#ifndef SCANN_HASHES_INTERNAL_ASYMMETRIC_HASHING_LUT16_H_
#define SCANN_HASHES_INTERNAL_ASYMMETRIC_HASHING_LUT16_H_
#include <cstdint>
#include <vector>
#include "scann/base/restrict_allowlist.h"
#include "scann/data_format/dataset.h"
#include "scann/hashes/internal/asymmetric_hashing_postprocess.h"
#include "scann/hashes/internal/lut16_interface.h"
#include "scann/hashes/internal/write_distances_to_topn.h"
#include "scann/utils/top_n_amortized_constant.h"
#include "scann/utils/types.h"
namespace research_scann {
namespace asymmetric_hashing_internal {
using TopFixedPointNeighbors = TopNeighbors<int32_t>;
template <typename TopN, typename PostprocessedDistance, typename Postprocess>
void GetNeighborsViaAsymmetricDistanceLUT16WithInt32Accumulator2(
ConstSpan<uint8_t> lookup, DatapointIndex dataset_size,
ConstSpan<uint8_t> packed_dataset,
const RestrictAllowlist* whitelist_or_null,
PostprocessedDistance max_distance, const Postprocess& postprocess,
TopN* top_items);
template <typename TopN, typename PostprocessedDistance, typename Postprocess>
void GetNeighborsViaAsymmetricDistanceLUT16WithInt16Accumulator2(
ConstSpan<uint8_t> lookup, DatapointIndex dataset_size,
ConstSpan<uint8_t> packed_dataset,
const RestrictAllowlist* whitelist_or_null,
PostprocessedDistance max_distance, const Postprocess& postprocess,
TopN* top_items);
template <size_t kNumQueries, typename TopN, typename PostprocessedDistance,
typename Postprocess>
void GetNeighborsViaAsymmetricDistanceLUT16WithInt16AccumulatorBatched2(
array<ConstSpan<int8_t>, kNumQueries> lookups, DatapointIndex dataset_size,
ConstSpan<uint8_t> packed_dataset,
array<const RestrictAllowlist*, kNumQueries> restrict_whitelists_or_null,
array<PostprocessedDistance, kNumQueries> max_distances,
const Postprocess& postprocess, array<TopN*, kNumQueries> top_items);
template <size_t kNumQueries, typename TopN, typename PostprocessedDistance,
typename Postprocess>
void GetNeighborsViaAsymmetricDistanceLUT16WithInt32AccumulatorBatched2(
array<ConstSpan<int8_t>, kNumQueries> lookups, DatapointIndex dataset_size,
ConstSpan<uint8_t> packed_dataset,
array<const RestrictAllowlist*, kNumQueries> restrict_whitelists_or_null,
array<PostprocessedDistance, kNumQueries> max_distances,
const Postprocess& postprocess, array<TopN*, kNumQueries> top_items);
template <typename TopN, typename PostprocessedDistance, typename Postprocess>
void GetNeighborsViaAsymmetricDistanceLUT16WithInt32Accumulator2(
ConstSpan<uint8_t> lookup, DatapointIndex dataset_size,
ConstSpan<uint8_t> packed_dataset,
const RestrictAllowlist* whitelist_or_null,
PostprocessedDistance max_distance, const Postprocess& postprocess,
TopN* top_items) {
const size_t num_32dp_simd_iters = DivRoundUp(dataset_size, 32);
unique_ptr<int32_t[]> distances(new int32_t[32 * num_32dp_simd_iters]);
size_t num_blocks = lookup.size() / 16;
LUT16Interface::GetDistances(packed_dataset.data(), num_32dp_simd_iters,
num_blocks, lookup.data(), distances.get());
if (std::is_same<Postprocess, IdentityPostprocessFunctor>::value &&
std::is_same<PostprocessedDistance, int32_t>::value &&
max_distance >=
static_cast<int32_t>(numeric_limits<int8_t>::max() * num_blocks)) {
max_distance = numeric_limits<int32_t>::max();
}
WriteDistancesToTopN(whitelist_or_null, max_distance,
ConstSpan<int32_t>(distances.get(), dataset_size),
postprocess, top_items);
}
template <typename TopN, typename PostprocessedDistance, typename Postprocess>
void GetNeighborsViaAsymmetricDistanceLUT16WithInt16Accumulator2(
ConstSpan<uint8_t> lookup, DatapointIndex dataset_size,
ConstSpan<uint8_t> packed_dataset,
const RestrictAllowlist* whitelist_or_null,
PostprocessedDistance max_distance, const Postprocess& postprocess,
TopN* top_items) {
if (max_distance > numeric_limits<int16_t>::max()) {
max_distance = numeric_limits<int16_t>::max();
} else if (max_distance < numeric_limits<int16_t>::min()) {
return;
}
const size_t num_32dp_simd_iters = DivRoundUp(dataset_size, 32);
unique_ptr<int16_t[]> distances(new int16_t[32 * num_32dp_simd_iters]);
size_t num_blocks = lookup.size() / 16;
LUT16Interface::GetDistances(packed_dataset.data(), num_32dp_simd_iters,
num_blocks, lookup.data(), distances.get());
return WriteDistancesToTopN(whitelist_or_null, max_distance,
ConstSpan<int16_t>(distances.get(), dataset_size),
postprocess, top_items);
}
template <size_t kNumQueries, typename TopN, typename PostprocessedDistance,
typename DistT, typename Postprocess>
void GetNeighborsViaAsymmetricDistanceLUT16BatchedImpl(
array<ConstSpan<uint8_t>, kNumQueries> lookups, DatapointIndex dataset_size,
ConstSpan<uint8_t> packed_dataset,
array<const RestrictAllowlist*, kNumQueries> restrict_whitelists_or_null,
array<PostprocessedDistance, kNumQueries> max_distances,
const Postprocess& postprocess, array<TopN*, kNumQueries> top_items) {
for (auto& lookup : lookups) {
DCHECK_EQ(lookup.size(), lookups[0].size());
}
bool all_thresholds_too_small = true;
for (auto& max_dist : max_distances) {
if (max_dist > numeric_limits<DistT>::max()) {
max_dist = numeric_limits<DistT>::max();
}
if (max_dist >= numeric_limits<DistT>::min()) {
all_thresholds_too_small = false;
}
}
if (all_thresholds_too_small) return;
const size_t num_32dp_simd_iters = DivRoundUp(dataset_size, 32);
array<const uint8_t*, kNumQueries> lookup_ptrs;
array<DistT*, kNumQueries> distances;
array<unique_ptr<DistT[]>, kNumQueries> distances_storage;
for (size_t i = 0; i < kNumQueries; ++i) {
distances_storage[i].reset(new DistT[32 * num_32dp_simd_iters]);
distances[i] = distances_storage[i].get();
lookup_ptrs[i] = lookups[i].data();
}
const size_t num_blocks = lookups[0].size() / 16;
LUT16Interface::GetDistances(packed_dataset.data(), num_32dp_simd_iters,
num_blocks, lookup_ptrs, distances);
for (size_t i = 0; i < kNumQueries; ++i) {
WriteDistancesToTopN(restrict_whitelists_or_null[i],
static_cast<int32_t>(max_distances[i]),
ConstSpan<DistT>(distances[i], dataset_size),
postprocess, top_items[i]);
}
}
template <size_t kNumQueries, typename TopN, typename PostprocessedDistance,
typename Postprocess>
void GetNeighborsViaAsymmetricDistanceLUT16WithInt16AccumulatorBatched2(
array<ConstSpan<uint8_t>, kNumQueries> lookups, DatapointIndex dataset_size,
ConstSpan<uint8_t> packed_dataset,
array<const RestrictAllowlist*, kNumQueries> restrict_whitelists_or_null,
array<PostprocessedDistance, kNumQueries> max_distances,
const Postprocess& postprocess, array<TopN*, kNumQueries> top_items) {
GetNeighborsViaAsymmetricDistanceLUT16BatchedImpl<
kNumQueries, TopN, PostprocessedDistance, int16_t, Postprocess>(
lookups, dataset_size, packed_dataset, restrict_whitelists_or_null,
max_distances, postprocess, top_items);
}
template <size_t kNumQueries, typename TopN, typename PostprocessedDistance,
typename Postprocess>
void GetNeighborsViaAsymmetricDistanceLUT16WithInt32AccumulatorBatched2(
array<ConstSpan<uint8_t>, kNumQueries> lookups, DatapointIndex dataset_size,
ConstSpan<uint8_t> packed_dataset,
array<const RestrictAllowlist*, kNumQueries> restrict_whitelists_or_null,
array<PostprocessedDistance, kNumQueries> max_distances,
const Postprocess& postprocess, array<TopN*, kNumQueries> top_items) {
GetNeighborsViaAsymmetricDistanceLUT16BatchedImpl<
kNumQueries, TopN, PostprocessedDistance, int32_t, Postprocess>(
lookups, dataset_size, packed_dataset, restrict_whitelists_or_null,
max_distances, postprocess, top_items);
}
} }
#endif