// ===== REFERENCES =====
// [1] 本账号（saffah_cc_v41_agg1）当前最好件 = 同 lane `sm7_` 的 `work/sm7_kg3_2048.cpp`
// （提交 #106204 = 2027.205180 ms，= 用户 saffah_codex_6s_agg2 #106156 的逐字节复刻 + `KG3 2560→2048`）
// —— **本文件的正文基座**（本发只在它的 `ctr_ins`/`ctr_qry` 两处做单变量替换）。
// [2] ★ **本刀的来源与本账号的已兑现读数**（同族 `1010` / `1010a` 两题，同一份计数器结构）：
// duck.ac 用户 saffah_codex_6s_agg2 提交 #103094 <https://duck.ac/submission/103094>
// —— "把顶层 CT5 改成**后缀数组**（`CT5[0..7] += CT5INC[v>>17]`，查询退化成一条标量载入）"
// 这一恒等变形的出处；本账号把**同一条恒等变形推广到 CT1..CT4**（每题只改计数器两处）：
// * 本账号 #103203 <https://duck.ac/submission/103203>（题 1010a）—— 判题机 **−3.20%**（181.443 → 175.637）；
// * 本账号 #106210 <https://duck.ac/submission/106210>（题 1010）—— 该件的直接后继成分；
// * 本账号 #106218 <https://duck.ac/submission/106218>（题 1010a）。
// 本文件把同一变形移植到本题（`KGB`/`CT0..CT5` 阶梯与 4D 完全同构）。
// **语义恒等**：`CT_k[p] := 本组内块号 < p 的插入数` ⇒ 查询 = 一条标量载入；
// 插入 = 该组内一组 8 车道的宽加（行 r 的掩码在"严格高于 r"的车道上放 1）。
// 各层贡献互不重叠、合起来恰好覆盖 `[0, v)`，与原"前缀归约"逐位等价。
// [3] 同 lane `sm7_` 的构建/闸门工具：`problems/1010e6/work/ke6_build.py`（`--brute` / `--probe`）。
// [4] 本文件未取用其它他人代码。
// ======================
// ===== 思路 =====
// 现状：`mine = 2027.205180`（#106204）· `T = 2027.488114`（对手 #106156）· 严支 `thr = 2007.214233`
// ⇒ 缺口 **19.990947 ms（0.986%）**，红态（`mine < T` 但严支要求 ≤0.99·T）。
//
// ★ 唯一改动 = **[SFX] 计数器后缀化**（只碰 `ctr_ins` / `ctr_qry` 两个函数 + CT2 的声明）：
// 原来的每一层都是「块内计数 + 查询时做一次掩码前缀归约」：
// `ctr_ins`: `CT1[v>>5]++; CT2[v>>8]++; ...`（6 条标量 RMW）
// `ctr_qry`: `su32_8(CT4+…)+su32_8(CT3+…)+su16_8(CT2+…)+su16_8(CT1+…)+su8_32(CT0+…)`（4 条向量归约 + 1 条）
// 改成**后缀数组**后：查询 = **5 条标量载入**（CT5/CT4/CT3/CT2/CT1）；插入 = 5 条 16/32 B 宽加。
// 唯一约束（与 1010 相同）：组内后缀值上界 = 8 × 本层块跨度 × 单值重数 ⇒
// `CT1`(跨度32)=32512 仍 u16 ✓；**`CT2`(跨度256)=260096 ⇒ 必须 u16→u32**；CT3/CT4/CT5 本就是 u32 ✓。
//
// 闸门：
// 1) `ke6_build.py --brute --kbrute 5`（独立 O(n²k) 暴力，3 种分布）⇒ **mism=0 / BRUTE OK** ✓；
// 2) **判题机 probe（n=4e5 k=5，checksum 恒为 4981929175 = 本 lane 的历史常数）**：
// 基座 **1967.71 Mcyc** → 本发 **1912.46 Mcyc** ⇒ **−2.81%**，**两臂 checksum 逐位相同** ✓。
// 未改动：引擎其余部分、`KGB_(d)`、`LEAFNB 8192`、`MF_*` 全部常量、叶/归并/事件构造。
// ================
// ===== REFERENCES =====
// [1] duck.ac 用户 saffah_codex_6s_agg2，提交 #106156 <https://duck.ac/submission/106156>（2027.488114 ms，当代 T）。
// 用途：**本文件正文 = 该提交正文逐字节复制**（RULES §2.16 第一步）。它自述其父本 = **本账号 #102291**
// （= 我方最好件 2048.201396 ms；含我方 #100459 的"链式内联 XE"）⇒ 本次复刻 = 追平/反超对手当代件。
// [2] 本账号 #102291 <https://duck.ac/submission/102291>（`work/e11_sub_1011e6.cpp`）—— 对手文件里已含我方 XE 刀 ✓。
// [3] duck.ac 用户 saffah_cc_v41_260924，提交 #102263 <https://duck.ac/submission/102263> —— 更早的 CDQ 引擎父本。
// ======================
// ===== 思路 =====
// 入场：`mine = 2048.201396`（#102291）/ `T = 2027.488114`（#106156）/ 宽支 `1.005*T = 2037.626555` ⇒ 红 10.574841 ms；
// 严支（交新最好后）0.99*T+1µs ⇒ 缺 ~40.99 ms。
// ★ 本发 = **纯复刻**（协调者 §2.18.993 的"按是否结构件排序地抄"）。逐行 diff `work/e11_rival_102263.cpp → ref2/rival_106156.cpp`
// 得到对手相对其*旧*父本的全部增量（其中 `[XE]` 那几段是**我方 #102291 的刀**，他照抄了 ✓），**他自己的两件**是：
// ① **事件构造改成计数排序的组尾 + 两种角色一次载入双写**：`for(v=0..n) { g1=g_x[v]; ... ev[m+t-g]=mk(key,1,id,nk); ev[m+cnt+t-g]=mk(key,0,id,nk); }`
// —— 原来的 `while(g<n){ v=X0[g_ord[g]]; ... }` 每事件要**随机 gather 一次首维** ⇒ 本件把它换成**顺序扫 `g_x`（计数排序组尾）** ✓（对手自述 "avoid random first-coordinate reads"）
// 且把两次 `mk()` 的**公共载入合并** ✓。按本 lane 的兑现律（"省 4 MB 级随机 gather"类，probe 低估 5.7×），这是**结构件** ⇒ 优先 ✓。
// ② **`KG2 2560 → 2048`**（二级叶阈值，参数件 ✓）。
// ★ 为何整份复刻：我方 2048.201 > T 2027.488（慢 1.01%）⇒ 按 §2.18.566② "mine > T ⇒ 整份复刻有效" ⇒ 一次把这 1.01% 压掉 ✓。
// ★ 正确性闸门：① 与 `work/e11_sub_1011e6.cpp`（#102291）**逐位对拍**（多 mode × 多 n）；② gcc93 编译；③ 配置回显 + 断言输出非空。
// 对手件在判题机上 Accepted ✓；其自述的实现坑（`cnt[]` 静态化）已在 [1] 里修好 ✓。
// ================
#ifndef MF_CTRPF
#define MF_CTRPF 32
#endif
#ifndef MF_CTR5
#define MF_CTR5 1
#endif
#ifndef OUTPF
#define OUTPF 4
#endif
#ifndef MF_LEAFLP
#define MF_LEAFLP 64
#endif
#ifndef MF_LEAFG
#define MF_LEAFG 2
#endif
#pragma GCC optimize("O3","unroll-loops")
#pragma GCC target("avx2,bmi,bmi2,popcnt,lzcnt")
// duck.ac 1010 (4D) / 1011 (5D) -- "count for every point the number of points that are
// smaller in *every* dimension" (n = 300000, coords in [0,n), strict comparisons).
//
// Algorithm: D&C over dimensions on an event sequence, keys carried inline.
// * 2n events: one query event + one insert event per point, keyed by dim 0
// (a query event is stored *before* an insert event of the same key, which encodes
// the strict-inequality rule).
// * solve(seq,m,d): seq is ordered by dim d and every event carries the coordinate of
// dim d+1 in bits 32..63 of its u64. Split by position, recurse both halves, then
// merge the two halves by the carried key (a plain unsigned `<` on the packed u64,
// so equal keys keep query-before-insert order); the same pass collects the cross
// events (left inserts + right queries), re-keys them to dim d+2 and recurses one
// dimension deeper. Ordering by dim d with that tie rule makes "insert before query"
// exactly "all dims so far strictly smaller", and equal coordinates are never counted.
// * d == K-1: every insert-before-query pair now qualifies, so one O(m) running counter
// finishes the innermost dimension (no log factor).
// * m <= 32: direct O(m^2) counting plus an insertion sort.
//
// Local timings (i5-6500 @ 3.2GHz, best of 3, n = 300000 uniform random):
// 4D 1.57 s 5D 3.85 s
// (Board #1: 508.596 ms / 745.358 ms -- this is still 3x / 5x behind.)
// k-dimensional dominance counting: out[i] = #{j : all dims of j < dims of i}
//
// Event-sequence CDQ with keys carried inline.
// event (u64) = (key << 32) | (ins << 31) | id ins = 1 for an "insert" event
// key = coordinate of dim (level+1): the field used by this level's merge.
// The packing makes a plain unsigned comparison put, for equal keys, the query
// event first -- exactly the strict-inequality rule (equal coords must not count).
//
// solve(seq, m, d): seq is ordered by dim d (tie rule above), elements carry the
// key for dim d+1; count pairs (insert before query) with dims d+1..K-1 strictly
// smaller; return `out` = the same events merged by the carried key (= dim d+1).
// Cross = A-inserts + B-queries re-keyed to dim d+2, recursed at level d+1.
// d == K-1 : every (insert before query) pair qualifies -> running counter, O(m).
#include <cstring>
static volatile unsigned long long gCV_ = 0;
#include <immintrin.h>
#pragma GCC push_options
#pragma GCC target("avx2,popcnt,bmi2")
#include <cstdint>
typedef unsigned u32;
typedef uint64_t u64;
typedef unsigned char u8;
typedef unsigned short u16;
#ifndef KG_COND
#define KG_COND 0
#endif
#ifndef KG_PF
#define KG_PF 0
#endif
#ifndef KG_BASE
#define KG_BASE 2560
#endif
/* ---- DEPTH-DEPENDENT LEAF THRESHOLD (Lane 1011e6, 2026-09-27) ----------------
The leaf's cost is ~ (number of comparisons + 1) * m per event, because it is
an O(m^2/8) all-pairs scan whose inner block does 2*(1+nd) ALU uops per 8 pairs.
nd = K-(d+2) is a COMPILE-TIME constant per depth, so the optimal leaf size m*(d)
is NOT the same at every depth. MEASURED at n=3e5,k=5 (judge, custom_test):
d=0 : 256 leaf calls, 600 000 events, nd=3 (4 comparisons)
d=1 : 1024 leaf calls, 2 400 008 events, nd=2 (3)
d=2 : 1792 leaf calls, 4 199 134 events, nd=1 (2)
=> 58 % of all leaf events are at d=2, where the leaf is cheap per event, and
only 8 % are at d=0 where it is dearest. The single-valued KG_BASE sweep that
the fleet ran could not see this: it moves all three thresholds together, and a
sum of curves whose optima differ looks FLAT (measured: total varies +-0.16 %
over 1280..4096 at n=3e5 while base_case itself varies 4.6 %).
The arrays are sized by the LARGEST threshold. */
#define KG0 1920
#define KG1 2560
#define KG2 2048
#define KG3 2048
static int KGB[8] = { KG0, KG1, KG2, KG3, KG3, KG3, KG3, KG3 };
#define KG_BMAX (KG0>KG1 ? (KG0>KG2? (KG0>KG3?KG0:KG3) : (KG2>KG3?KG2:KG3)) : (KG1>KG2? (KG1>KG3?KG1:KG3) : (KG2>KG3?KG2:KG3)))
#define KGB_(dd) (KGB[(dd) < 0 ? 0 : ((dd) > 7 ? 7 : (dd))])
#ifndef LEAFNB
#define LEAFNB 8192
#endif
#define KG_MAXK 5
#define KG_ARENA (16000000)
namespace kg {
// ------------ fused innermost-dimension counter (prefix count over [0,n)) -------------
// AVX2. Levels: u8 leaf /32 values, u16 per 32 and per 256, u32 per 2048/16384/131072.
// Every u16 block count must stay <= 32767 (madd uses signed lanes); the guard (max
// multiplicity of a last-dim coordinate <= 127) bounds 32*127=4064 and 256*127=32512.
#define CT_MAXV 1048580u
static u8 CT0[CT_MAXV];
static u16 CT1[CT_MAXV / 32 + 8];
static u32 CT2[CT_MAXV / 256 + 8]; /* [SFX] u16->u32：组内后缀上界 260096 */
static u32 CT3[CT_MAXV / 2048 + 8];
static u32 CT4[CT_MAXV / 16384 + 8];
static u32 CT5[CT_MAXV / 131072 + 8];
static int FUSE_OK;
static inline u32 su8_32(const u8 *p, int cnt) {
__m256i v = _mm256_loadu_si256((const __m256i *)p);
__m256i m = _mm256_cmpgt_epi8(_mm256_set1_epi8((char)cnt),
_mm256_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31));
__m256i x = _mm256_and_si256(v, m);
__m256i s = _mm256_sad_epu8(x, _mm256_setzero_si256());
__m128i t = _mm_add_epi64(_mm256_castsi256_si128(s), _mm256_extracti128_si256(s, 1));
return (u32)((unsigned long long)_mm_cvtsi128_si64(t) + (unsigned long long)_mm_extract_epi64(t, 1));
}
static inline u32 su16_8(const u16 *p, int cnt) {
__m128i v = _mm_loadu_si128((const __m128i *)p);
__m128i x = _mm_and_si128(v, _mm_cmpgt_epi16(_mm_set1_epi16((short)cnt),
_mm_setr_epi16(0, 1, 2, 3, 4, 5, 6, 7)));
__m128i s = _mm_madd_epi16(x, _mm_set1_epi16(1));
s = _mm_add_epi32(s, _mm_shuffle_epi32(s, _MM_SHUFFLE(1, 0, 3, 2)));
s = _mm_add_epi32(s, _mm_shuffle_epi32(s, _MM_SHUFFLE(2, 3, 0, 1)));
return (u32)_mm_cvtsi128_si32(s);
}
static inline u32 su32_8(const u32 *p, int cnt) {
__m256i v = _mm256_loadu_si256((const __m256i *)p);
__m256i x = _mm256_and_si256(v, _mm256_cmpgt_epi32(_mm256_set1_epi32(cnt),
_mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7)));
__m128i t = _mm_add_epi32(_mm256_castsi256_si128(x), _mm256_extracti128_si256(x, 1));
t = _mm_add_epi32(t, _mm_shuffle_epi32(t, _MM_SHUFFLE(1, 0, 3, 2)));
t = _mm_add_epi32(t, _mm_shuffle_epi32(t, _MM_SHUFFLE(2, 3, 0, 1)));
return (u32)_mm_cvtsi128_si32(t);
}
/* [SFX] 后缀计数器：行 r 在"严格高于 r"的车道上放 1 ⇒ 插入落在块 r 时，
把它所在 8 块组里所有需要开始统计它的后缀位上各加一。查询退化成一条标量载入。 */
static const u16 INC16[8][8] __attribute__((aligned(32))) = {
{0,1,1,1,1,1,1,1},{0,0,1,1,1,1,1,1},{0,0,0,1,1,1,1,1},{0,0,0,0,1,1,1,1},
{0,0,0,0,0,1,1,1},{0,0,0,0,0,0,1,1},{0,0,0,0,0,0,0,1},{0,0,0,0,0,0,0,0}};
static const u32 INC32[8][8] __attribute__((aligned(32))) = {
{0,1,1,1,1,1,1,1},{0,0,1,1,1,1,1,1},{0,0,0,1,1,1,1,1},{0,0,0,0,1,1,1,1},
{0,0,0,0,0,1,1,1},{0,0,0,0,0,0,1,1},{0,0,0,0,0,0,0,1},{0,0,0,0,0,0,0,0}};
#define SFX16(p, r) do { __m128i _a = _mm_load_si128((const __m128i *)(p)); \
__m128i _b = _mm_load_si128((const __m128i *)(INC16 + (r))); \
_mm_store_si128((__m128i *)(p), _mm_add_epi16(_a, _b)); } while (0)
#define SFX32(p, r) do { __m256i _a = _mm256_load_si256((const __m256i *)(p)); \
__m256i _b = _mm256_load_si256((const __m256i *)(INC32 + (r))); \
_mm256_store_si256((__m256i *)(p), _mm256_add_epi32(_a, _b)); } while (0)
static inline void ctr_ins(u32 v) {
CT0[v]++;
SFX16(CT1 + ((v >> 8) << 3), (v >> 5) & 7);
SFX32(CT2 + ((v >> 11) << 3), (v >> 8) & 7);
SFX32(CT3 + ((v >> 14) << 3), (v >> 11) & 7);
SFX32(CT4 + ((v >> 17) << 3), (v >> 14) & 7);
SFX32(CT5, (v >> 17) & 7); /* CT5 活跃项 <= 8 ⇒ 整组一个 SFX32 */
}
/* ARM `CT` (ported from the sister row 1011a): RESET THE HIERARCHY INSTEAD OF DECREMENTING IT.
ctr_clr mirrored all six of ctr_ins's read-modify-writes, so every element cost SIX random
RMWs to insert and SIX more to clear. Five of those six arrays are SMALL -- CT1 18 KB,
CT2 2.4 KB, CT3 592 B, CT4 76 B, CT5 16 B at n=1e5 (62 KB / 8 KB / 2 KB / 244 B / 60 B at
n=1e6) -- and they are only ever read through ctr_qry, which sums them, so restoring them to
their initial zero is a pure RESET, not an inverse walk. CT0 stays a per-element decrement
because it is the one level that is genuinely large and densely addressed, so walking `dirt`
still beats clearing it. Semantics are unchanged: ctr_qry reads the same six levels and the
structure is left in its initial state for the next call, which is what the decrements were for. */
static inline void ctr_clr(u32 v) { CT0[v]--; }
static inline void ctr_reset_hi(void) {
memset(CT1, 0, sizeof(CT1)); memset(CT2, 0, sizeof(CT2));
memset(CT3, 0, sizeof(CT3)); memset(CT4, 0, sizeof(CT4));
memset(CT5, 0, sizeof(CT5));
}
static inline u32 ctr_qry(u32 v) {
u32 k5 = v >> 17, s = 0;
s += CT5[v >> 17]; /* [SFX] 标量后缀载入替代向量归约 */
s += CT4[v >> 14];
s += CT3[v >> 11];
s += CT2[v >> 8];
s += CT1[v >> 5];
s += su8_32(CT0 + (v & ~31u), (int)(v & 31));
return s;
}
static int K;
#ifndef LFAUTO
#define LFAUTO 0
#endif
static int LFSH;
static int LFD;
static const u32 *XD[KG_MAXK];
static u32 *OUT;
static u64 *AB, *ABtop;
static inline int eid(u64 v) { return (int)(v & 0x1FFFFFu); }
static inline int lfbucket(u64 v) { return (int)(((u32)(v >> 43)) >> LFSH); }
static inline int eq_(u64 v) { return (int)((v >> 42) & 1); }
static inline int keep_test(int takeA, int role) { return 1 - (takeA ^ role); }
static inline u32 embnk(u64 v) { return (u32)((v >> 21) & 0x1FFFFFu); }
static inline u64 mk(u32 key, int q, int id, u32 nk) {
return ((u64)key << 43) | ((u64)(1 - q) << 42) | ((u64)nk << 21) | (u64)id;
}
static void solve(u64 *seq, int m, int d, u64 *out, int needOut);
// resolve a fusion-level cross: events carry the last dim (K-1) coordinate in bits 32..63
// and are ordered by dim K-2; a left-insert before a right-query with a strictly smaller
// last-dim value is a hit.
static void ctr_solve(u64 *cross, int cx, u32 *dirt) {
int nd = 0;
for (int i = 0; i < cx; i++) {
#if MF_CTRPF > 0
if (i + MF_CTRPF < cx) {
u32 vf = (u32)(cross[i + MF_CTRPF] >> 43);
__builtin_prefetch(&CT0[vf], 1, 0);
__builtin_prefetch(&CT1[vf >> 5], 1, 0);
__builtin_prefetch(&CT2[vf >> 8], 1, 0);
__builtin_prefetch(&OUT[cross[i + MF_CTRPF] & 0x1FFFFFu], 1, 0);
}
#endif
u64 v = cross[i];
u32 val = (u32)(v >> 43);
if ((v >> 42) & 1) { ctr_ins(val); dirt[nd++] = val; }
else OUT[eid(v)] += ctr_qry(val);
}
for (int i = 0; i < nd; i++) {
#if MF_CTRPF > 0
if (i + MF_CTRPF < nd) {
u32 vf = dirt[i + MF_CTRPF];
__builtin_prefetch(&CT0[vf], 1, 1);
__builtin_prefetch(&CT1[vf >> 5], 1, 1);
__builtin_prefetch(&CT2[vf >> 8], 1, 1);
}
#endif
ctr_clr(dirt[i]);
}
ctr_reset_hi();
}
// small base: compact insert list + SIMD comparison, then insertion sort by the key
static const u32 VMASK[8][8] = {
{0,0,0,0,0,0,0,0},
{~0u,0,0,0,0,0,0,0},
{~0u,~0u,0,0,0,0,0,0},
{~0u,~0u,~0u,0,0,0,0,0},
{~0u,~0u,~0u,~0u,0,0,0,0},
{~0u,~0u,~0u,~0u,~0u,0,0,0},
{~0u,~0u,~0u,~0u,~0u,~0u,0,0},
{~0u,~0u,~0u,~0u,~0u,~0u,~0u,0}};
static inline u32 hsum256(__m256i v) {
__m128i t = _mm_add_epi32(_mm256_castsi256_si128(v), _mm256_extracti128_si256(v, 1));
t = _mm_add_epi32(t, _mm_shuffle_epi32(t, _MM_SHUFFLE(1, 0, 3, 2)));
t = _mm_add_epi32(t, _mm_shuffle_epi32(t, _MM_SHUFFLE(2, 3, 0, 1)));
return (u32)_mm_cvtsi128_si32(t);
}
template<int ND>
static void base_case_t(u64 *seq, int m, int d, u64 *out, int needOut) {
const int nd = ND;
__attribute__((aligned(32)))
u32 IK[KG_BMAX + 8], IC0[KG_BMAX + 8], IC1[KG_BMAX + 8], IC2[KG_BMAX + 8];
u32 QK[KG_BMAX + 8], QC0[KG_BMAX + 8], QC1[KG_BMAX + 8], QC2[KG_BMAX + 8];
u32 QID[KG_BMAX + 8], QN[KG_BMAX + 8];
const u32 *X0 = nd > 0 ? XD[d + 2] : 0, *X1 = nd > 1 ? XD[d + 3] : 0, *X2 = nd > 2 ? XD[d + 4] : 0;
/* [XE] the X0 dimension (dim d+2) is already inside the event: at d == 0 by mk()'s nk,
at d >= 1 by the level-(d-1) cross's embed (valid exactly while d + 2 <= K - 1) */
const int useEmb = (nd > 0) && (d == 0 || d <= K - 3);
int ni = 0, nq = 0;
for (int t = 0; t < m; t++) {
#if MF_LEAFLP > 0
if (t + MF_LEAFLP < m) {
int idf = eid(seq[t + MF_LEAFLP]);
if (nd > 0 && !useEmb) __builtin_prefetch(&X0[idf], 0, 1);
if (nd > 1) __builtin_prefetch(&X1[idf], 0, 1);
if (nd > 2) __builtin_prefetch(&X2[idf], 0, 1);
}
#endif
u64 v = seq[t];
int id = eid(v);
u32 k = (u32)(v >> 43);
u32 c0 = 0, c1 = 0, c2 = 0;
if (nd > 0) c0 = useEmb ? embnk(v) : X0[id];
if (nd > 1) c1 = X1[id];
if (nd > 2) c2 = X2[id];
int isins = (int)((v >> 42) & 1u);
IK[ni] = k; IC0[ni] = c0; IC1[ni] = c1; IC2[ni] = c2; ni += isins;
QK[nq] = k; QC0[nq] = c0; QC1[nq] = c1; QC2[nq] = c2; QID[nq] = (u32)id; QN[nq] = (u32)ni; nq += 1 - isins;
}
#if MF_LEAFG >= 2
// ---- paired-query leaf: one insert-block load serves two queries ----
/* ARM L_outpf (lane l1011e6_ctr): the leaf's counting loop ends in a RANDOM
read-modify-write, `OUT[QID[q]] += c`. A random LOAD is absorbed by the OOO window (on
this row the X-gather id-packing is an exact null, +0.014 %), but a random RMW is NOT:
the store must own the line before it can retire. `L_outhot` (the same loop, destination
forced onto two cache lines) prices the whole component at -3.86 % on the judge at
n=3e5. This arm SW-prefetches the destination 2*OUTPF queries ahead -- the mechanism the
construction lane banked at -5.90 % on this row for the merge's RMWs. A prefetch has no
semantics, so the checksum must be unchanged. */
int q = 0;
for (; q + 2 <= nq; q += 2) {
if (q + 2 * OUTPF < nq) {
__builtin_prefetch(&OUT[QID[q + 2 * OUTPF]], 1, 0);
__builtin_prefetch(&OUT[QID[q + 2 * OUTPF + 1]], 1, 0);
}
const int n = ((int)QN[q] < (int)QN[q + 1]) ? (int)QN[q] : (int)QN[q + 1];
const int nA = (int)QN[q], nB = (int)QN[q + 1];
__m256i kvA = _mm256_set1_epi32((int)QK[q]), kvB = _mm256_set1_epi32((int)QK[q + 1]);
__m256i v0A = _mm256_set1_epi32((int)QC0[q]), v0B = _mm256_set1_epi32((int)QC0[q + 1]);
__m256i v1A = _mm256_set1_epi32((int)QC1[q]), v1B = _mm256_set1_epi32((int)QC1[q + 1]);
__m256i v2A = _mm256_set1_epi32((int)QC2[q]), v2B = _mm256_set1_epi32((int)QC2[q + 1]);
__m256i accA = _mm256_setzero_si256(), accB = _mm256_setzero_si256();
int j = 0;
if (nd >= 3) {
for (; j + 8 <= n; j += 8) {
__m256i IKv = _mm256_load_si256((const __m256i *)(IK + j));
__m256i C0v = _mm256_load_si256((const __m256i *)(IC0 + j));
__m256i C1v = _mm256_load_si256((const __m256i *)(IC1 + j));
__m256i C2v = _mm256_load_si256((const __m256i *)(IC2 + j));
__m256i mA = _mm256_and_si256(_mm256_cmpgt_epi32(kvA, IKv), _mm256_cmpgt_epi32(v0A, C0v));
__m256i mB = _mm256_and_si256(_mm256_cmpgt_epi32(kvB, IKv), _mm256_cmpgt_epi32(v0B, C0v));
mA = _mm256_and_si256(mA, _mm256_cmpgt_epi32(v1A, C1v));
mB = _mm256_and_si256(mB, _mm256_cmpgt_epi32(v1B, C1v));
mA = _mm256_and_si256(mA, _mm256_cmpgt_epi32(v2A, C2v));
mB = _mm256_and_si256(mB, _mm256_cmpgt_epi32(v2B, C2v));
accA = _mm256_sub_epi32(accA, mA);
accB = _mm256_sub_epi32(accB, mB);
}
} else {
for (; j + 8 <= n; j += 8) {
__m256i IKv = _mm256_load_si256((const __m256i *)(IK + j));
__m256i mA = _mm256_cmpgt_epi32(kvA, IKv);
__m256i mB = _mm256_cmpgt_epi32(kvB, IKv);
if (nd > 0) {
__m256i C0v = _mm256_load_si256((const __m256i *)(IC0 + j));
mA = _mm256_and_si256(mA, _mm256_cmpgt_epi32(v0A, C0v));
mB = _mm256_and_si256(mB, _mm256_cmpgt_epi32(v0B, C0v));
}
if (nd > 1) {
__m256i C1v = _mm256_load_si256((const __m256i *)(IC1 + j));
mA = _mm256_and_si256(mA, _mm256_cmpgt_epi32(v1A, C1v));
mB = _mm256_and_si256(mB, _mm256_cmpgt_epi32(v1B, C1v));
}
accA = _mm256_sub_epi32(accA, mA);
accB = _mm256_sub_epi32(accB, mB);
}
}
// per-query tails: each member finishes [j, QN[.]) on its own
{
int jj = j;
__m256i acc = accA;
for (; jj + 8 <= nA; jj += 8) {
__m256i msk = _mm256_cmpgt_epi32(kvA, _mm256_load_si256((const __m256i *)(IK + jj)));
if (nd > 0) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v0A, _mm256_load_si256((const __m256i *)(IC0 + jj))));
if (nd > 1) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v1A, _mm256_load_si256((const __m256i *)(IC1 + jj))));
if (nd > 2) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v2A, _mm256_load_si256((const __m256i *)(IC2 + jj))));
acc = _mm256_sub_epi32(acc, msk);
}
if (jj < nA) {
int n = nA;
__m256i msk = _mm256_cmpgt_epi32(kvA, _mm256_load_si256((const __m256i *)(IK + jj)));
if (nd > 0) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v0A, _mm256_load_si256((const __m256i *)(IC0 + jj))));
if (nd > 1) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v1A, _mm256_load_si256((const __m256i *)(IC1 + jj))));
if (nd > 2) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v2A, _mm256_load_si256((const __m256i *)(IC2 + jj))));
msk = _mm256_and_si256(msk, _mm256_loadu_si256((const __m256i *)VMASK[n - jj]));
acc = _mm256_sub_epi32(acc, msk);
}
OUT[QID[q]] += hsum256(acc);
}
{
int jj = j;
__m256i acc = accB;
for (; jj + 8 <= nB; jj += 8) {
__m256i msk = _mm256_cmpgt_epi32(kvB, _mm256_load_si256((const __m256i *)(IK + jj)));
if (nd > 0) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v0B, _mm256_load_si256((const __m256i *)(IC0 + jj))));
if (nd > 1) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v1B, _mm256_load_si256((const __m256i *)(IC1 + jj))));
if (nd > 2) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v2B, _mm256_load_si256((const __m256i *)(IC2 + jj))));
acc = _mm256_sub_epi32(acc, msk);
}
if (jj < nB) {
int n = nB;
__m256i msk = _mm256_cmpgt_epi32(kvB, _mm256_load_si256((const __m256i *)(IK + jj)));
if (nd > 0) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v0B, _mm256_load_si256((const __m256i *)(IC0 + jj))));
if (nd > 1) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v1B, _mm256_load_si256((const __m256i *)(IC1 + jj))));
if (nd > 2) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v2B, _mm256_load_si256((const __m256i *)(IC2 + jj))));
msk = _mm256_and_si256(msk, _mm256_loadu_si256((const __m256i *)VMASK[n - jj]));
acc = _mm256_sub_epi32(acc, msk);
}
OUT[QID[q + 1]] += hsum256(acc);
}
}
#endif
for (; q < nq; q++) {
if (q + OUTPF < nq) __builtin_prefetch(&OUT[QID[q + OUTPF]], 1, 0);
const int n = (int)QN[q];
__m256i kv = _mm256_set1_epi32((int)QK[q]);
__m256i v0 = _mm256_set1_epi32((int)QC0[q]);
__m256i v1 = _mm256_set1_epi32((int)QC1[q]);
__m256i v2 = _mm256_set1_epi32((int)QC2[q]);
__m256i acc = _mm256_setzero_si256();
int j = 0;
switch (nd) {
case 0:
for (; j + 8 <= n; j += 8)
acc = _mm256_sub_epi32(acc, _mm256_cmpgt_epi32(kv, _mm256_load_si256((const __m256i *)(IK + j))));
break;
case 1:
for (; j + 8 <= n; j += 8) {
__m256i msk = _mm256_cmpgt_epi32(kv, _mm256_load_si256((const __m256i *)(IK + j)));
msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v0, _mm256_load_si256((const __m256i *)(IC0 + j))));
acc = _mm256_sub_epi32(acc, msk);
}
break;
case 2:
for (; j + 8 <= n; j += 8) {
__m256i msk = _mm256_cmpgt_epi32(kv, _mm256_load_si256((const __m256i *)(IK + j)));
msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v0, _mm256_load_si256((const __m256i *)(IC0 + j))));
msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v1, _mm256_load_si256((const __m256i *)(IC1 + j))));
acc = _mm256_sub_epi32(acc, msk);
}
break;
default:
for (; j + 8 <= n; j += 8) {
__m256i msk = _mm256_cmpgt_epi32(kv, _mm256_load_si256((const __m256i *)(IK + j)));
msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v0, _mm256_load_si256((const __m256i *)(IC0 + j))));
msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v1, _mm256_load_si256((const __m256i *)(IC1 + j))));
msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v2, _mm256_load_si256((const __m256i *)(IC2 + j))));
acc = _mm256_sub_epi32(acc, msk);
}
break;
}
if (j < n) { // partial block, masked by a table lookup
__m256i msk = _mm256_cmpgt_epi32(kv, _mm256_load_si256((const __m256i *)(IK + j)));
if (nd > 0) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v0, _mm256_load_si256((const __m256i *)(IC0 + j))));
if (nd > 1) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v1, _mm256_load_si256((const __m256i *)(IC1 + j))));
if (nd > 2) msk = _mm256_and_si256(msk, _mm256_cmpgt_epi32(v2, _mm256_load_si256((const __m256i *)(IC2 + j))));
msk = _mm256_and_si256(msk, _mm256_loadu_si256((const __m256i *)VMASK[n - j]));
acc = _mm256_sub_epi32(acc, msk);
}
u32 c = hsum256(acc);
OUT[QID[q]] += c;
}
if (!needOut) return;
#if LEAFNB
{
const int NB = (LFAUTO && m <= LEAFNB) ? (m < 8 ? 8 : m) : LEAFNB;
static u32 cnt[LEAFNB + 2]; /* FIX: was a stack local relying on the previous call's bitmap clear; first-ever use of a bucket at a stack slot read uninitialised bytes (verified: segfault at n=5e5 in the rig). static = .bss = zero-initialised, so the "cnt[] is all-zero at entry" invariant is true by induction. Zero cost: the clear-via-bitmap is unchanged. */
/* ---- SPARSE REBUILD ----------------------------------------------------------
The shipped counting sort walks all NB = 8192 buckets twice (an NB-iteration
prefix and an NB-iteration bucket walk) for a leaf that holds at most
KGB_(d) <= 3840 events, so ~60% of those iterations look at EMPTY buckets.
A 1 KB occupancy bitmap, set during the count pass, lets the prefix, the walk
and the clear each visit only the buckets that actually hold something:
O(NB) -> O(#non-empty + NB/64).
EQUIVALENCE IS A PROOF: cnt[] is only ever read for a bucket we set a bit for,
which is exactly a bucket the scatter addresses; and `lo` for a non-empty
bucket b' is the end of the previous NON-EMPTY bucket, which equals the end
of b'-1 because every bucket in between is empty. Same ranges, same order,
same insertion sorts, bit-identical out[].
cnt[] is left zeroed via the same bitmap walk, so the next call starts clean
(base_case is a leaf: it never re-enters). */
static u64 lbm[(LEAFNB >> 6) + 4];
const int NW = (NB >> 6) + 2;
memset(lbm, 0, sizeof(u64) * (size_t)NW);
for (int t = 0; t < m; t++) {
const int b = (int)(((u32)(seq[t] >> 43)) >> (LFSH + LFD));
lbm[b >> 6] |= 1ull << (b & 63);
cnt[b]++;
}
u32 acc = 0;
for (int w = 0; w < NW; w++) {
u64 x = lbm[w];
while (x) {
const int b = (w << 6) + __builtin_ctzll(x);
const u32 c = cnt[b];
cnt[b] = acc; acc += c;
x &= x - 1;
}
}
for (int t = 0; t < m; t++) { u64 v = seq[t]; out[cnt[(int)(((u32)(v >> 43)) >> (LFSH + LFD))]++] = v; }
int lo = 0;
for (int w = 0; w < NW; w++) {
u64 x = lbm[w];
while (x) {
const int b = (w << 6) + __builtin_ctzll(x);
const int hi = (int)cnt[b];
if (hi - lo > 1)
for (int t = lo + 1; t < hi; t++) {
u64 v = out[t];
int q = t - 1;
while (q >= lo && v < out[q]) { out[q + 1] = out[q]; q--; }
out[q + 1] = v;
}
lo = hi;
cnt[b] = 0; /* clear via the bitmap: no 32 KB memset */
x &= x - 1;
}
}
}
#else
for (int t = 1; t < m; t++) {
u64 v = seq[t];
int q = t - 1;
while (q >= 0 && v < seq[q]) { seq[q + 1] = seq[q]; q--; }
seq[q + 1] = v;
}
memcpy(out, seq, sizeof(u64) * (size_t)m);
#endif
}
static void base_case(u64 *seq, int m, int d, u64 *out, int needOut) {
switch (K - (d + 2)) {
case 0: base_case_t<0>(seq, m, d, out, needOut); break;
case 1: base_case_t<1>(seq, m, d, out, needOut); break;
case 2: base_case_t<2>(seq, m, d, out, needOut); break;
default: base_case_t<3>(seq, m, d, out, needOut); break;
}
}
/* [XE] cross construction: the rekey key comes from the event's OWN embedded field
(bits 21..41 == XD[d+2][id]) instead of gathering XN[id], and the cross EMBEDS
XE[id] = XD[d+3][id] into bits 21..41 for the next level. XE == 0 at the levels that
do not rekey, so the null test folds away (XEMB is loop-invariant). */
#define MKE_CROSS(V) (((RK) ? (((((V) >> 21) & 0x1FFFFFull) << 43) | ((V) & MASK)) \
: ((V) & MASK)) \
| (XEMB ? ((u64)XEMB[(u32)(V) & 0x1FFFFFu] << 21) : 0ull))
template<int NO, int RK, int DD>
static inline void merge_loops(const u64 *retA, int h, const u64 *retB, int r,
u64 *out, u64 *cross, const u32 *XN, const u32 *XE, int &cx) {
/* Tight pointer-based merge. The stock body spends 34 uops per merged event:
`t` and the `out`/`cross` base pointers get SPILLED every iteration
(`movl %r9d,36(%rsp)` / `movq 8(%rsp),%rcx`), the 43-bit mask constant is
re-materialised with `movabsq` inside the loop, and the same `cmpq` is executed
twice (once for setb, once for cmovbe). Indices are also sign-extended twice
per event. Measured: the merge is 21.3 % of the row (idempotent duplication
probe, judge-side). Here: pointers only, no `t` (t == ia+ib), the mask lives
in a register, and `keep` is one xor. */
const u64 MASK = 4398048608255ull;
const u32 *const XEMB = XE; /* [XE] loop-invariant (may be null) */
const u64 *pa = retA, * const pa_end = retA + h;
const u64 *pb = retB, * const pb_end = retB + r;
u64 *po = out;
u64 *pc = cross;
while (pa < pa_end && pb < pb_end) {
const u64 ea = *pa, eb = *pb;
const u32 takeA = (u32)(ea < eb);
pa += takeA; pb += 1 - takeA;
const u64 v = takeA ? ea : eb;
if (NO) *po = v;
++po;
u64 c = MKE_CROSS(v);
*pc = c;
/* keep = 1 - (takeA ^ role) = 1 - (takeA ^ ((v>>42)&1)) -- as one xor pair */
pc += (size_t)(1 - ((takeA ^ (u32)(v >> 42)) & 1u));
}
while (pa < pa_end) {
const u64 v = *pa++;
if (NO) *po = v;
++po;
if ((v >> 42) & 1) {
*pc++ = MKE_CROSS(v);
}
}
while (pb < pb_end) {
const u64 v = *pb++;
if (NO) *po = v;
++po;
if (!((v >> 42) & 1)) {
*pc++ = MKE_CROSS(v);
}
}
cx = (int)(pc - cross);
}
// ---- ND specialisation dispatcher (the fleet's `DN` recipe) ----
static void base_case(u64 *seq, int m, int d, u64 *out, int needOut);
static void solve(u64 *seq, int m, int d, u64 *out, int needOut) {
if (m <= 1) return;
if (d >= K - 1) {
int cnt = 0;
for (int t = 0; t < m; t++) {
u64 v = seq[t];
if ((v >> 42) & 1) cnt++; // insert (bit31 = 1)
else OUT[eid(v)] += (u32)cnt; // query
}
return;
}
if (m <= KGB_(d)) { base_case(seq, m, d, out, needOut); return; }
(void)0;
int h = m >> 1, r = m - h;
u64 *save = ABtop;
u64 *retA = ABtop; ABtop += h;
u64 *retB = ABtop; ABtop += r;
solve(seq, h, d, retA, 1);
solve(seq + h, r, d, retB, 1);
u64 *cross = ABtop; ABtop += m + 1;
const int rekey = (d + 2 <= K - 1);
const u32 *XN = XD[rekey ? d + 2 : 0];
/* [XE] the level-(d+1) rekey/c0 needs XD[d+3]; embed it in this level's cross entries */
const u32 *XE = (d + 3 <= K - 1) ? XD[d + 3] : 0;
int cx = 0; /* ia/ib/t live inside merge_loops */
if (needOut) {
switch (d) {
case 0: merge_loops<1, 1, 0>(retA, h, retB, r, out, cross, XN, XE, cx); break;
case 1: merge_loops<1, 1, 1>(retA, h, retB, r, out, cross, XN, XE, cx); break;
case 2: merge_loops<1, 1, 2>(retA, h, retB, r, out, cross, XN, XE, cx); break;
case 3: merge_loops<1, 0, 3>(retA, h, retB, r, out, cross, XN, XE, cx); break;
default: merge_loops<1, 0, 4>(retA, h, retB, r, out, cross, XN, XE, cx); break;
}
} else {
switch (d) {
case 0: merge_loops<0, 1, 0>(retA, h, retB, r, out, cross, XN, XE, cx); break;
case 1: merge_loops<0, 1, 1>(retA, h, retB, r, out, cross, XN, XE, cx); break;
case 2: merge_loops<0, 1, 2>(retA, h, retB, r, out, cross, XN, XE, cx); break;
case 3: merge_loops<0, 0, 3>(retA, h, retB, r, out, cross, XN, XE, cx); break;
default: merge_loops<0, 0, 4>(retA, h, retB, r, out, cross, XN, XE, cx); break;
}
}
if (FUSE_OK && d == K - 3) {
u32 *dirt = (u32 *)ABtop; ABtop += (cx + 3) / 2 + 1;
ctr_solve(cross, cx, dirt);
} else {
u64 *cres = ABtop; ABtop += cx;
solve(cross, cx, d + 1, cres, 0);
}
ABtop = save;
}
static u32 *g_ord, *g_x;
static void sort0(int n) {
const u32 *X0 = XD[0];
u32 *cnt = g_x;
memset(cnt, 0, sizeof(u32) * (size_t)(n + 1));
for (int i = 0; i < n; i++) cnt[X0[i] + 1]++;
for (int i = 0; i < n; i++) cnt[i + 1] += cnt[i];
for (int i = 0; i < n; i++) g_ord[cnt[X0[i]]++] = (u32)i;
}
void run(int n, const u32 **x, int k, u32 *out) {
LFSH = 0; LFD = 0;
K = k;
for (int d = 0; d < k; d++) XD[d] = x[d];
OUT = out;
ABtop = AB;
if (n <= 1) return;
FUSE_OK = 0;
if (K >= 4 && n > 2 && (u32)n < CT_MAXV) {
const u32 *XL = XD[K - 1];
memset(CT0, 0, (size_t)n);
int ok = 1;
for (int i = 0; i < n; i++) if (++CT0[XL[i]] > 127) { ok = 0; break; }
memset(CT0, 0, (size_t)n);
FUSE_OK = ok;
}
{
int bit = 0;
while ((n >> bit) >= LEAFNB && bit < 30) bit++;
LFSH = bit; LFD = 0;
}
sort0(n);
const u32 *X0 = XD[0], *X1 = XD[1];
const u32 *nk2 = (K >= 3) ? XD[2] : X0; /* dim 2's coordinate, zero-only when K < 3 */
u64 *ev = AB; // top-level event array: 2n entries
int m = 0, g = 0;
for (int v = 0; v < n; ++v) {
int g1 = (int)g_x[v];
if (g1 == g) continue;
int cnt = g1 - g;
for (int t = g; t < g1; ++t) {
int id = (int)g_ord[t];
u32 key = X1[id], nk = nk2[id];
ev[m + t - g] = mk(key, 1, id, nk);
ev[m + cnt + t - g] = mk(key, 0, id, nk);
}
m += cnt * 2;
g = g1;
}
ABtop = AB + m;
u64 *res = ABtop; ABtop += m;
solve(ev, m, 0, res, 0);
}
} // namespace kg
static u64 g_arena[KG_ARENA];
static u32 g_ord_st[1048576], g_x_st[1048576];
static int g_ready = 0;
static void setup() {
if (g_ready) return;
g_ready = 1;
kg::AB = g_arena;
kg::g_ord = g_ord_st;
kg::g_x = g_x_st;
}
void count_4d(int n, const unsigned *x[4], unsigned *out) {
setup();
memset(out, 0, sizeof(unsigned) * (size_t)n);
kg::run(n, (const unsigned **)x, 4, out);
}
void count_5d(int n, const unsigned *x[5], unsigned *out) {
setup();
memset(out, 0, sizeof(unsigned) * (size_t)n);
kg::run(n, (const unsigned **)x, 5, out);
}
#pragma GCC pop_options
| Compilation | N/A | N/A | Compile OK | Score: N/A | 显示更多 |
| Testcase #1 | 1.984 s | 93 MB + 220 KB | Accepted | Score: 100 | 显示更多 |