在Java 8或9中只调用Integer。bitCount。

这是一个有助于了解您的微架构的问题。我只是在gcc 4.3.3下用-O3编译的两个变量使用c++内联来计时，以消除函数调用开销，十亿次迭代，保持所有计数的运行总和，以确保编译器不删除任何重要的东西，使用rdtsc计时(精确的时钟周期)。

inline int pop2(unsigned x, unsigned y)
{
    x = x - ((x >> 1) & 0x55555555);
    y = y - ((y >> 1) & 0x55555555);
    x = (x & 0x33333333) + ((x >> 2) & 0x33333333);
    y = (y & 0x33333333) + ((y >> 2) & 0x33333333);
    x = (x + (x >> 4)) & 0x0F0F0F0F;
    y = (y + (y >> 4)) & 0x0F0F0F0F;
    x = x + (x >> 8);
    y = y + (y >> 8);
    x = x + (x >> 16);
    y = y + (y >> 16);
    return (x+y) & 0x000000FF;
}

未经修改的黑客喜悦需要122亿周期。我的并行版本(计算的比特数是它的两倍)的运行周期为13.0千兆周期。在2.4GHz的酷睿双核上，两者总共消耗了10.5秒。在这个时钟频率下，25千兆周期= 10秒多一点，所以我相信我的计时是正确的。

这与指令依赖链有关，这对算法非常不利。通过使用一对64位寄存器，我几乎可以再次将速度提高一倍。事实上，如果我聪明一点，早点加上x+y，我就可以减少一些移位。64位版本做了一些小的调整，结果是相同的，但又增加了一倍的比特数。

对于128位SIMD寄存器，这是另一个因素，SSE指令集通常也有聪明的快捷方式。

没有理由让代码特别透明。该算法界面简单，可在多处在线引用，并能通过全面的单元测试。偶然发现它的程序员甚至可能学到一些东西。这些位操作在机器级别上是非常自然的。

好吧，我决定搁置调整后的64位版本。对于这个sizeof(unsigned long) == 8

inline int pop2(unsigned long x, unsigned long y)
{
    x = x - ((x >> 1) & 0x5555555555555555);
    y = y - ((y >> 1) & 0x5555555555555555);
    x = (x & 0x3333333333333333) + ((x >> 2) & 0x3333333333333333);
    y = (y & 0x3333333333333333) + ((y >> 2) & 0x3333333333333333);
    x = (x + (x >> 4)) & 0x0F0F0F0F0F0F0F0F;
    y = (y + (y >> 4)) & 0x0F0F0F0F0F0F0F0F;
    x = x + y; 
    x = x + (x >> 8);
    x = x + (x >> 16);
    x = x + (x >> 32); 
    return x & 0xFF;
}

这看起来是对的(不过我没有仔细测试)。现在计时结果是10.70亿周期/ 14.1亿周期。后面的数字加起来是1280亿比特，相当于这台机器运行了5.9秒。非并行版本稍微加快了一点，因为我在64位模式下运行，它更喜欢64位寄存器，而不是32位寄存器。

让我们看看这里是否有更多的OOO管道。这有点复杂，所以我实际上测试了一些。每一项单独加起来是64，所有项加起来是256。

inline int pop4(unsigned long x, unsigned long y, 
                unsigned long u, unsigned long v)
{
  enum { m1 = 0x5555555555555555, 
         m2 = 0x3333333333333333, 
         m3 = 0x0F0F0F0F0F0F0F0F, 
         m4 = 0x000000FF000000FF };

    x = x - ((x >> 1) & m1);
    y = y - ((y >> 1) & m1);
    u = u - ((u >> 1) & m1);
    v = v - ((v >> 1) & m1);
    x = (x & m2) + ((x >> 2) & m2);
    y = (y & m2) + ((y >> 2) & m2);
    u = (u & m2) + ((u >> 2) & m2);
    v = (v & m2) + ((v >> 2) & m2);
    x = x + y; 
    u = u + v; 
    x = (x & m3) + ((x >> 4) & m3);
    u = (u & m3) + ((u >> 4) & m3);
    x = x + u; 
    x = x + (x >> 8);
    x = x + (x >> 16);
    x = x & m4; 
    x = x + (x >> 32);
    return x & 0x000001FF;
}

我兴奋了一会儿，但结果是gcc在-O3上玩内联的把戏，尽管我在一些测试中没有使用内联关键字。当我让gcc玩把戏时，对pop4()的十亿次调用需要12.56 gigacycles，但我确定它是将参数折叠为常量表达式。更实际的数字似乎是19.6gc，以实现30%的加速。我的测试循环现在看起来像这样，确保每个参数足够不同，以阻止gcc耍花招。

   hitime b4 = rdtsc(); 
   for (unsigned long i = 10L * 1000*1000*1000; i < 11L * 1000*1000*1000; ++i) 
      sum += pop4 (i,  i^1, ~i, i|1); 
   hitime e4 = rdtsc(); 

2560亿比特加起来在8.17秒内过去了。根据16位表查找的基准测试，3200万比特的计算结果为1.02秒。不能直接比较，因为另一个工作台没有给出时钟速度，但看起来我已经把64KB表版本的鼻涕打出来了，这首先是L1缓存的悲惨使用。

更新:决定做明显的和创建pop6()通过增加四个重复的行。结果是22.8gc, 3840亿比特在9.5秒内加起来。所以还有20%现在是800毫秒，320亿比特。

我总是在竞争性编程中使用它，它很容易写，而且效率很高:

#include <bits/stdc++.h>

using namespace std;

int countOnes(int n) {
    bitset<32> b(n);
    return b.count();
}

天真的解决方案

时间复杂度为O(no。n的比特数)

int countSet(unsigned int n)
{
    int res=0;
    while(n!=0){
      res += (n&1);
      n >>= 1;      // logical right shift, like C unsigned or Java >>>
    }
   return res;
}

Brian Kerningam的算法

时间复杂度为O(n中设置位的个数)

int countSet(unsigned int n)
{
  int res=0;
  while(n != 0)
  {
    n = (n & (n-1));
    res++;
  }
  return res;
} 

32位数字的查找表方法-在这种方法中，我们将32位数字分解为4个8位数字的块

时间复杂度为O(1)

static unsigned char table[256]; /* the table size is 256,
                        the number of values i&0xFF (8 bits) can have */

void initialize() //holds the number of set bits from 0 to 255
{
  table[0]=0;
  for(unsigned int i=1;i<256;i++)
     table[i]=(i&1)+table[i>>1];
}

int countSet(unsigned int n)
{
  // 0xff is hexadecimal representation of 8 set bits.
  int res=table[n & 0xff];
  n=n>>8;
  res=res+ table[n & 0xff];
  n=n>>8;
  res=res+ table[n & 0xff];
  n=n>>8;
  res=res+ table[n & 0xff];
  return res;
}

这是一个可移植的模块(ANSI-C)，它可以在任何架构上对每个算法进行基准测试。

你的CPU有9位字节?目前它实现了2个算法，K&R算法和一个字节查找表。查找表的平均速度比K&R算法快3倍。如果有人能想出办法使“黑客的喜悦”算法可移植，请随意添加它。

#ifndef _BITCOUNT_H_
#define _BITCOUNT_H_

/* Return the Hamming Wieght of val, i.e. the number of 'on' bits. */
int bitcount( unsigned int );

/* List of available bitcount algorithms.  
 * onTheFly:    Calculate the bitcount on demand.
 *
 * lookupTalbe: Uses a small lookup table to determine the bitcount.  This
 * method is on average 3 times as fast as onTheFly, but incurs a small
 * upfront cost to initialize the lookup table on the first call.
 *
 * strategyCount is just a placeholder. 
 */
enum strategy { onTheFly, lookupTable, strategyCount };

/* String represenations of the algorithm names */
extern const char *strategyNames[];

/* Choose which bitcount algorithm to use. */
void setStrategy( enum strategy );

#endif

.

#include <limits.h>

#include "bitcount.h"

/* The number of entries needed in the table is equal to the number of unique
 * values a char can represent which is always UCHAR_MAX + 1*/
static unsigned char _bitCountTable[UCHAR_MAX + 1];
static unsigned int _lookupTableInitialized = 0;

static int _defaultBitCount( unsigned int val ) {
    int count;

    /* Starting with:
     * 1100 - 1 == 1011,  1100 & 1011 == 1000
     * 1000 - 1 == 0111,  1000 & 0111 == 0000
     */
    for ( count = 0; val; ++count )
        val &= val - 1;

    return count;
}

/* Looks up each byte of the integer in a lookup table.
 *
 * The first time the function is called it initializes the lookup table.
 */
static int _tableBitCount( unsigned int val ) {
    int bCount = 0;

    if ( !_lookupTableInitialized ) {
        unsigned int i;
        for ( i = 0; i != UCHAR_MAX + 1; ++i )
            _bitCountTable[i] =
                ( unsigned char )_defaultBitCount( i );

        _lookupTableInitialized = 1;
    }

    for ( ; val; val >>= CHAR_BIT )
        bCount += _bitCountTable[val & UCHAR_MAX];

    return bCount;
}

static int ( *_bitcount ) ( unsigned int ) = _defaultBitCount;

const char *strategyNames[] = { "onTheFly", "lookupTable" };

void setStrategy( enum strategy s ) {
    switch ( s ) {
    case onTheFly:
        _bitcount = _defaultBitCount;
        break;
    case lookupTable:
        _bitcount = _tableBitCount;
        break;
    case strategyCount:
        break;
    }
}

/* Just a forwarding function which will call whichever version of the
 * algorithm has been selected by the client 
 */
int bitcount( unsigned int val ) {
    return _bitcount( val );
}

#ifdef _BITCOUNT_EXE_

#include <stdio.h>
#include <stdlib.h>
#include <time.h>

/* Use the same sequence of pseudo random numbers to benmark each Hamming
 * Weight algorithm.
 */
void benchmark( int reps ) {
    clock_t start, stop;
    int i, j;
    static const int iterations = 1000000;

    for ( j = 0; j != strategyCount; ++j ) {
        setStrategy( j );

        srand( 257 );

        start = clock(  );

        for ( i = 0; i != reps * iterations; ++i )
            bitcount( rand(  ) );

        stop = clock(  );

        printf
            ( "\n\t%d psudoe-random integers using %s: %f seconds\n\n",
              reps * iterations, strategyNames[j],
              ( double )( stop - start ) / CLOCKS_PER_SEC );
    }
}

int main( void ) {
    int option;

    while ( 1 ) {
        printf( "Menu Options\n"
            "\t1.\tPrint the Hamming Weight of an Integer\n"
            "\t2.\tBenchmark Hamming Weight implementations\n"
            "\t3.\tExit ( or cntl-d )\n\n\t" );

        if ( scanf( "%d", &option ) == EOF )
            break;

        switch ( option ) {
        case 1:
            printf( "Please enter the integer: " );
            if ( scanf( "%d", &option ) != EOF )
                printf
                    ( "The Hamming Weight of %d ( 0x%X ) is %d\n\n",
                      option, option, bitcount( option ) );
            break;
        case 2:
            printf
                ( "Please select number of reps ( in millions ): " );
            if ( scanf( "%d", &option ) != EOF )
                benchmark( option );
            break;
        case 3:
            goto EXIT;
            break;
        default:
            printf( "Invalid option\n" );
        }

    }

 EXIT:
    printf( "\n" );

    return 0;
}

#endif

一个快速的c#解决方案，使用预先计算的字节位计数表，并根据输入大小进行分支。

public static class BitCount
{
    public static uint GetSetBitsCount(uint n)
    {
        var counts = BYTE_BIT_COUNTS;
        return n <= 0xff ? counts[n]
             : n <= 0xffff ? counts[n & 0xff] + counts[n >> 8]
             : n <= 0xffffff ? counts[n & 0xff] + counts[(n >> 8) & 0xff] + counts[(n >> 16) & 0xff]
             : counts[n & 0xff] + counts[(n >> 8) & 0xff] + counts[(n >> 16) & 0xff] + counts[(n >> 24) & 0xff];
    }

    public static readonly uint[] BYTE_BIT_COUNTS =
    {
        0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4,
        1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
        1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
        2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
        1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
        2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
        2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
        3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
        1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
        2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
        2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
        3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
        2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
        3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
        3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
        4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8
    };
}