代码

void sort(unsigned *a, int n)
{
    unsigned *b=new unsigned[n];
	unsigned *ptr0[256],*ptr8[256],*ptr16[256],*ptr24[256];
	unsigned cnt0[256], cnt8[256], cnt16[256], cnt24[256];
    unsigned *p;
    for (int i = 0; i < 256; ++i){
        cnt0[i] = 0;
        cnt8[i] = 0;
        cnt16[i] = 0;
        cnt24[i] = 0;
    }
    for (int i = 0; i < n;i+=4){
        ++cnt0[a[i]&255], ++cnt8[a[i]>>8&255], ++cnt16[a[i]>>16&255], ++cnt24[a[i]>>24];
        ++cnt0[a[i+1]&255], ++cnt8[a[i+1]>>8&255], ++cnt16[a[i+1]>>16&255], ++cnt24[a[i+1]>>24];
        ++cnt0[a[i+2]&255], ++cnt8[a[i+2]>>8&255], ++cnt16[a[i+2]>>16&255], ++cnt24[a[i+2]>>24];
        ++cnt0[a[i+3]&255], ++cnt8[a[i+3]>>8&255], ++cnt16[a[i+3]>>16&255], ++cnt24[a[i+3]>>24];
    }
    ptr0[0]=b-1,ptr8[0]=a-1,ptr16[0]=b-1,ptr24[0]=a-1;
    for (int i = 1; i < 256; ++i){
        ptr0[i] = ptr0[i-1]+cnt0[i-1];
		ptr8[i] = ptr8[i-1]+cnt8[i-1];
		ptr16[i] = ptr16[i-1]+cnt16[i-1];
		ptr24[i] = ptr24[i-1]+cnt24[i-1];
	}
#define countingSort(a,k) \
for (unsigned *p = a+n; p>=a+4; p -= 4){\
    *++ptr##k[p[-1]>>k&255]=p[-1];\
    *++ptr##k[p[-2]>>k&255]=p[-2];\
    *++ptr##k[p[-3]>>k&255]=p[-3];\
    *++ptr##k[p[-4]>>k&255]=p[-4];\
}
	countingSort(a,0);
	countingSort(b,8);
	countingSort(a,16);
	countingSort(b,24);
	delete[] b;
}

评测结果