代码

#pragma GCC optimize("Ofast,unroll-loops")
#pragma GCC target("avx2,tune=native")
typedef unsigned u32;
u32 cnt1[1 << 8], cnt2[1 << 8], cnt3[1 << 8], cnt4[1 << 8], ptr1[1<<8], ptr2[1<<8], ptr3[1<<8], ptr4[1<<8];
void sort(unsigned *a, int n)
{
	u32 *tmp = new u32[n];
	u32 *end= a + n;

	//#pragma unroll 
    for (register u32 *i = a; i != end; i++)
		++cnt1[*i & 0xff],
		++cnt2[(*i >> 8) & 0xff],
		++cnt3[(*i >> 16) & 0xff],
		++cnt4[*i >> 24];
		
	*ptr1 = *cnt1;
	*ptr2 = *cnt2;
	*ptr3 = *cnt3;
	*ptr4 = *cnt4;
	for(
		register u32
			*i = cnt1 + 1, *j = cnt2 + 1, *k = cnt3 + 1, *l = cnt4 + 1,
			*I = ptr1 + 1, *J = ptr2 + 1, *K = ptr3 + 1, *L = ptr4 + 1;
		i <= cnt1 + 0xff;
		++i, ++j, ++k, ++l,
		++I, ++J, ++K, ++L)
		*I = *(I - 1) + *i,
		*J = *(J - 1) + *j,
		*K = *(K - 1) + *k,
		*L = *(L - 1) + *l;

	for(register u32 *i= end - 1; i != a - 1; i-=8)
		tmp[--ptr1[*i & 0xff]] = *i,
		tmp[--ptr1[*(i - 1) & 0xff]] = *(i - 1),
		tmp[--ptr1[*(i - 2) & 0xff]] = *(i - 2),
		tmp[--ptr1[*(i - 3) & 0xff]] = *(i - 3),
		tmp[--ptr1[*(i - 4) & 0xff]] = *(i - 4),
		tmp[--ptr1[*(i - 5) & 0xff]] = *(i - 5),
		tmp[--ptr1[*(i - 6) & 0xff]] = *(i - 6),
		tmp[--ptr1[*(i - 7) & 0xff]] = *(i - 7);
	for(register u32 *i= tmp + n - 1; i != tmp - 1; i-=8)
		a[--ptr2[(*i >> 8) & 0xff]] = *i,
		a[--ptr2[(*(i - 1) >> 8) & 0xff]] = *(i - 1),
		a[--ptr2[(*(i - 2) >> 8) & 0xff]] = *(i - 2),
		a[--ptr2[(*(i - 3) >> 8) & 0xff]] = *(i - 3),
		a[--ptr2[(*(i - 4) >> 8) & 0xff]] = *(i - 4),
		a[--ptr2[(*(i - 5) >> 8) & 0xff]] = *(i - 5),
		a[--ptr2[(*(i - 6) >> 8) & 0xff]] = *(i - 6),
		a[--ptr2[(*(i - 7) >> 8) & 0xff]] = *(i - 7);
	for(register u32 *i= end - 1; i != a - 1; i-=8)
		tmp[--ptr3[(*i >> 16) & 0xff]] = *i,
		tmp[--ptr3[(*(i - 1) >> 16) & 0xff]] = *(i - 1),
		tmp[--ptr3[(*(i - 2) >> 16) & 0xff]] = *(i - 2),
		tmp[--ptr3[(*(i - 3) >> 16) & 0xff]] = *(i - 3),
		tmp[--ptr3[(*(i - 4) >> 16) & 0xff]] = *(i - 4),
		tmp[--ptr3[(*(i - 5) >> 16) & 0xff]] = *(i - 5),
		tmp[--ptr3[(*(i - 6) >> 16) & 0xff]] = *(i - 6),
		tmp[--ptr3[(*(i - 7) >> 16) & 0xff]] = *(i - 7);
	for(register u32 *i= tmp + n - 1; i != tmp - 1; i-=8)
		a[--ptr4[*i >> 24]] = *i,
		a[--ptr4[*(i - 1) >> 24]] = *(i - 1),
		a[--ptr4[*(i - 2) >> 24]] = *(i - 2),
		a[--ptr4[*(i - 3) >> 24]] = *(i - 3),
		a[--ptr4[*(i - 4) >> 24]] = *(i - 4),
		a[--ptr4[*(i - 5) >> 24]] = *(i - 5),
		a[--ptr4[*(i - 6) >> 24]] = *(i - 6),
		a[--ptr4[*(i - 7) >> 24]] = *(i - 7);
	delete[] tmp;
}

评测结果