Levenshtein.hpp

/* SPDX-License-Identifier: MIT */
/* Copyright © 2022-present Max Bachmann */
 
#pragma once
#include <limits>
#include <rapidfuzz/details/Range.hpp>
#include <rapidfuzz/distance/Levenshtein_impl.hpp>
 
namespace rapidfuzz {
 
template <typename InputIt1, typename InputIt2>
size_t levenshtein_distance(InputIt1 first1, InputIt1 last1, InputIt2 first2, InputIt2 last2,
                            LevenshteinWeightTable weights = {1, 1, 1},
                            size_t score_cutoff = std::numeric_limits<size_t>::max(),
                            size_t score_hint = std::numeric_limits<size_t>::max())
{
    return detail::Levenshtein::distance(first1, last1, first2, last2, weights, score_cutoff, score_hint);
}
 
template <typename Sentence1, typename Sentence2>
size_t levenshtein_distance(const Sentence1& s1, const Sentence2& s2,
                            LevenshteinWeightTable weights = {1, 1, 1},
                            size_t score_cutoff = std::numeric_limits<size_t>::max(),
                            size_t score_hint = std::numeric_limits<size_t>::max())
{
    return detail::Levenshtein::distance(s1, s2, weights, score_cutoff, score_hint);
}
 
template <typename InputIt1, typename InputIt2>
size_t levenshtein_similarity(InputIt1 first1, InputIt1 last1, InputIt2 first2, InputIt2 last2,
                              LevenshteinWeightTable weights = {1, 1, 1}, size_t score_cutoff = 0,
                              size_t score_hint = 0)
{
    return detail::Levenshtein::similarity(first1, last1, first2, last2, weights, score_cutoff, score_hint);
}
 
template <typename Sentence1, typename Sentence2>
size_t levenshtein_similarity(const Sentence1& s1, const Sentence2& s2,
                              LevenshteinWeightTable weights = {1, 1, 1}, size_t score_cutoff = 0,
                              size_t score_hint = 0)
{
    return detail::Levenshtein::similarity(s1, s2, weights, score_cutoff, score_hint);
}
 
template <typename InputIt1, typename InputIt2>
double levenshtein_normalized_distance(InputIt1 first1, InputIt1 last1, InputIt2 first2, InputIt2 last2,
                                       LevenshteinWeightTable weights = {1, 1, 1}, double score_cutoff = 1.0,
                                       double score_hint = 1.0)
{
    return detail::Levenshtein::normalized_distance(first1, last1, first2, last2, weights, score_cutoff,
                                                    score_hint);
}
 
template <typename Sentence1, typename Sentence2>
double levenshtein_normalized_distance(const Sentence1& s1, const Sentence2& s2,
                                       LevenshteinWeightTable weights = {1, 1, 1}, double score_cutoff = 1.0,
                                       double score_hint = 1.0)
{
    return detail::Levenshtein::normalized_distance(s1, s2, weights, score_cutoff, score_hint);
}
 
template <typename InputIt1, typename InputIt2>
double levenshtein_normalized_similarity(InputIt1 first1, InputIt1 last1, InputIt2 first2, InputIt2 last2,
                                         LevenshteinWeightTable weights = {1, 1, 1},
                                         double score_cutoff = 0.0, double score_hint = 0.0)
{
    return detail::Levenshtein::normalized_similarity(first1, last1, first2, last2, weights, score_cutoff,
                                                      score_hint);
}
 
template <typename Sentence1, typename Sentence2>
double levenshtein_normalized_similarity(const Sentence1& s1, const Sentence2& s2,
                                         LevenshteinWeightTable weights = {1, 1, 1},
                                         double score_cutoff = 0.0, double score_hint = 0.0)
{
    return detail::Levenshtein::normalized_similarity(s1, s2, weights, score_cutoff, score_hint);
}
 
template <typename InputIt1, typename InputIt2>
Editops levenshtein_editops(InputIt1 first1, InputIt1 last1, InputIt2 first2, InputIt2 last2,
                            size_t score_hint = std::numeric_limits<size_t>::max())
{
    return detail::levenshtein_editops(detail::make_range(first1, last1), detail::make_range(first2, last2),
                                       score_hint);
}
 
template <typename Sentence1, typename Sentence2>
Editops levenshtein_editops(const Sentence1& s1, const Sentence2& s2,
                            size_t score_hint = std::numeric_limits<size_t>::max())
{
    return detail::levenshtein_editops(detail::make_range(s1), detail::make_range(s2), score_hint);
}
 
#ifdef RAPIDFUZZ_SIMD
namespace experimental {
template <int MaxLen>
struct MultiLevenshtein : public detail::MultiDistanceBase<MultiLevenshtein<MaxLen>, size_t, 0,
                                                           std::numeric_limits<int64_t>::max()> {
private:
    friend detail::MultiDistanceBase<MultiLevenshtein<MaxLen>, size_t, 0,
                                     std::numeric_limits<int64_t>::max()>;
    friend detail::MultiNormalizedMetricBase<MultiLevenshtein<MaxLen>, size_t>;
 
    RAPIDFUZZ_CONSTEXPR_CXX14 static size_t get_vec_size()
    {
#    ifdef RAPIDFUZZ_AVX2
        using namespace detail::simd_avx2;
#    else
        using namespace detail::simd_sse2;
#    endif
        RAPIDFUZZ_IF_CONSTEXPR (MaxLen <= 8)
            return native_simd<uint8_t>::size;
        else RAPIDFUZZ_IF_CONSTEXPR (MaxLen <= 16)
            return native_simd<uint16_t>::size;
        else RAPIDFUZZ_IF_CONSTEXPR (MaxLen <= 32)
            return native_simd<uint32_t>::size;
        else RAPIDFUZZ_IF_CONSTEXPR (MaxLen <= 64)
            return native_simd<uint64_t>::size;
 
        static_assert(MaxLen <= 64, "expected MaxLen <= 64");
    }
 
    static size_t find_block_count(size_t count)
    {
        size_t vec_size = get_vec_size();
        size_t simd_vec_count = detail::ceil_div(count, vec_size);
        return detail::ceil_div(simd_vec_count * vec_size * MaxLen, 64);
    }
 
public:
    MultiLevenshtein(size_t count, LevenshteinWeightTable aWeights = {1, 1, 1})
        : input_count(count), PM(find_block_count(count) * 64), weights(aWeights)
    {
        str_lens.resize(result_count());
        if (weights.delete_cost != 1 || weights.insert_cost != 1 || weights.replace_cost > 2)
            throw std::invalid_argument("unsupported weights");
    }
 
    size_t result_count() const
    {
        size_t vec_size = get_vec_size();
        size_t simd_vec_count = detail::ceil_div(input_count, vec_size);
        return simd_vec_count * vec_size;
    }
 
    template <typename Sentence1>
    void insert(const Sentence1& s1_)
    {
        insert(detail::to_begin(s1_), detail::to_end(s1_));
    }
 
    template <typename InputIt1>
    void insert(InputIt1 first1, InputIt1 last1)
    {
        auto len = std::distance(first1, last1);
        int block_pos = static_cast<int>((pos * MaxLen) % 64);
        auto block = (pos * MaxLen) / 64;
        assert(len <= MaxLen);
 
        if (pos >= input_count) throw std::invalid_argument("out of bounds insert");
 
        str_lens[pos] = static_cast<size_t>(len);
        for (; first1 != last1; ++first1) {
            PM.insert(block, *first1, block_pos);
            block_pos++;
        }
        pos++;
    }
 
private:
    template <typename InputIt2>
    void _distance(size_t* scores, size_t score_count, const detail::Range<InputIt2>& s2,
                   size_t score_cutoff = std::numeric_limits<size_t>::max()) const
    {
        if (score_count < result_count())
            throw std::invalid_argument("scores has to have >= result_count() elements");
 
        auto scores_ = detail::make_range(scores, scores + score_count);
        RAPIDFUZZ_IF_CONSTEXPR (MaxLen == 8)
            detail::levenshtein_hyrroe2003_simd<uint8_t>(scores_, PM, str_lens, s2, score_cutoff);
        else RAPIDFUZZ_IF_CONSTEXPR (MaxLen == 16)
            detail::levenshtein_hyrroe2003_simd<uint16_t>(scores_, PM, str_lens, s2, score_cutoff);
        else RAPIDFUZZ_IF_CONSTEXPR (MaxLen == 32)
            detail::levenshtein_hyrroe2003_simd<uint32_t>(scores_, PM, str_lens, s2, score_cutoff);
        else RAPIDFUZZ_IF_CONSTEXPR (MaxLen == 64)
            detail::levenshtein_hyrroe2003_simd<uint64_t>(scores_, PM, str_lens, s2, score_cutoff);
    }
 
    template <typename InputIt2>
    size_t maximum(size_t s1_idx, const detail::Range<InputIt2>& s2) const
    {
        return detail::levenshtein_maximum(str_lens[s1_idx], s2.size(), weights);
    }
 
    size_t get_input_count() const noexcept
    {
        return input_count;
    }
 
    size_t input_count;
    size_t pos = 0;
    detail::BlockPatternMatchVector PM;
    std::vector<size_t> str_lens;
    LevenshteinWeightTable weights;
};
} /* namespace experimental */
#endif /* RAPIDFUZZ_SIMD */
 
template <typename CharT1>
struct CachedLevenshtein : public detail::CachedDistanceBase<CachedLevenshtein<CharT1>, size_t, 0,
                                                             std::numeric_limits<int64_t>::max()> {
    template <typename Sentence1>
    explicit CachedLevenshtein(const Sentence1& s1_, LevenshteinWeightTable aWeights = {1, 1, 1})
        : CachedLevenshtein(detail::to_begin(s1_), detail::to_end(s1_), aWeights)
    {}
 
    template <typename InputIt1>
    CachedLevenshtein(InputIt1 first1, InputIt1 last1, LevenshteinWeightTable aWeights = {1, 1, 1})
        : s1(first1, last1), PM(detail::make_range(first1, last1)), weights(aWeights)
    {}
 
private:
    friend detail::CachedDistanceBase<CachedLevenshtein<CharT1>, size_t, 0,
                                      std::numeric_limits<int64_t>::max()>;
    friend detail::CachedNormalizedMetricBase<CachedLevenshtein<CharT1>>;
 
    template <typename InputIt2>
    size_t maximum(const detail::Range<InputIt2>& s2) const
    {
        return detail::levenshtein_maximum(s1.size(), s2.size(), weights);
    }
 
    template <typename InputIt2>
    size_t _distance(const detail::Range<InputIt2>& s2, size_t score_cutoff, size_t score_hint) const
    {
        if (weights.insert_cost == weights.delete_cost) {
            /* when insertions + deletions operations are free there can not be any edit distance */
            if (weights.insert_cost == 0) return 0;
 
            /* uniform Levenshtein multiplied with the common factor */
            if (weights.insert_cost == weights.replace_cost) {
                // max can make use of the common divisor of the three weights
                size_t new_score_cutoff = detail::ceil_div(score_cutoff, weights.insert_cost);
                size_t new_score_hint = detail::ceil_div(score_hint, weights.insert_cost);
                size_t dist = detail::uniform_levenshtein_distance(PM, detail::make_range(s1), s2,
                                                                   new_score_cutoff, new_score_hint);
                dist *= weights.insert_cost;
 
                return (dist <= score_cutoff) ? dist : score_cutoff + 1;
            }
            /*
             * when replace_cost >= insert_cost + delete_cost no substitutions are performed
             * therefore this can be implemented as InDel distance multiplied with the common factor
             */
            else if (weights.replace_cost >= weights.insert_cost + weights.delete_cost) {
                // max can make use of the common divisor of the three weights
                size_t new_max = detail::ceil_div(score_cutoff, weights.insert_cost);
                size_t dist = detail::indel_distance(PM, detail::make_range(s1), s2, new_max);
                dist *= weights.insert_cost;
                return (dist <= score_cutoff) ? dist : score_cutoff + 1;
            }
        }
 
        return detail::generalized_levenshtein_distance(detail::make_range(s1), s2, weights, score_cutoff);
    }
 
    std::vector<CharT1> s1;
    detail::BlockPatternMatchVector PM;
    LevenshteinWeightTable weights;
};
 
#ifdef RAPIDFUZZ_DEDUCTION_GUIDES
template <typename Sentence1>
explicit CachedLevenshtein(const Sentence1& s1_, LevenshteinWeightTable aWeights = {
                                                     1, 1, 1}) -> CachedLevenshtein<char_type<Sentence1>>;
 
template <typename InputIt1>
CachedLevenshtein(InputIt1 first1, InputIt1 last1,
                  LevenshteinWeightTable aWeights = {1, 1, 1}) -> CachedLevenshtein<iter_value_t<InputIt1>>;
#endif
 
} // namespace rapidfuzz