競技プログラミング用 C++ ライブラリ
components/fps_output.hpp · GitHub で見る
共通ページ
Linux x86_64 / GCC 用の内部部品です。呼び出し側が AVX2 対応を実行時に確認します。
| 操作 | 内容・返り値 | 計算量 |
|---|---|---|
prepare_fps_output(out,total) |
出力配列に total 個分の領域を予約 | 再確保なし: \(O(1)\)、再確保時: 既存要素数に比例+確保時間 |
append_fps_words<montgomery>(out,in,n) |
内部ワードを係数へ変換して追加 | 予約済み: \(O(n)\)、再確保時: 既存要素数+n に比例 |
fps_detail::prepare_fps_output(out,total) は結果の最終サイズ total を収容する容量を確保します。既に容量が十分なら何もしません。新しく容量を増やし、これから書く範囲が 131072 語以上の場合、その範囲に完全に含まれる 2 MiB ページへ MADV_HUGEPAGE、4 KiB ページへ MADV_POPULATE_WRITE を指定します。既存の要素は変更せず、サイズも増やしません。これは OS のページ配置・事前準備の助言であり、オブジェクトの生成・書き換えは行いません。未対応の定数はコンパイル時に省略し、システムコールが失敗した場合は通常の割り当てのまま進みます。
大きな結果を複数回に分けて追加する呼び出し側は、追加の前に最終サイズで一度呼びます。append_fps_words 自身も、容量を増やす必要がある場合にこの処理を使います。
fps_detail::append_fps_words<montgomery>(out,in,n) は、in の \(n\) 個の unsigned ワードを std::vector<Mint> の末尾へ追加します。montgomery=true なら入力は内部 Montgomery 表現 \([0,2p)\)、偽なら通常の剰余 \([0,p)\) とし、後者は Montgomery 表現へ変換します。ここで \(p=998244353\) です。
入力は出力と重ねません。入力にはちょうど \(n\) 個の要素があればよく、アラインメントや padding は要求しません。\(n=0\) なら何もせず、入力が null でも構いません。結果の長さが out.max_size() を超える場合は std::length_error を送出します。
64 個の生存する Mint オブジェクトへコピー・変換してから、vector の通常の insert で追加します。vector のサイズを超えた領域への書き込みや、オブジェクトの寿命を省く操作は行いません。
追加処理の時間は \(O(n)\)、変換バッファは \(O(1)\) 空間です。容量の増加と OS の事前ページ準備は、新しい総容量を \(N\) として \(O(N)\) 時間・追加空間を使います。既存の \(m\) 要素を移す再確保には追加で \(O(m)\) 時間が必要です。
#ifndef CP_COMPONENTS_FPS_OUTPUT_HPP
#define CP_COMPONENTS_FPS_OUTPUT_HPP
#include "mint_arithmetic.hpp"
#include <array>
#include <cstring>
#include <cstdint>
#include <stdexcept>
#include <type_traits>
#include <vector>
#if defined(__linux__) && defined(__x86_64__) && defined(__GNUC__) && !defined(__clang__)
#include <immintrin.h>
#include <sys/mman.h>
#pragma GCC push_options
_Pragma("GCC target(\"avx2\")")
#pragma GCC optimize("O3")
namespace fps_detail {
inline void prepare_fps_output(std::vector<mint> &out, std::size_t n) {
if (n <= out.capacity()) return;
out.reserve(n);
if (n - out.size() < (1u << 17)) return;
const auto first = reinterpret_cast<std::uintptr_t>(out.data() + out.size());
const auto end = reinterpret_cast<std::uintptr_t>(out.data() + n);
[[maybe_unused]] const auto advise = [first, end](std::uintptr_t alignment, int advice) {
const auto lo = (first + alignment - 1) & ~(alignment - 1), hi = end & ~(alignment - 1);
if (lo < hi) static_cast<void>(madvise(reinterpret_cast<void *>(lo), hi - lo, advice));
};
#if defined(MADV_HUGEPAGE)
advise(2u * 1024 * 1024, MADV_HUGEPAGE);
#endif
#if defined(MADV_POPULATE_WRITE)
advise(4096, MADV_POPULATE_WRITE);
#endif
}
template <bool montgomery>
inline void append_fps_words(std::vector<mint> &out, const unsigned *in, std::size_t n) {
static_assert(sizeof(mint) == sizeof(unsigned) && std::is_trivially_copyable_v<mint>);
if (n == 0) return;
if (n > out.max_size() - out.size()) throw std::length_error("fps output");
prepare_fps_output(out, out.size() + n);
std::array<mint, 64> batch;
std::size_t i = 0;
for (; i + 64 <= n; i += 64) {
if constexpr (montgomery) std::memcpy(static_cast<void *>(batch.data()), in + i, sizeof(batch));
else {
#pragma GCC unroll 8
for (int j = 0; j < 64; j += 8) {
const auto x = mint_detail::Ops::mul8(_mm256_loadu_si256(reinterpret_cast<const __m256i *>(in + i + j)), _mm256_set1_epi32(932051910u));
std::memcpy(static_cast<void *>(batch.data() + j), &x, sizeof(x));
}
}
out.insert(out.end(), batch.begin(), batch.end());
}
for (; i < n; ++i) {
if constexpr (montgomery) out.emplace_back(mint::word(in[i]));
else out.emplace_back(mint::raw(in[i]));
}
}
}
#pragma GCC pop_options
#endif
#endif
#line 1 "components/fps_output.hpp"
#line 1 "components/mint_arithmetic.hpp"
#line 1 "math/mint.hpp"
#include <cassert>
#include <istream>
#include <ostream>
#include <type_traits>
#if defined(__GNUC__) && !defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
#pragma GCC push_options
#pragma GCC optimize("O3")
#endif
namespace mint_detail {
template <class I>
constexpr bool mint_integral = std::is_integral_v<I>
#if defined(__SIZEOF_INT128__)
|| std::is_same_v<I, __int128> || std::is_same_v<I, unsigned __int128>
#endif
;
struct [[gnu::warn_unused]] Mint {
using mint = Mint;
unsigned v = 0;
[[nodiscard]] static constexpr unsigned mod() { return 998244353; }
[[nodiscard, gnu::always_inline]] static unsigned reduce(unsigned long long x) {
return (x + (unsigned long long)(unsigned(x) * 998244351u) * mod()) >> 32;
}
[[gnu::always_inline]] Mint() = default;
template <class I, std::enable_if_t<mint_integral<I>, int> = 0>
Mint(I x) {
long long r = x % (long long)mod();
if (r < 0) r += mod();
v = reduce((unsigned long long)r * 932051910);
}
[[nodiscard, gnu::always_inline]] static Mint raw(unsigned x) {
Mint a;
a.v = reduce((unsigned long long)x * 932051910);
return a;
}
[[nodiscard, gnu::always_inline]] static Mint word(unsigned x) {
Mint a;
a.v = x;
return a;
}
[[nodiscard, gnu::always_inline]] unsigned val() const {
unsigned r = reduce(v);
return r < mod() ? r : r - mod();
}
friend std::istream &operator>>(std::istream &is, Mint &x) {
long long value;
if (is >> value) x = Mint(value);
return is;
}
friend std::ostream &operator<<(std::ostream &os, const Mint &x) {
return os << x.val();
}
[[gnu::always_inline]] Mint &operator+=(Mint b) {
v += b.v;
if (v >= 2 * mod()) v -= 2 * mod();
return *this;
}
[[gnu::always_inline]] Mint &operator-=(Mint b) {
v -= b.v;
if (v >> 31) v += 2 * mod();
return *this;
}
[[gnu::always_inline]] Mint &operator*=(Mint b) {
v = reduce((unsigned long long)v * b.v);
return *this;
}
Mint &operator/=(Mint b) { return *this *= b.inv(); }
Mint &operator++() { return *this += Mint(1); }
Mint &operator--() { return *this -= Mint(1); }
Mint operator++(int) {
Mint a = *this;
++*this;
return a;
}
Mint operator--(int) {
Mint a = *this;
--*this;
return a;
}
[[nodiscard, gnu::always_inline]] Mint operator+() const { return *this; }
[[nodiscard, gnu::always_inline]] Mint operator-() const { return word(v ? 2 * mod() - v : 0); }
[[nodiscard, gnu::always_inline]] Mint operator+(Mint b) const { return Mint(*this) += b; }
[[nodiscard, gnu::always_inline]] Mint operator-(Mint b) const { return Mint(*this) -= b; }
[[nodiscard, gnu::always_inline]] Mint operator*(Mint b) const { return Mint(*this) *= b; }
[[nodiscard]] Mint operator/(Mint b) const { return Mint(*this) /= b; }
[[nodiscard, gnu::always_inline]] bool operator==(Mint b) const { return (v < mod() ? v : v - mod()) == (b.v < mod() ? b.v : b.v - mod()); }
[[nodiscard, gnu::always_inline]] bool operator!=(Mint b) const { return !(*this == b); }
template <class I, std::enable_if_t<mint_integral<I>, int> = 0>
[[nodiscard]] friend Mint operator+(I a, Mint b) { return Mint(a) + b; }
template <class I, std::enable_if_t<mint_integral<I>, int> = 0>
[[nodiscard]] friend Mint operator-(I a, Mint b) { return Mint(a) - b; }
template <class I, std::enable_if_t<mint_integral<I>, int> = 0>
[[nodiscard]] friend Mint operator*(I a, Mint b) { return Mint(a) * b; }
template <class I, std::enable_if_t<mint_integral<I>, int> = 0>
[[nodiscard]] friend Mint operator/(I a, Mint b) { return Mint(a) / b; }
template <class I, std::enable_if_t<mint_integral<I>, int> = 0>
[[nodiscard]] friend bool operator==(I a, Mint b) { return Mint(a) == b; }
template <class I, std::enable_if_t<mint_integral<I>, int> = 0>
[[nodiscard]] friend bool operator!=(I a, Mint b) { return Mint(a) != b; }
[[nodiscard]] Mint pow(long long n) const {
assert(n >= 0);
if (!n) return Mint(1);
Mint r(1), a = *this;
for (; n > 1; n >>= 1, a *= a)
if (n & 1) r *= a;
return r * a;
}
[[nodiscard, gnu::noinline]] Mint inv() const {
assert(*this != Mint(0));
Mint a2 = *this * *this, a3 = a2 * *this;
Mint a7 = a3 * a3 * *this;
Mint a31 = a7;
for (int i = 0; i < 2; ++i) a31 *= a31;
a31 *= a3;
Mint r = a7;
for (int i = 0; i < 3; ++i) r *= r;
r *= a3;
r *= r;
for (int j = 0; j < 4; ++j) {
for (int i = 0; i < 5; ++i) r *= r;
r *= a31;
}
for (int i = 0; i < 3; ++i) r *= r;
return r * a7;
}
};
} // namespace mint_detail
using mint = mint_detail::Mint;
using fps_mint = mint;
#if defined(__GNUC__) && !defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
#pragma GCC pop_options
#endif
#line 5 "components/mint_arithmetic.hpp"
#if defined(__x86_64__) || defined(__i386__)
#include <immintrin.h>
#endif
#if defined(__GNUC__) && !defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
#pragma GCC push_options
#pragma GCC optimize("O3")
#endif
namespace mint_detail {
struct Ops {
using T = mint;
typedef unsigned int U __attribute__((may_alias));
static constexpr U mod = 998244353, twice = 2 * mod, one = 301989884;
static U reduce(unsigned long long x) { return T::reduce(x); }
static U mul(U a, U b) { return reduce((unsigned long long)a * b); }
static U add(U a, U b) {
U c = a + b;
return c < twice ? c : c - twice;
}
static U sub(U a, U b) {
U c = a - b;
return c < twice ? c : c + twice;
}
static U in(U a) { return mul(a, 932051910); }
static U out(U a) {
a = reduce(a);
return a < mod ? a : a - mod;
}
static U quotient(U r) { return ((unsigned long long)r << 32) / mod; }
static U fixed_mul(U a, U r, U q) {
return a * r - U(((unsigned long long)a * q) >> 32) * mod;
}
#if defined(__x86_64__) || defined(__i386__)
static bool vectorized() {
#if defined(__AVX2__)
return true;
#else
static const bool available = __builtin_cpu_supports("avx2");
return available;
#endif
}
using I8 = __m256i;
__attribute__((target("avx2"))) static I8 load8(const U *a) {
return _mm256_loadu_si256(reinterpret_cast<const I8 *>(a));
}
__attribute__((target("avx2"))) static void store8(U *a, I8 x) {
_mm256_storeu_si256(reinterpret_cast<I8 *>(a), x);
}
__attribute__((target("avx2"), always_inline)) static inline I8 reduce8(I8 x, I8 y) {
I8 r = _mm256_set1_epi32(998244351), p = _mm256_set1_epi32(mod);
x = _mm256_add_epi64(x, _mm256_mul_epu32(_mm256_mul_epu32(x, r), p));
y = _mm256_add_epi64(y, _mm256_mul_epu32(_mm256_mul_epu32(y, r), p));
return _mm256_blend_epi32(_mm256_srli_epi64(x, 32), y, 0xaa);
}
__attribute__((target("avx2"))) static I8 mul8(I8 a, I8 b) {
return reduce8(_mm256_mul_epu32(a, b), _mm256_mul_epu32(_mm256_srli_epi64(a, 32), _mm256_srli_epi64(b, 32)));
}
__attribute__((target("avx2"))) static I8 fixed_mul8(I8 a, I8 r, I8 q) {
I8 x = _mm256_srli_epi64(_mm256_mul_epu32(a, q), 32);
I8 y = _mm256_mul_epu32(_mm256_srli_epi64(a, 32), q);
I8 t = _mm256_blend_epi32(x, y, 0xaa);
return _mm256_sub_epi32(_mm256_mullo_epi32(a, r),
_mm256_mullo_epi32(t, _mm256_set1_epi32(mod)));
}
__attribute__((target("avx2"))) static I8 fixed_even8(I8 a, I8 r, I8 q) {
I8 t = _mm256_srli_epi64(_mm256_mul_epu32(a, q), 32);
return _mm256_sub_epi64(_mm256_mul_epu32(a, r),
_mm256_mul_epu32(t, _mm256_set1_epi32(mod)));
}
__attribute__((target("avx2"))) static I8 add8(I8 a, I8 b) {
I8 x = _mm256_add_epi32(a, b);
return _mm256_min_epu32(x, _mm256_sub_epi32(x, _mm256_set1_epi32(twice)));
}
__attribute__((target("avx2"))) static I8 sub8(I8 a, I8 b) {
I8 x = _mm256_sub_epi32(a, b);
return _mm256_min_epu32(x, _mm256_add_epi32(x, _mm256_set1_epi32(twice)));
}
#endif
};
}
#if defined(__GNUC__) && !defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
#pragma GCC pop_options
#endif
#line 5 "components/fps_output.hpp"
#include <array>
#include <cstring>
#include <cstdint>
#include <stdexcept>
#line 10 "components/fps_output.hpp"
#include <vector>
#if defined(__linux__) && defined(__x86_64__) && defined(__GNUC__) && !defined(__clang__)
#include <immintrin.h>
#include <sys/mman.h>
#pragma GCC push_options
_Pragma("GCC target(\"avx2\")")
#pragma GCC optimize("O3")
namespace fps_detail {
inline void prepare_fps_output(std::vector<mint> &out, std::size_t n) {
if (n <= out.capacity()) return;
out.reserve(n);
if (n - out.size() < (1u << 17)) return;
const auto first = reinterpret_cast<std::uintptr_t>(out.data() + out.size());
const auto end = reinterpret_cast<std::uintptr_t>(out.data() + n);
[[maybe_unused]] const auto advise = [first, end](std::uintptr_t alignment, int advice) {
const auto lo = (first + alignment - 1) & ~(alignment - 1), hi = end & ~(alignment - 1);
if (lo < hi) static_cast<void>(madvise(reinterpret_cast<void *>(lo), hi - lo, advice));
};
#if defined(MADV_HUGEPAGE)
advise(2u * 1024 * 1024, MADV_HUGEPAGE);
#endif
#if defined(MADV_POPULATE_WRITE)
advise(4096, MADV_POPULATE_WRITE);
#endif
}
template <bool montgomery>
inline void append_fps_words(std::vector<mint> &out, const unsigned *in, std::size_t n) {
static_assert(sizeof(mint) == sizeof(unsigned) && std::is_trivially_copyable_v<mint>);
if (n == 0) return;
if (n > out.max_size() - out.size()) throw std::length_error("fps output");
prepare_fps_output(out, out.size() + n);
std::array<mint, 64> batch;
std::size_t i = 0;
for (; i + 64 <= n; i += 64) {
if constexpr (montgomery) std::memcpy(static_cast<void *>(batch.data()), in + i, sizeof(batch));
else {
#pragma GCC unroll 8
for (int j = 0; j < 64; j += 8) {
const auto x = mint_detail::Ops::mul8(_mm256_loadu_si256(reinterpret_cast<const __m256i *>(in + i + j)), _mm256_set1_epi32(932051910u));
std::memcpy(static_cast<void *>(batch.data() + j), &x, sizeof(x));
}
}
out.insert(out.end(), batch.begin(), batch.end());
}
for (; i < n; ++i) {
if constexpr (montgomery) out.emplace_back(mint::word(in[i]));
else out.emplace_back(mint::raw(in[i]));
}
}
}
#pragma GCC pop_options
#endif