t i : Seq(num_blocks)) {
    DenseDataset<double> centers;
    vector<vector<DatapointIndex>> subpartitions;
    SCANN_RETURN_IF_ERROR(gmm.ComputeKmeansClustering(
        chunked_dataset[i], opts.config().num_clusters_per_block(), &centers,
        {.final_partitions = &subpartitions, .weights = weights}));

    for (size_t center_idx : IndicesOf(centers)) {
      SCANN_RETURN_IF_ERROR(VerifyAllFinite(centers[center_idx].values_span()));
      if (!opts.config().use_norm_biasing_correction()) continue;
      SCANN_ASSIGN_OR_RETURN(
          const double norm_bias_correction,
          ComputeNormBiasCorrection(chunked_dataset[i], centers[center_idx],
                                    subpartitions[center_idx]));
      SCANN_RET_CHECK(std::isfinite(norm_bias_correction))
          << norm_bias_correction;
      for (double& d : centers.mutable_data(center_idx)) {
        d *= norm_bias_correction;
      }
    }

    chunked_dataset[i].clear();
    chunked_dataset[i].ShrinkToFit();

    vector<uint32_t> centers_permutation(centers.size());
    std::iota(centers_permutation.begin(), centers_permutation.end(), 0U);
    std::stable_sort(centers_permutation.begin(), centers_permutation.end(),
                     [&subpartitions](uint32_t a, uint32_t b) {
                       return subpartitions[a].size() > subpartitions[b].size();
                     });

    constexpr size_t kAssumedCacheLineSize = 64;
    constexpr size_t kFloatsPerCacheLine =
        kAssumedCacheLineSize / sizeof(float);
    const uint64_t cache_lines_per_row =
        std::max(static_cast<size_t>(1), centers.size() / kFloatsPerCacheLine);
    const size_t num_rotate =
        ((i / 2) % cache_lines_per_row) * kFloatsPerCacheLine;
    std::rotate(centers_permutation.begin(),
                centers_permutation.begin() + num_rotate,
                centers_permutation.end());

    if (i & 1) {
      std::reverse(centers_permutation.begin(), centers_permutation.end());
    }

    for (uint32_t j : centers_permutation) {
      all_centers[i].AppendOrDie(centers[j], "");
    }
  }

  return std::move(all_centers);
}

template <typename T>
Status AhImpl<T>::IndexDatapoint(const DatapointPtr<T>& input,
                                 const ChunkingProjection<T>& projection,
                                 const DistanceMeasure& quantization_distance,
                                 ConstSpan<DenseDataset<FloatT>> centers,
                                 MutableSpan<uint8_t> result) {
  DCHECK(!centers.empty());
  ChunkedDatapoint<FloatT> projected;
  SCANN_RETURN_IF_ERROR(projection.ProjectInput(input, &projected));

  DCHECK_LE(projected.size(), result.size());

  vector<float> distances(centers[0].size());
  DCHECK_GE(distances.size(), 1);
  DCHECK_LE(distances.size(), 256);
  for (size_t i = 0; i < projected.size(); ++i) {
    DCHECK_EQ(centers[0].size(), centers[i].size());
    size_t closest = 0;
    const DatapointPtr<FloatT> projected_ptr = projected[i];
    const DenseDataset<FloatT>& cur_centers = centers[i];

    const size_t centers_size = cur_centers.size();

    if (projected_ptr.IsSparse()) {
      double closest_distance = numeric_limits<double>::infinity();
      for (size_t j = 0; j < centers_size; ++j) {
        const double distance = quantization_distance.GetDistanceHybrid(
            projected_ptr, cur_centers[j]);
        if (ABSL_PREDICT_FALSE(distance < closest_distance)) {
          closest_distance = distance;
          closest = j;
        }
      }
    } else {
      DCHECK_EQ(distances.size(), cur_centers.size());
      DenseDistanceOneToMany(quantization_distance, projected_ptr, cur_centers,
                             MutableSpan<float>(distances));
      auto min_it = std::min_element(distances.begin(), distances.end());
      closest = min_it - distances.begin();
    }
    result[i] = closest;
  }

  return OkStatus();
}

template <typename T>
Status AhImpl<T>::IndexDatapoint(const DatapointPtr<T>& input,
                                 const ChunkingProjection<T>& projection,
                                 const DistanceMeasure& quantization_distance,
                                 ConstSpan<DenseDataset<FloatT>> centers,
                                 Datapoint<uint8_t>* result) {
  DatapointIndex result_size = centers.size();
  DCHECK_EQ(result_size, projection.num_blocks());
  result->clear();
  result->mutable_values()->resize(result_size, 0);

  return AhImpl<T>::IndexDatapoint(input, projection, quantization_distance,
                                   centers, result->mutable_values_span());
}

namespace fallback {

double ComputeParallelResidualComponent(
    ConstSpan<uint8_t> quantized,
    ConstSpan<std::vector<SubspaceResidualStats>> residual_stats) {
  double result = 0.0;
  for (size_t subspace_idx : IndicesOf(quantized)) {
    const uint8_t cluster_idx = quantized[subspace_idx];
    result +=
        residual_stats[subspace_idx][cluster_idx].parallel_residual_component;
  }
  return result;
}

CoordinateDescentResult OptimizeSingleSubspace(
    ConstSpan<SubspaceResidualStats> cur_subspace_residual_stats,
    const uint8_t cur_center_idx, const double parallel_residual_component,
    const double parallel_cost_multiplier) {
  CoordinateDescentResult result;
  result.new_center_idx = cur_center_idx;
  result.new_parallel_residual_component = parallel_residual_component;
  const double old_subspace_residual_norm =
      cur_subspace_residual_stats[cur_center_idx].residual_norm;
  const double old_subspace_parallel_component =
      cur_subspace_residual_stats[cur_center_idx].parallel_residual_component;
  for (size_t new_center_idx : IndicesOf(cur_subspace_residual_stats)) {
    if (new_center_idx == cur_center_idx) continue;
    const SubspaceResidualStats& rs =
        cur_subspace_residual_stats[new_center_idx];
    const double new_parallel_residual_component =
        parallel_residual_component - old_subspace_parallel_component +
        rs.parallel_residual_component;
    const double parallel_norm_delta = Square(new_parallel_residual_component) -
                                       Square(parallel_residual_component);
    if (parallel_norm_delta > 0.0) continue;
    const double residual_norm_delta =
        rs.residual_norm - old_subspace_residual_norm;
    const double perpendicular_norm_delta =
        residual_norm_delta - parallel_norm_delta;
    const double cost_delta = parallel_cost_multiplier * parallel_norm_delta +
                              perpendicular_norm_delta;
    if (cost_delta < result.cost_delta) {
      result.new_center_idx = new_center_idx;
      result.cost_delta = cost_delta;
      result.new_parallel_residual_component = new_parallel_residual_component;
    }
  }
  return result;
}

Status ValidateNoiseShapingParams(double threshold, double eta) {
  if (std::isnan(eta) && std::isnan(threshold)) {
    return InvalidArgumentError(
        "Either threshold or eta must be specified for noise-shaped AH "
        "indexing.");
  }
  if (!std::isnan(eta) && !std::isnan(threshold)) {
    return InvalidArgumentError(
        "Threshold and eta may not both be specified for noise-shaped AH "
        "indexing.");
  }
  return OkStatus();
}

}  // namespace fallback

namespace {

using fallback::ComputeParallelResidualComponent;
using fallback::OptimizeSingleSubspace;
using fallback::SubspaceResidualStats;
using fallback::ValidateNoiseShapingParams;

template <typename T>
T Square(T x) {
  return x * x;
}

double ComputeParallelCostMultiplier(double t, double squared_l2_norm,
                                     DimensionIndex dims) {
  const double parallel_cost = Square(t) / squared_l2_norm;
  const double perpendicular_cost =
      (1.0 - Square(t) / squared_l2_norm) / (dims - 1.0);
  return parallel_cost / perpendicular_cost;
}

template <typename T>
SubspaceResidualStats ComputeResidualStatsForCluster(
    ConstSpan<T> maybe_residual_dptr, ConstSpan<T> original_dptr,
    double inv_norm, ConstSpan<FloatingTypeFor<T>> quantized) {
  DCHECK_EQ(maybe_residual_dptr.size(), quantized.size());
  const size_t dims = maybe_residual_dptr.size();
  SubspaceResidualStats result;
  for (size_t i : Seq(dims)) {
    const double residual_coordinate =
        static_cast<double>(maybe_residual_dptr[i]) -
        static_cast<double>(quantized[i]);
    result.residual_norm += Square(residual_coordinate);
    result.parallel_residual_component +=
        residual_coordinate * original_dptr[i] * inv_norm;
  }
  return result;
}

template <typename T>
StatusOr<vector<std::vector<SubspaceResidualStats>>> ComputeResidualStats(
    DatapointPtr<T> maybe_residual_dptr, DatapointPtr<T> original_dptr,
    ConstSpan<DenseDataset<FloatingTypeFor<T>>> centers,
    const ChunkingProjection<T>& projection) {
  const size_t num_subspaces = centers.size();
  DCHECK_GE(num_subspaces, 1);
  vector<std::vector<SubspaceResidualStats>> result(num_subspaces);
  vector<std::vector<SubspaceResidualStats>> residual_stats(num_subspaces);
  const size_t num_clusters_per_block = centers[0].size();

  using FloatT = FloatingTypeFor<T>;
  ChunkedDatapoint<FloatT> maybe_residual_dptr_chunked;
  ChunkedDatapoint<FloatT> original_dptr_chunked;
  SCANN_RETURN_IF_ERROR(projection.ProjectInput(maybe_residual_dptr,
                                                &maybe_residual_dptr_chunked));
  SCANN_RETURN_IF_ERROR(
      projection.ProjectInput(original_dptr, &original_dptr_chunked));
  SCANN_RET_CHECK_EQ(maybe_residual_dptr_chunked.size(), num_subspaces);
  SCANN_RET_CHECK_EQ(original_dptr_chunked.size(), num_subspaces);
  double chunked_norm = 0.0;
  for (size_t subspace_idx : Seq(num_subspaces)) {
    for (FloatT x : original_dptr_chunked[subspace_idx].values_span()) {
      chunked_norm += Square<double>(x);
    }
  }
  chunked_norm = std::sqrt(chunked_norm);
  double inverse_chunked_norm = 1.0 / chunked_norm;

  for (size_t subspace_idx : Seq(num_subspaces)) {
    auto& cur_subspace_residual_stats = residual_stats[subspace_idx];
    cur_subspace_residual_stats.resize(num_clusters_per_block);
    const DenseDataset<FloatingTypeFor<T>>& cur_subspace_centers =
        centers[subspace_idx];
    for (size_t cluster_idx : Seq(num_clusters_per_block)) {
      ConstSpan<FloatingTypeFor<T>> center =
          cur_subspace_centers[cluster_idx].values_span();
      ConstSpan<FloatT> maybe_residual_dptr_span =
          maybe_residual_dptr_chunked[subspace_idx].values_span();
      ConstSpan<FloatT> original_dptr_span =
          original_dptr_chunked[subspace_idx].values_span();
      cur_subspace_residual_stats[cluster_idx] = ComputeResidualStatsForCluster(
          maybe_residual_dptr_span, original_dptr_span, inverse_chunked_norm,
          center);
    }
  }
  return residual_stats;
}

void InitializeToMinResidualNorm(
    ConstSpan<std::vector<SubspaceResidualStats>> residual_stats,
    MutableSpan<uint8_t> result) {
  DCHECK_EQ(result.size(), residual_stats.size());
  for (size_t subspace_idx : IndicesOf(residual_stats)) {
    auto it = std::min_element(
        residual_stats[subspace_idx].begin(),
        residual_stats[subspace_idx].end(),
        [](const SubspaceResidualStats& a, const SubspaceResidualStats& b) {
          return a.residual_norm < b.residual_norm;
        });
    result[subspace_idx] = it - residual_stats[subspace_idx].begin();
  }
}

}  // namespace

template <typename T>
Status IndexDatapointNoiseShapedFallback(
    const DatapointPtr<T>& maybe_residual_dptr,
    const DatapointPtr<T>& original_dptr,
    const ChunkingProjection<T>& projection,
    ConstSpan<DenseDataset<FloatingTypeFor<T>>> centers, double threshold,
    double eta, MutableSpan<uint8_t> result) {
  SCANN_RET_CHECK_EQ(result.size(), centers.size());
  SCANN_RET_CHECK_EQ(maybe_residual_dptr.dimensionality(),
                     original_dptr.dimensionality());
  SCANN_RETURN_IF_ERROR(ValidateNoiseShapingParams(threshold, eta));
  SCANN_ASSIGN_OR_RETURN(
      auto residual_stats,
      ComputeResidualStats(maybe_residual_dptr, original_dptr, centers,
                           projection));

  const double parallel_cost_multiplier =
      std::isnan(eta) ? ComputeParallelCostMultiplier(
                            threshold, SquaredL2Norm(original_dptr),
                            original_dptr.dimensionality())
                      : eta;
  InitializeToMinResidualNorm(residual_stats, result);
  double parallel_residual_component =
      ComputeParallelResidualComponent(result, residual_stats);

  vector<uint16_t> subspace_idxs(result.size());
  std::iota(subspace_idxs.begin(), subspace_idxs.end(), 0U);
  vector<double> subspace_residual_norms(result.size());
  SCANN_RET_CHECK_EQ(residual_stats.size(), result.size());
  for (size_t subspace_idx : IndicesOf(result)) {
    const uint8_t cluster_idx = result[subspace_idx];
    DCHECK_LT(subspace_idx, residual_stats.size());
    DCHECK_LT(cluster_idx, residual_stats[subspace_idx].size());
    subspace_residual_norms[subspace_idx] =
        residual_stats[subspace_idx][cluster_idx].residual_norm;
  }
  std::vector<uint8_t> result_sorted(result.begin(), result.end());
  ZipSortBranchOptimized(
      std::greater<double>(), subspace_residual_norms.begin(),
      subspace_residual_norms.end(), result_sorted.begin(), result_sorted.end(),
      subspace_idxs.begin(), subspace_idxs.end());

  enum { kMaxRounds = 10 };
  bool cur_round_changes = true;
  for (int round = 0; cur_round_changes && round < kMaxRounds; ++round) {
    cur_round_changes = false;
    for (size_t i : IndicesOf(subspace_idxs)) {
      const size_t subspace_idx = subspace_idxs[i];
      DCHECK_LT(subspace_idx, residual_stats.size());
      ConstSpan<SubspaceResidualStats> cur_subspace_residual_stats =
          residual_stats[subspace_idx];
      const uint8_t cur_center_idx = result_sorted[i];
      auto subspace_result = OptimizeSingleSubspace(
          cur_subspace_residual_stats, cur_center_idx,
          parallel_residual_component, parallel_cost_multiplier);
      if (subspace_result.new_center_idx != cur_center_idx) {
        parallel_residual_component =
            subspace_result.new_parallel_residual_component;
        result_sorted[i] = subspace_result.new_center_idx;
        cur_round_changes = true;
      }
    }
  }

  double final_residual_norm = 0.0;
  for (size_t i : IndicesOf(result_sorted)) {
    const size_t subspace_idx = subspace_idxs[i];
    const uint8_t center_idx = result_sorted[i];
    DCHECK_LT(subspace_idx, residual_stats.size());
    DCHECK_LT(center_idx, residual_stats[subspace_idx].size());
    result[subspace_idx] = center_idx;
    final_residual_norm +=
        residual_stats[subspace_idx][center_idx].residual_norm;
  }
  return OkStatus();
}

template <typename T>
Status AhImpl<T>::IndexDatapointNoiseShaped(
    const DatapointPtr<T>& maybe_residual_dptr,
    const DatapointPtr<T>& original_dptr,
    const ChunkingProjection<T>& projection,
    ConstSpan<DenseDataset<FloatingTypeFor<T>>> centers, double threshold,
    double eta, MutableSpan<uint8_t> result) {
#ifdef __aarch64__
  if constexpr (std::is_same_v<T, float>) {
    if (RuntimeSupportsNeon()) {
      return neon::IndexDatapointNoiseShaped(maybe_residual_dptr, original_dptr,
                                             projection, centers, threshold,
                                             eta, result);
    }
  }
#endif
  return IndexDatapointNoiseShapedFallback(maybe_residual_dptr, original_dptr,
                                           projection, centers, threshold, eta,
                                           result);
}

template <typename T>
StatusOr<vector<float>> AhImpl<T>::CreateRawFloatLookupTable(
    const DatapointPtr<T>& query, const ChunkingProjection<T>& projection,
    const DistanceMeasure& lookup_distance,
    ConstSpan<DenseDataset<FloatT>> centers,
    ConstSpan<FloatT> block_transposed_centers,
    int32_t num_clusters_per_block) {
  ChunkedDatapoint<FloatT> projected;
  SCANN_RETURN_IF_ERROR(projection.ProjectInput(query, &projected));
  SCANN_RET_CHECK_EQ(centers.size(), projected.size());

  const auto lookup_distance_tag =
      lookup_distance.specially_optimized_distance_tag();
  const bool distance_supported_for_transpose = [&]() {
    switch (lookup_distance_tag) {
      case DistanceMeasure::L1:
      case DistanceMeasure::SQUARED_L2:
      case DistanceMeasure::DOT_PRODUCT:
      case DistanceMeasure::COSINE:
        return true;
      default:
        return false;
    }
  }();

  vector<float> result(num_clusters_per_block * projected.size());
  float* result_row_start = result.data();
  size_t centers_start_offset = 0;
  if constexpr (std::is_same_v<FloatT, float>) {
    if (!block_transposed_centers.empty() && distance_supported_for_transpose) {
      for (size_t i = 0; i < centers.size();
           ++i, result_row_start += num_clusters_per_block) {
        const DatapointPtr<FloatT> projected_ptr = projected[i];
        ConstSpan<float> cur_centers =
            MakeConstSpan(block_transposed_centers)
                .subspan(
                    centers_start_offset,
                    num_clusters_per_block * projected_ptr.dimensionality());
        centers_start_offset +=
            num_clusters_per_block * projected_ptr.dimensionality();
        DenseDistanceOneToManyBlockTransposed(
            lookup_distance_tag, projected_ptr, cur_centers,
            MutableSpan<float>(result_row_start, num_clusters_per_block));
      }
      return std::move(result);
    }
  }

  for (size_t i = 0; i < centers.size();
       ++i, result_row_start += num_clusters_per_block) {
    const DatapointPtr<FloatT> projected_ptr = projected[i];
    const DenseDataset<FloatT>& cur_centers = centers[i];
    if (lookup_distance_tag == DistanceMeasure::LIMITED_INNER_PRODUCT) {
      DenseDistanceOneToMany(
          DotProductDistance(), projected_ptr, cur_centers,
          MutableSpan<float>(result_row_start, num_clusters_per_block));
    } else {
      DenseDistanceOneToMany(
          lookup_distance, projected_ptr, cur_centers,
          MutableSpan<float>(result_row_start, num_clusters_per_block));
    }
  }

  return std::move(result);
}

namespace {
float ComputeMultiplierByQuantile(ConstSpan<float> raw_lookup, float quantile,
                                  int32_t max_integer_value) {
  const size_t k = raw_lookup.size() * (1.0 - quantile) + 1;
  if (k == 1) {
    const float max_abs_lookup_element = std::max(
        std::sqrt(numeric_limits<float>::epsilon()), MaxAbsValue(raw_lookup));
    return max_integer_value / max_abs_lookup_element;
  } else {
    DCHECK_LT(quantile, 1.0f);
    TopNAmortizedConstant<float> tn(k);
    for (auto& elem : raw_lookup) {
      tn.push(std::abs(elem));
    }
    return max_integer_value / tn.exact_bottom();
  }
}

template <typename TargetT>
inline TargetT SafeSaturatingCast(float f) {
  if (std::isnan(f)) {
    return 0;
  }
  if (f >= static_cast<float>(std::numeric_limits<TargetT>::max())) {
    return std::numeric_limits<TargetT>::max();
  }
  if (f <= static_cast<float>(std::numeric_limits<TargetT>::min())) {
    return std::numeric_limits<TargetT>::min();
  }
  return static_cast<TargetT>(f);
}

// scann-core: SafeSaturatingCast<int8_t>(f) (truncating) without branches,
// so that the conversion loop vectorizes: NaN -> 0, then clamping to
// [-128, 127] before truncating gives the same value for every input
// (f >= 127 -> 127, f <= -128 -> -128, anything in between truncates alike).
inline int8_t SaturatingTruncateToInt8(float f) {
  f = std::isnan(f) ? 0.0f : f;
  f = std::min(std::max(f, -128.0f), 127.0f);
  return static_cast<int8_t>(static_cast<int32_t>(f));
}

template <typename T, typename Lambda>
inline vector<T> ConvertLookupToFixedPointImpl(ConstSpan<float> raw_lookup,
                                               Lambda convert_to_int_lambda,
                                               float multiplier) {
  constexpr T kBias = FixedPointBias<T>();
  vector<T> result(raw_lookup.size());
  for (size_t i = 0; i < raw_lookup.size(); ++i) {
    result[i] = convert_to_int_lambda(raw_lookup[i] * multiplier) + kBias;
  }
  return result;
}

}  // namespace

template <typename T>
vector<T> ConvertLookupToFixedPoint(
    ConstSpan<float> raw_lookup,
    const AsymmetricHasherConfig::FixedPointLUTConversionOptions&
        conversion_options,
    float* multiplier) {
  DCHECK_GT(conversion_options.multiplier_quantile(), 0.0f);
  DCHECK_LE(conversion_options.multiplier_quantile(), 1.0f);
  using SignedT = make_signed_t<T>;
  *multiplier = ComputeMultiplierByQuantile(
      raw_lookup, conversion_options.multiplier_quantile(),
      numeric_limits<SignedT>::max());
  constexpr int kRound =
      AsymmetricHasherConfig::FixedPointLUTConversionOptions::ROUND;
  if (conversion_options.multiplier_quantile() == 1.0f) {
    if (conversion_options.float_to_int_conversion_method() == kRound) {
      return ConvertLookupToFixedPointImpl<T>(
          raw_lookup,
          [](float f) { return SafeSaturatingCast<SignedT>(std::round(f)); },
          *multiplier);
    } else if constexpr (std::is_same_v<SignedT, int8_t>) {
      return ConvertLookupToFixedPointImpl<T>(
          raw_lookup, [](float f) { return SaturatingTruncateToInt8(f); },
          *multiplier);
    } else {
      return ConvertLookupToFixedPointImpl<T>(
          raw_lookup, [](float f) { return SafeSaturatingCast<SignedT>(f); },
          *multiplier);
    }
  } else {
    if (conversion_options.float_to_int_conversion_method() == kRound) {
      return ConvertLookupToFixedPointImpl<T>(
          raw_lookup,
          [&](float f) { return SafeSaturatingCast<SignedT>(std::round(f)); },
          *multiplier);
    } else {
      return ConvertLookupToFixedPointImpl<T>(
          raw_lookup, [&](float f) { return SafeSaturatingCast<SignedT>(f); },
          *multiplier);
    }
  }
}

template vector<uint8_t> ConvertLookupToFixedPoint<uint8_t>(
    ConstSpan<float> raw_lookup,
    const AsymmetricHasherConfig::FixedPointLUTConversionOptions&,
    float* multiplier);
template vector<uint16_t> ConvertLookupToFixedPoint<uint16_t>(
    ConstSpan<float> raw_lookup,
    const AsymmetricHasherConfig::FixedPointLUTConversionOptions&,
    float* multiplier);

bool CanUseInt16Accumulator(ConstSpan<uint8_t> lookup_table,
                            size_t num_blocks) {
  const size_t num_centers_per_block = lookup_table.size() / num_blocks;
  DCHECK_EQ(lookup_table.size() % num_blocks, 0);
  if (num_centers_per_block != 16) return false;
  constexpr uint8_t kBias = FixedPointBias<uint8_t>();

  constexpr size_t kMinMin =
      numeric_limits<int16_t>::min() / numeric_limits<int8_t>::min();
  constexpr size_t kMaxMax =
      numeric_limits<int16_t>::max() / numeric_limits<int8_t>::max();
  constexpr size_t kGuaranteedToWork = (kMaxMax < kMinMin) ? kMaxMax : kMinMin;
  if (num_blocks <= kGuaranteedToWork) return true;

  int32_t sum_of_maxes = 0, sum_of_mins = 0;
  auto block_start = lookup_table.begin();
  for (size_t block = num_blocks; block != 0;
       --block, block_start += num_centers_per_block) {
    int8_t max_val = numeric_limits<int8_t>::min();
    int8_t min_val = numeric_limits<int8_t>::max();
    for (size_t i = 0; i < 16; ++i) {
      const int8_t unbiased = block_start[i] - kBias;
      max_val = std::max<int8_t>(unbiased, max_val);
      min_val = std::min<int8_t>(unbiased, min_val);
    }

    sum_of_mins += static_cast<int32_t>(min_val);
    sum_of_maxes += static_cast<int32_t>(max_val);
  }

  return sum_of_maxes <= numeric_limits<int16_t>::max() &&
         sum_of_mins >= numeric_limits<int16_t>::min();
}

vector<uint8_t> CreatePackedDataset(
    const DenseDataset<uint8_t>& hashed_database) {
  vector<uint8_t> packed_dataset;
  if (hashed_database.empty()) {
    return packed_dataset;
  }
  const DatapointIndex data_size = hashed_database.size();

  DimensionIndex num_blocks = hashed_database[0].nonzero_entries();
  packed_dataset.resize(num_blocks *
                        ((data_size + kNumDatapointsPerBlock - 1) &
                         (~(kNumDatapointsPerBlock - 1))) /
                        2);
  DatapointIndex k = 0;
  for (; k < data_size / kNumDatapointsPerBlock; ++k) {
    size_t start = k * kPackedDatasetBlockSize * num_blocks;
    for (size_t j = 0; j < num_blocks; ++j) {
      for (size_t m = 0; m < kPackedDatasetBlockSize; m++) {
        uint8_t u0 =
            hashed_database[k * kNumDatapointsPerBlock + m].values()[j];
        uint8_t u1 = hashed_database[k * kNumDatapointsPerBlock + m +
                                     kPackedDatasetBlockSize]
                         .values()[j];
        packed_dataset[start + j * kPackedDatasetBlockSize + m] =
            u1 * kPackedDatasetBlockSize + u0;
      }
    }
  }

  if (k * kNumDatapointsPerBlock < data_size) {
    size_t start = k * kPackedDatasetBlockSize * num_blocks;
    for (size_t j = 0; j < num_blocks; ++j) {
      for (size_t m = 0; m < kPackedDatasetBlockSize; m++) {
        DatapointIndex dp_idx = k * kNumDatapointsPerBlock + m;
        dp_idx = dp_idx >= data_size ? (data_size - 1) : dp_idx;
        uint8_t u0 = hashed_database[dp_idx].values()[j];

        dp_idx = k * kNumDatapointsPerBlock + m + kPackedDatasetBlockSize;
        dp_idx = dp_idx >= data_size ? (data_size - 1) : dp_idx;
        uint8_t u1 = hashed_database[dp_idx].values()[j];
        packed_dataset[start + j * kPackedDatasetBlockSize + m] =
            u1 * kPackedDatasetBlockSize + u0;
      }
    }
  }

  return packed_dataset;
}

template class UnrestrictedIndexIterator<6, IdentityPostprocessFunctor>;
template class UnrestrictedIndexIterator<6, AddBiasFunctor>;
template class UnrestrictedIndexIterator<6, LimitedInnerFunctor>;
template class PopulateDistancesIterator<6, IdentityPostprocessFunctor>;
template class PopulateDistancesIterator<6, AddBiasFunctor>;
template class PopulateDistancesIterator<6, LimitedInnerFunctor>;

SCANN_INSTANTIATE_TYPED_CLASS(, AhImpl);

// scann-core: Lut16DotProductLookupBuilder (see the header).
//
// For a dot-product lookup distance and a block of d <= 4 dimensions,
// CreateRawFloatLookupTable's DenseDistanceOneToMany(lookup_distance, q,
// centers, row) runs DenseAccumulatingDistanceMeasureOneToManyInternal
// (d < 8), whose arithmetic depends on the Highway static target's lane
// count L (`lanes_`): rows are taken in groups of three (rows 0-14 of 16),
// and for each
//   - L full lanes (only if d >= L), then L / 2 lanes (if d >= L / 2) are
//     accumulated as NegMulAdd(q, c, 0), i.e. -(q * c) rounded once
//     (whether or not the target fuses it: the addend is +0);
//   - the lanes are summed by ReduceSum: for L = 4, (a0 + a3) + (a1 + a2);
//     for L = 8, first a[i] + a[i + 4], then the same within the block;
//   - the remaining dimensions are added with the scalar
//     DotProductDistanceLambdas::AccTerm (acc - q * c, as the compiler
//     contracts it), one by one;
// and row 15 is DotProductDistance::GetDistanceDense. ComputeBlock()
// evaluates exactly that, 16 centers at a time from a dimension-major copy
// of the centers (the loops over rows vectorize without reassociating
// anything). The zero lanes' +0.0f terms are kept: they only matter for
// the sign of a zero, but they keep the evaluation literal.
namespace {

using DotLambdas = one_to_many_low_level::DotProductDistanceLambdas<float>;
constexpr size_t kLut16Centers = 16;
constexpr size_t kGroupedRows = 15;

size_t StaticTargetFloatLanes() {
  namespace hn = hwy::HWY_NAMESPACE;
  return hn::Lanes(hn::ScalableTag<float>());
}

}  // namespace

void Lut16DotProductLookupBuilder::ComputeBlock(
    size_t block, const float* query, float* __restrict__ result) const {
  const uint32_t dims = dims_[block];
  const float* c0 = columns_.data() + offsets_[block];
  const float* c1 = c0 + kLut16Centers;
  const float* c2 = c1 + kLut16Centers;
  const float* c3 = c2 + kLut16Centers;
  const float q0 = query[0];
  const float q1 = dims > 1 ? query[1] : 0.0f;
  const float q2 = dims > 2 ? query[2] : 0.0f;
  const float q3 = dims > 3 ? query[3] : 0.0f;
  const bool lane_sum = lanes_ == 4 || (lanes_ == 8 && dims == 4);
  if (!lane_sum) {
    // L / 2 > d: no vector lanes; ReduceSum of zeros, then all of the
    // dimensions through AccTerm.
    for (size_t r = 0; r < kGroupedRows; ++r) {
      float acc = 0.0f;
      acc = DotLambdas::AccTerm(acc, q0, c0[r]);
      if (dims > 1) acc = DotLambdas::AccTerm(acc, q1, c1[r]);
      if (dims > 2) acc = DotLambdas::AccTerm(acc, q2, c2[r]);
      if (dims > 3) acc = DotLambdas::AccTerm(acc, q3, c3[r]);
      result[r] = acc;
    }
  } else if (dims == 1) {
    // L = 4: nothing fills a vector lane.
    for (size_t r = 0; r < kGroupedRows; ++r) {
      result[r] = DotLambdas::AccTerm(0.0f, q0, c0[r]);
    }
  } else if (dims <= 3) {
    // L = 4: two lanes, then (for d = 3) one AccTerm.
    for (size_t r = 0; r < kGroupedRows; ++r) {
      const float p0 = q0 * c0[r];
      const float p1 = q1 * c1[r];
      const float a0 = 0.0f - p0;
      const float a1 = 0.0f - p1;
      result[r] = (a0 + 0.0f) + (a1 + 0.0f);
    }
    if (dims == 3) {
      for (size_t r = 0; r < kGroupedRows; ++r) {
        result[r] = DotLambdas::AccTerm(result[r], q2, c2[r]);
      }
    }
  } else if (lanes_ == 4) {
    for (size_t r = 0; r < kGroupedRows; ++r) {
      const float p0 = q0 * c0[r];
      const float p1 = q1 * c1[r];
      const float p2 = q2 * c2[r];
      const float p3 = q3 * c3[r];
      const float a0 = 0.0f - p0;
      const float a1 = 0.0f - p1;
      const float a2 = 0.0f - p2;
      const float a3 = 0.0f - p3;
      result[r] = (a0 + a3) + (a1 + a2);
    }
  } else {
    // L = 8, d = 4: four lanes of the lower block; the upper block is zero.
    for (size_t r = 0; r < kGroupedRows; ++r) {
      const float p0 = q0 * c0[r];
      const float p1 = q1 * c1[r];
      const float p2 = q2 * c2[r];
      const float p3 = q3 * c3[r];
      const float b0 = (0.0f - p0) + 0.0f;
      const float b1 = (0.0f - p1) + 0.0f;
      const float b2 = (0.0f - p2) + 0.0f;
      const float b3 = (0.0f - p3) + 0.0f;
      result[r] = (b0 + b3) + (b1 + b2);
    }
  }
  // Row 15: -DenseDotProduct(q, c), which on x86 is DenseDotProductSse4:
  // the products go to lanes (d = 4: 0-3; d = 2, 3: the first two to lanes
  // 2 and 3, a third one to lane 0; d = 1: lane 0), each added to +0, and
  // two hadds sum them as (l0 + l1) + (l2 + l3). The double result negated
  // and cast back to float is exact.
#if defined(__x86_64__)
  {
    const size_t r = kGroupedRows;
    float l0 = 0.0f, l1 = 0.0f, l2 = 0.0f, l3 = 0.0f;
    if (dims == 4) {
      l0 = 0.0f + q0 * c0[r];
      l1 = 0.0f + q1 * c1[r];
      l2 = 0.0f + q2 * c2[r];
      l3 = 0.0f + q3 * c3[r];
    } else if (dims >= 2) {
      l0 = 0.0f + 0.0f;
      l1 = 0.0f + 0.0f;
      l2 = 0.0f + q0 * c0[r];
      l3 = 0.0f + q1 * c1[r];
      if (dims == 3) l0 += q2 * c2[r];
    } else {
      l0 += q0 * c0[r];
    }
    result[r] = -((l0 + l1) + (l2 + l3));
  }
#else
  static const DotProductDistance* const kDot = new DotProductDistance();
  result[kGroupedRows] = static_cast<float>(kDot->GetDistanceDense(
      DatapointPtr<float>(nullptr, query, dims, dims),
      centers_[block][kGroupedRows]));
#endif
}

bool Lut16DotProductLookupBuilder::Accepts(
    const ChunkedDatapoint<float>& projected) const {
  if (projected.size() != dims_.size()) return false;
  for (size_t b = 0; b < dims_.size(); ++b) {
    const DatapointPtr<float> chunk = projected[b];
    if (!chunk.IsDense() || chunk.dimensionality() != dims_[b] ||
        chunk.nonzero_entries() != dims_[b]) {
      return false;
    }
  }
  return true;
}

void Lut16DotProductLookupBuilder::Compute(
    const ChunkedDatapoint<float>& projected, MutableSpan<float> result) const {
  DCHECK_EQ(result.size(), kLut16Centers * dims_.size());
  for (size_t b = 0; b < dims_.size(); ++b) {
    ComputeBlock(b, projected[b].values(), result.data() + kLut16Centers * b);
  }
}

void Lut16DotProductLookupBuilder::ComputeContiguous(
    const float* query, MutableSpan<float> result) const {
  DCHECK_EQ(result.size(), kLut16Centers * dims_.size());
  for (size_t b = 0; b < dims_.size(); ++b) {
    ComputeBlock(b, query, result.data() + kLut16Centers * b);
    query += dims_[b];
  }
}

std::unique_ptr<const Lut16DotProductLookupBuilder>
Lut16DotProductLookupBuilder::Create(ConstSpan<DenseDataset<float>> centers) {
  if (centers.empty()) return nullptr;
  const size_t lanes = StaticTargetFloatLanes();
  if (lanes != 4 && lanes != 8 && lanes != 16) return nullptr;
  std::unique_ptr<Lut16DotProductLookupBuilder> result(
      new Lut16DotProductLookupBuilder());
  result->lanes_ = lanes;
  result->centers_ = centers;
  for (const DenseDataset<float>& block : centers) {
    const size_t dims = block.dimensionality();
    if (block.size() != kLut16Centers || !block.IsDense() || dims < 1 ||
        dims > 4) {
      return nullptr;
    }
    result->dims_.push_back(dims);
    result->total_dims_ += dims;
    result->offsets_.push_back(result->columns_.size());
    for (size_t d = 0; d < dims; ++d) {
      for (size_t r = 0; r < kLut16Centers; ++r) {
        result->columns_.push_back(block[r].values()[d]);
      }
    }
  }

  // The check: the generic computation and ComputeBlock on the same inputs,
  // bit for bit. Inputs: random values over many magnitudes (products
  // underflow and cancel), the centers themselves and their negations,
  // zeros of both signs.
  std::mt19937 rng(0x5ca77);
  std::normal_distribution<float> normal;
  std::uniform_real_distribution<float> log_scale(-12.0f, 12.0f);
  const DotProductDistance dot;
  std::array<float, kLut16Centers> expected, actual;
  std::array<float, 4> query;
  for (size_t b = 0; b < centers.size(); ++b) {
    const size_t dims = result->dims_[b];
    auto check = [&]() {
      DenseDistanceOneToMany(
          dot, DatapointPtr<float>(nullptr, query.data(), dims, dims),
          centers[b], MakeMutableSpan(expected));
      result->ComputeBlock(b, query.data(), actual.data());
      return std::memcmp(expected.data(), actual.data(),
                         sizeof(expected)) == 0;
    };
    for (int t = 0; t < 64; ++t) {
      const float scale = std::pow(10.0f, t < 32 ? 0.0f : log_scale(rng));
      for (size_t d = 0; d < dims; ++d) query[d] = normal(rng) * scale;
      if (!check()) return nullptr;
    }
    for (size_t r = 0; r < kLut16Centers; ++r) {
      for (float sign : {1.0f, -1.0f}) {
        for (size_t d = 0; d < dims; ++d) {
          query[d] = sign * centers[b][r].values()[d];
        }
        if (!check()) return nullptr;
      }
    }
    for (float zero : {0.0f, -0.0f}) {
      std::fill(query.begin(), query.end(), zero);
      if (!check()) return nullptr;
    }
  }
  return result;
}

}  // namespace asymmetric_hashing_internal
}  // namespace research_scann