How to count the number of set bits in a 32-bit in

Question

8 bits representing the number 7 look like this:

00000111

Three bits are set.

What are algorithms to determine the number of set bits in a 32-bit integer?

Answer 1

I wrote a fast bitcount macro for RISC machines in about 1990. It does not use advanced arithmetic (multiplication, division, %), memory fetches (way too slow), branches (way too slow), but it does assume the CPU has a 32-bit barrel shifter (in other words, >> 1 and >> 32 take the same amount of cycles.) It assumes that small constants (such as 6, 12, 24) cost nothing to load into the registers, or are stored in temporaries and reused over and over again.

With these assumptions, it counts 32 bits in about 16 cycles/instructions on most RISC machines. Note that 15 instructions/cycles is close to a lower bound on the number of cycles or instructions, because it seems to take at least 3 instructions (mask, shift, operator) to cut the number of addends in half, so log_2(32) = 5, 5 x 3 = 15 instructions is a quasi-lowerbound.

#define BitCount(X,Y)           \
                Y = X - ((X >> 1) & 033333333333) - ((X >> 2) & 011111111111); \
                Y = ((Y + (Y >> 3)) & 030707070707); \
                Y =  (Y + (Y >> 6)); \
                Y = (Y + (Y >> 12) + (Y >> 24)) & 077;

Here is a secret to the first and most complex step:

input output
AB    CD             Note
00    00             = AB
01    01             = AB
10    01             = AB - (A >> 1) & 0x1
11    10             = AB - (A >> 1) & 0x1

so if I take the 1st column (A) above, shift it right 1 bit, and subtract it from AB, I get the output (CD). The extension to 3 bits is similar; you can check it with an 8-row boolean table like mine above if you wish.

• Don Gillies

Answer 2

I found an implementation of bit counting in an array with using of SIMD instruction (SSSE3 and AVX2). It has in 2-2.5 times better performance than if it will use __popcnt64 intrinsic function.

SSSE3 version:

#include <smmintrin.h>
#include <stdint.h>

const __m128i Z = _mm_set1_epi8(0x0);
const __m128i F = _mm_set1_epi8(0xF);
//Vector with pre-calculated bit count:
const __m128i T = _mm_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);

uint64_t BitCount(const uint8_t * src, size_t size)
{
    __m128i _sum =  _mm128_setzero_si128();
    for (size_t i = 0; i < size; i += 16)
    {
        //load 16-byte vector
        __m128i _src = _mm_loadu_si128((__m128i*)(src + i));
        //get low 4 bit for every byte in vector
        __m128i lo = _mm_and_si128(_src, F);
        //sum precalculated value from T
        _sum = _mm_add_epi64(_sum, _mm_sad_epu8(Z, _mm_shuffle_epi8(T, lo)));
        //get high 4 bit for every byte in vector
        __m128i hi = _mm_and_si128(_mm_srli_epi16(_src, 4), F);
        //sum precalculated value from T
        _sum = _mm_add_epi64(_sum, _mm_sad_epu8(Z, _mm_shuffle_epi8(T, hi)));
    }
    uint64_t sum[2];
    _mm_storeu_si128((__m128i*)sum, _sum);
    return sum[0] + sum[1];
}

AVX2 version:

#include <immintrin.h>
#include <stdint.h>

const __m256i Z = _mm256_set1_epi8(0x0);
const __m256i F = _mm256_set1_epi8(0xF);
//Vector with pre-calculated bit count:
const __m256i T = _mm256_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4, 
                                   0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);

uint64_t BitCount(const uint8_t * src, size_t size)
{
    __m256i _sum =  _mm256_setzero_si256();
    for (size_t i = 0; i < size; i += 32)
    {
        //load 32-byte vector
        __m256i _src = _mm256_loadu_si256((__m256i*)(src + i));
        //get low 4 bit for every byte in vector
        __m256i lo = _mm256_and_si256(_src, F);
        //sum precalculated value from T
        _sum = _mm256_add_epi64(_sum, _mm256_sad_epu8(Z, _mm256_shuffle_epi8(T, lo)));
        //get high 4 bit for every byte in vector
        __m256i hi = _mm256_and_si256(_mm256_srli_epi16(_src, 4), F);
        //sum precalculated value from T
        _sum = _mm256_add_epi64(_sum, _mm256_sad_epu8(Z, _mm256_shuffle_epi8(T, hi)));
    }
    uint64_t sum[4];
    _mm256_storeu_si256((__m256i*)sum, _sum);
    return sum[0] + sum[1] + sum[2] + sum[3];
}

Answer 3

The function you are looking for is often called the "sideways sum" or "population count" of a binary number. Knuth discusses it in pre-Fascicle 1A, pp11-12 (although there was a brief reference in Volume 2, 4.6.3-(7).)

The locus classicus is Peter Wegner's article "A Technique for Counting Ones in a Binary Computer", from the Communications of the ACM, Volume 3 (1960) Number 5, page 322. He gives two different algorithms there, one optimized for numbers expected to be "sparse" (i.e., have a small number of ones) and one for the opposite case.

Answer 4

The Hacker's Delight bit-twiddling becomes so much clearer when you write out the bit patterns.

unsigned int bitCount(unsigned int x)
{
  x = ((x >> 1) & 0b01010101010101010101010101010101)
     + (x       & 0b01010101010101010101010101010101);
  x = ((x >> 2) & 0b00110011001100110011001100110011)
     + (x       & 0b00110011001100110011001100110011); 
  x = ((x >> 4) & 0b00001111000011110000111100001111)
     + (x       & 0b00001111000011110000111100001111); 
  x = ((x >> 8) & 0b00000000111111110000000011111111)
     + (x       & 0b00000000111111110000000011111111); 
  x = ((x >> 16)& 0b00000000000000001111111111111111)
     + (x       & 0b00000000000000001111111111111111); 
  return x;
}

The first step adds the even bits to the odd bits, producing a sum of bits in each two. The other steps add high-order chunks to low-order chunks, doubling the chunk size all the way up, until we have the final count taking up the entire int.

Answer 5

I got bored, and timed a billion iterations of three approaches. Compiler is gcc -O3. CPU is whatever they put in the 1st gen Macbook Pro.

Fastest is the following, at 3.7 seconds:

static unsigned char wordbits[65536] = { bitcounts of ints between 0 and 65535 };
static int popcount( unsigned int i )
{
    return( wordbits[i&0xFFFF] + wordbits[i>>16] );
}

Second place goes to the same code but looking up 4 bytes instead of 2 halfwords. That took around 5.5 seconds.

Third place goes to the bit-twiddling 'sideways addition' approach, which took 8.6 seconds.

Fourth place goes to GCC's __builtin_popcount(), at a shameful 11 seconds.

The counting one-bit-at-a-time approach was waaaay slower, and I got bored of waiting for it to complete.

So if you care about performance above all else then use the first approach. If you care, but not enough to spend 64Kb of RAM on it, use the second approach. Otherwise use the readable (but slow) one-bit-at-a-time approach.

It's hard to think of a situation where you'd want to use the bit-twiddling approach.

Edit: Similar results here.

Answer 6

unsigned int count_bit(unsigned int x)
{
  x = (x & 0x55555555) + ((x >> 1) & 0x55555555);
  x = (x & 0x33333333) + ((x >> 2) & 0x33333333);
  x = (x & 0x0F0F0F0F) + ((x >> 4) & 0x0F0F0F0F);
  x = (x & 0x00FF00FF) + ((x >> 8) & 0x00FF00FF);
  x = (x & 0x0000FFFF) + ((x >> 16)& 0x0000FFFF);
  return x;
}

Let me explain this algorithm.

This algorithm is based on Divide and Conquer Algorithm. Suppose there is a 8bit integer 213(11010101 in binary), the algorithm works like this(each time merge two neighbor blocks):

+-------------------------------+
| 1 | 1 | 0 | 1 | 0 | 1 | 0 | 1 |  <- x
|  1 0  |  0 1  |  0 1  |  0 1  |  <- first time merge
|    0 0 1 1    |    0 0 1 0    |  <- second time merge
|        0 0 0 0 0 1 0 1        |  <- third time ( answer = 00000101 = 5)
+-------------------------------+