代码

typedef unsigned int u32;
typedef unsigned short u16;
#define N 10000
void sort(u32 *__restrict__ a, int n) {
    (void)n;
    static u32 tmp[N];
    u16 c1[256], c2[256], c3[256], c4[256];
    for (int i=0;i<256;i++){c1[i]=c2[i]=c3[i]=c4[i]=0;}
    for (int i=0; i<N; i+=4) {
        u32 x0=a[i], x1=a[i+1], x2=a[i+2], x3=a[i+3];
        c1[x0&255]++; c2[(x0>>8)&255]++; c3[(x0>>16)&255]++; c4[x0>>24]++;
        c1[x1&255]++; c2[(x1>>8)&255]++; c3[(x1>>16)&255]++; c4[x1>>24]++;
        c1[x2&255]++; c2[(x2>>8)&255]++; c3[(x2>>16)&255]++; c4[x2>>24]++;
        c1[x3&255]++; c2[(x3>>8)&255]++; c3[(x3>>16)&255]++; c4[x3>>24]++;
    }
    for (int i=1;i<256;i++){c1[i]+=c1[i-1];c2[i]+=c2[i-1];c3[i]+=c3[i-1];c4[i]+=c4[i-1];}
    for (int i=N-1; i>=0; i-=4) {
        u32 x0=a[i],x1=a[i-1],x2=a[i-2],x3=a[i-3];
        u32 k0=x0&255,k1=x1&255,k2=x2&255,k3=x3&255;
        u32 p0=c1[k0]; c1[k0]=(u16)(p0-1); tmp[p0-1]=x0;
        u32 p1=c1[k1]; c1[k1]=(u16)(p1-1); tmp[p1-1]=x1;
        u32 p2=c1[k2]; c1[k2]=(u16)(p2-1); tmp[p2-1]=x2;
        u32 p3=c1[k3]; c1[k3]=(u16)(p3-1); tmp[p3-1]=x3;
    }
    for (int i=N-1; i>=0; i-=4) {
        u32 x0=tmp[i],x1=tmp[i-1],x2=tmp[i-2],x3=tmp[i-3];
        u32 k0=(x0>>8)&255,k1=(x1>>8)&255,k2=(x2>>8)&255,k3=(x3>>8)&255;
        u32 p0=c2[k0]; c2[k0]=(u16)(p0-1); a[p0-1]=x0;
        u32 p1=c2[k1]; c2[k1]=(u16)(p1-1); a[p1-1]=x1;
        u32 p2=c2[k2]; c2[k2]=(u16)(p2-1); a[p2-1]=x2;
        u32 p3=c2[k3]; c2[k3]=(u16)(p3-1); a[p3-1]=x3;
    }
    for (int i=N-1; i>=0; i-=4) {
        u32 x0=a[i],x1=a[i-1],x2=a[i-2],x3=a[i-3];
        u32 k0=(x0>>16)&255,k1=(x1>>16)&255,k2=(x2>>16)&255,k3=(x3>>16)&255;
        u32 p0=c3[k0]; c3[k0]=(u16)(p0-1); tmp[p0-1]=x0;
        u32 p1=c3[k1]; c3[k1]=(u16)(p1-1); tmp[p1-1]=x1;
        u32 p2=c3[k2]; c3[k2]=(u16)(p2-1); tmp[p2-1]=x2;
        u32 p3=c3[k3]; c3[k3]=(u16)(p3-1); tmp[p3-1]=x3;
    }
    for (int i=N-1; i>=0; i-=4) {
        u32 x0=tmp[i],x1=tmp[i-1],x2=tmp[i-2],x3=tmp[i-3];
        u32 k0=x0>>24,k1=x1>>24,k2=x2>>24,k3=x3>>24;
        u32 p0=c4[k0]; c4[k0]=(u16)(p0-1); a[p0-1]=x0;
        u32 p1=c4[k1]; c4[k1]=(u16)(p1-1); a[p1-1]=x1;
        u32 p2=c4[k2]; c4[k2]=(u16)(p2-1); a[p2-1]=x2;
        u32 p3=c4[k3]; c4[k3]=(u16)(p3-1); a[p3-1]=x3;
    }
}

评测结果