#ifndef SCANN_UTILS_CROWDING_MULTI_DIMENSIONAL_H_
#define SCANN_UTILS_CROWDING_MULTI_DIMENSIONAL_H_
#include <bitset>
#include <cstdint>
#include <optional>
#include "absl/container/inlined_vector.h"
#include "absl/strings/string_view.h"
#include "scann/data_format/features.pb.h"
#include "scann/utils/common.h"
#include "scann/utils/types.h"
namespace research_scann {
class CrowdingMultiDimensional {
public:
static constexpr absl::string_view kQuotaWeightDimensionName = "quota_weight";
CrowdingMultiDimensional(
ConstSpan<int64_t> datapoint_index_to_crowding_attribute,
ConstSpan<std::string> crowding_dimension_names);
void SetQuota(absl::string_view dimension, int quota);
void SetQuota(absl::string_view dimension, int64_t crowding_attribute,
int quota);
[[nodiscard]] bool Add(DatapointIndex index);
private:
struct DimensionData {
std::array<int32_t, 64 / 4> small_attribute_quotas;
int32_t default_quota;
absl::flat_hash_map<int64_t, int32_t> large_attribute_quotas;
DimensionData();
void SetQuota(int32_t value);
int32_t& MutableQuota(int64_t crowding_attribute);
};
ConstSpan<int64_t> datapoint_index_to_crowding_attribute_;
ConstSpan<std::string> crowding_dimension_names_;
static constexpr int kExpectedMaxNumDimensions = 8;
std::optional<int> weight_dimension_index_;
std::bitset<32> used_dimensions_mask_ = 0;
absl::InlinedVector<DimensionData, kExpectedMaxNumDimensions> dimension_;
};
class CrowdingAttributesGuard {
public:
absl::Status Append(const GenericFeatureVector::Crowding& crowding,
vector<int64_t>& destination,
vector<std::string>* dimension_names);
private:
absl::Status AppendSingleDimensional(
const GenericFeatureVector::Crowding& crowding,
vector<int64_t>& destination);
absl::Status AppendMultiDimensional(
const GenericFeatureVector::Crowding& crowding,
vector<int64_t>& destination, vector<std::string>* dimension_names);
enum class State { kEmpty, kSingleDimensional, kMultiDimensional };
State state_ = State::kEmpty;
};
}
#endif