代码

#include <string.h>
typedef unsigned int u32;
typedef unsigned char u8;

#define NB 65536
#define LINE 32

static u32 tmp[134217728 + NB * LINE] __attribute__((aligned(64)));
static u32 staging[NB * LINE] __attribute__((aligned(64)));
static u32 cnt[NB];
static u32 rbase[NB];
static u32 pbase[NB];
static u32 off[NB];
static u8  pos[NB];
static u32 scratch2[65536] __attribute__((aligned(64)));
static u32 bc[256];
static u32 bc2[256];

static void scatter_top16(u32 *src, u32 *dst, int n) {
    int i, j;
    for (j = 0; j < NB; j++) cnt[j] = 0;
    int n8 = n & ~7;
    for (i = 0; i < n8; i += 8) {
        cnt[src[i]>>16]++; cnt[src[i+1]>>16]++; cnt[src[i+2]>>16]++; cnt[src[i+3]>>16]++;
        cnt[src[i+4]>>16]++; cnt[src[i+5]>>16]++; cnt[src[i+6]>>16]++; cnt[src[i+7]>>16]++;
    }
    for (; i < n; i++) cnt[src[i] >> 16]++;
    { u32 t = 0, p = 0; for (j = 0; j < NB; j++) { u32 c = cnt[j]; rbase[j] = t; pbase[j] = p; t += c; p = (p + c + 31) & ~31u; } }
    for (j = 0; j < NB; j++) { off[j] = pbase[j]; pos[j] = 0; }
#define S1(X) do { u32 x_ = (X); u32 b_ = x_ >> 16; u32 p_ = pos[b_]; \
    staging[b_ * LINE + p_] = x_; p_++; \
    if (p_ == LINE) { u32 go_ = off[b_]; memcpy(dst + go_, staging + b_ * LINE, 128); off[b_] = go_ + LINE; p_ = 0; } \
    pos[b_] = (u8)p_; } while (0)
    int n8s = n & ~7;
    for (i = 0; i < n8s; i += 8) {
        S1(src[i]); S1(src[i+1]); S1(src[i+2]); S1(src[i+3]);
        S1(src[i+4]); S1(src[i+5]); S1(src[i+6]); S1(src[i+7]);
    }
    for (; i < n; i++) S1(src[i]);
#undef S1
    for (j = 0; j < NB; j++) { u32 p = pos[j]; if (p > 0) memcpy(dst + off[j], staging + j * LINE, p * 4); }
}

static void sort16(u32 *src, u32 *dst, int m) {
    int i, j;
    for (j = 0; j < 256; j++) bc[j] = 0;
    int m8 = m & ~7;
    for (i = 0; i < m8; i += 8) {
        bc[src[i]&255]++; bc[src[i+1]&255]++; bc[src[i+2]&255]++; bc[src[i+3]&255]++;
        bc[src[i+4]&255]++; bc[src[i+5]&255]++; bc[src[i+6]&255]++; bc[src[i+7]&255]++;
    }
    for (; i < m; i++) bc[src[i] & 255]++;
    for (j = 1; j < 256; j++) bc[j] += bc[j-1];
    for (j = 0; j < 256; j++) bc2[j] = 0;
    int rm = m & ~7;
    for (i = m - 1; i >= rm; i--) { u32 x = src[i]; scratch2[--bc[x & 255]] = x; bc2[(x >> 8) & 255]++; }
    for (i = rm - 1; i >= 0; i -= 8) {
        u32 x0=src[i],x1=src[i-1],x2=src[i-2],x3=src[i-3],x4=src[i-4],x5=src[i-5],x6=src[i-6],x7=src[i-7];
        scratch2[--bc[x0&255]]=x0; bc2[(x0>>8)&255]++;
        scratch2[--bc[x1&255]]=x1; bc2[(x1>>8)&255]++;
        scratch2[--bc[x2&255]]=x2; bc2[(x2>>8)&255]++;
        scratch2[--bc[x3&255]]=x3; bc2[(x3>>8)&255]++;
        scratch2[--bc[x4&255]]=x4; bc2[(x4>>8)&255]++;
        scratch2[--bc[x5&255]]=x5; bc2[(x5>>8)&255]++;
        scratch2[--bc[x6&255]]=x6; bc2[(x6>>8)&255]++;
        scratch2[--bc[x7&255]]=x7; bc2[(x7>>8)&255]++;
    }
    for (j = 1; j < 256; j++) bc2[j] += bc2[j-1];
    for (i = m - 1; i >= rm; i--) { u32 x = scratch2[i]; dst[--bc2[(x >> 8) & 255]] = x; }
    for (i = rm - 1; i >= 0; i -= 8) {
        u32 x0=scratch2[i],x1=scratch2[i-1],x2=scratch2[i-2],x3=scratch2[i-3],x4=scratch2[i-4],x5=scratch2[i-5],x6=scratch2[i-6],x7=scratch2[i-7];
        dst[--bc2[(x0>>8)&255]]=x0;
        dst[--bc2[(x1>>8)&255]]=x1;
        dst[--bc2[(x2>>8)&255]]=x2;
        dst[--bc2[(x3>>8)&255]]=x3;
        dst[--bc2[(x4>>8)&255]]=x4;
        dst[--bc2[(x5>>8)&255]]=x5;
        dst[--bc2[(x6>>8)&255]]=x6;
        dst[--bc2[(x7>>8)&255]]=x7;
    }
}

void sort(unsigned *aa, int n) {
    u32 *a = (u32*)aa;
    int j;
    scatter_top16(a, tmp, n);
    for (j = 0; j < NB; j++) {
        int m = (int)cnt[j];
        if (m > 0) sort16(tmp + pbase[j], a + rbase[j], m);
    }
}

评测结果