// count_3d with fast I/O via DuckInfo auxv (getauxval).
#include <cstddef>
#include <cstring>
#include <algorithm>
#include <cstdarg>
#include <cstdio>
#include <sys/auxv.h>
#include <stdint.h>
struct DuckInfo {
uint64_t abi_version;
const char *stdin_ptr; uint64_t stdin_size;
char *stdout_ptr; uint64_t stdout_limit; uint64_t stdout_size;
char *stderr_ptr; uint64_t stderr_limit; uint64_t stderr_size;
const char *IB_ptr; uint64_t IB_limit;
char *OB_ptr; uint64_t OB_limit;
uint64_t tsc_frequency;
} __attribute__((packed));
static const char* IN_P;
static const char* IN_E;
static char* OUT_P;
static char* OUT_START;
static DuckInfo* DI;
static inline void io_init() {
DI = (DuckInfo*)getauxval(0x6b637564);
IN_P = DI->stdin_ptr;
IN_E = DI->stdin_ptr + DI->stdin_size;
OUT_P = DI->stdout_ptr;
OUT_START = DI->stdout_ptr;
}
static inline unsigned read_uint() {
const char* p = IN_P;
while (p < IN_E && (*p == ' ' || *p == '\n' || *p == '\t' || *p == '\r')) p++;
unsigned v = 0;
while (p < IN_E && *p >= '0' && *p <= '9') { v = v * 10u + (unsigned)(*p - '0'); p++; }
IN_P = p;
return v;
}
static int vscan_core(const char* fmt, va_list ap) {
int ret = 0;
for (const char* f = fmt; *f; f++) {
if (*f == ' ' || *f == '\n' || *f == '\t' || *f == '\r') continue;
if (*f != '%') {
while (IN_P < IN_E && (*IN_P == ' ' || *IN_P == '\n' || *IN_P == '\t' || *IN_P == '\r')) IN_P++;
if (IN_P < IN_E && *IN_P == *f) IN_P++;
else break;
continue;
}
f++;
int isl = 0, isll = 0;
while (*f == 'l') { isll = isl; isl = 1; f++; }
unsigned v = read_uint();
if (*f == 'd' || *f == 'i') {
if (isll) *va_arg(ap, long long*) = (long long)v;
else if (isl) *va_arg(ap, long*) = (long)v;
else *va_arg(ap, int*) = (int)v;
ret++;
} else if (*f == 'u') {
if (isll) *va_arg(ap, unsigned long long*) = (unsigned long long)v;
else if (isl) *va_arg(ap, unsigned long*) = (unsigned long)v;
else *va_arg(ap, unsigned*) = v;
ret++;
}
}
return ret;
}
extern "C" int scanf(const char* fmt, ...) {
if (!DI) io_init();
va_list ap; va_start(ap, fmt);
int r = vscan_core(fmt, ap);
va_end(ap);
return r;
}
extern "C" int fscanf(FILE* f, const char* fmt, ...) {
if (!DI) io_init();
va_list ap; va_start(ap, fmt);
int r = vscan_core(fmt, ap);
va_end(ap);
return r;
}
static inline void put_u(unsigned v) {
char buf[12];
char* p = buf + 12;
do { *--p = (char)('0' + (v % 10)); v /= 10; } while (v);
while (p < buf + 12) *OUT_P++ = *p++;
}
static inline void put_ull(unsigned long long v) {
char buf[24];
char* p = buf + 24;
do { *--p = (char)('0' + (v % 10)); v /= 10; } while (v);
while (p < buf + 24) *OUT_P++ = *p++;
}
static int vprint_core(const char* fmt, va_list ap) {
for (const char* f = fmt; *f; f++) {
if (*f != '%') { *OUT_P++ = *f; continue; }
f++;
int isl = 0, isll = 0;
while (*f == 'l') { isll = isl; isl = 1; f++; }
if (*f == 'd' || *f == 'i') {
if (isll) { long long v = va_arg(ap, long long); if (v < 0) { *OUT_P++ = '-'; v = -v; } put_ull((unsigned long long)v); }
else if (isl) { long v = va_arg(ap, long); if (v < 0) { *OUT_P++ = '-'; v = -v; } put_u((unsigned)v); }
else { int v = va_arg(ap, int); if (v < 0) { *OUT_P++ = '-'; v = -v; } put_u((unsigned)v); }
} else if (*f == 'u') {
if (isll) put_ull(va_arg(ap, unsigned long long));
else if (isl) put_u((unsigned)va_arg(ap, unsigned long));
else put_u(va_arg(ap, unsigned int));
} else if (*f == 'c') {
*OUT_P++ = (char)va_arg(ap, int);
} else if (*f == 's') {
const char* s = va_arg(ap, const char*);
while (*s) *OUT_P++ = *s++;
} else if (*f == '%') {
*OUT_P++ = '%';
}
}
return 0;
}
extern "C" int printf(const char* fmt, ...) {
if (!DI) io_init();
va_list ap; va_start(ap, fmt);
int r = vprint_core(fmt, ap);
va_end(ap);
DI->stdout_size = (uint64_t)(OUT_P - OUT_START);
return r;
}
extern "C" int fprintf(FILE* f, const char* fmt, ...) {
if (!DI) io_init();
va_list ap; va_start(ap, fmt);
int r = vprint_core(fmt, ap);
va_end(ap);
DI->stdout_size = (uint64_t)(OUT_P - OUT_START);
return r;
}
// ================= count_3d core =================
namespace {
int N;
const unsigned *X, *Y, *Z;
unsigned *OUT;
int *ord, *ordz, *tmp, *bit, *cnt, *zrank;
unsigned short *mark;
unsigned short cur;
const int BASE = 32;
static inline void bit_add(int i, int v) {
for (; i <= N; i += i & -i) {
if (mark[i] != cur) { mark[i] = cur; bit[i] = 0; }
bit[i] += v;
}
}
static inline int bit_sum(int i) {
int s = 0;
for (; i > 0; i -= i & -i) {
if (mark[i] == cur) s += bit[i];
}
return s;
}
static void sort_by_x(int n) {
for (int i = 0; i < n; i++) cnt[i] = 0;
for (int i = 0; i < n; i++) cnt[X[i]]++;
int acc = 0;
for (int v = 0; v < n; v++) { int c = cnt[v]; cnt[v] = acc; acc += c; }
for (int i = 0; i < n; i++) ord[cnt[X[i]]++] = i;
}
static void sort_y_range(int l, int r) {
std::sort(ord + l, ord + r, [](int a, int b){ return Y[a] < Y[b]; });
}
static void sort_z_range(int l, int r) {
std::sort(ordz + l, ordz + r, [](int a, int b){ return Z[a] < Z[b]; });
}
void cdq(int l, int r) {
if (r - l <= 1) return;
if (r - l <= BASE) {
for (int i = l; i < r; i++) {
unsigned xi = X[ord[i]], yi = Y[ord[i]], zi = Z[ord[i]];
for (int j = i + 1; j < r; j++) {
if (xi < X[ord[j]] && yi < Y[ord[j]] && zi < Z[ord[j]]) OUT[ord[j]]++;
}
}
sort_y_range(l, r);
sort_z_range(l, r);
return;
}
int m = (l + r) >> 1;
int gs = m;
while (gs > l && X[ord[gs - 1]] == X[ord[gs]]) gs--;
int ge = m;
while (ge + 1 < r && X[ord[ge + 1]] == X[ord[ge]]) ge++;
if (gs > l) m = gs;
else if (ge + 1 < r) m = ge + 1;
else { sort_y_range(l, r); sort_z_range(l, r); return; }
cdq(l, m);
cdq(m, r);
{
int i = l, j = m, k = l;
while (i < m && j < r) {
if (Z[ordz[i]] <= Z[ordz[j]]) tmp[k++] = ordz[i++];
else tmp[k++] = ordz[j++];
}
while (i < m) tmp[k++] = ordz[i++];
while (j < r) tmp[k++] = ordz[j++];
unsigned prev = Z[tmp[l]];
int rk = 0;
zrank[tmp[l]] = 0;
for (int t = l + 1; t < r; t++) {
if (Z[tmp[t]] != prev) { rk++; prev = Z[tmp[t]]; }
zrank[tmp[t]] = rk;
}
memcpy(ordz + l, tmp + l, (size_t)(r - l) * sizeof(int));
}
if (++cur == 0) { memset(mark, 0, (size_t)(N + 1) * sizeof(unsigned short)); cur = 1; }
{
int i = l, j = m, k = l;
while (i < m && j < r) {
if (Y[ord[i]] < Y[ord[j]]) {
bit_add(zrank[ord[i]] + 1, 1);
tmp[k++] = ord[i++];
} else {
OUT[ord[j]] += (unsigned)bit_sum(zrank[ord[j]]);
tmp[k++] = ord[j++];
}
}
while (i < m) { bit_add(zrank[ord[i]] + 1, 1); tmp[k++] = ord[i++]; }
while (j < r) { OUT[ord[j]] += (unsigned)bit_sum(zrank[ord[j]]); tmp[k++] = ord[j++]; }
memcpy(ord + l, tmp + l, (size_t)(r - l) * sizeof(int));
}
}
}
void count_3d(int n, const unsigned *x, const unsigned *y, const unsigned *z, unsigned *out) {
N = n; X = x; Y = y; Z = z; OUT = out;
ord = new int[n];
ordz = new int[n];
tmp = new int[n];
bit = new int[n + 1];
cnt = new int[n + 1];
zrank = new int[n];
mark = new unsigned short[n + 1];
memset(mark, 0, (size_t)(n + 1) * sizeof(unsigned short));
cur = 0;
for (int i = 0; i < n; i++) ord[i] = i;
sort_by_x(n);
for (int i = 0; i < n; i++) ordz[i] = ord[i];
cdq(0, n);
delete[] ord;
delete[] ordz;
delete[] tmp;
delete[] bit;
delete[] cnt;
delete[] zrank;
delete[] mark;
}
| Compilation | N/A | N/A | Compile OK | Score: N/A | 显示更多 |
| Testcase #1 | 4.111 ms | 312 KB | Accepted | Score: 100 | 显示更多 |