競技プログラミング用 C++ ライブラリ
components/sequence_compare.hpp · GitHub で見る
共通ページ
#include "lib/components/sequence_compare.hpp"
#include <string_view>
std::string_view a = "abcde", b = "abcxy";
int prefix = sequence_compare_detail::prefix(a.data(), b.data(), 5); // 3
int suffix = sequence_compare_detail::suffix(a.data() + a.size(),
b.data() + b.size(), 5); // 0
2 つの配列の一致部分を調べる内部部品です。Run 列挙・LCS・編集距離で共有しています。文字列には char、整数列にはその要素型のポインターを渡します。返り値と limit は要素数です。
| 操作 | 内容・返り値 | 計算量 |
|---|---|---|
sequence_compare_detail::prefix(a,b,limit) |
a[0] と b[0] から前方に比較した一致長 |
\(O(\mathrm{limit})\) |
sequence_compare_detail::suffix(a_end,b_end,limit) |
末尾の次の位置から後方に比較した一致長 | \(O(\mathrm{limit})\) |
prefix<false>(a,b,limit) / suffix<false>(a_end,b_end,limit)
|
AVX2 を使わずに同じ比較を実行 | \(O(\mathrm{limit})\) |
limit >= 0 とし、前方では [a,a+limit)・[b,b+limit)、後方では [a_end-limit,a_end)・[b_end-limit,b_end) が有効な配列範囲になるように指定してください。limit == 0 では読み込まずに 0 を返し、型付きの null ポインターも渡せます。配列の外側に余分な領域を確保する必要はありません。
要素は == で比較します。浮動小数点の正負のゼロは等しい値として扱います。
要素比較を \(O(1)\) として、時間は最悪 \(O(\mathrm{limit})\)、追加空間は \(O(1)\) です。
#ifndef SEQUENCE_COMPARE_HPP
#define SEQUENCE_COMPARE_HPP
#include <bit>
#include <cstddef>
#include <cstdint>
#include <cstring>
#include <type_traits>
#if defined(__x86_64__)
#include <immintrin.h>
#endif
namespace sequence_compare_detail {
#if defined(__x86_64__)
__attribute__((target("avx2"))) inline std::size_t forward_avx2(const unsigned char *a, const unsigned char *b, std::size_t size) {
std::size_t i = 0;
for (; size - i >= 32; i += 32) {
auto x = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(a + i));
auto y = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(b + i));
auto difference = ~std::uint32_t(_mm256_movemask_epi8(_mm256_cmpeq_epi8(x, y)));
if (difference) return i + std::countr_zero(difference);
}
while (i < size && a[i] == b[i]) ++i;
return i;
}
__attribute__((target("avx2"))) inline std::size_t backward_avx2(const unsigned char *a, const unsigned char *b, std::size_t size) {
std::size_t i = 0;
for (; size - i >= 32; i += 32) {
auto x = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(a - i - 32));
auto y = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(b - i - 32));
auto difference = ~std::uint32_t(_mm256_movemask_epi8(_mm256_cmpeq_epi8(x, y)));
if (difference) return i + std::countl_zero(difference);
}
while (i < size && a[-std::ptrdiff_t(i) - 1] == b[-std::ptrdiff_t(i) - 1]) ++i;
return i;
}
#endif
template<bool Vectorize = true, class T>
inline int prefix(const T *first, const T *second, int limit) {
if (!limit) return 0;
int matched = 0;
if constexpr (std::is_integral_v<T> && std::has_unique_object_representations_v<T>) {
const auto *a = reinterpret_cast<const unsigned char *>(first);
const auto *b = reinterpret_cast<const unsigned char *>(second);
std::size_t offset = 0, size = std::size_t(limit) * sizeof(T);
while (size - offset >= 8) {
std::uint64_t x, y;
std::memcpy(&x, a + offset, 8);
std::memcpy(&y, b + offset, 8);
auto difference = x ^ y;
if (difference) {
int bits = std::endian::native == std::endian::little ? std::countr_zero(difference) : std::countl_zero(difference);
return static_cast<int>((offset + bits / 8) / sizeof(T));
}
offset += 8;
#if defined(__x86_64__)
if constexpr (Vectorize) if (size - offset >= 64 && __builtin_cpu_supports("avx2"))
return static_cast<int>((offset + forward_avx2(a + offset, b + offset, size - offset)) / sizeof(T));
#endif
}
matched = static_cast<int>(offset / sizeof(T));
}
while (matched < limit && first[matched] == second[matched]) ++matched;
return matched;
}
template<bool Vectorize = true, class T>
inline int suffix(const T *first_end, const T *second_end, int limit) {
if (!limit) return 0;
int matched = 0;
if constexpr (std::is_integral_v<T> && std::has_unique_object_representations_v<T>) {
const auto *a = reinterpret_cast<const unsigned char *>(first_end);
const auto *b = reinterpret_cast<const unsigned char *>(second_end);
std::size_t offset = 0, size = std::size_t(limit) * sizeof(T);
while (size - offset >= 8) {
std::uint64_t x, y;
std::memcpy(&x, a - offset - 8, 8);
std::memcpy(&y, b - offset - 8, 8);
auto difference = x ^ y;
if (difference) {
int bits = std::endian::native == std::endian::little ? std::countl_zero(difference) : std::countr_zero(difference);
return static_cast<int>((offset + bits / 8) / sizeof(T));
}
offset += 8;
#if defined(__x86_64__)
if constexpr (Vectorize) if (size - offset >= 64 && __builtin_cpu_supports("avx2"))
return static_cast<int>((offset + backward_avx2(a - offset, b - offset, size - offset)) / sizeof(T));
#endif
}
matched = static_cast<int>(offset / sizeof(T));
}
while (matched < limit && first_end[-matched - 1] == second_end[-matched - 1]) ++matched;
return matched;
}
}
#endif
#line 1 "components/sequence_compare.hpp"
#include <bit>
#include <cstddef>
#include <cstdint>
#include <cstring>
#include <type_traits>
#if defined(__x86_64__)
#include <immintrin.h>
#endif
namespace sequence_compare_detail {
#if defined(__x86_64__)
__attribute__((target("avx2"))) inline std::size_t forward_avx2(const unsigned char *a, const unsigned char *b, std::size_t size) {
std::size_t i = 0;
for (; size - i >= 32; i += 32) {
auto x = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(a + i));
auto y = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(b + i));
auto difference = ~std::uint32_t(_mm256_movemask_epi8(_mm256_cmpeq_epi8(x, y)));
if (difference) return i + std::countr_zero(difference);
}
while (i < size && a[i] == b[i]) ++i;
return i;
}
__attribute__((target("avx2"))) inline std::size_t backward_avx2(const unsigned char *a, const unsigned char *b, std::size_t size) {
std::size_t i = 0;
for (; size - i >= 32; i += 32) {
auto x = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(a - i - 32));
auto y = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(b - i - 32));
auto difference = ~std::uint32_t(_mm256_movemask_epi8(_mm256_cmpeq_epi8(x, y)));
if (difference) return i + std::countl_zero(difference);
}
while (i < size && a[-std::ptrdiff_t(i) - 1] == b[-std::ptrdiff_t(i) - 1]) ++i;
return i;
}
#endif
template<bool Vectorize = true, class T>
inline int prefix(const T *first, const T *second, int limit) {
if (!limit) return 0;
int matched = 0;
if constexpr (std::is_integral_v<T> && std::has_unique_object_representations_v<T>) {
const auto *a = reinterpret_cast<const unsigned char *>(first);
const auto *b = reinterpret_cast<const unsigned char *>(second);
std::size_t offset = 0, size = std::size_t(limit) * sizeof(T);
while (size - offset >= 8) {
std::uint64_t x, y;
std::memcpy(&x, a + offset, 8);
std::memcpy(&y, b + offset, 8);
auto difference = x ^ y;
if (difference) {
int bits = std::endian::native == std::endian::little ? std::countr_zero(difference) : std::countl_zero(difference);
return static_cast<int>((offset + bits / 8) / sizeof(T));
}
offset += 8;
#if defined(__x86_64__)
if constexpr (Vectorize) if (size - offset >= 64 && __builtin_cpu_supports("avx2"))
return static_cast<int>((offset + forward_avx2(a + offset, b + offset, size - offset)) / sizeof(T));
#endif
}
matched = static_cast<int>(offset / sizeof(T));
}
while (matched < limit && first[matched] == second[matched]) ++matched;
return matched;
}
template<bool Vectorize = true, class T>
inline int suffix(const T *first_end, const T *second_end, int limit) {
if (!limit) return 0;
int matched = 0;
if constexpr (std::is_integral_v<T> && std::has_unique_object_representations_v<T>) {
const auto *a = reinterpret_cast<const unsigned char *>(first_end);
const auto *b = reinterpret_cast<const unsigned char *>(second_end);
std::size_t offset = 0, size = std::size_t(limit) * sizeof(T);
while (size - offset >= 8) {
std::uint64_t x, y;
std::memcpy(&x, a - offset - 8, 8);
std::memcpy(&y, b - offset - 8, 8);
auto difference = x ^ y;
if (difference) {
int bits = std::endian::native == std::endian::little ? std::countl_zero(difference) : std::countr_zero(difference);
return static_cast<int>((offset + bits / 8) / sizeof(T));
}
offset += 8;
#if defined(__x86_64__)
if constexpr (Vectorize) if (size - offset >= 64 && __builtin_cpu_supports("avx2"))
return static_cast<int>((offset + backward_avx2(a - offset, b - offset, size - offset)) / sizeof(T));
#endif
}
matched = static_cast<int>(offset / sizeof(T));
}
while (matched < limit && first_end[-matched - 1] == second_end[-matched - 1]) ++matched;
return matched;
}
}