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
// Modified in 2026 by Elias Benali (@ebenali) and TheCleaners for
// scann-core (a derived work of ScaNN, not an official Google product);
// see NOTICE.
#include <cstdint>

#include "absl/base/prefetch.h"
#include "scann/hashes/internal/lut16_args.h"
#include "scann/hashes/internal/lut16_avx2.h"
#include "scann/oss_wrappers/scann_bits.h"
#include "scann/utils/common.h"

#ifdef __x86_64__

#include "scann/utils/bits.h"
#include "scann/utils/intrinsics/avx2.h"

namespace research_scann {
namespace asymmetric_hashing_internal {
namespace {

SCANN_AVX2_INLINE Avx2<int16_t> CombineAvxLanes(const Avx2<int16_t>& a,
                                                const Avx2<int16_t>& b) {
  constexpr uint8_t kDestLoEqALo = 0x00;
  constexpr uint8_t kDestLoEqAHi = 0x01;
  constexpr uint8_t kDestHiEqBLo = 0x20;
  constexpr uint8_t kDestHiEqBHi = 0x30;
  constexpr uint8_t t1spec = (kDestLoEqALo + kDestHiEqBHi);
  constexpr uint8_t t2spec = (kDestLoEqAHi + kDestHiEqBLo);
  Avx2<int16_t> term0 = _mm256_permute2x128_si256(*a, *b, t1spec);
  Avx2<int16_t> term1 = _mm256_permute2x128_si256(*a, *b, t2spec);
  return term0 + term1;
}

SCANN_AVX2_INLINE Avx2<int16_t> PostprocessAccumulatorPair(
    const Avx2<int16_t>& even_plus_tag_along_bits, const Avx2<int16_t>& odd) {
  Avx2<int16_t> even = even_plus_tag_along_bits - (odd << 8);

  Avx2<int16_t> lo_per_lane = _mm256_unpacklo_epi16(*even, *odd);
  Avx2<int16_t> hi_per_lane = _mm256_unpackhi_epi16(*even, *odd);

  return CombineAvxLanes(lo_per_lane, hi_per_lane);
}

template <size_t size, typename T>
SCANN_INLINE array<T, size> ToLocalArray(ConstSpan<T> span) {
  DCHECK_EQ(span.size(), size);
  array<T, size> result;
  std::copy(span.begin(), span.begin() + size, result.begin());
  return result;
}

SCANN_INLINE ssize_t ComputeSmartPrefetchIndex(size_t num_blocks,
                                               size_t num_32dp_simd_iters) {
  return std::min(0l, static_cast<ssize_t>(
                          DivRoundUp(kPrefetchBytesAhead / 16, num_blocks)) -
                          static_cast<ssize_t>(num_32dp_simd_iters));
}

enum class PrefetchLookahead { kOff, kSeq, kNonSeq };
enum class PrefetchHint { kT0, kNta };

constexpr PrefetchLookahead kStrategyToLookahead[] = {
    PrefetchLookahead::kOff,
    PrefetchLookahead::kSeq,
    PrefetchLookahead::kNonSeq,
    PrefetchLookahead::kNonSeq,
};

template <PrefetchHint kHint>
inline void Prefetch(const void* addr) {
  if constexpr (kHint == PrefetchHint::kT0) {
    absl::PrefetchToLocalCache(addr);
  } else {
    absl::PrefetchToLocalCacheNta(addr);
  }
}

template <size_t kNumQueries, PrefetchLookahead kLookahead, PrefetchHint kHint>
SCANN_AVX2_INLINE Avx2<int16_t, kNumQueries, 2> Avx2LUT16BottomLoop(
    const uint8_t* data_start, array<const uint8_t*, kNumQueries> lookup_starts,
    const DimensionIndex num_blocks, const uint8_t* prefetch_start) {
  static_assert(kNumQueries <= 3,
                "Register spilling happens when kNumQueries > 3");
  Avx2<int16_t, kNumQueries, 4> int16_accums = avx2::Zeros();
  const Avx2<uint8_t> sign7 = 0x0F;

  DimensionIndex num_unroll_iter = num_blocks / 2;
  for (; num_unroll_iter != 0; --num_unroll_iter) {
    constexpr uint32_t kPointsPerIter = 32;
    if constexpr (kLookahead == PrefetchLookahead::kSeq) {
      Prefetch<kHint>(data_start + kPrefetchBytesAhead);
    } else if constexpr (kLookahead == PrefetchLookahead::kNonSeq) {
      Prefetch<kHint>(prefetch_start);
      prefetch_start += kPointsPerIter;
    }

    auto mask = Avx2<uint8_t>::Load(data_start);
    data_start += kPointsPerIter;

    Avx2<uint8_t> mask0 = mask & sign7;

    Avx2<uint8_t> mask1 = Avx2<uint8_t>((Avx2<uint16_t>(mask) >> 4)) & sign7;

    for (size_t j : Seq(kNumQueries)) {
      const Avx2<uint8_t> dict = Avx2<uint8_t>::Load(lookup_starts[j]);
      lookup_starts[j] += kPointsPerIter;
      const Avx2<uint8_t> res0 = _mm256_shuffle_epi8(*dict, *mask0);
      const Avx2<uint8_t> res1 = _mm256_shuffle_epi8(*dict, *mask1);

      int16_accums[j][0] += Avx2<int16_t>(Avx2<uint16_t>(res0));
      int16_accums[j][1] += Avx2<int16_t>(Avx2<uint16_t>(res0) >> 8);
      int16_accums[j][2] += Avx2<int16_t>(Avx2<uint16_t>(res1));
      int16_accums[j][3] += Avx2<int16_t>(Avx2<uint16_t>(res1) >> 8);
    }
  }

  Avx2<int16_t, kNumQueries, 2> results;
  for (size_t j : Seq(kNumQueries)) {
    results[j][0] =
        PostprocessAccumulatorPair(int16_accums[j][0], int16_accums[j][1]);
    results[j][1] =
        PostprocessAccumulatorPair(int16_accums[j][2], int16_accums[j][3]);
  }

  const bool has_odd_block = num_blocks & 1;
  if (has_odd_block) {
    const Sse4<uint8_t> sign7 = 0x0F;
    const Sse4<uint8_t> mask = Sse4<uint8_t>::Load(data_start);
    const Sse4<uint8_t> mask0 = mask & sign7;
    const Sse4<uint8_t> mask1 =
        Sse4<uint8_t>((Sse4<uint16_t>(mask) >> 4)) & sign7;

    for (size_t j : Seq(kNumQueries)) {
      auto dict = Sse4<uint8_t>::Load(lookup_starts[j]);
      Sse4<uint8_t> val0 = _mm_shuffle_epi8(*dict, *mask0);
      Sse4<uint8_t> val1 = _mm_shuffle_epi8(*dict, *mask1);
      results[j][0] += Avx2<int16_t>(_mm256_cvtepu8_epi16(*val0));
      results[j][1] += Avx2<int16_t>(_mm256_cvtepu8_epi16(*val1));
    }
  }

  const Avx2<int16_t> total_bias = num_blocks * 128;
  for (size_t j : Seq(kNumQueries)) {
    results[j] -= total_bias;
  }
  return results;
}

template <size_t kBottomLevelBatchSize, size_t kNumQueries>
SCANN_AVX2_INLINE array<const uint8_t*, kBottomLevelBatchSize>
MakeBottomLevelBatchLookupArray(
    array<const uint8_t*, kNumQueries> mid_level_lookups, size_t start) {
  DCHECK_LE(start + kBottomLevelBatchSize, kNumQueries);
  array<const uint8_t*, kBottomLevelBatchSize> result;
  for (size_t j : Seq(kBottomLevelBatchSize)) {
    result[j] = mid_level_lookups[start + j];
  }
  return result;
}

template <size_t kNumQueries, PrefetchLookahead kLookahead, PrefetchHint kHint>
SCANN_AVX2_INLINE Avx2<int16_t, kNumQueries, 2> Avx2LUT16MiddleLoop(
    const uint8_t* data_start, array<const uint8_t*, kNumQueries> lookup_starts,
    const DimensionIndex num_blocks, const uint8_t* prefetch_start) {
  constexpr size_t kSizeB = (kNumQueries == 1) ? 1 : 2;
  constexpr size_t kNumBCases[] = {0, 2, 1};
  constexpr size_t kNumB = (kNumQueries == 1) ? 1 : kNumBCases[kNumQueries % 3];

  constexpr size_t kRemaining = kNumQueries - kNumB * kSizeB;
  static_assert(kRemaining % 3 == 0, "");

  constexpr size_t kSizeA = 3;
  constexpr size_t kNumA = kRemaining / 3;

  Avx2<int16_t, kNumQueries, 2> result;
  for (size_t j : Seq(kNumA)) {
    const size_t start = j * kSizeA;
    auto bottom_level_lookups =
        MakeBottomLevelBatchLookupArray<kSizeA>(lookup_starts, start);
    auto acc = Avx2LUT16BottomLoop<kSizeA, kLookahead, kHint>(
        data_start, bottom_level_lookups, num_blocks, prefetch_start);
    for (size_t jj : Seq(kSizeA)) {
      result[start + jj] = acc[jj];
    }
  }

  for (size_t j : Seq(kNumB)) {
    const size_t start = kNumA * kSizeA + j * kSizeB;
    auto bottom_level_lookups =
        MakeBottomLevelBatchLookupArray<kSizeB>(lookup_starts, start);
    auto acc = Avx2LUT16BottomLoop<kSizeB, kLookahead, kHint>(
        data_start, bottom_level_lookups, num_blocks, prefetch_start);
    for (size_t jj : Seq(kSizeB)) {
      result[start + jj] = acc[jj];
    }
  }
  return result;
}

template <size_t kNumQueries, PrefetchStrategy kPrefetch>
SCANN_AVX2_INLINE Avx2<int16_t, kNumQueries, 2> PrefetchDispatcher(
    const uint8_t* data_start, array<const uint8_t*, kNumQueries> lookups,
    const DimensionIndex num_blocks, const uint8_t* next_partition,
    ssize_t* __restrict__ next_prefetch_idx) {
  if constexpr (kPrefetch == PrefetchStrategy::kSmart ||
                kPrefetch == PrefetchStrategy::kSmartT0) {
    constexpr PrefetchHint kHint = (kPrefetch == PrefetchStrategy::kSmart)
                                       ? PrefetchHint::kNta
                                       : PrefetchHint::kT0;
    // scann-core: upstream also took this branch for the last partition,
    // where next_partition is null, and did pointer arithmetic on the null
    // pointer (undefined behavior; UBSan: "applying non-zero offset to null
    // pointer"). Prefetch sequentially instead when there's no next one.
    if ((*next_prefetch_idx) >= 0 && next_partition != nullptr) {
      const uint8_t* prefetch_start =
          next_partition + (*next_prefetch_idx) * 16 * num_blocks;
      (*next_prefetch_idx)++;
      return Avx2LUT16MiddleLoop<kNumQueries, PrefetchLookahead::kNonSeq,
                                 kHint>(data_start, lookups, num_blocks,
                                        prefetch_start);
    } else {
      (*next_prefetch_idx)++;
      return Avx2LUT16MiddleLoop<kNumQueries, PrefetchLookahead::kSeq, kHint>(
          data_start, lookups, num_blocks, nullptr);
    }
  } else {
    constexpr PrefetchLookahead kLookahead =
        kStrategyToLookahead[static_cast<size_t>(kPrefetch)];
    return Avx2LUT16MiddleLoop<kNumQueries, kLookahead, PrefetchHint::kT0>(
        data_start, lookups, num_blocks, nullptr);
  }
}

template <size_t kNumQueries, PrefetchStrategy kPrefetch>
SCANN_AVX2_INLINE Avx2<int32_t, kNumQueries, 4> Avx2LUT16MiddleLoopInt32(
    const uint8_t* data_start, array<const uint8_t*, kNumQueries> lookup_starts,
    const DimensionIndex num_blocks, const uint8_t* next_partition,
    ssize_t* __restrict__ next_prefetch_idx) {
  Avx2<int32_t, kNumQueries, 4> int32_accumulators = avx2::Zeros();
  for (DimensionIndex k = 0; k < num_blocks;) {
    DimensionIndex reaccumulate_limit = std::min(num_blocks - k, uint64_t{256});

    auto int16_accums = [&]() SCANN_AVX2_INLINE_LAMBDA {
      if constexpr (kPrefetch == PrefetchStrategy::kSmart ||
                    kPrefetch == PrefetchStrategy::kSmartT0) {
        constexpr PrefetchHint kHint = (kPrefetch == PrefetchStrategy::kSmart)
                                           ? PrefetchHint::kNta
                                           : PrefetchHint::kT0;
        // scann-core: as in PrefetchDispatcher, no next partition to
        // prefetch from on the last one.
        if ((*next_prefetch_idx) >= 0 && next_partition != nullptr) {
          const uint8_t* prefetch_start =
              next_partition + (*next_prefetch_idx) * 16 * num_blocks + k * 16;
          return Avx2LUT16MiddleLoop<kNumQueries, PrefetchLookahead::kNonSeq,
                                     kHint>(data_start, lookup_starts,
                                            reaccumulate_limit, prefetch_start);
        } else {
          return Avx2LUT16MiddleLoop<kNumQueries, PrefetchLookahead::kSeq,
                                     kHint>(data_start, lookup_starts,
                                            reaccumulate_limit, nullptr);
        }
      } else {
        constexpr PrefetchLookahead kLookahead =
            kStrategyToLookahead[static_cast<size_t>(kPrefetch)];
        return Avx2LUT16MiddleLoop<kNumQueries, kLookahead, PrefetchHint::kT0>(
            data_start, lookup_starts, reaccumulate_limit, nullptr);
      }
    }();
    data_start += 16 * reaccumulate_limit;
    k += reaccumulate_limit;
    for (size_t j : Seq(kNumQueries)) {
      int32_accumulators[j] += int16_accums[j].template ExpandTo<int32_t>();
      lookup_starts[j] += 16 * reaccumulate_limit;
    }
  }
  if constexpr (kPrefetch == PrefetchStrategy::kSmart ||
                kPrefetch == PrefetchStrategy::kSmartT0) {
    (*next_prefetch_idx)++;
  }
  return int32_accumulators;
}

}  // namespace

template <size_t kNumQueries, PrefetchStrategy kPrefetch>
SCANN_AVX2_OUTLINE void LUT16Avx2<kNumQueries, kPrefetch>::GetInt16Distances(
    LUT16Args<int16_t> args) {
  const uint8_t* packed_dataset = args.packed_dataset;
  const uint8_t* next_partition = args.next_partition;
  const size_t num_32dp_simd_iters = args.num_32dp_simd_iters;
  const size_t num_blocks = args.num_blocks;
  auto lookups = ToLocalArray<kNumQueries>(args.lookups);
  auto distances = ToLocalArray<kNumQueries>(args.distances);

  ssize_t next_prefetch_idx =
      ComputeSmartPrefetchIndex(num_blocks, num_32dp_simd_iters);
  for (DatapointIndex k : Seq(num_32dp_simd_iters)) {
    const uint8_t* data_start = packed_dataset + k * 16 * num_blocks;
    auto int16_accums = PrefetchDispatcher<kNumQueries, kPrefetch>(
        data_start, lookups, num_blocks, next_partition, &next_prefetch_idx);
    for (size_t j : Seq(kNumQueries)) {
      int16_accums[j].Store(distances[j] + 32 * k);
    }
  }
}

template <size_t kNumQueries, PrefetchStrategy kPrefetch>
SCANN_AVX2_OUTLINE void LUT16Avx2<kNumQueries, kPrefetch>::GetInt32Distances(
    LUT16Args<int32_t> args) {
  const uint8_t* packed_dataset = args.packed_dataset;
  const uint8_t* next_partition = args.next_partition;
  const size_t num_32dp_simd_iters = args.num_32dp_simd_iters;
  const size_t num_blocks = args.num_blocks;
  auto lookups = ToLocalArray<kNumQueries>(args.lookups);
  auto distances = ToLocalArray<kNumQueries>(args.distances);

  ssize_t next_prefetch_idx =
      ComputeSmartPrefetchIndex(num_blocks, num_32dp_simd_iters);
  for (DatapointIndex k : Seq(num_32dp_simd_iters)) {
    const uint8_t* data_start = packed_dataset + k * 16 * num_blocks;
    auto int32_accumulators = Avx2LUT16MiddleLoopInt32<kNumQueries, kPrefetch>(
        data_start, lookups, num_blocks, next_partition, &next_prefetch_idx);
    for (size_t j : Seq(kNumQueries)) {
      int32_accumulators[j].Store(distances[j] + 32 * k);
    }
  }
}

template <size_t kNumQueries, PrefetchStrategy kPrefetch>
SCANN_AVX2_OUTLINE void LUT16Avx2<kNumQueries, kPrefetch>::GetFloatDistances(
    LUT16Args<float> args, ConstSpan<float> inv_fp_multipliers) {
  const uint8_t* packed_dataset = args.packed_dataset;
  const uint8_t* next_partition = args.next_partition;
  const size_t num_32dp_simd_iters = args.num_32dp_simd_iters;
  const size_t num_blocks = args.num_blocks;
  auto lookups = ToLocalArray<kNumQueries>(args.lookups);
  auto distances = ToLocalArray<kNumQueries>(args.distances);
  auto mults = ToLocalArray<kNumQueries>(inv_fp_multipliers);

  ssize_t next_prefetch_idx =
      ComputeSmartPrefetchIndex(num_blocks, num_32dp_simd_iters);
  for (DatapointIndex k : Seq(num_32dp_simd_iters)) {
    const uint8_t* data_start = packed_dataset + k * 16 * num_blocks;
    Avx2<int32_t, kNumQueries, 4> int32_accums =
        Avx2LUT16MiddleLoopInt32<kNumQueries, kPrefetch>(
            data_start, lookups, num_blocks, next_partition,
            &next_prefetch_idx);
    for (size_t j : Seq(kNumQueries)) {
      (int32_accums[j].template ConvertTo<float>() * Avx2<float>(mults[j]))
          .Store(distances[j] + 32 * k);
    }
  }
}

namespace {
template <size_t kNumQueries, PrefetchStrategy kPrefetch, typename TopN>
SCANN_AVX2_INLINE void GetTopInt16DistancesImpl(
    LUT16ArgsTopN<int16_t, TopN> args) {
  const uint8_t* packed_dataset = args.packed_dataset;
  const uint8_t* next_partition = args.next_partition;
  const size_t num_32dp_simd_iters = args.num_32dp_simd_iters;
  const size_t num_blocks = args.num_blocks;
  auto lookups = ToLocalArray<kNumQueries>(args.lookups);
  const DatapointIndex first_dp_index = args.first_dp_index;
  const uint32_t final_mask = GetFinalMask32(args.num_datapoints);
  DCHECK_EQ(num_32dp_simd_iters, DivRoundUp(args.num_datapoints, 32));

  Avx2<int16_t> simd_thresholds[kNumQueries];
  for (size_t j : Seq(kNumQueries)) {
    const int16_t int16_threshold = args.fast_topns[j]->epsilon();
    simd_thresholds[j] = int16_threshold;
  }

  typename TopN::Mutator topn_mutators[kNumQueries];
  for (size_t j : Seq(kNumQueries)) {
    args.fast_topns[j]->AcquireMutator(&topn_mutators[j]);
  }

  int16_t distances_buffer[32];
  auto restrict_whitelist_ptrs =
      args.template GetRestrictWhitelistPtrs<kNumQueries>();
  ssize_t next_prefetch_idx =
      ComputeSmartPrefetchIndex(num_blocks, num_32dp_simd_iters);
  for (DatapointIndex k : Seq(num_32dp_simd_iters)) {
    bool can_skip_batch = true;
    for (size_t j : Seq(kNumQueries)) {
      if (restrict_whitelist_ptrs[j] == nullptr ||
          restrict_whitelist_ptrs[j][k] != 0) {
        can_skip_batch = false;
        break;
      }
    }
    if (can_skip_batch) continue;

    const uint8_t* data_start = packed_dataset + k * 16 * num_blocks;
    auto int16_accums = PrefetchDispatcher<kNumQueries, kPrefetch>(
        data_start, lookups, num_blocks, next_partition, &next_prefetch_idx);
    for (size_t j : Seq(kNumQueries)) {
      auto compute_push_mask = [&]() SCANN_AVX2_INLINE_LAMBDA {
        return GetComparisonMask(int16_accums[j] < simd_thresholds[j]);
      };
      uint32_t push_mask = compute_push_mask();

      if (!push_mask) continue;

      int16_accums[j].Store(distances_buffer);

      if (k == num_32dp_simd_iters - 1) {
        push_mask &= final_mask;
      }
      if (restrict_whitelist_ptrs[j]) {
        push_mask &= restrict_whitelist_ptrs[j][k];
      }

      while (push_mask) {
        const int offset = bits::FindLSBSetNonZero(push_mask);
        push_mask &= (push_mask - 1);
        const DatapointIndex dp_idx = first_dp_index + 32 * k + offset;
        DCHECK(
            !restrict_whitelist_ptrs[j] ||
            args.restrict_whitelists[j].IsWhitelisted(dp_idx - first_dp_index))
            << dp_idx;
        const int16_t distance = distances_buffer[offset];
        const bool needs_collection = topn_mutators[j].Push(dp_idx, distance);
        if (ABSL_PREDICT_FALSE(needs_collection)) {
          topn_mutators[j].GarbageCollect();

          simd_thresholds[j] = topn_mutators[j].epsilon();

          push_mask &= compute_push_mask();
        }
      }
    }
  }
}
}  // namespace

template <size_t kNumQueries, PrefetchStrategy kPrefetch>
SCANN_AVX2_OUTLINE void LUT16Avx2<kNumQueries, kPrefetch>::GetTopInt16Distances(
    LUT16ArgsTopN<int16_t> args) {
  return GetTopInt16DistancesImpl<kNumQueries, kPrefetch>(std::move(args));
}

SCANN_AVX2_INLINE int16_t GetInt16Threshold(float float_threshold) {
  constexpr float kMinValue = std::numeric_limits<int16_t>::min();
  constexpr float kMaxValue = std::numeric_limits<int16_t>::max();
  return std::max(kMinValue, std::min(kMaxValue, float_threshold));
}

SCANN_AVX2_INLINE int32_t GetInt32Threshold(float float_threshold) {
  constexpr float kMinValue = std::numeric_limits<int32_t>::min();

  static const float kMaxValue =
      std::nextafterf(std::numeric_limits<int32_t>::max(), 0);
  return std::max(kMinValue, std::min(kMaxValue, float_threshold));
}

namespace {
template <size_t kNumQueries, PrefetchStrategy kPrefetch, bool kHasAnyPredicate,
          bool kWithSpilling = false, bool kUseInt32Accumulator = false,
          typename TopN>
SCANN_AVX2_INLINE void GetTopFloatDistancesImpl(
    LUT16ArgsTopN<float, TopN> args) {
  static_assert(
      !kHasAnyPredicate || !kWithSpilling,
      "Spilling is incompatible with datapoint translation predicates.");
  const uint8_t* packed_dataset = args.packed_dataset;
  const uint8_t* next_partition = args.next_partition;
  const size_t num_32dp_simd_iters = args.num_32dp_simd_iters;
  const size_t num_blocks = args.num_blocks;
  auto lookups = ToLocalArray<kNumQueries>(args.lookups);
  const DatapointIndex first_dp_index = args.first_dp_index;
  const uint32_t final_mask = GetFinalMask32(args.num_datapoints);
  DCHECK_EQ(num_32dp_simd_iters, DivRoundUp(args.num_datapoints, 32));

  auto biases = ToLocalArray<kNumQueries>(args.biases);
  Avx2<float> simd_biases[kNumQueries];
  for (size_t j : Seq(kNumQueries)) {
    simd_biases[j] = biases[j];
  }

  auto mults = ToLocalArray<kNumQueries>(args.fixed_point_multipliers);
  Avx2<float> inv_mults[kNumQueries];
  for (size_t j : Seq(kNumQueries)) {
    inv_mults[j] = 1.0 / mults[j];
  }

  using ThresholdT = std::conditional_t<kUseInt32Accumulator, int32_t, int16_t>;
  Avx2<ThresholdT> simd_thresholds[kNumQueries];
  for (size_t j : Seq(kNumQueries)) {
    const float epsilon = args.fast_topns[j]->epsilon();
    const float float_threshold = (epsilon - biases[j]) * mults[j];
    if constexpr (kUseInt32Accumulator) {
      simd_thresholds[j] = GetInt32Threshold(float_threshold);
    } else {
      simd_thresholds[j] = GetInt16Threshold(float_threshold);
    }
  }

  typename TopN::Mutator topn_mutators[kNumQueries];
  for (size_t j : Seq(kNumQueries)) {
    args.fast_topns[j]->AcquireMutator(&topn_mutators[j]);
  }

  float distances_buffer[32];
  auto restrict_whitelist_ptrs =
      args.template GetRestrictWhitelistPtrs<kNumQueries>();
  ssize_t next_prefetch_idx =
      ComputeSmartPrefetchIndex(num_blocks, num_32dp_simd_iters);
  for (DatapointIndex k : Seq(num_32dp_simd_iters)) {
    bool can_skip_batch = true;
    for (size_t j : Seq(kNumQueries)) {
      if (restrict_whitelist_ptrs[j] == nullptr ||
          restrict_whitelist_ptrs[j][k] != 0) {
        can_skip_batch = false;
        break;
      }
    }
    if (can_skip_batch) continue;

    const uint8_t* data_start = packed_dataset + k * 16 * num_blocks;
    auto accums = [&]() SCANN_AVX2_INLINE_LAMBDA {
      if constexpr (kUseInt32Accumulator) {
        return Avx2LUT16MiddleLoopInt32<kNumQueries, kPrefetch>(
            data_start, lookups, num_blocks, next_partition,
            &next_prefetch_idx);
      } else {
        return PrefetchDispatcher<kNumQueries, kPrefetch>(
            data_start, lookups, num_blocks, next_partition,
            &next_prefetch_idx);
      }
    }();

    for (size_t j : Seq(kNumQueries)) {
      auto compute_push_mask = [&]() SCANN_AVX2_INLINE_LAMBDA {
        return GetComparisonMask(accums[j] < simd_thresholds[j]);
      };
      uint32_t push_mask = compute_push_mask();

      if (!push_mask) continue;

      auto fvals = [&]() SCANN_AVX2_INLINE_LAMBDA {
        if constexpr (kUseInt32Accumulator) {
          return accums[j].template ConvertTo<float>();
        } else {
          return accums[j]
              .template ExpandTo<int32_t>()
              .template ConvertTo<float>();
        }
      }();

      (fvals * inv_mults[j] + simd_biases[j]).Store(distances_buffer);

      if (k == num_32dp_simd_iters - 1) {
        push_mask &= final_mask;
      }
      if (restrict_whitelist_ptrs[j]) {
        push_mask &= restrict_whitelist_ptrs[j][k];
      }
      if (kHasAnyPredicate) {
        if (args.batch_filter_predicate) {
          push_mask =
              args.batch_filter_predicate(first_dp_index + 32 * k, push_mask,
                                          args.datapoint_translation_predicate);
        }
      }

      while (push_mask) {
        const int offset = bits::FindLSBSetNonZero(push_mask);
        push_mask &= (push_mask - 1);
        DatapointIndex dp_idx;
        if constexpr (!kWithSpilling) {
          dp_idx = first_dp_index + 32 * k + offset;
        }
        DCHECK(
            !restrict_whitelist_ptrs[j] ||
            args.restrict_whitelists[j].IsWhitelisted(dp_idx - first_dp_index))
            << dp_idx;
        if (kHasAnyPredicate) {
          if (args.datapoint_translation_predicate) {
            dp_idx = args.datapoint_translation_predicate(dp_idx);
          }
        }

        DCHECK_LE(distances_buffer[offset],
                  topn_mutators[j].epsilon() + (*inv_mults[j])[0]);
        const bool needs_gc = topn_mutators[j].PushNoEpsilonCheck(
            dp_idx, distances_buffer[offset]);
        if (ABSL_PREDICT_FALSE(needs_gc)) {
          topn_mutators[j].GarbageCollect();

          const float new_epsilon = topn_mutators[j].epsilon();
          const float float_threshold = (new_epsilon - biases[j]) * mults[j];

          if constexpr (kUseInt32Accumulator) {
            simd_thresholds[j] = GetInt32Threshold(float_threshold);
          } else {
            simd_thresholds[j] = GetInt16Threshold(float_threshold);
          }

          push_mask &= compute_push_mask();
        }
      }
    }
  }
}
}  // namespace

template <size_t kNumQueries, PrefetchStrategy kPrefetch, bool kHasAnyPredicate,
          bool kWithSpilling = false, typename TopN>
SCANN_AVX2_INLINE void GetTopFloatDistancesChooseAccum(
    LUT16ArgsTopN<float, TopN> args) {
  const bool use_int32 = args.num_blocks >= 256;
  if (use_int32) {
    GetTopFloatDistancesImpl<kNumQueries, kPrefetch, kHasAnyPredicate,
                             kWithSpilling, true>(std::move(args));
  } else {
    GetTopFloatDistancesImpl<kNumQueries, kPrefetch, kHasAnyPredicate,
                             kWithSpilling, false>(std::move(args));
  }
}

template <size_t kNumQueries, PrefetchStrategy kPrefetch>
SCANN_AVX2_OUTLINE void LUT16Avx2<kNumQueries, kPrefetch>::GetTopFloatDistances(
    LUT16ArgsTopN<float> args) {
  if (args.batch_filter_predicate || args.datapoint_translation_predicate) {
    GetTopFloatDistancesChooseAccum<kNumQueries, kPrefetch, true, false>(
        std::move(args));
  } else {
    GetTopFloatDistancesChooseAccum<kNumQueries, kPrefetch, false, false>(
        std::move(args));
  }
}

}  // namespace asymmetric_hashing_internal
}  // namespace research_scann

#endif